We build Agents. You build the Future.

● B2B offer — Compliance & regulatory watch

Regulatory watch: the text spotted, dated, linked to your processes

A text changes, and the question is never “what moved?” but “does this concern us, where, and by when?”. Your agent spots the change in the corpora you open to it, dates it, separates a genuine amendment from a re-publication, links it to your entities, processes and contracts, and hands you the impact analysis ready for review. Data hosted in France; the AI model is called at its publisher, except with the Sovereign model option. With the 100% Sovereign package, your regulatory map does not leave your walls. Your lawyer decides — on a complete file, in minutes.

Discuss your project → See the prices

Updated on 05/11/2026

Deployed in a few weeks

Regulatory watch agent · hosted in France

Fictional simulation — not a client result

What changed this month that concerns us?

Eleven texts picked up, three concern you. The other eight are re-publications with no content change — identical hashes.

The first one — when does it apply?

Published on 20 August, applicable on 1 October. Both dates are on file separately, with the passage that carries them. The plan is set on the second.

Data hosted in France · sovereign model as an option

Sources opened under your mandate

Hosted in FranceAI model called at its publisher, except with the Sovereign model option

Protected regulatory map

Every alert citedThe source passage, never a summary alone

The lawyer decidesThe agent prepares, the expert characterises

Two dates, never onePublication and application kept apart

For an organisation established outside the European Union, the law of its own country applies; at its request, we put equivalent safeguards in place (human review before any action, a log of the criteria, information for the people concerned).

✦ In brief

Your agent reads the official corpora you open to it, compares versions by hash to set aside re-publications, extracts each requirement with its source passage, matches it against your declared scope, and hands you a quantified file: what concerns you, what is missing to decide, and the plan ready to arbitrate. The legal characterisation stays signed by your expert — that is what makes it stand up.

100 %

hosted in France in the target architecture

watch and impact-analysis modules included

file states traced, from discovery to closure

legal characterisation issued without expert sign-off

The first two figures describe the target architecture and the scope of the offer, not a production measurement. The next two describe an engine rule, verified by the repository’s tests.

Indicative estimate, to be confirmed by a pilot on your real files (general terms of sale, article 4).

The offer

Three plans, a single price list

The price list is common to the whole catalogue. Sector variants point to this same entry: they create no separate pricing.

Agility

Setup + controlled subscription

5,805 € ^{excl. VAT} setup

then 461 € ^{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
Hosting in France, a dedicated and isolated resource

Order →

The simplest Serenity

All inclusive, no setup fee

781 € ^{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
Hosting in France, managed end to end

Order →

100% Sovereign

On site, you own it

8,320 € ^{excl. VAT} setup

then 568 € ^{excl. VAT}/month · + hardware from 2,491 € (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.

Hardware installed on your premises (you own it)
French / European AI models run locally
Secure remote maintenance (Pro support included)

Order →

Included in the package: hosting and AI consumption up to the volume of your capacity tier, in documents, exchanges, tokens and storage; beyond that, re-invoiced at real cost with no margin, tracked in your client area as usage statements are issued. 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 for the package chosen, in France by default, 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.

See the full price list →

Estimate my gains in 1 minute →

What the agent does

Twelve modules, from open source to signed plan

All twelve are included in the core offer. The four neighbouring agents take over where the work changes object, and the boundary is written on every card.

Included in your agent The 12 capabilities essential to this promise are included, at no extra cost.

From 461 € excl. VAT / month

Source register

Every corpus you open is declared: authority, jurisdiction, topic, access method, frequency, owner and fallback. A source that goes down becomes visible.

Official collection

Each document arrives with its identifier, canonical address, date and hash. That hash is what later separates a real change from a mere re-publication.

Versions and consolidation

Initial text, amendments, corrigenda, consolidated versions and guidance are linked to one another. You read a history, not a pile of files.

Scope taxonomy

Your entities, activities, products, territories, thresholds and regimes are structured once and signed off by you. That declaration is what makes applicability computable.

Change detection

Article, annex, definition, threshold, date and status are compared one by one. Addition, removal, replacement, clarification and plain re-publication are filed separately.

Requirements and obligations

For each candidate requirement: actor, action, object, condition, date, frequency — and the exact source passage. Without a citation, the extraction is rejected.

Applicability matrix

The requirement meets your declared scope. You see the facts that match, those that are missing, the exemptions, and the precise question to put to your expert.

Impact mapping

The obligation is linked to your policies, processes, products, contracts, data, systems, controls and training. Impact becomes a list of names, not a hunch.

Gap analysis

The versioned requirement is set against your declared state and your evidence. Gap, missing evidence, partial control and unconfirmed non-applicability never get mixed.

Compliance action plan

Actions, owners, dependencies, deadlines, effort, priority and expected evidence are prepared and quantified. You decide on a written plan, not on a blank page.

Review and distribution

Challenge, sign-off, targeted subscription, bilingual impact note, read receipt and review date. You know who received the alert, and who acted on it.

Evidence and steering

Snapshots, citations, decisions, versions, mappings, plans and closures are kept. Coverage, lead time, false positives and open gaps are measured on your own data.

Controls and safeguards These 12 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.

Tie every document to an authority declared in the source register and record its canonical address — without them, no alert goes out Record the official identifier and the issuing jurisdiction; two jurisdictions are never merged into a single file Establish the version read and its status — a draft and a text in force are not filed in the same place Carry the dates of publication, of entry into force and of application separately; no deadline is computed without citing the computation rule Check whether a corrigendum or a later consolidated version exists before any impact analysis Compute and keep the hash of the collected document — that is what separates a real change from a re-publication Have entities, activities, territories and regimes signed off by the scope owner: they are the basis of all applicability Record the organisation’s value against the threshold cited, with its basis of calculation and its reference period Attach the exact source passage — a requirement extracted without a citation is rejected, however plausible it is File the evidence that establishes the declared internal state; a declaration without evidence remains a declaration Have applicability and characterisation settled by the lawyer or the compliance officer named on the file Collect the recipients’ read receipt and set the review date of the impact note

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.

Regulatory control

Applying an ALREADY established rulebook to your files, one by one, and returning qualified findings is the business of the regulatory control agent. This agent works upstream: it establishes the rule, its version, its date and its applicability, and then hands it over. The two chain in that order, and never the other way round.

Regulatory control from 371 € excl. VAT / month Discover the agent →

AI governance & AI systems register

Keeping the inventory of your AI systems, their model-version-deployment lineage, their evaluations and their incidents belongs to the AI governance agent. This agent keeps no asset inventory: it says that a text has moved, extracts the requirement with its passage, and hands over the question of which systems are affected.

AI governance & AI systems register from 407 € excl. VAT / month Discover the agent →

GDPR rights & breaches

Handling an individual's request — access, erasure, objection — or a personal data breach, with its clock and its notification, belongs to the rights agent. This agent opens no individual case: when a new text touches data protection, it draws out the impact on your processes and hands the case over.

GDPR rights & breaches from 517 € excl. VAT / month Discover the agent →

End-to-end contracts

Negotiating a clause, running a contract lifecycle and tracking the deadlines of your 213 contracts belong to the contracts agent. This agent drafts the clause that a regulatory gap calls for, in the wording of your existing contracts, and stops there: the negotiation starts from that text and belongs to you.

End-to-end contracts from 362 € excl. VAT / month Discover the agent →

Does your need fall outside this?

In 30 minutes we identify the most relevant agent — without oversizing the project.

Build your agent

Demonstration

See the agent at work

4 everyday 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 company

Groupe Ternoise — design, manufacture and maintenance of industrial lifting and handling equipment

Sector: Lifting and handling equipment manufacturing (NACE 28.22) — seven companies, one active holding, two plants and four maintenance branches
Headcount: 610 employees, including ONE part-time group compliance manager and two lawyers shared with procurement; no dedicated regulatory watch role
Market served: Food processing, logistics and port operations; France, Belgium and Morocco — two of the seven companies are established outside France
Order of magnitude: 41 documented processes, 213 live supplier contracts, 9 AI tools in service, a quarterly compliance committee
Tools already in place: A document management system, a process repository kept in a spreadsheet, a contract base inside the ERP — the agent reads the DEPOSITED corpus and repositories, it connects to none of these tools
Who decides what: The group general counsel characterises and signs; the compliance manager works the file through and keeps the evidence; each process owner answers for their own gap; the compliance committee arbitrates the plan and its deadlines
Room for improvement: The second-quarter review left 6 of 11 texts unlinked for lack of time; two deadlines were discovered less than three weeks ahead; no group evidence file can be replayed beyond six months

Groupe Ternoise is not short of alerts: it gets them from trade bodies, from its law firm and from two newsletters. What it lacks is the work that comes after — checking that a text really changed, isolating the requirement, matching it against seven companies that share neither size nor regime, finding which process and which contract it touches, and gathering the evidence of what the group already does. The agent runs on local inference on a machine at head office and reads the corpus the compliance manager deposits, with its date and its hash. It compares, extracts, links, quantifies and drafts; the general counsel characterises and signs, each process owner answers for their gap, and the compliance committee arbitrates the plan. The four tabs follow the September 2026 review, from the corpus deposit to the impact analysis handed to the committee.

This company, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real client.

Regulatory watch agent · hosted in France

Fictional simulation — not a client result

Your September corpus is processed: 11 documents deposited on 2 September, all 11 entered in the source register with their identifier, canonical address, date and hash. Eight are re-publications with no change of content. Three genuinely moved, and two concern you.
A re-publication is a document an authority puts back online without altering a letter of the text: new address, new file, same content. It costs half a day of reading every quarter and teaches nothing.
The source register as you approved it on 14 March, and the age of each corpus at collection time:
· a European corpus you opened under your own mandate — deposited 02/09 at 06:40, age 9 hours, against the 72-hour threshold you set;
· a national corpus deposited by your law firm — deposited 02/09 at 07:15, age 8 hours;
· guidance from two sector authorities — deposited 01/09, age 32 hours;
· a publisher's newsletter — filed as commentary, named owner, and it never grounds an alert on its own.
The three real changes, in the order in which they reach you:
· VR-2026-091 — the calendar for the "high-risk" obligations of the artificial intelligence regulation has moved. You run nine AI tools: this is the file of the month, priority 1.
· VR-2026-093 — the electronic invoicing regime. The text does not change; it is your scope that is unsettled on one company. Priority 2.
· VR-2026-088 — a proposal out for consultation on machinery safety. Non-final text, status displayed on every screen where it appears, tracked since March. Priority 3.
What I suggest: start with the version comparison on file VR-2026-091 — you will see where the conclusion comes from before you read it. collection-log_september-2026.pdf11 documents, 4 declared corpora, ages recorded, 8 re-publications
⛓ Sourced · corpus deposited 02/09/2026, source register approved 14/03/2026, 11 hashes computed

How do you know the other eight changed nothing? We have been caught out before.

By an article-by-article comparison, not by the file date — which is precisely the surprise you are describing.
What the comparison covers, item by item: article, annex, definition, threshold, date and status are matched one against one, between the consolidated version you held and the one that has just come in. Additions, removals, replacements, clarifications and plain re-publications are filed separately. Across the 8 re-publications, the hash of the normative content is identical to the comma: what changed is the file address, the layout, or a non-normative footnote.
The figure I publish about myself: at the June review, the authority re-published a document at a new address altering nothing but an ANNEX — the body of the text was identical to the comma, and I filed it as a re-publication. It was your lawyer who opened the annex, and that is exactly what the adversarial review is there for.
What I did with that, and it is the only part that counts: the change-detection rule was corrected on 3 July, then replayed across twelve months of your corpus — 147 documents. It found 3 further cases of the same defect, all three in annexes, all three now open as files. None carried a deadline that had already passed, and I checked that document by document.
What it costs you today: the 8 re-publications are set aside in 4 minutes of machine time, and the comparison trail stays available. Human reading concentrates on the 3 that moved. version-comparison_VR-2026-091.pdfArticle by article, 8 re-publications set aside, 3 changes retained
⛓ Sourced · hash comparison of the 11 documents, detection rule v4 of 03/07/2026, replay across 147 documents

The first one, the AI text — when does it apply?

Two dates, and they do not say the same thing — that confusion is the most expensive one in regulatory watch.
What the deposited passage carries, and I reproduce it in the attached extract:
· the obligations for high-risk systems apply on 2 December 2027, a date moved by Regulation (EU) 2026/1744 amending Regulation (EU) 2024/1689;
· the Article 50 transparency obligation has applied since 2 August 2026 — that is, for the past five weeks.
Here is the real result of the month, and it runs the other way round: the postponement gives you room on the programme you had started, and it uncovers an obligation that is already in force and was not yet on anyone's plan. Your plan was built on the earlier high-risk date; it re-phases without losing anything. Article 50, on the other hand, has been running since 2 August.
What the version comparison shows exactly: Article 50 did not move by a letter between the two consolidated versions. What moved is the calendar around it — and a text that shifts a neighbouring date changes the way everything else reads.
Where I stand, and what I need from you: the file is open, both dates are carried separately with the passage that grounds them, and the characterisation is waiting for your general counsel: she is the one who states that the text applies, and that is what makes the statement stand up. I hand her the complete file; she decides in minutes. deposited-corpus-extract_VR-2026-091.pdfInput document · source passage, hash, two dates kept apart
⛓ Sourced · extract from deposited corpus VR-2026-091, Reg. (EU) 2024/1689 Art. 50 applicable since 02/08/2026, high-risk calendar at 02/12/2027 (Reg. (EU) 2026/1744)

Fictional simulation — not a client result

Nine requirements and obligations are extracted from file VR-2026-091, and each carries its source passage. A tenth candidate was rejected: it was plausible, and its quotation traced back to no passage.
A requirement extracted without a quotation is rejected by the engine, however plausible it looks. That is a versioned rule, not an intention — and it is what makes a file hold up a year later.
What each requirement carries, and nothing less: the actor, the action, the object, the condition, the date, the frequency, the exact passage, and the consolidated version it was extracted from.
The three that weigh most on you:
· EX-04 — informing people exposed to an AI system (Article 50 transparency, applicable since 2 August 2026). Actor: the deployer. Frequency: at every interaction. It touches your customer service assistant and your warranty claim triage tool.
· EX-07 — reinforced obligations for systems falling under Annex III point 4, employment and worker management, applicable on 2 December 2027. It touches your CV screening tool and your maintenance operator assessment tool.
· EX-09 — AI literacy: the people who operate these systems must have a sufficient level of knowledge. Actor: provider and deployer. It touches 74 people at your company.
What I suggest: move to the applicability matrix, where these nine requirements meet your seven companies — that is where the work starts paying. extracted-requirements_VR-2026-091.pdf9 requirements quoted, 1 rejected without quotation, actor and date for each
⛓ Sourced · 9 requirements extracted with passage and version, 1 candidate rejected for want of a quotation, corpus VR-2026-091

Seven companies, two outside France. Are we all concerned in the same way?

No, and that is exactly what the applicability matrix computes — on the scope taxonomy you declared once, in February, and have kept approved since.
What your taxonomy carries: seven entities, their activities, products, territories, headcount thresholds and regimes. That declaration is what makes applicability computable rather than arguable.
What crossing the 9 requirements with the 7 entities gives, and I hand you the detail:
· 34 applicable crossings — the declared facts match the condition in the text, and I name which ones;
· 21 non-applicable crossings, with the fact that rules them out: the Moroccan company is neither established in the Union nor places a system on the Union market — the exclusion is reasoned, not assumed;
· 7 crossings under an exemption to confirm, with the precise question to ask and the person to ask;
· 1 crossing in unknown scope, and I come to it.
What the matrix shows, and what no table brought into view: your CV screening tool is operated by the holding on behalf of all seven companies. One entity therefore carries the obligation for the whole group, and it is not the one doing the hiring. That moves the owner of the gap, and it is a linkage conclusion, not an opinion: it comes out of the intra-group services agreement you deposited, article 3.
What I suggest next: the crossing in unknown scope — the only place in the matrix where I stop, and I will tell you exactly on what. applicability-matrix_9-requirements-7-entities.pdf63 crossings, 34 applicable, 21 ruled out with reasons, 7 exemptions, 1 unknown
⛓ Sourced · scope taxonomy approved 12/02/2026, 63 crossings computed, deposited intra-group services agreement (art. 3)

What is this unknown scope?

It is Ternoise Services, and file VR-2026-093 — electronic invoicing. The question is not one of law but of fact: SME or mid-cap, and I do not hold the figure that settles it.
What is certain, and depends on no threshold: receiving an electronic invoice has been mandatory for every VAT-registered business established in France since 1 September 2026, with no size condition. All seven of your companies were concerned on the same day. Five are registered and connected to an approved platform; I have collected the confirmations. Two are not yet, and registration is not retroactive.
What does depend on a threshold: issuing. Large companies and mid-caps since 1 September 2026; SMEs and micro-businesses on 1 September 2027. Four of your companies are classified without ambiguity. Ternoise Services sits on the line.
The precise figure that is missing, and the person who holds it: the balance sheet total for the year ended 31 December 2025 for Ternoise Services, held by your chief financial officer. Headcount and turnover are already on file; the third term is missing.
What the file says in the meantime, word for word: unknown scope — the passage carrying the threshold is quoted, the missing figure is named, the person to ask is named. It says neither "applicable" nor "not applicable", and that is what keeps a false certainty out of your plan.
What I have prepared so this settles in one reply: the question is drafted, it runs to three lines, it names the value expected and its closing date. As soon as the value comes back, applicability recomputes without redoing the work already done — the other 62 crossings are not replayed. unknown-scope_ternoise-services.pdfThreshold passage quoted, missing figure named, question drafted
⛓ Sourced · corpus VR-2026-093, receiving mandatory since 01/09/2026 with no size condition, issuing mid-caps 01/09/2026 and SMEs 01/09/2027, deposited entity repository

Fictional simulation — not a client result

The impact mapping is done: the 9 requirements touch 38 named objects in your repository. Impact stops being an intuition and becomes a list your owners can read.
What the linkage produced, item by item:
· 6 policies — including your digital tools charter, which mentions no AI system at all;
· 9 processes out of the 41 documented — hiring, annual appraisal, warranty claim handling, customer service;
· 2 products — the customer service assistant and the diagnostic module embedded in your overhead cranes;
· 7 contracts out of 213 — six AI component suppliers and one recruitment provider;
· 4 datasets and 5 systems, with the version in service;
· 3 existing controls and 2 training courses already in the plan.
The gap analysis, matched against your declared state and your evidence: 11 gaps, and I do not mix them. 4 confirmed gaps — the requirement applies, nothing is in place. 3 missing pieces of evidence — the control exists, its evidence was never filed. 3 partial controls — in place in two companies out of seven. 1 non-applicability to confirm. These four families never blur into one another: a control without evidence is not an absent control, and treating it as one would make you rebuild what already exists.
The most urgent item, and it was not in the plan: informing people exposed to your customer service assistant. The obligation has been running since 2 August; the gap has been open for five weeks, and it closes with a notice and a configuration change.
What I suggest: look at the 7 contracts, because that is where the question "how much work?" really lands. impact-mapping_38-objects.pdf38 objects touched, 11 gaps in 4 families, owners named
⛓ Sourced · process repository deposited 02/09, contract base deposited 02/09, group evidence register

We are not going to reopen 213 contract amendments for this.

Seven, not 213 — and I read all 213 so that I could tell you so.
How the sorting was done, and it guesses nothing: I looked through your deposited contracts for clauses covering an AI component, a model supply, automated processing of applications, or a technical documentation duty. 206 contracts carry none, and the attachment gives you the reason for exclusion contract by contract — a contract set aside without a reason gets read again next year, and the work has to be redone.
The 7 retained, and exactly what they lack:
· 4 AI component supply contracts: no duty to hand you the model's technical documentation, and no duty to notify you of a version change. That is the costliest gap: without notification, your register is wrong the day the supplier updates.
· 2 contracts do carry a documentation clause, but with no deadline — an obligation without a deadline does not get performed.
· 1 recruitment services contract: the provider operates the screening tool, and the contract does not say who answers for what.
What I have written, ready for review: a draft clause in two versions — a four-line short form for simple amendments, and a long form for renegotiating the recruitment contract, which expires on 31 January 2027. Both use the vocabulary of your existing contracts, not textbook wording: I read your last 12 amendments for that.
What I recommend, quantified: put the short form into the 3 contracts falling due within eight months — the clause rides in with the renewal, with no negotiation opened. For the other 4, a simple amendment. The negotiation starts from there; it belongs to you. contract-analysis-and-draft-clause.pdf213 read, 206 excluded with reasons, 7 retained, clause in 2 versions
⛓ Sourced · 213 contracts read, 206 excluded with reasons, 12 recent amendments read for vocabulary

And the compliance action plan — how long does it run?

Four milestones, and the first closes this week. The plan is written, costed and allocated; what is missing is the committee's arbitration — not the work.
The compliance action plan, milestone by milestone:
· Milestone 1 — the obligation already in force, before 30 September. An information notice on the customer service assistant and on warranty claim triage. Estimated effort: 6 hours, owner: the customer service manager. Evidence expected: a dated screenshot and the configuration version.
· Milestone 2 — the contracts, before 31 March 2027. 3 clauses at renewal, 4 amendments. Owner: procurement. Effort: 22 hours, dependency: your counsel's approval of the clause.
· Milestone 3 — AI literacy for the 74 people, before 30 June 2027. Two of your existing courses already cover 31 people; 43 people and one module remain. Effort: 34 hours.
· Milestone 4 — the file for the Annex III point 4 systems, before 30 September 2027, for an obligation applicable on 2 December 2027. Effort: 96 hours, owner: the compliance manager, with the two business owners.
The figure that makes the plan liveable: 158 hours in total, which is more than four weeks of work on a 35-hour week, spread over fifteen months. The move to 2 December 2027 is what makes that spread possible — without it, milestone 4 fell in the same quarter as milestone 2.
What belongs to the committee, and to nobody else: approving the effort figures, arbitrating between milestone 3 and milestone 4, and accepting or refusing to push milestone 2 beyond March. The file is ready for the committee on 15 September; the arbitration takes one sitting. compliance-action-plan_4-milestones.pdf4 milestones, 158 hours, owners, dependencies, expected evidence
✎ Framework · plan costed on a 35-hour week, owners and expected evidence named, arbitration with the compliance committee

Fictional simulation — not a client result

The impact analysis is ready for the committee on 15 September. It runs to nine pages, it is bilingual, and every one of its findings points back to the passage that grounds it.
What it carries, in the order the committee will read it:
· what changed — the calendar that moved, with both dates and their passage;
· what concerns you — 34 applicable crossings out of 63, and the 21 ruled out with their reason;
· what it touches — 38 named objects, owners against each;
· the gap — 11 gaps in four families, with what already exists and must not be rebuilt;
· the plan — 4 milestones, 158 hours, fifteen months;
· what stays open — one unknown scope, seven exemptions to confirm, and the question drafted for each.
The conclusion, as it is written: the postponement of the high-risk calendar loosens the heaviest milestone by fifteen months; it does however uncover an obligation applicable since 2 August 2026 on two of your products, and that gap closes in six hours.
What gives it weight in front of an inspector: every statement carries its source, its version and its consultation date. A statement lacking any of the three would not be in the document — that is an engine rule, checked at generation.
What belongs to your general counsel: the characterisation, at the head of the document, under her name and date. The document goes out marked "proposal not approved" until she signs, on screen and in exports. She signs on a complete file, in session. impact-analysis_VR-2026-091.pdfOutput document · 9 pages, every finding sourced, characterisation to sign
⛓ Sourced · impact analysis generated 05/09/2026, 9 pages, every finding tied to its passage and version

Who read it, and how do I prove that a year from now in front of an inspector?

Review and distribution are traced end to end, and that is precisely the part usually rebuilt after the fact — therefore badly.
What went out, and to whom: the impact note is addressed to 9 named recipients, selected by targeted subscription on the objects touched — a process owner only receives what touches their processes. 7 read receipts as of 5 September, 2 reminders scheduled for the 8th. Every send carries its date, its recipient and the version distributed.
What the challenge round brought, before distribution: the file goes through a contradictory review. Your customer service manager corrected a fact — the assistant is not deployed on the telephone channel, contrary to what the deposited repository stated. One crossing moved from applicable to non-applicable, the correction is dated, signed in her name, and the previous version stays available. That is what makes the file defensible: not that it was right first time, but that you can see who corrected what, and when.
What you will replay a year from now, in order: the deposited document and its hash, the version comparison, the requirement and its passage, the crossing and the scope facts that grounded it, the contradictory correction, the distribution and its receipts, the committee decision, the evidence filed. Eight links, all dated, all timestamped.
The review date is already set: 5 March 2027 for this file, or earlier if a new version of the text enters the corpus. You will not have to remember it — the file comes back up on its own. distribution-and-evidence-chain.pdf9 recipients, 7 receipts, 1 contradictory correction, 8 links
⛓ Sourced · 9 recipients, 7 receipts as of 05/09, contradictory review of 04/09, 8-link evidence chain

And you — what do you handle on your own, without asking us?

Three moves, all three reversible in a word. I would rather name them than let you find them.
· I open a file as soon as a version comparison shows a real change, without waiting for the quarterly review. And the reverse holds: if a later consolidated version cancels the change, I close the file myself, with the trail.
· I remind an owner who has not acknowledged the note, once, after three days. I stop there: a second reminder is an escalation, and an escalation carries a name — yours.
· I mark "stale source" on any corpus older than the threshold you set, and I take its documents out of the review count. The move undoes itself at the next deposit, with nothing to re-enter.
Everything else waits for a signature, and the list fits in one attachment: characterising a text as applicable, declaring a control compliant, accepting a gap, fixing a deadline, distributing a note, closing a file. Six acts, six signatures, each with a name, a date and a reason.
The steering dashboard, over your last twelve months: source coverage 100 % of the corpora you opened; median delay between a text being deposited and its file being linked: 2 days, against 34 days at the March review; false positives 4 % after the July correction, against 11 % before; open gaps: 11, of which 1 with a deadline inside thirty days. Evidence and steering are built as the decisions are taken, not at the end.
Where all this runs: on a machine at head office, on sovereign AI (100% Sovereign package) — local inference, hosting in France; your regulatory map and your gaps do not leave the group. No direct link to an outside corpus is in service, and I announce none: such a link is chosen, documented and tested before it is promised. What is sold here is what works today — you deposit, I work, and you know at every moment how old what I am showing you is. who-decides-what-and-steering-dashboard.pdf3 reversible moves, 6 signed acts, 4 indicators over 12 months
✎ Framework · three reversible moves, six signed acts, steering dashboard over 12 months

Data hosted in France · sovereign model as an option

Your case is not here? That is exactly what a 30-minute conversation is for. Discuss your project →

How it works

The stages of your AI agent project

Audit & scoping

30 minutes to target the use case with the best return.

Quote or direct sign-up

A catalogue offer is bought online; a specific need gets a costed quote.

Design

We design the agent and its guardrails.

Integration & testing

We connect your tools to the agent, which is itself hosted in France.

Rollout

Going live and training your team.

Operation

Continuous supervision and improvement.

Context

What does an AI agent bring to your regulatory watch?

The cost of a watch is not the reading: it is the linking. Knowing a text moved takes a minute; knowing whether it concerns you takes a week.

! The real cost is not spotting, it is linking

Raw alerts are not in short supply: feeds, newsletters, portals. What is missing is the work that follows — checking the text actually changed, isolating the requirement, testing it against your entities and thresholds, finding which contract and which process it touches, and gathering the evidence of what you already do. That work is done by hand, it does not delegate easily, and it decides whether the alert was worth anything.

✓ Our answer

Your agent takes that whole chain and makes it traceable. It compares versions by hash, so a re-publication disturbs no one. It extracts each requirement with the passage that carries it — an extraction without a citation is rejected, however plausible. It matches it against the scope you declared once, and when a fact is missing, it names which one and who to ask rather than concluding. On the texts that concern you, it goes as far as the quantified plan, the draft clause and the bilingual impact note. Your expert signs — in minutes, on a complete file.

Trust

With 100% Sovereign, your regulatory map does not leave your walls

Knowing which texts you watch and where your gaps are is sensitive information. It is handled as such.

Local inference

With the 100% Sovereign package, the agent runs on a machine of your own: neither the texts examined, nor your map, nor your gaps leave the network.

Isolated resource

With French hosting, your environment is dedicated: no watch data is shared between organisations.

Encryption

Collected documents, mappings and evidence are encrypted at rest and in transit.

Role-based access

Watch manager, lawyer, process owner, auditor: each sees what their role allows, and nothing more.

Audit logging

Each state change carries its evidence, its actor and its timestamp. The chain replays a year later.

Ingested content, never an instruction

A collected text is data. An instruction hidden inside it commands nothing: the engine runs only its versioned rules.

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 encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
Roles and permissions are configured and accepted for the identities and systems actually connected.
The events logged, their content, their retention period and who may access them are defined for the deployment chosen.

Official corpora are not connected in advance: you decide which ones to open, and each is declared in the source register with its authority, access method and reuse terms. See our confidentiality commitment.

Time

Where the time moves

Reading does not disappear: it concentrates where a human decides.

Sorting alerts and re-publications

Today · reading everything

Hash comparison

Linking to scope and contracts

Today · manual search

Match prepared

Legal decision and signature

Today · file to reconstruct

Complete file to review

These proportions illustrate a shift in workload, not a measurement: they come from no pilot and constitute no commitment. Legal decision time does not shrink — it is spent on a file that is already assembled.

Indicative estimate, to be confirmed by a pilot on your real files (general terms of sale, article 4).

Our commitment

Our commitments

With 100% Sovereign, your map stays with youData hosted in France on an isolated resource; the AI model is called at its publisher, except with the Sovereign model option. With that option, no watch data entrusted to a third party, none used to train a model.

No requirement without its passageA requirement extracted without a citation of its source is rejected by the engine, however plausible. That is a rule, not an intention.

Characterisation is signedThe agent prepares, quantifies and justifies. Declaring that a text applies or that a control is compliant binds your organisation: that belongs to the expert you appoint.

No invented deadlineIf the source carries no application date, the agent says so and gives the missing computation rule. It never invents one.

Going further

Related resources

AI agents for business →Our full offer Sovereign AI hosted in France →Our confidentiality commitment

Frequently asked questions

Your questions, our answers

How does the agent tell an official source from a commentary?

Through the source register, which you approve. Each corpus is declared with its authority, jurisdiction and access method; a document arrives with its canonical address, date and hash. A publisher’s commentary may be tracked, but it is filed as commentary and never grounds an alert on its own.

Can it decide that a text applies to our organisation?

It performs the full match and hands it back in minutes: the passage carrying the requirement, the facts of your scope that align, those that are missing, the exemptions and the threshold calculation laid out. The legal characterisation is signed — and that is what makes it stand up. Your lawyer decides on a finished file, not a blank page.

Does it handle consolidated versions and corrigenda?

Yes, and this is where it gives back the most time. Initial text, amendments, corrigenda, consolidated versions and guidance are linked, with their dates. A re-publication whose content hash has not moved is identified as such and disturbs no one.

How does it link a change to our processes and contracts?

Through the mapping you declare once: policies, processes, products, contracts, data, systems, controls and training. Each retained obligation points to the affected objects, by name. On a contract, the agent goes as far as a draft clause ready for review — the negotiation starts from there.

What does it do when a threshold or exemption is unknown?

It says so, and that is a result in itself. The file moves to unknown scope with the citation of the passage carrying the threshold, the precise missing data point and the person to ask. As soon as the value comes back, applicability is computed without redoing the work already done.

Can it declare a control compliant?

It produces the file that allows the declaration to be made: versioned requirement, declared internal state, submitted evidence and its age, quantified gap where there is one. The declaration binds the organisation and belongs to the compliance officer — on this file, it takes minutes.

How does it track consultations and non-final texts?

It tracks them, dates them and compares them with the version in force, with their status shown on every screen where they appear. This is when a contribution costs the least: the agent prepares the points to argue and the gap they would create for you.

How does it prove a team received and then acted on the alert?

Each impact note goes to named recipients, with a read receipt and a review date. The file keeps the dispatch, the receipt, the decision taken and the accepted evidence. A year later, you can replay the whole chain in front of an auditor.

Going further

The agents that work alongside this one

Regulatory control →Your rules applied file by file Competitive & press watch →Market information, alongside the law Corporate legal →Research, contracts and case law

Let's talk

Which text slipped past you last quarter?

Open one corpus and one scope to us, and we will hand back the full mapping on your own texts: what concerns you, what is missing to decide, and the plan ready to arbitrate.

Write to us