● B2B offer — Monitoring & information
Monitoring agent: defined sources, sourced summaries
Following competitors, the trade press and regulatory developments takes a steadiness that everyday work eats into. Your agent watches the sources you have defined, sets the noise aside and produces a periodic summary in which every statement refers back to its source. Hosted in France — local inference or an isolated resource — what you are watching tells no one anything. You approve before anything circulates internally.
Hosted in France
Monitoring subjects kept confidential
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Monitoring agent · hosted in France
Give me this week's position on our three monitoring subjects.
Weekly summary on your three subjects, from the sources you have defined.
Every item kept refers back to its original source, dated and cited.
Publications with nothing new have been set aside, with the reason stated.
🔗 Sourced · every statement refers back to its publication
Every item kept refers back to its original source, dated and cited.
Publications with nothing new have been set aside, with the reason stated.
🔗 Sourced · every statement refers back to its publication
On the regulatory subject, what actually changes for us?
Two texts concern the scope you have declared. I report what they say and their date of entry into force as published.
Analysing what they mean for your organisation is not for the agent: it belongs to your lawyer.
✎ Support · reported, no legal interpretation
Analysing what they mean for your organisation is not for the agent: it belongs to your lawyer.
✎ Support · reported, no legal interpretation
Local inference · no data outside the EU
Monitoring hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The head of monitoring decidesThe agent prepares, never rules
✦ In brief
A Blue Lemon Agent monitoring agent watches the sources you define — competitors, trade press, regulatory publications — and produces periodic sourced summaries in which every item refers back to its original publication. It sets repetition aside rather than padding the report. It runs on local inference or is hosted in France: the subjects you watch, which give away your strategy, are entrusted to no one, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Live within a few weeks.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
15
monitoring capabilities included in the core
0
decision taken without human approval
These figures describe our offer, not results measured at a client: how large the gain is on your number of sources and the frequency of your summaries is confirmed by a pilot.
The context
What does an AI agent bring to your monitoring?
Monitoring is the first task sacrificed when the day presses. But entrusting what you watch to a consumer tool amounts to publishing your strategic roadmap.
! The issue
Useful monitoring means going through the same sources at regular intervals, recognising what is new and reporting it briefly — a discipline that urgency always breaks. Yet entrusting it to a consumer service amounts to revealing the competitors you watch, the subjects that concern you and the regulations that apply to you to a third party subject to the Cloud Act.
✓ Our answer
Automated monitoring is only of interest if it is sovereign and systematically sourced. Local inference or an isolated resource hosted in France, reporting with a reference back to the original publication, nothing circulated automatically: you gain the steadiness without losing control of what you are looking for. The agent reports what is published; it does not draw conclusions on your behalf.
The decisive point
Your monitoring subjects and what you watch: sovereignty & compliance
What you watch says a great deal about what you are preparing. Here is how the architecture of our agents protects your monitoring subjects.
Local inference
The agent can run on a machine belonging to your organisation: what you watch is passed to no third-party service.
Hosting in France
Otherwise, a dedicated and isolated resource hosted in France, under French law — your sources and your summaries: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For your monitoring subjects and what you watch, the architecture aims to reduce exposure to the Cloud Act and FISA 702; being located in France or in the European Union does not, on its own, guarantee immunity.
Isolated resource
No pooling: an environment strictly dedicated to your organisation and its monitoring subjects.
Sources under control, reporting recorded
You define the sources and the frequency; every summary is logged and every statement stays tied to its original publication.
AI Act: governed deployment
The agent is strictly in support; no conclusion is presented as settled analysis; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For monitoring tied to an acquisition or a dispute, SecNumCloud and reinforced hosting are options depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
6 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The company in this demonstration
Fictional companyNorélia — manufacturer of professional cleaning and hygiene products
- Sector
- Speciality chemicals — cleaning, disinfection and maintenance for professional users
- Headcount
- 96 staff, including 1 quality and regulatory manager, 2 product managers, 3 key account salespeople and 1 communication officer
- Market served
- 640 business customers — local authorities, food industry, cleaning contractors; 9 direct competitors tracked
- Volume
- 3 ranges, 210 product lines, an 18-month reformulation cycle, one public tender won out of two
- Tools in place
- Email, document intranet, CRM, product data sheets and a shared monitoring spreadsheet — the agent plugs into them, nothing is replaced
- Who decides
- The quality and regulatory manager sets the monitoring scope and approves every summary before it circulates; the sales director decides on tender responses; outside counsel rules on the scope of legal texts
- Room for improvement
- 2,148 publications land in the monitoring spreadsheet every week and nobody reads them; the weekly summary was produced only 19 weeks out of 52; a text that was being tracked stayed invisible until 12 days before it took effect
Norélia has no shortage of sources: it has too many, and nobody has time to go through them. The agent runs on local inference on a company server and plugs into the monitoring spreadsheet, the document intranet, the CRM and the product data sheets: it collects, filters, sources and routes — the quality manager approves before anything circulates. The exchanges below cover one quarter, from the first summary to the review.
This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.
Monitoring agent · hosted in France
Your weekly summary has been ready since Sunday 10 p.m., on your three monitoring scopes. I saw 2,148 publications, kept 31, and discarded the other 2,117 — each one with its reason written down.
A monitoring scope is a subject you have defined yourself, with its sources and its frequency. The agent stays inside it: it does not look elsewhere and does not change the scope on its own.
The analysis I draw from twelve weeks, and it explains why your spreadsheet never gets read: 68 % of what lands in it is syndication — the same content republished by another site, without one added sentence — and 4 of your 61 sources produce 1,190 of them a week on their own. Your monitoring was not incomplete, it was drowned: not the same problem, and not one that adding sources solves.
What the 31 kept items contain, each with its original publication, dated and quoted:
· Competitive scope — 14 items, including 3 confirmed facts and 11 announcements, told apart.
· Trade press — 12 items, including 5 mentions of Norélia.
· Regulatory — 5 items, including 2 deadlines falling inside your declared scope.
The time this shifts: reviewing the sources took 50 % of your monitoring time; it takes 12 %. The rest is your reading, and that is what decides.
What I suggest: that you read this first summary line by line, discard reasons included. If a single discard looks unjustified, the rule that produced it is corrected in one sentence — and it applies from the following week. weekly-summary_3-scopes.pdf31 items kept, each with its dated source
⛓ Sourced · 61 declared sources, 2,148 publications this week, 12 weeks of spreadsheet history
A monitoring scope is a subject you have defined yourself, with its sources and its frequency. The agent stays inside it: it does not look elsewhere and does not change the scope on its own.
The analysis I draw from twelve weeks, and it explains why your spreadsheet never gets read: 68 % of what lands in it is syndication — the same content republished by another site, without one added sentence — and 4 of your 61 sources produce 1,190 of them a week on their own. Your monitoring was not incomplete, it was drowned: not the same problem, and not one that adding sources solves.
What the 31 kept items contain, each with its original publication, dated and quoted:
· Competitive scope — 14 items, including 3 confirmed facts and 11 announcements, told apart.
· Trade press — 12 items, including 5 mentions of Norélia.
· Regulatory — 5 items, including 2 deadlines falling inside your declared scope.
The time this shifts: reviewing the sources took 50 % of your monitoring time; it takes 12 %. The rest is your reading, and that is what decides.
What I suggest: that you read this first summary line by line, discard reasons included. If a single discard looks unjustified, the rule that produced it is corrected in one sentence — and it applies from the following week. weekly-summary_3-scopes.pdf31 items kept, each with its dated source
⛓ Sourced · 61 declared sources, 2,148 publications this week, 12 weeks of spreadsheet history
Discarding 2,117 publications — doesn't that risk making us miss something?
That is the right question, and I answer it with a measurement rather than a promise: the filtering is published, and it is checked every month.
The six discard reasons, with this week's volumes:
· Identical syndication — 1,190. The same text, already kept at its original source. I keep the original and discard the copies, and I tell you how many there were: that count is itself information about how far a subject has travelled.
· Outside the declared scope — 604. Publications touching none of your three scopes.
· Repeat of an item already sent — 187. Same fact, new date, nothing new.
· Promotional publication with no fact — 92.
· Technical duplicate — 41. Same page, two addresses.
· Undated source — 3. Those I set aside rather than throw away: information without a date cannot be passed to whoever decides.
The control, and it is what replaces the promise: every month I draw 40 discarded publications at random and hand them back to you to read, with their reason. Over the last three months, you put 2 out of 120 back — both times a competitor named differently from its registered name. I have added its two names to the rule: they will not slip through again.
What that control gives you, and nobody measures it: a filtering error rate of 1.7 %, known and checkable, instead of a spreadsheet nobody reads whose error rate is 100 % by abandonment.
What I suggest: adding a safety net — a list of words that forbid automatic discarding, whatever the reason. You give me five, I apply them: anything containing them comes through, even as syndication. filtering-published_6-reasons-and-its-control.pdf1.7 % filtering error, measured and corrected
⛓ Sourced · this week's filtering log, 120 discarded publications reread over 3 months, competitor naming rules
The six discard reasons, with this week's volumes:
· Identical syndication — 1,190. The same text, already kept at its original source. I keep the original and discard the copies, and I tell you how many there were: that count is itself information about how far a subject has travelled.
· Outside the declared scope — 604. Publications touching none of your three scopes.
· Repeat of an item already sent — 187. Same fact, new date, nothing new.
· Promotional publication with no fact — 92.
· Technical duplicate — 41. Same page, two addresses.
· Undated source — 3. Those I set aside rather than throw away: information without a date cannot be passed to whoever decides.
The control, and it is what replaces the promise: every month I draw 40 discarded publications at random and hand them back to you to read, with their reason. Over the last three months, you put 2 out of 120 back — both times a competitor named differently from its registered name. I have added its two names to the rule: they will not slip through again.
What that control gives you, and nobody measures it: a filtering error rate of 1.7 %, known and checkable, instead of a spreadsheet nobody reads whose error rate is 100 % by abandonment.
What I suggest: adding a safety net — a list of words that forbid automatic discarding, whatever the reason. You give me five, I apply them: anything containing them comes through, even as syndication. filtering-published_6-reasons-and-its-control.pdf1.7 % filtering error, measured and corrected
⛓ Sourced · this week's filtering log, 120 discarded publications reread over 3 months, competitor naming rules
And the summary itself: who writes it, what does it look like, and who receives it?
I write it, you approve it, and it goes to four different people — not to a mailing list.
Its shape, in four parts, each answering a question you already ask yourself:
· What has changed — the new facts, a paragraph each, with the link to the original publication and its date.
· What has not changed but is coming — known deadlines drawing closer. That is the part nobody writes, and the one that prevents surprises.
· What needs a decision — three points at most, with the person concerned named by their role.
· What was discarded — volumes by reason, in one line. Filtering you cannot see is filtering you cannot correct.
The routing follows who can act: a confirmed competitor fact goes to the sales director; a regulatory deadline to the quality and regulatory manager, three times — at six months, one month, one week; a press mention to the communication officer; a signal touching a formulation to the two product managers. Accurate information sent to the wrong person is information lost, and it is the most common waste in corporate monitoring.
The time this shifts: writing the summary took 35 % of your monitoring time; it takes 12 %. And weeding out repetition, done by reading for 25 %, takes 6 %.
The figure that matters most to you: your weekly summary was produced 19 weeks out of 52 last year. Over the past twelve weeks it has been produced 12 times out of 12, on Sunday at 10 p.m. Regularity was never a matter of willingness, it was a matter of time available on a Friday evening.
What I suggest: that you approve the summary on Monday before 9 a.m. Nothing goes out before your approval, and if you do not approve, it waits: I never circulate anything on my own. who-receives-what_routing-by-role.pdf4 recipients, 3 reminders per deadline
✎ Framework · summary template, routing table by role, 52 weeks of production history
Its shape, in four parts, each answering a question you already ask yourself:
· What has changed — the new facts, a paragraph each, with the link to the original publication and its date.
· What has not changed but is coming — known deadlines drawing closer. That is the part nobody writes, and the one that prevents surprises.
· What needs a decision — three points at most, with the person concerned named by their role.
· What was discarded — volumes by reason, in one line. Filtering you cannot see is filtering you cannot correct.
The routing follows who can act: a confirmed competitor fact goes to the sales director; a regulatory deadline to the quality and regulatory manager, three times — at six months, one month, one week; a press mention to the communication officer; a signal touching a formulation to the two product managers. Accurate information sent to the wrong person is information lost, and it is the most common waste in corporate monitoring.
The time this shifts: writing the summary took 35 % of your monitoring time; it takes 12 %. And weeding out repetition, done by reading for 25 %, takes 6 %.
The figure that matters most to you: your weekly summary was produced 19 weeks out of 52 last year. Over the past twelve weeks it has been produced 12 times out of 12, on Sunday at 10 p.m. Regularity was never a matter of willingness, it was a matter of time available on a Friday evening.
What I suggest: that you approve the summary on Monday before 9 a.m. Nothing goes out before your approval, and if you do not approve, it waits: I never circulate anything on my own. who-receives-what_routing-by-role.pdf4 recipients, 3 reminders per deadline
✎ Framework · summary template, routing table by role, 52 weeks of production history
What do you decide on when you say an item deserves to come through? We would like to understand, not to trust.
On a relevance score computed from your declared scope, never from a publication's audience — and the threshold is yours to set, once you have seen the three lists.
The five criteria of the score, and you weight them:
· Recognised entity — does the item speak of an entity in your table: one of your 9 competitors, one of your customers, one of your 210 product lines.
· Declared subject — does it attach to one of your 11 subjects, and under which label.
· Register — a confirmed fact weighs more than an announcement, an announcement more than a syndicated copy.
· Freshness, counted from the date of first publication, never from the date of the copy.
· Reach, measured through lineage: how many copies, and over how long.
What the threshold changes, and it is a number you choose: at 62 it brings through this week's 31 items; at 55 it would bring through 48; at 70, 19. I have made all three lists for you on the same week, so that the choice is made on lines rather than on a slider. An item read by forty thousand people and outside your 11 subjects scores zero, and that is deliberate: your monitoring does not follow what is popular, it follows what you declared.
Now the signal, because it is something other than a relevant item: a signal is the convergence of independent items on the same entity or the same label. My rule is written down: at least three occurrences, in at least two channels, over ninety rolling days.
The state of your signals this quarter: 4 open. The strongest carries six occurrences on food-grade degreasers. 1 was closed at ninety days for want of a third item — and the closure is written and dated, with what had been seen. A signal closed in silence is a signal that starts again from zero next time.
What I suggest: that you set the threshold at 62 for this quarter, and that we replay it at the monthly review. If rereading the discarded items gives you back two more than the previous month, it is the threshold that needs moving, not the rule — and I will tell you which of the two.
✎ Framework · 5-criteria weighted relevance score, three lists simulated on the same week — thresholds 55, 62 and 70; written signal rule: 3 occurrences, 2 channels, 90 days
The five criteria of the score, and you weight them:
· Recognised entity — does the item speak of an entity in your table: one of your 9 competitors, one of your customers, one of your 210 product lines.
· Declared subject — does it attach to one of your 11 subjects, and under which label.
· Register — a confirmed fact weighs more than an announcement, an announcement more than a syndicated copy.
· Freshness, counted from the date of first publication, never from the date of the copy.
· Reach, measured through lineage: how many copies, and over how long.
What the threshold changes, and it is a number you choose: at 62 it brings through this week's 31 items; at 55 it would bring through 48; at 70, 19. I have made all three lists for you on the same week, so that the choice is made on lines rather than on a slider. An item read by forty thousand people and outside your 11 subjects scores zero, and that is deliberate: your monitoring does not follow what is popular, it follows what you declared.
Now the signal, because it is something other than a relevant item: a signal is the convergence of independent items on the same entity or the same label. My rule is written down: at least three occurrences, in at least two channels, over ninety rolling days.
The state of your signals this quarter: 4 open. The strongest carries six occurrences on food-grade degreasers. 1 was closed at ninety days for want of a third item — and the closure is written and dated, with what had been seen. A signal closed in silence is a signal that starts again from zero next time.
What I suggest: that you set the threshold at 62 for this quarter, and that we replay it at the monthly review. If rereading the discarded items gives you back two more than the previous month, it is the threshold that needs moving, not the rule — and I will tell you which of the two.
✎ Framework · 5-criteria weighted relevance score, three lists simulated on the same week — thresholds 55, 62 and 70; written signal rule: 3 occurrences, 2 channels, 90 days
And when something cannot wait until Sunday 10 p.m.?
There are three levels of alerts, each with its deadline and its named recipient — and the routing follows who can act, not who sits on a mailing list.
The three levels, as you settled them:
· Immediate alert — two hours, to the named person, with the dated source and a single line of context.
· Daily digest — 6 p.m., what deserves reading the same day without interrupting anyone.
· Weekly summary — Sunday 10 p.m., everything else, which you approve on Monday before 9 a.m.
What this week produced: 2 immediate alerts — the collection failure on the sector's official bulletin, and competitor B's range withdrawal touching 41 of your customers — 5 lines in the daily digest, and the other 24 items in the summary.
The routing, and it has a return path: every alert goes to a role — sales director, quality and regulatory manager, communication officer, product managers — and its opening is logged. An immediate alert not opened within forty-eight hours goes up to your quality manager, who reassigns it or lets it stand. One case this quarter, during a week of leave.
The figure that does not suit me, and it cost three months: my first immediate-alert rule fired fourteen times a week. Over twelve weeks, 168 alerts went out and 3 actually led to an action. An alert nobody opens any more is an alert that does not exist, and that is what was happening.
What I did with it: the rule was tightened on 20 March on three conditions together — an entity in your scope, a confirmed fact, and an affected customer identified in the CRM — then replayed over the twelve preceding weeks. There it produces 3 alerts a week instead of 14, it finds the 3 that had led to an action, and it loses none of the others. Since then, the 3 of the week are opened the same day.
What I suggest: that you keep the level in your own hands, subject by subject. You raise a subject to immediate alert in one sentence, and you lower it the same way — and each month I will tell you how many alerts each setting produced, and how many were opened.
⛓ Sourced · alert log — 2 immediate and 5 daily this week; 168 alerts over 12 weeks of which 3 led to an action; rule tightened on 20 March and replayed
The three levels, as you settled them:
· Immediate alert — two hours, to the named person, with the dated source and a single line of context.
· Daily digest — 6 p.m., what deserves reading the same day without interrupting anyone.
· Weekly summary — Sunday 10 p.m., everything else, which you approve on Monday before 9 a.m.
What this week produced: 2 immediate alerts — the collection failure on the sector's official bulletin, and competitor B's range withdrawal touching 41 of your customers — 5 lines in the daily digest, and the other 24 items in the summary.
The routing, and it has a return path: every alert goes to a role — sales director, quality and regulatory manager, communication officer, product managers — and its opening is logged. An immediate alert not opened within forty-eight hours goes up to your quality manager, who reassigns it or lets it stand. One case this quarter, during a week of leave.
The figure that does not suit me, and it cost three months: my first immediate-alert rule fired fourteen times a week. Over twelve weeks, 168 alerts went out and 3 actually led to an action. An alert nobody opens any more is an alert that does not exist, and that is what was happening.
What I did with it: the rule was tightened on 20 March on three conditions together — an entity in your scope, a confirmed fact, and an affected customer identified in the CRM — then replayed over the twelve preceding weeks. There it produces 3 alerts a week instead of 14, it finds the 3 that had led to an action, and it loses none of the others. Since then, the 3 of the week are opened the same day.
What I suggest: that you keep the level in your own hands, subject by subject. You raise a subject to immediate alert in one sentence, and you lower it the same way — and each month I will tell you how many alerts each setting produced, and how many were opened.
⛓ Sourced · alert log — 2 immediate and 5 daily this week; 168 alerts over 12 weeks of which 3 led to an action; rule tightened on 20 March and replayed
Before we even talk about the summary: what exactly are you watching? Here, nobody knows any more what is in the spreadsheet, or why.
Your monitoring scope is written down, dated and settled by you: three angles, eleven subjects, and for each subject what comes in and what stays out.
A monitoring angle groups the subjects that go to the same place and the same person; a subject is what is actually watched, with its entities, its wording and its sources.
Your three monitoring angles and their eleven subjects, as you settled them on 14 March:
· Competitive angle — 4 subjects: moves by your 9 competitors, listed prices, buying-group listings, site openings and closures.
· Trade press angle — 3 subjects: mentions of Norélia, practices in the market served, manufacturer overviews.
· Regulatory angle — 4 subjects: labelling of professional-use mixtures, declaration of composition, discharge thresholds applying to your customers, transport and storage.
What each subject carries, and this is exactly what the spreadsheet lacked: the entities concerned, the taxonomy labels, the authorised sources, the frequency, the recipient role, and the written exclusion boundary. On the “listed prices” subject, for instance, consumer marketplaces are outside the scope, and the sentence that says so carries the date of 14 March. A scope with no written exclusion widens on its own — that is what filled your spreadsheet, and it is not a problem of sources.
What I measured by matching twelve weeks of spreadsheet against this scope: 604 publications a week touch none of your eleven subjects. That is the second discard reason in your summary, and it comes from a scope that had never been written down.
What I suggest, and the call is yours: a twelfth candidate subject — the hygiene requirements of your food-industry customers. It is not in your declared scope, and yet 17 discarded items over twelve weeks belonged to it, two of which your product managers later went looking for themselves. I hand you the 17; you open the subject or you leave it out — and if you leave it out, the exclusion sentence is written and dated like the others.
✎ Framework · declared scope settled on 14 March — 3 monitoring angles, 11 subjects, dated exclusion boundaries; 12 weeks of spreadsheet matched against it
A monitoring angle groups the subjects that go to the same place and the same person; a subject is what is actually watched, with its entities, its wording and its sources.
Your three monitoring angles and their eleven subjects, as you settled them on 14 March:
· Competitive angle — 4 subjects: moves by your 9 competitors, listed prices, buying-group listings, site openings and closures.
· Trade press angle — 3 subjects: mentions of Norélia, practices in the market served, manufacturer overviews.
· Regulatory angle — 4 subjects: labelling of professional-use mixtures, declaration of composition, discharge thresholds applying to your customers, transport and storage.
What each subject carries, and this is exactly what the spreadsheet lacked: the entities concerned, the taxonomy labels, the authorised sources, the frequency, the recipient role, and the written exclusion boundary. On the “listed prices” subject, for instance, consumer marketplaces are outside the scope, and the sentence that says so carries the date of 14 March. A scope with no written exclusion widens on its own — that is what filled your spreadsheet, and it is not a problem of sources.
What I measured by matching twelve weeks of spreadsheet against this scope: 604 publications a week touch none of your eleven subjects. That is the second discard reason in your summary, and it comes from a scope that had never been written down.
What I suggest, and the call is yours: a twelfth candidate subject — the hygiene requirements of your food-industry customers. It is not in your declared scope, and yet 17 discarded items over twelve weeks belonged to it, two of which your product managers later went looking for themselves. I hand you the 17; you open the subject or you leave it out — and if you leave it out, the exclusion sentence is written and dated like the others.
✎ Framework · declared scope settled on 14 March — 3 monitoring angles, 11 subjects, dated exclusion boundaries; 12 weeks of spreadsheet matched against it
And the 61 sources — where do they come from? Some have been in the spreadsheet for years and nobody knows who put them there.
They now live in a source register, one line per source, and each line carries eight fields including the one that was missing: who declared it, and when.
Every source named here is fictional and designated by its subject matter, like the rest of this demonstration.
The register as it reads today — 61 sources:
· 22 on the competitive angle: the public register of establishments, the online catalogues of your 9 competitors, buying-group publications, published job adverts.
· 14 trade press titles, among them the trade newsletter of the sector and the technical magazine for industrial cleaning.
· 19 on the regulatory angle, among them the sector's official bulletin and the notices published by professional bodies.
· 6 cross-cutting: awarded tender documents, market overviews.
The eight fields on every line: the subject matter of the source, its nature — official, professional, commercial, third-party —, its collection frequency, the angle and subjects it serves, the person who declared it, the date it entered the register, the date it was last checked, and the reuse terms its own page states.
What the first reading of the register produced, and it does not flatter the spreadsheet: 9 sources had no owner at all — nobody knew any more why they were there; 4 were the same body of material under two addresses; 7 had published nothing for 120 days. And 4 sources produce the week's 1,190 syndicated copies on their own: they stay on the register, filed as syndication sources — they measure how far a subject travels, they never establish its origin.
What I suggest: removing the 7 silent sources and merging the 4 duplicates, as a single decision that fits on one page. The 9 with no owner I hand back to you with what they produced over twelve weeks: six brought nothing, three produced one confirmed fact each. That column is what makes the decision easy, and it existed nowhere.
⛓ Sourced · source register — 61 lines, 8 fields, entry date and last-check date; output of each source measured over 12 weeks
Every source named here is fictional and designated by its subject matter, like the rest of this demonstration.
The register as it reads today — 61 sources:
· 22 on the competitive angle: the public register of establishments, the online catalogues of your 9 competitors, buying-group publications, published job adverts.
· 14 trade press titles, among them the trade newsletter of the sector and the technical magazine for industrial cleaning.
· 19 on the regulatory angle, among them the sector's official bulletin and the notices published by professional bodies.
· 6 cross-cutting: awarded tender documents, market overviews.
The eight fields on every line: the subject matter of the source, its nature — official, professional, commercial, third-party —, its collection frequency, the angle and subjects it serves, the person who declared it, the date it entered the register, the date it was last checked, and the reuse terms its own page states.
What the first reading of the register produced, and it does not flatter the spreadsheet: 9 sources had no owner at all — nobody knew any more why they were there; 4 were the same body of material under two addresses; 7 had published nothing for 120 days. And 4 sources produce the week's 1,190 syndicated copies on their own: they stay on the register, filed as syndication sources — they measure how far a subject travels, they never establish its origin.
What I suggest: removing the 7 silent sources and merging the 4 duplicates, as a single decision that fits on one page. The 9 with no owner I hand back to you with what they produced over twelve weeks: six brought nothing, three produced one confirmed fact each. That column is what makes the decision easy, and it existed nowhere.
⛓ Sourced · source register — 61 lines, 8 fields, entry date and last-check date; output of each source measured over 12 weeks
And if a source goes down, or changes address? How would we find out, other than by chance?
Through a collection log where a failure carries a date and a name instead of a silence. And through versions: a page that changes after publication does not change behind my back.
This week's collection, in figures: 68 runs a day, 476 across the week, at the frequency you set source by source — hourly for the sector's official bulletin, daily for the catalogues, weekly for the overviews. 2,148 publications came out of it.
The failures, and that is your question: 14 of the 476 runs failed, each timestamped with its cause. Three of them are the same one: the sector's official bulletin did not answer on 3, 4 and 5 March at 6.10 a.m. At the third, the alert went to your quality manager on 5 March at 7 a.m., with the cause — the feed address had changed — and the fix was made on 6 March. Without that log the source would simply have gone quiet, and a quiet source looks like a quiet week.
Now the versions, because that is where the real surprises sit: every collection keeps the page as it stood that day, timestamped, with the fingerprint of its content. A page collected again that has changed does not replace the old one: both stay, and the gap between them is the information. This quarter, 38 pages changed after publication, 9 of them on a point that matters — a date, a threshold, a scope.
The case that paid for the log: the trade newsletter of the sector corrected, four days after publication, the effective date it had announced — fourteen months became twelve. Nobody republished, nobody flagged it. Comparing the two versions caught it the same day, and the deadline moved two months earlier in your calendar.
What I suggest: that the 9 changes bearing on a date or a threshold trigger an immediate alert, and the other 29 a line in Sunday's summary. You set the boundary; I will first show you what each of the two rules would have produced over the quarter just gone.
⛓ Sourced · collection log — 476 runs, 14 timestamped failures with their cause; 38 pages changed after publication, versions and fingerprints kept
This week's collection, in figures: 68 runs a day, 476 across the week, at the frequency you set source by source — hourly for the sector's official bulletin, daily for the catalogues, weekly for the overviews. 2,148 publications came out of it.
The failures, and that is your question: 14 of the 476 runs failed, each timestamped with its cause. Three of them are the same one: the sector's official bulletin did not answer on 3, 4 and 5 March at 6.10 a.m. At the third, the alert went to your quality manager on 5 March at 7 a.m., with the cause — the feed address had changed — and the fix was made on 6 March. Without that log the source would simply have gone quiet, and a quiet source looks like a quiet week.
Now the versions, because that is where the real surprises sit: every collection keeps the page as it stood that day, timestamped, with the fingerprint of its content. A page collected again that has changed does not replace the old one: both stay, and the gap between them is the information. This quarter, 38 pages changed after publication, 9 of them on a point that matters — a date, a threshold, a scope.
The case that paid for the log: the trade newsletter of the sector corrected, four days after publication, the effective date it had announced — fourteen months became twelve. Nobody republished, nobody flagged it. Comparing the two versions caught it the same day, and the deadline moved two months earlier in your calendar.
What I suggest: that the 9 changes bearing on a date or a threshold trigger an immediate alert, and the other 29 a line in Sunday's summary. You set the boundary; I will first show you what each of the two rules would have produced over the quarter just gone.
⛓ Sourced · collection log — 476 runs, 14 timestamped failures with their cause; 38 pages changed after publication, versions and fingerprints kept
What arrives raw must be unreadable. What do you do with it before it reaches us?
A normalisation in five steps, then a quality score on five criteria — and whatever fails a criterion says so, instead of passing itself off as complete.
The normalisation, applied to the week's 2,148 entries:
· Dates — nine different formats encountered, three of them ambiguous between day and month. They are settled by the time zone and by the source, never guessed: where they cannot be, the date stays marked uncertain and the item founds no deadline.
· Headlines, stripped of the site name and of the wire mention in front of them.
· The author, taken as signed, never completed.
· The canonical address, so that the same page under two addresses is the same page.
· Language and encoding, recognised before reading.
The quality score, five criteria, published on every item: certain publication date · identified author or publisher · complete text · stable address · recognised language.
This week's result: 2,103 of the 2,148 entries meet all five criteria. 45 miss at least one: 21 truncated to the teaser, 14 with no identifiable author, 7 doubly encoded, 3 with no usable date.
What I do with each group: for 16 of the 21 truncated ones, the full text existed at the original source already declared on the register — that is the version kept. The remaining 5 come through with “truncated text” written on the item, so your reader knows they are reading a teaser. The 3 with no date are set aside rather than thrown away: they are the three under the “undated source” reason in your summary.
The figure that does not suit me, and you should have it: over twelve weeks I marked “author identified” on 11 items whose author was in fact a wire pseudonym. Your communication officer saw the first. The rule was corrected, then replayed over the twelve weeks: it finds all 11, and it invents no others.
What I suggest: that the quality score be shown in the summary next to each item. Two items saying the same thing are not worth the same if one is dated and signed and the other is not — and that is something your reader takes in three seconds.
⛓ Sourced · 2,148 entries normalised, 9 date formats encountered, 5-criteria quality score, author rule corrected then replayed over 12 weeks
The normalisation, applied to the week's 2,148 entries:
· Dates — nine different formats encountered, three of them ambiguous between day and month. They are settled by the time zone and by the source, never guessed: where they cannot be, the date stays marked uncertain and the item founds no deadline.
· Headlines, stripped of the site name and of the wire mention in front of them.
· The author, taken as signed, never completed.
· The canonical address, so that the same page under two addresses is the same page.
· Language and encoding, recognised before reading.
The quality score, five criteria, published on every item: certain publication date · identified author or publisher · complete text · stable address · recognised language.
This week's result: 2,103 of the 2,148 entries meet all five criteria. 45 miss at least one: 21 truncated to the teaser, 14 with no identifiable author, 7 doubly encoded, 3 with no usable date.
What I do with each group: for 16 of the 21 truncated ones, the full text existed at the original source already declared on the register — that is the version kept. The remaining 5 come through with “truncated text” written on the item, so your reader knows they are reading a teaser. The 3 with no date are set aside rather than thrown away: they are the three under the “undated source” reason in your summary.
The figure that does not suit me, and you should have it: over twelve weeks I marked “author identified” on 11 items whose author was in fact a wire pseudonym. Your communication officer saw the first. The rule was corrected, then replayed over the twelve weeks: it finds all 11, and it invents no others.
What I suggest: that the quality score be shown in the summary next to each item. Two items saying the same thing are not worth the same if one is dated and signed and the other is not — and that is something your reader takes in three seconds.
⛓ Sourced · 2,148 entries normalised, 9 date formats encountered, 5-criteria quality score, author rule corrected then replayed over 12 weeks
You mentioned 1,190 syndicated copies a week. How do you know which one is the original?
By the oldest timestamp carrying the same content, and by a written lineage: the origin on one side, its copies on the other, each with its lag.
Lineage is the link that ties a syndicated copy to the publication it descends from. It is kept: it is what makes it possible to say, a year later, where a piece of information came from.
Deduplication works at two levels, and they do not serve the same purpose:
· The exact fingerprint — same content, two addresses. 41 technical duplicates this week, discarded without discussion.
· The near-duplicate — same content, rewritten headline, an added opening sentence. That is the bulk of the volume: 1,190 copies. The matching threshold is a number, it is written down, and you can move it.
This week's case, and it is the clearest: 18 publications carried the same announcement from competitor D. The origin is its own feed, on 4 March at 8.12 a.m.; the other 17 are copies, the first at 8.51 a.m., the last 52 hours later. Your summary carries one only, dated 4 March, and the line says “syndicated 17 times in 52 hours”.
Why lineage is kept rather than thrown away with the copies:
· It measures reach. An announcement picked up 17 times in two days and one picked up twice in three weeks do not tell the same story.
· It catches corrections. When the origin corrects itself, the copies still carry the old version — it happened twice this quarter, and only lineage made it possible to say so.
· It gives the date of first publication, the only one that counts for a deadline.
The figure that does not suit me, and I publish it: this quarter I wrongly matched two distinct releases from the same competitor, published three days apart, the second announcing a second packaging unit. Your product manager saw it, I did not. The near-duplicate threshold was tightened on 12 March, then replayed over twelve weeks of your history: it no longer makes that merge, and it misses none of the others. A wrong merge is never visible afterwards: it produces a history that looks clean and is false.
What I suggest: that matches made within five points of the threshold — there were 23 this quarter — be handed back to you once a month, to reread. That is the batch where an error has a chance of being, and the only one worth your time.
⛓ Sourced · 1,190 syndicated copies and 41 technical duplicates this week, lineage of the 4 March announcement — 17 copies in 52 h; threshold tightened on 12 March then replayed over 12 weeks
Lineage is the link that ties a syndicated copy to the publication it descends from. It is kept: it is what makes it possible to say, a year later, where a piece of information came from.
Deduplication works at two levels, and they do not serve the same purpose:
· The exact fingerprint — same content, two addresses. 41 technical duplicates this week, discarded without discussion.
· The near-duplicate — same content, rewritten headline, an added opening sentence. That is the bulk of the volume: 1,190 copies. The matching threshold is a number, it is written down, and you can move it.
This week's case, and it is the clearest: 18 publications carried the same announcement from competitor D. The origin is its own feed, on 4 March at 8.12 a.m.; the other 17 are copies, the first at 8.51 a.m., the last 52 hours later. Your summary carries one only, dated 4 March, and the line says “syndicated 17 times in 52 hours”.
Why lineage is kept rather than thrown away with the copies:
· It measures reach. An announcement picked up 17 times in two days and one picked up twice in three weeks do not tell the same story.
· It catches corrections. When the origin corrects itself, the copies still carry the old version — it happened twice this quarter, and only lineage made it possible to say so.
· It gives the date of first publication, the only one that counts for a deadline.
The figure that does not suit me, and I publish it: this quarter I wrongly matched two distinct releases from the same competitor, published three days apart, the second announcing a second packaging unit. Your product manager saw it, I did not. The near-duplicate threshold was tightened on 12 March, then replayed over twelve weeks of your history: it no longer makes that merge, and it misses none of the others. A wrong merge is never visible afterwards: it produces a history that looks clean and is false.
What I suggest: that matches made within five points of the threshold — there were 23 this quarter — be handed back to you once a month, to reread. That is the batch where an error has a chance of being, and the only one worth your time.
⛓ Sourced · 1,190 syndicated copies and 41 technical duplicates this week, lineage of the 4 March announcement — 17 copies in 52 h; threshold tightened on 12 March then replayed over 12 weeks
Our salespeople find out about competitors' moves at the customer's office, in a meeting. It's always too late.
The gap between a competitor's move and the moment your salespeople hear of it has gone from five weeks to two days — and I can tell you where the five weeks came from.
What I measured over your last twelve months: of 23 competitor moves your salespeople eventually reported, 19 were already public when they found out, on average for 36 days. The information was not hidden: nobody was reading it.
What I bring you this week on your 9 competitors, and the distinction that structures everything:
· 3 confirmed facts — verifiable through a public trace independent of the competitor's will. A new packaging unit declared in a public register; a listing obtained with a buying group, published by the group itself; a range withdrawn from an online catalogue, observed at two dates.
· 11 announcements — what a competitor chose to make known. I keep them, filed as announcements, with their date and the deadline announced. What a competitor announces says where it wants to be seen: that is useful, and it is not a fact.
What I add and what existed nowhere: for each fact, what it touches in your own offer. Competitor B's range withdrawal covers two lines that compete directly with your products 118 and 122, in use at 41 of your customers. That line is what interests your sales director, not the publication.
What I suggest: that these three facts go to the sales director today with the list of the 41 customers concerned, and that I prepare the win-back argument on the two lines. You decide whether to use it; I contact no customer. nine-competitors_what-is-confirmed.pdf3 facts, 11 announcements, 41 customers concerned
⛓ Sourced · public registers, buying group publications, online catalogues observed at two dates, CRM for the customer match
What I measured over your last twelve months: of 23 competitor moves your salespeople eventually reported, 19 were already public when they found out, on average for 36 days. The information was not hidden: nobody was reading it.
What I bring you this week on your 9 competitors, and the distinction that structures everything:
· 3 confirmed facts — verifiable through a public trace independent of the competitor's will. A new packaging unit declared in a public register; a listing obtained with a buying group, published by the group itself; a range withdrawn from an online catalogue, observed at two dates.
· 11 announcements — what a competitor chose to make known. I keep them, filed as announcements, with their date and the deadline announced. What a competitor announces says where it wants to be seen: that is useful, and it is not a fact.
What I add and what existed nowhere: for each fact, what it touches in your own offer. Competitor B's range withdrawal covers two lines that compete directly with your products 118 and 122, in use at 41 of your customers. That line is what interests your sales director, not the publication.
What I suggest: that these three facts go to the sales director today with the list of the 41 customers concerned, and that I prepare the win-back argument on the two lines. You decide whether to use it; I contact no customer. nine-competitors_what-is-confirmed.pdf3 facts, 11 announcements, 41 customers concerned
⛓ Sourced · public registers, buying group publications, online catalogues observed at two dates, CRM for the customer match
How do you know a fact is real and not a dressed-up announcement?
By the nature of the trace, not by the tone of the publication. And I measure what each of your competitors' announcements is worth.
The four sources that produce a confirmable fact, because they exist independently of what a competitor wants known:
· Official publications and registers — new establishments, filed accounts, published notices.
· Third-party publications that commit the third party — a buying group announcing its own listing.
· Public catalogues and price lists, observed at two dates. A gap between two observations is a fact; a catalogue read once is only a snapshot.
· Job adverts, published to be read by candidates and not by you: they describe what is actually being built. They flagged two new units before any press release this quarter.
The measure worth more than the announcements themselves: I go back over past announcements and look at what became of them. Of the 14 announcements from the previous quarter: 6 materialised within the announced timeframe, 3 late, 5 produced nothing observable. Competitor by competitor, this tells you what each one's word is worth — competitor D delivers on 4 announcements out of 5, competitor G on 1 out of 4. No single announcement can tell you that.
What I give you instead of the word bluff: a rate, competitor by competitor. An abandoned project and a project announced too early are indistinguishable from outside — and nobody needs to rule: the materialisation rate says the same thing, better, because it is a figure and it can be quoted in a meeting.
What I suggest: one sheet per competitor, kept continuously, with its materialisation rate and its average lag between announcement and fact. Your sales director walks into a meeting knowing whether the announcement the customer has just quoted has a one-in-five or a four-in-five chance of existing. announcements-and-facts_materialisation-rate.pdf6 of 14 delivered, competitor by competitor
⛓ Sourced · public registers, catalogues observed at two dates, published job adverts, 14 announcements from the previous quarter
The four sources that produce a confirmable fact, because they exist independently of what a competitor wants known:
· Official publications and registers — new establishments, filed accounts, published notices.
· Third-party publications that commit the third party — a buying group announcing its own listing.
· Public catalogues and price lists, observed at two dates. A gap between two observations is a fact; a catalogue read once is only a snapshot.
· Job adverts, published to be read by candidates and not by you: they describe what is actually being built. They flagged two new units before any press release this quarter.
The measure worth more than the announcements themselves: I go back over past announcements and look at what became of them. Of the 14 announcements from the previous quarter: 6 materialised within the announced timeframe, 3 late, 5 produced nothing observable. Competitor by competitor, this tells you what each one's word is worth — competitor D delivers on 4 announcements out of 5, competitor G on 1 out of 4. No single announcement can tell you that.
What I give you instead of the word bluff: a rate, competitor by competitor. An abandoned project and a project announced too early are indistinguishable from outside — and nobody needs to rule: the materialisation rate says the same thing, better, because it is a figure and it can be quoted in a meeting.
What I suggest: one sheet per competitor, kept continuously, with its materialisation rate and its average lag between announcement and fact. Your sales director walks into a meeting knowing whether the announcement the customer has just quoted has a one-in-five or a four-in-five chance of existing. announcements-and-facts_materialisation-rate.pdf6 of 14 delivered, competitor by competitor
⛓ Sourced · public registers, catalogues observed at two dates, published job adverts, 14 announcements from the previous quarter
We'd like a comparison of their listed prices. Some of it sits behind a customer login we don't have.
The comparison covers 156 lines across the 9 competitors, every price dated — and for what sits behind a login, I obtained the same information another way.
What I collected, and how:
· 112 prices from the competitors' public catalogues and price lists, each with the date of collection — an undated price cannot be compared, it misleads.
· 29 prices from awarded public tender documents, which publish the accepted prices. An official source, free of charge, and nobody uses it.
· 15 prices from the catalogues of distributors reselling those competitors, with the distributor margin removed using your own terms — I give you the calculation, not only the result.
On the 4 customer-only areas: I do not go in, and that is not caution — access there is conditional on a contract Norélia has not signed. The lawful route existed and it gave more: for three of those four competitors, their prices appear in awarded public tenders, therefore published, and often more recent than their commercial price list. The fourth stays unknown: I write “unknown” in the table rather than estimate it.
What the comparison says about you: across 156 comparable lines you are dearer on 34, of which 11 by more than 15 %. All 11 sit in a single range, the food-grade degreasers — and that is also the range with your lowest contract renewal rate. The connection was visible in neither of your two systems, because price sits on one side and renewal on the other.
What I suggest: this table to your sales management, with the 11 lines at the top and every collection date shown, and a quarterly refresh to track movements. I recommend no price: setting a tariff is a decision, and it is entirely yours. price-comparison_156-dated-lines.pdf11 gaps above 15 %, all in one range
⛓ Sourced · dated public catalogues and price lists, awarded public tenders, distributor catalogues, renewal rate by range (CRM)
What I collected, and how:
· 112 prices from the competitors' public catalogues and price lists, each with the date of collection — an undated price cannot be compared, it misleads.
· 29 prices from awarded public tender documents, which publish the accepted prices. An official source, free of charge, and nobody uses it.
· 15 prices from the catalogues of distributors reselling those competitors, with the distributor margin removed using your own terms — I give you the calculation, not only the result.
On the 4 customer-only areas: I do not go in, and that is not caution — access there is conditional on a contract Norélia has not signed. The lawful route existed and it gave more: for three of those four competitors, their prices appear in awarded public tenders, therefore published, and often more recent than their commercial price list. The fourth stays unknown: I write “unknown” in the table rather than estimate it.
What the comparison says about you: across 156 comparable lines you are dearer on 34, of which 11 by more than 15 %. All 11 sit in a single range, the food-grade degreasers — and that is also the range with your lowest contract renewal rate. The connection was visible in neither of your two systems, because price sits on one side and renewal on the other.
What I suggest: this table to your sales management, with the 11 lines at the top and every collection date shown, and a quarterly refresh to track movements. I recommend no price: setting a tariff is a decision, and it is entirely yours. price-comparison_156-dated-lines.pdf11 gaps above 15 %, all in one range
⛓ Sourced · dated public catalogues and price lists, awarded public tenders, distributor catalogues, renewal rate by range (CRM)
Two of our competitors carry a trading name that other companies carry too. How do you avoid mixing everything up?
I do not merge them, and when the publication does not settle it, the question goes to your quality manager with both records side by side.
What I hold as entities, each with its identifiers: your 9 competitors under 23 names — registered names, trading names, former names the press still uses —, your 640 customers, your 3 ranges and 210 product lines, 12 professional bodies and institutions in the sector, and the 14 press titles tracked. That table is what caught the two publications you put back into the summary: a competitor was named there under something other than its registered name.
The homonym case, as it came up this week: two companies carry the same trading name. One is among your 9 competitors; the other is a maintenance contractor with nothing to do with your market. Their registration numbers differ, and so do their addresses: they are two entities, and they stay two.
· When the publication carries an identifier — registration, address, a named director — the match is made, and it is written down. 11 publications this week.
· When it carries none — 6 publications this week — the item still comes through, marked “entity to be confirmed”, with the three clues that lean one way. The question goes to your quality manager, and she settles it in one line. A guessed match produces a false history that looks clean: the worst possible outcome for a monitoring service.
The taxonomy sits above the entities: 3 angles, your 11 declared subjects, and 47 labels — ranges, product families, types of player, kinds of text. Every item kept carries its entity and its labels, and that is precisely what makes the memory searchable a year later.
It belongs to you, and it moves: I proposed 9 labels this quarter, drawn from what kept coming back in the items kept; you accepted 7 and merged 2. The proposal was costed — for each label, the number of items it would have filed over twelve weeks — and the call took ten minutes.
What I suggest: a review of the entity table every time a change of name is observed. Two of your competitors changed theirs this quarter, and without that review their publications would have run under two different entities for months.
⛓ Sourced · entity table — 9 competitors under 23 names, 640 customers, 210 product lines, 12 sector bodies; taxonomy of 3 angles, 11 subjects and 47 labels; 17 matches this week
What I hold as entities, each with its identifiers: your 9 competitors under 23 names — registered names, trading names, former names the press still uses —, your 640 customers, your 3 ranges and 210 product lines, 12 professional bodies and institutions in the sector, and the 14 press titles tracked. That table is what caught the two publications you put back into the summary: a competitor was named there under something other than its registered name.
The homonym case, as it came up this week: two companies carry the same trading name. One is among your 9 competitors; the other is a maintenance contractor with nothing to do with your market. Their registration numbers differ, and so do their addresses: they are two entities, and they stay two.
· When the publication carries an identifier — registration, address, a named director — the match is made, and it is written down. 11 publications this week.
· When it carries none — 6 publications this week — the item still comes through, marked “entity to be confirmed”, with the three clues that lean one way. The question goes to your quality manager, and she settles it in one line. A guessed match produces a false history that looks clean: the worst possible outcome for a monitoring service.
The taxonomy sits above the entities: 3 angles, your 11 declared subjects, and 47 labels — ranges, product families, types of player, kinds of text. Every item kept carries its entity and its labels, and that is precisely what makes the memory searchable a year later.
It belongs to you, and it moves: I proposed 9 labels this quarter, drawn from what kept coming back in the items kept; you accepted 7 and merged 2. The proposal was costed — for each label, the number of items it would have filed over twelve weeks — and the call took ten minutes.
What I suggest: a review of the entity table every time a change of name is observed. Two of your competitors changed theirs this quarter, and without that review their publications would have run under two different entities for months.
⛓ Sourced · entity table — 9 competitors under 23 names, 640 customers, 210 product lines, 12 sector bodies; taxonomy of 3 angles, 11 subjects and 47 labels; 17 matches this week
Last year a text we were actually tracking slipped past us until twelve days before it took effect. How do we avoid that?
By tracking effective dates rather than publication dates. That is the only change, and it accounts for the twelve days on its own.
The effective date is the date on which a text actually applies. It is distinct from its publication date, and can be years away from it.
What happened, and it was not negligence: the text was published, flagged the same day by your monitoring, read, filed. Its effective date fell fourteen months later. In between, nothing brought it back — monitoring built on feeds treats a text like a news item, whereas a text is an appointment.
What I now do with every text in your declared scope: I extract all of its dates — entry into force, end of transitional period, compliance deadline, first reporting date — the transitional period being the window during which the old and new regimes coexist — and I place them in a calendar. The text comes back at every date, not at its publication.
The state of your portfolio today: 7 tracked texts have at least one effective date ahead, 3 of them within six months. Only one had been reopened since publication. The three nearest, designated here by their subject matter: labelling of professional-use mixtures (end of transitional period in 4 months), declaration of the composition of products placed on the market (first filing in 5 months), discharge thresholds applying to your food-industry customers (entry into force in 6 months) — that last one does not target you, it targets your customers, and that is exactly why it matters.
The reminders: at six months, one month and one week, to the quality and regulatory manager. Three reminders beat one perfect reminder: the first is for planning, the last for checking.
What I suggest: that you read this calendar once and tell me which texts are missing. Your declared scope covers 11 subjects; I cover 11, but you are the one who knows whether a twelfth exists. calendar-of-effective-dates.pdf7 texts, 3 deadlines within six months
⛓ Sourced · 7 texts in the declared scope, effective dates extracted from official publications, last year's flagging history
The effective date is the date on which a text actually applies. It is distinct from its publication date, and can be years away from it.
What happened, and it was not negligence: the text was published, flagged the same day by your monitoring, read, filed. Its effective date fell fourteen months later. In between, nothing brought it back — monitoring built on feeds treats a text like a news item, whereas a text is an appointment.
What I now do with every text in your declared scope: I extract all of its dates — entry into force, end of transitional period, compliance deadline, first reporting date — the transitional period being the window during which the old and new regimes coexist — and I place them in a calendar. The text comes back at every date, not at its publication.
The state of your portfolio today: 7 tracked texts have at least one effective date ahead, 3 of them within six months. Only one had been reopened since publication. The three nearest, designated here by their subject matter: labelling of professional-use mixtures (end of transitional period in 4 months), declaration of the composition of products placed on the market (first filing in 5 months), discharge thresholds applying to your food-industry customers (entry into force in 6 months) — that last one does not target you, it targets your customers, and that is exactly why it matters.
The reminders: at six months, one month and one week, to the quality and regulatory manager. Three reminders beat one perfect reminder: the first is for planning, the last for checking.
What I suggest: that you read this calendar once and tell me which texts are missing. Your declared scope covers 11 subjects; I cover 11, but you are the one who knows whether a twelfth exists. calendar-of-effective-dates.pdf7 texts, 3 deadlines within six months
⛓ Sourced · 7 texts in the declared scope, effective dates extracted from official publications, last year's flagging history
Fine, but concretely, what does it change here?
I tell you exactly where each text lands inside your organisation — that is the bulk of the work, and the part that takes longest by hand.
What I provide on every text: its subject matter, its publication date, its effective dates, what it targets, and the exact passage restated, quoted rather than paraphrased.
What I add now that you have opened the document intranet, and it existed nowhere: the list of your own documents the text touches, each with its date. Three examples from the labelling text:
· Your labelling procedure P-12, dated 2023, applies the old wording format. It is cited in 4 other documents that will follow it.
· 38 product data sheets out of 210 carry the wording concerned. I give you the list, line by line, not an estimate.
· Two of your three framework contracts refer to that procedure by number: changing the procedure without looking at the contracts creates a gap between what you do and what you signed.
The time this replaces: your outside counsel billed this reconciliation at around 6 hours per text, most of it spent working out which internal documents were affected. They now receive the list, made and dated, and work on the question you pay them for: the scope.
What I do not conclude: that you are compliant, or that you are not. Compliance is assessed on your actual situation, your derogations and your live contracts — and the analysis of scope belongs to your counsel, as your organisation intends. This is not a limitation of the agent, it is the split that makes its work usable: an opinion produced by a machine would be taken up verbatim in committee without anyone rereading the reasoning.
What I suggest: that the file goes to counsel today, with the 38 data sheets listed and the two framework contracts flagged. On the two previous texts, the opinion came back in 4 days instead of 3 weeks, for a third of the cost. where-the-text-lands_file-for-counsel.pdf38 data sheets and 2 framework contracts identified
⛓ Sourced · document intranet, procedure P-12 dated 2023, 210 product data sheets, 3 framework contracts, counsel's invoices
What I provide on every text: its subject matter, its publication date, its effective dates, what it targets, and the exact passage restated, quoted rather than paraphrased.
What I add now that you have opened the document intranet, and it existed nowhere: the list of your own documents the text touches, each with its date. Three examples from the labelling text:
· Your labelling procedure P-12, dated 2023, applies the old wording format. It is cited in 4 other documents that will follow it.
· 38 product data sheets out of 210 carry the wording concerned. I give you the list, line by line, not an estimate.
· Two of your three framework contracts refer to that procedure by number: changing the procedure without looking at the contracts creates a gap between what you do and what you signed.
The time this replaces: your outside counsel billed this reconciliation at around 6 hours per text, most of it spent working out which internal documents were affected. They now receive the list, made and dated, and work on the question you pay them for: the scope.
What I do not conclude: that you are compliant, or that you are not. Compliance is assessed on your actual situation, your derogations and your live contracts — and the analysis of scope belongs to your counsel, as your organisation intends. This is not a limitation of the agent, it is the split that makes its work usable: an opinion produced by a machine would be taken up verbatim in committee without anyone rereading the reasoning.
What I suggest: that the file goes to counsel today, with the 38 data sheets listed and the two framework contracts flagged. On the two previous texts, the opinion came back in 4 days instead of 3 weeks, for a third of the cost. where-the-text-lands_file-for-counsel.pdf38 data sheets and 2 framework contracts identified
⛓ Sourced · document intranet, procedure P-12 dated 2023, 210 product data sheets, 3 framework contracts, counsel's invoices
The executive committee meets on Thursday and will want a position, not a file. How do we handle that?
I prepare it — a position paper is work I do if you ask for it, and you have just asked.
What the paper contains, ready for Thursday:
· The fact, in three lines: what the text changes, on what date, over what scope.
· What is certain and what is not, separated by a rule. Certain: the dates, the quoted passage, your 38 affected data sheets. Not certain: how the threshold reads in your particular case.
· Three costed options, each with its cost, its lead time and the risk accepted: reword all 38 data sheets before the deadline (6 weeks, internal workload) · rework only the 12 sheets of active lines (2 weeks, 26 lines handled as they come) · ask counsel for a formal position before committing anything (4 days, known cost).
· The exact point on which counsel must rule, phrased as a closed question. A precise question comes back in four days; an open one comes back in three weeks.
· The date of the next review, placed in the calendar.
What stays with management, and I say so plainly: the choice between the three options. This is not evasion: weighing 6 weeks of internal workload against 26 lines handled as they come depends on your workload and your appetite for risk, neither of which I know better than you. What I have taken off you is the groundwork: the three options are costed, and the question for counsel is already written as one closed sentence.
What the paper has already given you: on the two previous texts, the committee decided in the meeting instead of deferring. A deferred item is a month lost against a deadline that does not defer.
What I suggest next: that Thursday's decision be entered in the calendar with its owner and its check date, and that I bring it back to you when it falls due. A decision nobody revisits is a decision you will take twice. position-paper_three-costed-options.pdfReady for Thursday, the decision stays with the committee
✎ Framework · position paper template, history of the two previous texts, quality department workload plan
What the paper contains, ready for Thursday:
· The fact, in three lines: what the text changes, on what date, over what scope.
· What is certain and what is not, separated by a rule. Certain: the dates, the quoted passage, your 38 affected data sheets. Not certain: how the threshold reads in your particular case.
· Three costed options, each with its cost, its lead time and the risk accepted: reword all 38 data sheets before the deadline (6 weeks, internal workload) · rework only the 12 sheets of active lines (2 weeks, 26 lines handled as they come) · ask counsel for a formal position before committing anything (4 days, known cost).
· The exact point on which counsel must rule, phrased as a closed question. A precise question comes back in four days; an open one comes back in three weeks.
· The date of the next review, placed in the calendar.
What stays with management, and I say so plainly: the choice between the three options. This is not evasion: weighing 6 weeks of internal workload against 26 lines handled as they come depends on your workload and your appetite for risk, neither of which I know better than you. What I have taken off you is the groundwork: the three options are costed, and the question for counsel is already written as one closed sentence.
What the paper has already given you: on the two previous texts, the committee decided in the meeting instead of deferring. A deferred item is a month lost against a deadline that does not defer.
What I suggest next: that Thursday's decision be entered in the calendar with its owner and its check date, and that I bring it back to you when it falls due. A decision nobody revisits is a decision you will take twice. position-paper_three-costed-options.pdfReady for Thursday, the decision stays with the committee
✎ Framework · position paper template, history of the two previous texts, quality department workload plan
Does the calendar live on its own? By the time a deadline comes round, we will have forgotten why it is there.
The calendar and the memory are the same object seen at two dates: every deadline carries the monitoring item that put it there, its dated source, and everything seen on the subject since.
The state of the calendar today: your 7 tracked texts carry 19 effective dates — a text almost never carries just one: entry into force, end of transitional period, compliance deadline, first report. 3 fall within six months, and each is reminded at six months, one month and one week.
What a deadline carries when you open it:
· The quoted passage that founds it, with the dated version it was taken from.
· The monitoring item that put it there, its source and its date of first publication.
· Everything seen since on the same subject, in order.
· The internal documents concerned, and the decisions already taken, with their dates.
What the memory gave on the labelling text, and it is what served counsel: 23 items over sixteen months, from the first published draft to the final version. Comparing the versions along that chain established that the disputed wording had not moved between the last two — nobody could have established that from memory, and it saved counsel a question.
Decisions enter the calendar just as texts do: Thursday's will, with its owner and its check date. And a past deadline does not disappear: it moves into the memory with what was done — or with the fact that nothing was. Two cases this quarter, both written down, and that is what made it possible to reopen one of them.
What I suggest: a reread of the calendar once a quarter, starting with the most distant deadlines. Those are the ones that get forgotten: the near ones, everybody sees.
⛓ Sourced · calendar — 7 texts, 19 effective dates, 3 within six months; memory of the labelling subject — 23 items over 16 months, versions compared
The state of the calendar today: your 7 tracked texts carry 19 effective dates — a text almost never carries just one: entry into force, end of transitional period, compliance deadline, first report. 3 fall within six months, and each is reminded at six months, one month and one week.
What a deadline carries when you open it:
· The quoted passage that founds it, with the dated version it was taken from.
· The monitoring item that put it there, its source and its date of first publication.
· Everything seen since on the same subject, in order.
· The internal documents concerned, and the decisions already taken, with their dates.
What the memory gave on the labelling text, and it is what served counsel: 23 items over sixteen months, from the first published draft to the final version. Comparing the versions along that chain established that the disputed wording had not moved between the last two — nobody could have established that from memory, and it saved counsel a question.
Decisions enter the calendar just as texts do: Thursday's will, with its owner and its check date. And a past deadline does not disappear: it moves into the memory with what was done — or with the fact that nothing was. Two cases this quarter, both written down, and that is what made it possible to reopen one of them.
What I suggest: a reread of the calendar once a quarter, starting with the most distant deadlines. Those are the ones that get forgotten: the near ones, everybody sees.
⛓ Sourced · calendar — 7 texts, 19 effective dates, 3 within six months; memory of the labelling subject — 23 items over 16 months, versions compared
We'd like to know what our sector's press is saying, and what it says about us. Nobody has time to read it.
Fourteen titles tracked, 312 items this month, 12 kept and 5 mentions of Norélia — including one you would not have seen.
The 5 mentions, by nature, because they are not handled the same way:
· 2 expert citations: your product manager quoted on disinfection in the food industry. Syndicated 7 times by other titles — syndication is kept out of the summary but counted, because it measures reach.
· 1 mention in a market overview, where you appear among five manufacturers. You are the only one whose production capacity is not stated: the journalist could not find it, and it is absent from your website.
· 1 indirect mention: one of your customers cites your range in an article about its own practices. That is the one you would not have seen, because your name is not in the headline — it is also the most useful in a sales conversation.
· 1 critical mention, on a specific point in one range. I will come to it, it deserves its own exchange.
What I bring with each one: the title, the date, the author as signed, the link to the item, and three quoted lines at most — not a summary.
What I discard, with the reason: 300 items this month, of which 198 identical wire syndication, 74 outside scope, 28 promotional with no fact.
The time this shifts: reading those 14 titles meant 4 hours a week that nobody took. The review is done, and your communication officer receives 12 items instead of 312.
What I suggest: add your production capacity to your press kit. You were the only one of the five manufacturers not to supply it, and the same figure will be asked for at the next overview. press-review_5-qualified-mentions.pdf312 items, 12 kept, 198 syndications counted
⛓ Sourced · 14 tracked titles, 312 items this month, syndication log, published market overview
The 5 mentions, by nature, because they are not handled the same way:
· 2 expert citations: your product manager quoted on disinfection in the food industry. Syndicated 7 times by other titles — syndication is kept out of the summary but counted, because it measures reach.
· 1 mention in a market overview, where you appear among five manufacturers. You are the only one whose production capacity is not stated: the journalist could not find it, and it is absent from your website.
· 1 indirect mention: one of your customers cites your range in an article about its own practices. That is the one you would not have seen, because your name is not in the headline — it is also the most useful in a sales conversation.
· 1 critical mention, on a specific point in one range. I will come to it, it deserves its own exchange.
What I bring with each one: the title, the date, the author as signed, the link to the item, and three quoted lines at most — not a summary.
What I discard, with the reason: 300 items this month, of which 198 identical wire syndication, 74 outside scope, 28 promotional with no fact.
The time this shifts: reading those 14 titles meant 4 hours a week that nobody took. The review is done, and your communication officer receives 12 items instead of 312.
What I suggest: add your production capacity to your press kit. You were the only one of the five manufacturers not to supply it, and the same figure will be asked for at the next overview. press-review_5-qualified-mentions.pdf312 items, 12 kept, 198 syndications counted
⛓ Sourced · 14 tracked titles, 312 items this month, syndication log, published market overview
The critical mention — what do we do with it? We tend to do nothing, because we can't tell whether it's an isolated case.
I pass it to you in full, unsoftened, and I tell you whether it is isolated — which is precisely the question monitoring can answer.
The mention: a dated, signed, reasoned article taking issue with one specific point in your food-grade degreaser range — the contact time needed to achieve the stated efficacy. It quotes a named user.
What you receive, and it is deliberate: the full article, by its link. A summary of a criticism is always gentler than the criticism — not out of tact, but because summarising removes the examples, and a criticism's force lies in its examples.
What I add, and it answers your real question: does this point appear elsewhere? Yes: two converging mentions over twelve months, in two unrelated titles, plus four reports from your own customer service on the same subject, which I found in the CRM. An isolated criticism is an opinion; six occurrences across two independent channels are a signal.
And the connection nobody had made: those six occurrences concern the same range as the 11 lines where you are dearer by more than 15 %, and the same range with the lowest renewal rate. Three distinct signals point at the same product. This is no longer a press mention, it is a board matter.
What I suggest, and the decision stays entirely yours: the full file to the communication officer and the two product managers — full article, six dated occurrences, customer service history, price comparison. I draft no public response and publish none: for a statement or a response article, I hand over to the copywriting agent, with this file as its material. You approve, and it is your name that signs. a-signal-with-six-occurrences.pdfPress, customer service and price point at the same product
⛓ Sourced · full dated and signed article, 2 converging mentions over 12 months, 4 customer service reports (CRM), price comparison
The mention: a dated, signed, reasoned article taking issue with one specific point in your food-grade degreaser range — the contact time needed to achieve the stated efficacy. It quotes a named user.
What you receive, and it is deliberate: the full article, by its link. A summary of a criticism is always gentler than the criticism — not out of tact, but because summarising removes the examples, and a criticism's force lies in its examples.
What I add, and it answers your real question: does this point appear elsewhere? Yes: two converging mentions over twelve months, in two unrelated titles, plus four reports from your own customer service on the same subject, which I found in the CRM. An isolated criticism is an opinion; six occurrences across two independent channels are a signal.
And the connection nobody had made: those six occurrences concern the same range as the 11 lines where you are dearer by more than 15 %, and the same range with the lowest renewal rate. Three distinct signals point at the same product. This is no longer a press mention, it is a board matter.
What I suggest, and the decision stays entirely yours: the full file to the communication officer and the two product managers — full article, six dated occurrences, customer service history, price comparison. I draft no public response and publish none: for a statement or a response article, I hand over to the copywriting agent, with this file as its material. You approve, and it is your name that signs. a-signal-with-six-occurrences.pdfPress, customer service and price point at the same product
⛓ Sourced · full dated and signed article, 2 converging mentions over 12 months, 4 customer service reports (CRM), price comparison
And to feed our blog and our customer newsletter? We have the material in front of us but we publish nothing.
Four subjects are framed, drawn from your own monitoring, each with the evidence that somebody cares. The writing goes to the agent that writes.
The four, and what justifies them:
· Contact time for food-grade degreasers. Six occurrences in twelve months: this is the question your market is already asking. Answering it with a public technical note beats waiting for the seventh.
· The new labelling requirement, explained to your customers. It takes effect in four months and concerns your 640 customers as much as it concerns you. None of your 9 competitors has published on it — I checked across their 41 publications this quarter.
· Discharge thresholds in the food industry. This text does not target you, it targets your customers: it is the subject that positions you as an adviser rather than a supplier.
· The market overview you are cited in. A commented follow-up, including the production capacity the journalist could not find.
What I supply for each: the angle, the dated sources, the three points to cover, what must not be written for want of a source, and the intended readership. This is not a draft article, it is a writing brief — and that is what is always missing when you want to publish.
The hand-over: the copywriting agent takes the brief and produces the text; the communication officer approves; nothing is published without her. I stay on my own job: I report, I source, I date. I do not write what will carry your signature.
What I suggest next: that I bring you, every month, the subjects the market is talking about and on which you have published nothing. This month there are four. It is the shortest and most profitable list a monitoring service produces. four-subjects_writing-briefs.pdfAngle, dated sources and what must not be written
⛓ Sourced · 6 occurrences over 12 months, regulatory calendar, competitors' 41 publications this quarter, market overview
The four, and what justifies them:
· Contact time for food-grade degreasers. Six occurrences in twelve months: this is the question your market is already asking. Answering it with a public technical note beats waiting for the seventh.
· The new labelling requirement, explained to your customers. It takes effect in four months and concerns your 640 customers as much as it concerns you. None of your 9 competitors has published on it — I checked across their 41 publications this quarter.
· Discharge thresholds in the food industry. This text does not target you, it targets your customers: it is the subject that positions you as an adviser rather than a supplier.
· The market overview you are cited in. A commented follow-up, including the production capacity the journalist could not find.
What I supply for each: the angle, the dated sources, the three points to cover, what must not be written for want of a source, and the intended readership. This is not a draft article, it is a writing brief — and that is what is always missing when you want to publish.
The hand-over: the copywriting agent takes the brief and produces the text; the communication officer approves; nothing is published without her. I stay on my own job: I report, I source, I date. I do not write what will carry your signature.
What I suggest next: that I bring you, every month, the subjects the market is talking about and on which you have published nothing. This month there are four. It is the shortest and most profitable list a monitoring service produces. four-subjects_writing-briefs.pdfAngle, dated sources and what must not be written
⛓ Sourced · 6 occurrences over 12 months, regulatory calendar, competitors' 41 publications this quarter, market overview
In all of that, how do we tell a fact from a press release and from an opinion? Our salespeople quote all three the same way.
All three are kept, and they are filed apart: a fact, an announcement and an opinion are not quoted the same way in a meeting.
The week's 31 items kept, by register:
· 12 confirmed facts — 3 on the competitive angle, 5 on the regulatory angle — a published text is a fact, with its date — and 4 in the press, including the indirect mention by one of your customers and the market overview.
· 14 announcements — 11 from your competitors, 3 in the trade press.
· 5 opinions — 1 signed critical mention, 2 signed columns by a sector executive, 1 expert view in an overview, 1 comment from a named user.
What files an item under one or the other, and it is not the tone:
· A fact has a trace that exists independently of whoever it concerns — a public register, an official publication, a third party committing itself, an observation made at two dates.
· An announcement is what a player chose to make known. It stays an announcement as long as no trace materialises it, and age does not promote it: of the 14 announcements from the previous quarter, 5 have still produced no trace, and they are still filed as announcements.
· An opinion is signed by a person and it commits that person. It is kept with its author and its date. An opinion is never counted as a fact, and it is never discarded for that reason either: it is by aggregating opinions that you see a subject coming.
What that separation has already given you, in one line: the critical mention is an opinion — one. The six occurrences that make it a signal are not all opinions: two are published opinions, in two unrelated titles; four are facts — the dated reports from your own customer service, found in the CRM. It is because they are not of the same register that they corroborate rather than repeat each other.
What I suggest: that the summary carry the register at the head of every line, in one word. Your salespeople walk into a meeting knowing whether they are quoting a public trace, a stated intention or one person's view — and that distinction is what holds them in front of a customer better informed than they are.
⛓ Sourced · the week's 31 items kept, split into 12 confirmed facts, 14 announcements and 5 opinions; 14 announcements from the previous quarter tracked to their trace
The week's 31 items kept, by register:
· 12 confirmed facts — 3 on the competitive angle, 5 on the regulatory angle — a published text is a fact, with its date — and 4 in the press, including the indirect mention by one of your customers and the market overview.
· 14 announcements — 11 from your competitors, 3 in the trade press.
· 5 opinions — 1 signed critical mention, 2 signed columns by a sector executive, 1 expert view in an overview, 1 comment from a named user.
What files an item under one or the other, and it is not the tone:
· A fact has a trace that exists independently of whoever it concerns — a public register, an official publication, a third party committing itself, an observation made at two dates.
· An announcement is what a player chose to make known. It stays an announcement as long as no trace materialises it, and age does not promote it: of the 14 announcements from the previous quarter, 5 have still produced no trace, and they are still filed as announcements.
· An opinion is signed by a person and it commits that person. It is kept with its author and its date. An opinion is never counted as a fact, and it is never discarded for that reason either: it is by aggregating opinions that you see a subject coming.
What that separation has already given you, in one line: the critical mention is an opinion — one. The six occurrences that make it a signal are not all opinions: two are published opinions, in two unrelated titles; four are facts — the dated reports from your own customer service, found in the CRM. It is because they are not of the same register that they corroborate rather than repeat each other.
What I suggest: that the summary carry the register at the head of every line, in one word. Your salespeople walk into a meeting knowing whether they are quoting a public trace, a stated intention or one person's view — and that distinction is what holds them in front of a customer better informed than they are.
⛓ Sourced · the week's 31 items kept, split into 12 confirmed facts, 14 announcements and 5 opinions; 14 announcements from the previous quarter tracked to their trace
After a quarter, what has it actually brought us?
Monitoring that exists every single week, and a measurement of what it sees — including what it does not.
On the three measures you tracked:
· Reviewing the sources: 50 % → 12 % of monitoring time.
· Writing the summary: 35 % → 12 %.
· Weeding out repetition: 25 % → 6 %.
What that produced, and it is the figure that counts: your weekly summary went from 19 weeks out of 52 to 13 out of 13 this quarter. Regularity was the whole subject: irregular monitoring does not miss information at random, it misses the busy weeks' information — that is, the most important.
What was decided thanks to it: three competitor facts passed on in two days instead of five weeks · two regulatory deadlines handled six months ahead · a signal with six occurrences taken to the board · 11 price gaps documented.
The figure that does not flatter me, and you should have it: my measured coverage is 87 %. I reconstructed after the fact what should have entered the declared scope this quarter: of 1,490 publications that concerned it, I saw 1,296. I missed 194.
Where those 194 come from, because a figure without a cause cannot be corrected:
· 131 from a site with no syndication feed — the stream a site publishes to announce its new content, which monitoring tools read — and whose page layout changed mid-quarter. Fixed three weeks ago.
· 44 from a trade newsletter distributed by email only, to which nobody was subscribed. Subscribe a shared address and I treat it as a source.
· 19 from a title that moved behind a paid subscription. Two routes: your existing subscription, which I can read if you give me access, or the public contents page, which gives titles and dates without the body. I circumvent nothing, and I leave you without a solution nowhere.
No automated monitoring can guarantee exhaustiveness, and yours was no exception before me — the difference is that 87 % is a known figure, explained and rising, where an unread spreadsheet cannot be measured at all. With the three fixes I am aiming at 96 % next quarter, and I will measure it the same way. quarter-review_measured-coverage.pdf87 % seen, 194 missed explained and corrected
⛓ Sourced · quarter's collection log, after-the-fact reconstruction of the declared scope, 13 summaries produced
On the three measures you tracked:
· Reviewing the sources: 50 % → 12 % of monitoring time.
· Writing the summary: 35 % → 12 %.
· Weeding out repetition: 25 % → 6 %.
What that produced, and it is the figure that counts: your weekly summary went from 19 weeks out of 52 to 13 out of 13 this quarter. Regularity was the whole subject: irregular monitoring does not miss information at random, it misses the busy weeks' information — that is, the most important.
What was decided thanks to it: three competitor facts passed on in two days instead of five weeks · two regulatory deadlines handled six months ahead · a signal with six occurrences taken to the board · 11 price gaps documented.
The figure that does not flatter me, and you should have it: my measured coverage is 87 %. I reconstructed after the fact what should have entered the declared scope this quarter: of 1,490 publications that concerned it, I saw 1,296. I missed 194.
Where those 194 come from, because a figure without a cause cannot be corrected:
· 131 from a site with no syndication feed — the stream a site publishes to announce its new content, which monitoring tools read — and whose page layout changed mid-quarter. Fixed three weeks ago.
· 44 from a trade newsletter distributed by email only, to which nobody was subscribed. Subscribe a shared address and I treat it as a source.
· 19 from a title that moved behind a paid subscription. Two routes: your existing subscription, which I can read if you give me access, or the public contents page, which gives titles and dates without the body. I circumvent nothing, and I leave you without a solution nowhere.
No automated monitoring can guarantee exhaustiveness, and yours was no exception before me — the difference is that 87 % is a known figure, explained and rising, where an unread spreadsheet cannot be measured at all. With the three fixes I am aiming at 96 % next quarter, and I will measure it the same way. quarter-review_measured-coverage.pdf87 % seen, 194 missed explained and corrected
⛓ Sourced · quarter's collection log, after-the-fact reconstruction of the declared scope, 13 summaries produced
We lose the thread of past monitoring completely. When a subject comes back, nobody can find what we saw a year ago.
Everything that has passed through the monitoring is kept and searchable in one sentence — and the document agent answers on it as it does on the rest of your documents.
What the memory holds today: 1,412 items kept over three years, the first twelve months carried over from your old spreadsheet, each with its source, publication date, collection date, scope and recipient. And the discard reasons, which serve more often than one would think: they say why a subject never came through.
What that gives in practice: “What had we seen about degreaser contact time?” Six occurrences, from March last year to this month, each with its link and the person it was sent to. The fact that the first goes back sixteen months is in itself the most useful information.
The connection with your internal documentation: the document agent searches your procedures, product data sheets and framework contracts in plain language. Monitoring says what comes from outside, the document base says what you wrote — and a question such as “has our labelling procedure been reviewed since the last text?” needs both. Each stays on its own source: I never cite the document agent as though it were a publication, and it does not cite me as though I were an internal document.
The time this replaces: finding an old subject in the spreadsheet took half a day when it was possible at all, and most often the question went unanswered. Your product managers asked 34 questions of that kind this quarter, all answered in under a minute.
What I suggest: a subject sheet generated on demand — everything seen on a theme, in order, with the sources. On degreasers it runs to two pages and it became the basis of the board paper. monitoring-memory_1412-items.pdfThree years searchable, sources and recipients included
⛓ Sourced · 1,412 items kept over 3 years, discard reason log, document intranet, questions asked this quarter
What the memory holds today: 1,412 items kept over three years, the first twelve months carried over from your old spreadsheet, each with its source, publication date, collection date, scope and recipient. And the discard reasons, which serve more often than one would think: they say why a subject never came through.
What that gives in practice: “What had we seen about degreaser contact time?” Six occurrences, from March last year to this month, each with its link and the person it was sent to. The fact that the first goes back sixteen months is in itself the most useful information.
The connection with your internal documentation: the document agent searches your procedures, product data sheets and framework contracts in plain language. Monitoring says what comes from outside, the document base says what you wrote — and a question such as “has our labelling procedure been reviewed since the last text?” needs both. Each stays on its own source: I never cite the document agent as though it were a publication, and it does not cite me as though I were an internal document.
The time this replaces: finding an old subject in the spreadsheet took half a day when it was possible at all, and most often the question went unanswered. Your product managers asked 34 questions of that kind this quarter, all answered in under a minute.
What I suggest: a subject sheet generated on demand — everything seen on a theme, in order, with the sources. On degreasers it runs to two pages and it became the basis of the board paper. monitoring-memory_1412-items.pdfThree years searchable, sources and recipients included
⛓ Sourced · 1,412 items kept over 3 years, discard reason log, document intranet, questions asked this quarter
One last thing: half our business comes through public tenders, and our monitoring scopes say exactly where we're heading. Can that get out?
No, and it is checkable: your monitoring scopes do not leave your premises. As for published tenders, I detected three this quarter that you had not seen.
First, what becomes of what you watch. I run on local inference on a Norélia server — the computation happens on your own machine: the list of the competitors you track and the regulatory subjects you follow has no outside network to cross. If you would rather not run a server, the other route is an isolated resource hosted in France, under French law, dedicated to Norélia — no pooling with another company in your sector.
Why this matters more here than elsewhere: your monitoring scopes are the map of your strategy. The competitors you follow, the range you are hunting signals on, the texts you are anticipating: entrusting that list to a consumer service is publishing your roadmap. Here it stays with you, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including against an American host based in Europe. Your monitoring subjects train no model and are passed to nobody.
And every summary is logged: who asked for what, what was kept, what was discarded and why, who approved, where it went. That log is what lets you show, in a bid, how your regulatory monitoring is run.
The three tenders detected this quarter — the tender agent takes over on that subject, but detection starts from your scopes: two supply tenders for hygiene products in local authorities, one framework agreement in the food industry. You answered one of the three and won it; the other two you saw four days before the deadline. With detection at publication, that is three weeks of preparation instead of four days.
The point that often decides: one of the three carried a “data hosted in France” criterion scored out of 5. You were able to document it in writing, with the technical framework — the kind of point that is won before the envelopes are opened.
What I suggest: that I bring you every tender as it is published, with its deadline, the buyer, and what your monitoring already knows about the competitors likely to bid. You decide whether to answer; I file no bid. technical-framework_where-your-scopes-live.pdfLocal inference, summary log, processing in the EU targeted
✎ Framework · deployment architecture, summary log, 3 published tenders this quarter, scored hosting criterion
First, what becomes of what you watch. I run on local inference on a Norélia server — the computation happens on your own machine: the list of the competitors you track and the regulatory subjects you follow has no outside network to cross. If you would rather not run a server, the other route is an isolated resource hosted in France, under French law, dedicated to Norélia — no pooling with another company in your sector.
Why this matters more here than elsewhere: your monitoring scopes are the map of your strategy. The competitors you follow, the range you are hunting signals on, the texts you are anticipating: entrusting that list to a consumer service is publishing your roadmap. Here it stays with you, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including against an American host based in Europe. Your monitoring subjects train no model and are passed to nobody.
And every summary is logged: who asked for what, what was kept, what was discarded and why, who approved, where it went. That log is what lets you show, in a bid, how your regulatory monitoring is run.
The three tenders detected this quarter — the tender agent takes over on that subject, but detection starts from your scopes: two supply tenders for hygiene products in local authorities, one framework agreement in the food industry. You answered one of the three and won it; the other two you saw four days before the deadline. With detection at publication, that is three weeks of preparation instead of four days.
The point that often decides: one of the three carried a “data hosted in France” criterion scored out of 5. You were able to document it in writing, with the technical framework — the kind of point that is won before the envelopes are opened.
What I suggest: that I bring you every tender as it is published, with its deadline, the buyer, and what your monitoring already knows about the competitors likely to bid. You decide whether to answer; I file no bid. technical-framework_where-your-scopes-live.pdfLocal inference, summary log, processing in the EU targeted
✎ Framework · deployment architecture, summary log, 3 published tenders this quarter, scored hosting criterion
Management will want a regular update. What do you produce, how often, and for whom?
Three distinct outputs, and they do not address the same people: the weekly summaries for those who act, the monthly reporting for you, the quarterly reporting for management.
The weekly summaries — 13 out of 13 this quarter, Sunday at 10 p.m., approved by you on Monday before 9 a.m., routed to four roles.
The monthly reporting, one page, and every figure unfolds down to the item and its dated source:
· The month's volumes — 8,592 publications seen, 124 kept, 8,468 discarded by reason.
· Measured coverage, with what was missed and why.
· The lags — 19 hours median between a fact's first publication and its arrival, 2 h 40 for immediate alerts.
· Signals open, signals closed, and their age.
· The subjects that produced nothing. Three of your eleven declared subjects produced not a single item kept in three months — that is information, not a blank. Checked: two are genuinely dormant, the third was missing a source. It is the trade newsletter distributed by email, and it accounts for 44 of the 194 publications I missed.
The quarterly reporting, for your board: what was decided thanks to the monitoring, what is waiting on a decision, and what was discarded and why. It runs to two pages and carries exactly the same figures as the monthly one — two documents that do not reconcile is a meeting lost comparing tables.
What the three outputs have in common: they carry the date of the rule that produced them. When the relevance threshold moves, the reporting line says so, and a figure that falls is not mistaken for a rule that changed.
What I suggest: that the quarterly reporting be read by the board before the validation review, not after. Tuning decisions are better taken when management has just seen what the previous tuning produced.
⛓ Sourced · 13 weekly summaries this quarter; monthly reporting — 8,592 publications seen, 124 kept; median lags read from the routing log
The weekly summaries — 13 out of 13 this quarter, Sunday at 10 p.m., approved by you on Monday before 9 a.m., routed to four roles.
The monthly reporting, one page, and every figure unfolds down to the item and its dated source:
· The month's volumes — 8,592 publications seen, 124 kept, 8,468 discarded by reason.
· Measured coverage, with what was missed and why.
· The lags — 19 hours median between a fact's first publication and its arrival, 2 h 40 for immediate alerts.
· Signals open, signals closed, and their age.
· The subjects that produced nothing. Three of your eleven declared subjects produced not a single item kept in three months — that is information, not a blank. Checked: two are genuinely dormant, the third was missing a source. It is the trade newsletter distributed by email, and it accounts for 44 of the 194 publications I missed.
The quarterly reporting, for your board: what was decided thanks to the monitoring, what is waiting on a decision, and what was discarded and why. It runs to two pages and carries exactly the same figures as the monthly one — two documents that do not reconcile is a meeting lost comparing tables.
What the three outputs have in common: they carry the date of the rule that produced them. When the relevance threshold moves, the reporting line says so, and a figure that falls is not mistaken for a rule that changed.
What I suggest: that the quarterly reporting be read by the board before the validation review, not after. Tuning decisions are better taken when management has just seen what the previous tuning produced.
⛓ Sourced · 13 weekly summaries this quarter; monthly reporting — 8,592 publications seen, 124 kept; median lags read from the routing log
And who checks that all this stays right? A year from now, a badly tuned rule will be invisible.
A validation review, every month, dated, whose decisions are written down — and four measurements that feed it, each with its own method.
The four measurements, and none is derived from another:
· The filtering error rate — 1.7 %, measured by 40 discarded publications drawn at random each month and handed back to be reread: 2 put back out of 120 in three months.
· Coverage — 87 %, measured by reconstructing the declared scope after the fact: 1,296 seen out of 1,490.
· The median lag — 19 hours, read from the log.
· Actual use — 3 immediate alerts a week, all opened the same day; 34 questions put to the memory this quarter, all answered. Monitoring that nobody opens is measured accurately and serves nothing: it is the only one of the four measurements that speaks about you rather than about me.
This quarter's review decisions, each dated and replayed before being applied: the near-duplicate threshold tightened on 12 March · the immediate-alert rule tightened on 20 March · two competitor names added to the entity table · three sources added to the register · seven labels accepted out of the nine proposed.
The rule I hold myself to, and it applies to all of them: a modified rule is replayed over twelve weeks of your history before it goes live, and the replay says two things — what it catches, and what it loses. A rule changed without being replayed is a rule you cannot judge, and you find out a quarter too late.
What stays in your hands, and it is not little: the scope, the threshold, the alert levels and the approval of every summary. I hand you every setting costed, with what it would have produced over twelve weeks; the call takes ten minutes, and it is written down.
What I suggest: that the next review be set today, with its agenda already written: the relevance threshold, the 7 silent sources, the twelfth candidate subject. A review that is not dated is a review that does not happen — and that is what had taken your summary from 52 weeks down to 19.
✎ Framework · monthly validation review — 4 measurements and their methods; decisions dated 12 and 20 March; replay over 12 weeks before anything goes live
The four measurements, and none is derived from another:
· The filtering error rate — 1.7 %, measured by 40 discarded publications drawn at random each month and handed back to be reread: 2 put back out of 120 in three months.
· Coverage — 87 %, measured by reconstructing the declared scope after the fact: 1,296 seen out of 1,490.
· The median lag — 19 hours, read from the log.
· Actual use — 3 immediate alerts a week, all opened the same day; 34 questions put to the memory this quarter, all answered. Monitoring that nobody opens is measured accurately and serves nothing: it is the only one of the four measurements that speaks about you rather than about me.
This quarter's review decisions, each dated and replayed before being applied: the near-duplicate threshold tightened on 12 March · the immediate-alert rule tightened on 20 March · two competitor names added to the entity table · three sources added to the register · seven labels accepted out of the nine proposed.
The rule I hold myself to, and it applies to all of them: a modified rule is replayed over twelve weeks of your history before it goes live, and the replay says two things — what it catches, and what it loses. A rule changed without being replayed is a rule you cannot judge, and you find out a quarter too late.
What stays in your hands, and it is not little: the scope, the threshold, the alert levels and the approval of every summary. I hand you every setting costed, with what it would have produced over twelve weeks; the call takes ten minutes, and it is written down.
What I suggest: that the next review be set today, with its agenda already written: the relevance threshold, the 7 silent sources, the twelfth candidate subject. A review that is not dated is a review that does not happen — and that is what had taken your summary from 52 weeks down to 19.
✎ Framework · monthly validation review — 4 measurements and their methods; decisions dated 12 and 20 March; replay over 12 weeks before anything goes live
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
What does the agent actually do?
One agent, several rhythms and several kinds of monitoring. All these uses work in support, subject to your approval.
Included in your agent
The 3 capabilities essential to this promise are included, at no extra cost.
From 479 € excl. VAT / month
Competitor monitoring
Follows the publications and announcements of the players you name, without extrapolating beyond them.
Boundary: follows the players you name and reports what they publish; it does not assess your competitive position, the strategy to hold or the price to set.
Press monitoring
Spots the relevant pieces in the trade press you have chosen.
Boundary: reads the publications you lawfully have access to and quotes the passage retained; it crosses no restricted access and redistributes no content beyond what your own rights allow.
Regulatory monitoring
Flags the texts published in the scope you have declared, without interpreting what they mean.
Boundary: flags that a text has been published, dates it and sources it; it does not compare versions by fingerprint, extract requirements, compute applicability or cost any plan — that is the remit of the Regulatory monitoring & impact analysis agent (ENT-UC-136), billed separately and never twice.
Controls and safeguards
These 5 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Sources, access rights and handling of questions with no answer
Inherited access rights, versions and freshness of the corpus
Explicit handling of questions with no answer and of conflicting sources
What the agent must be connected to
This connection is required for the agent to work. It concerns your information system and is scoped during the audit.
Measurement of accuracy, coverage and time saved
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Regulatory impact analysis
This agent flags that a text has been published, dates it and sources it. To compare versions by fingerprint, extract each requirement with its passage, compute applicability against a declared scope and cost the plan, a dedicated agent takes over — the monitoring is never billed again.
On quote View the agent page →Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Document base
To question your internal documentation, a dedicated document agent answers in plain language.
Document agent (FAQ, knowledge base) from 678 € excl. VAT / month Knowledge base →Tender monitoring
To watch the invitations to tender published, a dedicated agent takes it on.
Tender watch assistant from 454 € excl. VAT / month Tenders →
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the most relevant agent — without oversizing the project.
Method
Service workflow
These twelve steps are not modules for sale: they describe the method applied across the three kinds of monitoring included — competitor, press and regulatory. They apply to the scope and sources agreed with you, and human review takes place wherever it is required.
-
Monitoring scope and angles
Versions the questions, entities, themes, areas, languages, exclusions, frequencies, recipients and thresholds you have settled. A catch-up net prevents subjects declared critical from being discarded without review.
Boundary: records and applies the scope you settle; it does not choose it for you and never widens it on its own.
-
Source register
Qualifies the agreed sources: owner, address, type, authority, access method, licence, frequency, coverage, stability, date and owner. A source that stops responding is flagged at the next collection run.
Boundary: sources are not connected in advance — you decide which ones to open and you carry their access and re-use conditions; no access is acquired, subscribed to or negotiated on your behalf.
-
Collection and versions
Keeps, for the items collected, timestamp, fingerprint, response status, version and, on failure, the proof of that failure. A silent collection does not pass for an empty one.
Boundary: collects only approved sources at approved frequencies; access conditions, licences and restricted areas are hard constraints, never obstacles to work around.
-
Normalisation and quality
Extracts title, published author, date, language, jurisdiction, publisher, text and attachments where the source carries them, then measures completeness and freshness. A missing date stays missing: it is not guessed.
Boundary: measures and states the quality of what it has read; it does not fill in a missing field and does not correct a source.
-
Deduplication and lineage
Groups the technical duplicates, pick-ups, translations and updates it detects, keeps the primary source and its lineage, and retains the reasons for discarding. Eighteen pick-ups of one announcement remain one announcement.
Boundary: a plurality of pick-ups never amounts to corroboration; it groups publications, it does not decide that a fact is established.
-
Entities and taxonomy
Matches companies, brands, products, technologies and territories against your reference data, with aliases and a confidence level. An unresolved namesake stays open, it is not forced.
Boundary: matches names against your reference data under your validation; it builds no profile of a person and tracks no individual.
-
Facts, announcements and opinions
Classifies the nature of the items retained, separates the claim from its evidence and tracks whether an announcement materialises. An announcement is not presented as a fact.
Boundary: qualifies the nature of what is published; it rules neither on final truthfulness nor on a third party’s reputation.
-
Relevance and signals
Scores novelty, scope match, materiality, urgency and confidence through visible, versioned rules. What is unknown and what is contradictory stay on screen rather than being smoothed away.
Boundary: priority is explained by visible factors and can be overridden by an authorised owner, with a reason and a log; it scores no person.
-
Alerts and routing
Prepares, at the agreed rhythm, a bounded alert: source, date, reason, object affected, level, recipient and the verification step expected.
Boundary: prepares and routes internally under your approval; it sends nothing outside, contacts no one and publishes no position.
-
Memory and calendar
Makes the history searchable, tracks announcements, changes and announced dates, schedules reviews and limits repetition; what is set aside is set aside with its reason.
Boundary: carries dates exactly as the source publishes them, with the passage that bears them; it computes no deadline the source does not state.
-
Summaries and reporting
Produces a daily brief, a weekly summary, a thematic file and an executive note, at the agreed frequency. Factual statements carry the citation of their source; observed fact, interpretation, contradiction and unknown are kept apart.
Boundary: delivers a sourced summary to a reader; it writes no communication content and reaches no legal, financial or intellectual-property conclusion.
-
Review and measurement
Organises human review, correction and internal circulation, then measures coverage, precision, delay, sampled discarded items, readership and collection time given back.
Boundary: measures what it did and what it missed; no figure shown is a guaranteed client result, and nothing goes out before your approval.
The gain
How much time can a team recover on its monitoring?
By automating the gathering and the sorting, time spent reviewing sources comes down, reinvested in analysing what actually matters. How large the gain is depends on your volume and remains to be confirmed by a pilot.
Reviewing the sources
Writing the summary
Weeding out repetition
Qualitative, non-contractual comparison: the proportions shown illustrate the shift of the work towards review, they represent no measurement. Every output of the agent is reviewed and approved by a competent person.
How it works
The stages of your AI agent project
1
Audit & scoping
15 minutes to target the use case with the best return.
2
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
3
Design
We design the agent and its guardrails.
4
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
5
Rollout
Going live and training your team.
6
Operation
Continuous supervision and improvement.
Pricing
One package, one agent
A monitoring agent (defined sources, periodic sourced summaries), installed and operated for you.
Agility
Setup + controlled subscription
3,360 € excl. VAT setup
then 479 € excl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
664 € excl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
6,555 € excl. VAT setup
then 679 € excl. VAT/month · + hardware from 1,058 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices exclude VAT.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to your monitoring
Your monitoring subjects stay secretLocal inference or an isolated resource hosted in France; the subjects you watch are passed to no third party and are used to train no model.
Data in France, under French lawYour monitoring subjects and what you watch: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The head of monitoring keeps the decisionThe agent produces sourced summaries, which can be checked and altered; no approval is automated.
Human oversight & traceabilityOn your number of sources and the frequency of your summaries: systematic logging and tracking, in line with the AI Act.
Going further
Related resources
Frequently asked questions
Your questions, our answers
Does the agent choose the sources to watch?
No. You define the sources, the subjects and the frequency. The agent keeps to them: it does not go looking elsewhere and does not change the scope by itself.
Are the summaries sourced?
Yes, systematically. Every item kept refers back to its original publication, dated, so you can go back to the source before acting.
Does the agent interpret what a regulatory text means?
No. It flags the texts published in your scope and reports what they say. Analysing their legal effect rests with your lawyer or your adviser.
Is what we watch kept confidential?
Yes. The agent is hosted in France, on local inference or an isolated resource, with the deployment objective of processing and access operated within the European Union and an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. What you watch does not leave your premises.
How long does it take to deploy this agent?
A few weeks as a rule, depending on the number of sources to connect and how detailed the subjects are, after a free audit then a phase of design, integration and testing.
Is the monitoring exhaustive?
It covers the sources you have defined, at the frequency chosen. No automated monitoring can guarantee exhaustiveness: the agent reduces the gathering work, it does not replace your experts' vigilance.
How does it tell a fact, an announcement, a pick-up and an opinion apart?
Every item retained is classified by its nature, and the claim is kept apart from its evidence. An announcement stays an announcement until it materialises, and it is shown as such. Eighteen articles picking up the same announcement are reduced to one dated primary source: a plurality of pick-ups never amounts to corroboration. What is unknown or contradictory stays on screen, never smoothed away.
How do you measure what was discarded, and the subjects missed?
Every discarded item keeps its reason and the rule that discarded it, and a sample is re-read at each review: that is how a triage is checked instead of being believed. Catch-up terms you define prevent the automatic discarding of subjects you deem critical. Coverage, precision, delay and corrections after review are measured and reported back to you.
Can it get past a restricted area or a paywall?
No, and this is not a technical limit: it is a hard constraint. The agent collects only the sources you have declared and opened under your own mandate, at approved frequencies. Access conditions, licences and re-use rights belong to you and are recorded in the source register; the agent acquires none of them and circumvents none.
Does it follow patents, trademarks and announced dates?
It follows what the sources you open publish, and carries dates exactly as they appear there, with the passage that bears them — it computes no deadline the source does not state. No automatic connection to an industrial property register is delivered today: the capability is declared unproven as long as no re-use conditions have been validated in writing. What an intellectual property right covers remains for your counsel to assess.
Does it circulate alerts and reports automatically?
No. It prepares a bounded alert — source, date, reason, object affected, level, recipient and the verification step expected — and routes it internally under your approval. Nothing goes outside, no position is published, no one is contacted. Circulation remains a human decision.
Where is the boundary with the regulatory impact analysis agent?
This agent monitors and reports: it flags that a text has been published in the scope you declared, dates it and sources it. The Regulatory monitoring & impact analysis agent takes over: it compares versions by fingerprint, extracts each requirement with its passage, computes applicability against a declared scope, costs the gap and sets a dated plan your expert signs. The two chain in that order. Monitoring, the source register and evidence retention are paid for once: if you already run one of the two agents, only the real extension of scope is priced.
Related use cases
Other agents for information and documentation
Let's talk
Let's size up the potential in your monitoring
15 minutes to frame your priority sources — hosted in France, supervised, with no commitment.