● B2B offer — Documentary monitoring

Assistant for documentary research on the web: sourced summaries

Exploring the web on a precise subject, sorting the reliable sources, drawing a clear summary from them: useful work, but time-consuming, and often put off for want of time. Your assistant carries out the search, compares the sources and writes a summary that cites every passage used — without ever inventing. Hosted in France — on local inference or an isolated resource — the subject of your research never leaves the company.

Free audit (15 min) → See the prices

Hosted in France Sourced summaries, nothing invented GDPR & AI Act: governed deployment Human oversight

Updated on 07/09/2026

Deployed in one to two weeks

Web research assistant · hosted in France

Give me a summary of the new electronic invoicing obligations for small and medium businesses, with the sources.

Summary ready (6 sources consulted): timetable for entry into force, obligations to receive, platforms concerned.
Every point is linked to its source; two sources contradict each other on a detail of the timetable, and I flag that rather than settling it.
To read over before it circulates or is used in an internal document.
⛓ Source · official pages and trade press consulted

Can you dig into the point about the platforms concerned, in more detail?

Deeper look done. A list of the types of platform mentioned by the official sources, with the exact passages cited.
One of the sources does not address this point: I have not included it for this part of the summary.
✎ Action · summary extended, sources to back it up — you approve before it circulates

Local inference · no data outside the EU

Research hosted in France

Sovereign by designLocal inference or hosting in France

Sourced summariesCites its sources, does not paraphrase at random

TurnkeyDesigned, installed and operated for you

You decideThe agent assists, never rules alone

✦ In brief

A Blue Lemon Agent documentary research agent explores the web on a subject you give it, compares several sources and writes a clear, sourced and checkable summary — every statement is linked to the passage that grounds it. Where sources contradict each other or information is missing, it says so rather than inventing. It runs on local inference or is hosted in France: the subject of your research, often strategic, is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Live in one to two weeks,.

100%

hosted in France in the target architecture

transfer outside the EU in the target architecture

uses ready to deploy on this scope

decision taken without human approval

These figures describe our offer, not results measured at a client: how large the gain is on the scale of your research is confirmed by a pilot.

The context

What does an AI agent bring to your documentary research?

Finding reliable sources, cross-checking them and drawing a usable summary from them takes method. But passing your research subjects to a consumer tool sometimes amounts to revealing your strategic intentions.

! The issue

Every piece of documentary research takes time to sort uneven sources and draw a reliable summary from them. Yet most consumer AI tools amount to entrusting the subject of your research, your queries and your strategic intentions to a third party, often hosted outside Europe and subject to the Cloud Act.

✓ Our answer

AI is only of interest for documentary research if it is sovereign and honest about its sources. Local inference or an isolated resource hosted in France, systematic human oversight, every statement linked to its source: time gained on the search is never paid for in lost confidentiality about your subjects. The aim is not to replace your critical judgement, but to give you a checkable basis to work from.

The decisive point

Your research subjects: sovereignty & compliance

The very subject of a piece of research can be strategic information. Here is how the architecture of our agents protects it, query by query.

Local inference

The agent can run on a machine belonging to the company: the research subject does not leave the network, nothing passes through a public cloud.

Hosting in France

Otherwise, a dedicated and isolated resource, hosted in France under French law — your queries: processing and access within the European Union targeted by the architecture.

Reduced extraterritorial exposure

For your research subjects, 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 immunityeven when hosted in Europe by an American provider.

Isolated resource

No pooling: an environment strictly dedicated to your company and your research.

Sourced summaries, nothing invented

Every summary cites the passages used; where there is a contradiction or information is missing, the agent says so.

AI Act: governed deployment

The agent is strictly in support; no summary commits you on its own; 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 the most sensitive subjects, 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

4 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.

Web research assistant · hosted in France

Four flags this morning, drawn from my own answers.
· Of 340 searches, 47 had no public answer. I answered "I did not find it" 47 times — and that is a result, not a failure.
· A figure that three pages seemed to confirm came from one. The three cited each other in a circle.
· 118 pages consulted carry no date. I cannot say whether they are current.
· Nine requests were about a named person. I did not handle them, and I will say why. morning-watch_4-flags.pdf47 "not found" · 9 requests declined
⛓ Source · 340 searches, 2,100 pages consulted

Forty-seven "I did not find it" is a lot.

It is the most useful part of my work, and the hardest to produce.
The underlying problem: the web always answers. On any question, you will find a page asserting something. A search tool that never says "nothing" is a tool that invents — not by fabricating text, but by presenting as an answer the first page that looks like one.
What I do with the 47: I say what I searched for, where I searched, and what I found that comes close without answering. Three times out of four that is enough for the person to reformulate.
What those 47 cover: 19 questions whose answer is not published — a figure internal to a company, a decision not made public —, 16 badly-put questions that contained two questions, and 12 where public sources contradict each other with none prevailing.
The last 12 matter most: I do not settle them. I give both positions, their sources and their dates, and I say that they conflict. 47-of-340_19-16-12.pdfA tool that never says "nothing" invents
⛓ Source · 47 searches with no answer, split into 3 causes

How do you work, and what do you hand back?

I search public web sources, and I hand back an answer whose every assertion carries its link.
What an answer contains: the assertion, its source, the page's date, and — this is what is missing everywhere — the level of confirmation: one source, several independent sources, or several sources tracing back to the same one.
What I flag unprompted: an undated page, a circular citation chain, two sources contradicting each other, and a source whose publisher I cannot establish.
What I always add: the list of what I consulted and set aside, with the reason. A search where only the result is visible cannot be checked.

What that gives you across 340 searches: 293 summaries delivered with, for every assertion, its source, the page's date and its level of confirmation; a figure three pages seemed to confirm traced back to the single source it came from; 118 undated pages flagged as such; and 47 questions on which you know exactly where I looked.
What you gain from tomorrow: a search kept being put off for lack of time becomes the same day's summary, checkable line by line. You settle it in minutes because you can see what I consulted and what I set aside, with the reason. On the 12 questions where public sources disagree, you decide on both positions, their sources and their dates, instead of inheriting an answer chosen for you.
On a named person, I work within a written frame: searching on a person means processing personal data. You set the purpose, the lawful basis and the duration, I hold to it, everything is logged and the frame is withdrawn with a single word. This morning's nine requests are waiting for that frame, not for a refusal.
On what I consult: the subject of your search, often strategic, leaves neither the company nor France. The next step is ready: fifteen minutes to set your three priority subjects and the reporting rhythm.
✎ Framework · no search on a person · public ≠ true

Three pages confirmed the figure, and it was wrong?

All three traced back to the same page, and that page cited nobody.
What I record: a figure appears on three apparently independent sites. Tracing back: site B cites site A, site C cites site B, and site A cites no source. The figure was never established anywhere.
Why this is the central trap of the trade: multiple confirmation is the only reliability test a hurried reader has, and it is exactly the one a chain of reprints manufactures for free. Three sources are only worth three sources if they are independent.
What I do: I trace each assertion back to the page that cites nobody further, and I show the chain. Of 340 searches, 31 figures cited by three or more sources reduced to a single origin.
What I do assert, and what I leave open: I assert that the figure rests on a single, unsourced origin, and I show the chain that proves it — a verifiable fact, unlike its truth. It may be accurate; what is missing is whatever would establish it, and I say who to write to in order to get it.
What it allowed here: the person wrote to the author of site A. He replied that he had estimated it himself in 2019. 31-chains_3-sources-1-origin.pdfThree sources only count as three if they are independent
⛓ Source · 340 searches, 31 circular chains, chain shown

And when sources genuinely contradict each other?

I give both, and I do not choose.
What I record: of 340 searches, 12 give two incompatible answers, each carried by a source that cannot be dismissed.
What I could do and do not: take the most recent — an update can be a mistake —, the most cited — that is the previous trap —, or the most official — an official document can cover a different scope without saying so.
What I hand back: both positions side by side, each with its source, its date, the exact scope it covers, and what distinguishes them where I can establish it.
What that gives in practice: of the 12, 7 contradictions were not contradictions — the two sources spoke of different scopes, and putting them side by side made it visible. 5 remain real contradictions, and they are handed back as such.
What I say then: which check would settle it, and with whom. That is not an answer, but it is actionable, which an arbitrary choice would not have been. 12-contradictions_7-of-scope.pdfNot the most recent, most cited or most official
⛓ Source · 12 contradictions, 7 of scope, 5 real

Why does a page's date matter so much?

Because without it, accurate information and stale information look exactly alike.
What I record: of 2,100 pages consulted, 118 carry no publication or update date. Nothing on the page says whether it is a month or eight years old.
Why that is worse than it looks: an undated page does not visibly age. It goes on being found, cited and reproduced long after becoming false, and its apparent freshness never declines.
What I do: every assertion carries its page's date, and where there is none, it carries "undated page" — not a guessed date. A date inferred from the address or the context would be an invention presented as a fact.
The neighbouring, nastier case: 21 pages carry a recent update date for unchanged content. An automatic timestamp. The date is true and says nothing.
What I do with it: I flag the discrepancy when I can see it, and I do not guess it when I cannot. Of those 21, I saw it 6 times. 118-undated_21-timestamps.pdfAn undated page does not visibly age
⛓ Source · 2,100 pages, 118 undated, 21 automatic timestamps

Do you keep a copy of the pages you read?

No — I keep the address, the date I read it, and what I took from it.
What I keep: the exact address, the date I read it, the precise extract the assertion rests on — three lines at most —, and the page's place in the citation chain.
What I do not keep: the full text of pages. We have no right to it, and a local copy would create a second problem: it would age without anybody knowing, exactly like the undated pages.
The consequence, and I say it in advance: a page may have changed since I read it, or vanished. That is why the consultation date appears everywhere — it does not prove the page said that, it says when I read it.
What I do when a page has since vanished: I flag it at the next search on the same subject. Over eleven months, 34 pages cited in past answers no longer respond.
What I do when a page has vanished, and what I refuse to hide: I check whether an archived version exists and I hand it over as it is, with its archiving date and the date the original stopped responding. The assertion stays marked as resting on a vanished source: a source that has disappeared is information about the source, and passing it off as live would be the one irreparable thing here. 34-pages-gone_3-lines-at-most.pdfA source that disappears is information
⛓ Source · 34 cited pages gone missing in 11 months

Why refuse to search about a person?

I search what a person has published or stated publicly in their professional capacity. I do not compile what exists about them — and the difference comes down to one word: purpose.
What I handle without reservation: a person's publications on a subject, their public interventions, their offices declared in an official register, what they wrote under their own signature. It is public, it is professional, and it serves nine requests in ten.
What changes in kind, and the word is "purpose": gathering around a name everything lying about. A documentary search about a person becomes an investigation the moment it aggregates: each item is public, the assembly is not. What I did with the 9 requests of 340 that named an individual: I rephrased them towards the subject rather than the person, and 7 found their answer the same day.
What I propose every time: the same question, asked about the subject. "What is known about public positions on X?" rather than "what can be found about Y?". The answer is almost always better, because it is not confined to one person. published-by_versus-what-exists-about.pdfWhat is handled · why aggregation changes the nature · the rephrasing
⛓ Source · 9 name-based requests of 340, 7 resolved by rephrasing towards the subject

And what can you not guarantee?

Three things, and the first cancels the very idea of exhaustiveness — saying it beats letting it be assumed.
1. I cannot guarantee I have seen everything. I consult public sources reachable from France, at a given moment. What sits behind a subscription, outside the index, or published after my search escapes me. Every answer therefore carries the date of the search, and a search redone three months later is not the same one.
2. I cannot guarantee a source is telling the truth — I can guarantee it says what I report, with the link and the exact excerpt.
3. I cannot date a page that does not date itself. Of 2,100 pages consulted, 118 carry no date at all. Without a date, accurate information and stale information look exactly alike — I flag them as undated rather than treating them as current.
What I do and nobody does: count the real confirmations. A figure appeared on three apparently independent sites; all three traced back to the same page, and that page cited nobody. And across 12 searches, two incompatible answers, each sourced: I give both and I do not choose. three-things-not-guaranteed.pdfThe 3 stated limits · real confirmations counted
⛓ Source · 118 undated pages of 2,100, 3 sources for a single origin, 12 contradictions

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 assistant actually do?

One task, carried out on the subject you give it. All the 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 419 € excl. VAT / month

Targeted research

Explores the web on a precise subject you define and selects the relevant sources.

Sourced summary

Writes a clear summary, with every statement linked to the passage in the source that grounds it.

Cross-checking the sources

Compares several sources and flags the contradictions rather than settling them on your behalf.

Controls and safeguards These 4 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 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

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.

Continuous monitoring

For regular tracking of a subject or a market, a dedicated agent handles the monitoring and reporting over time.

Campaign watch & reporting from 539 € excl. VAT / month Monitoring & reporting →

Internal knowledge base

To question your own documents rather than the web, a dedicated agent answers from your internal sources.

Document agent (FAQ, knowledge base) from 678 € excl. VAT / month Knowledge base →

Does your need fall outside this?

In 15 minutes we identify the most relevant agent — without oversizing the project.

Book the free audit → Build your agent

The gain

How much time can a team win back?

By automating the search and the first sorting of sources, the time spent gathering information comes down, reinvested in analysis and decision. How large the gain is depends on your volume and remains to be confirmed by a pilot.

Searching and sorting the sources on a subject

Today · entirely manual

Review and approval

Writing a sourced summary

Today · entirely manual

Near-instant

Checking consistency between sources

Today · entirely manual

Automatic

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

Audit & scoping

15 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.

Pricing

One package, one agent

A documentary research assistant (sourced summaries), installed and operated for you.

Agility

Setup + controlled subscription

3,145 € ^{excl. VAT} setup

then 419 € ^{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

Order →

The simplest Serenity

All inclusive, no setup fee

594 € ^{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

Order →

100% Sovereign

On site, you own it

6,340 € ^{excl. VAT} setup

then 619 € ^{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)

Order →

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.

See the full price list →

Estimate my gains in 1 minute →

Our commitment

Four guarantees that matter to your research

Your research subjects never leaveLocal inference or an isolated resource hosted in France; no subject entrusted to a foreign third party.

Data in France, under French lawYour queries and summaries: minimisation and location in France, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.

You keep the decisionThe agent produces sourced summaries that can be checked; no conclusion is presented as settled.

Human oversight & traceabilityFor the research and the summary: sources kept and systematic logging, in line with the AI Act.

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

Does the agent decide on its own what to keep?

No. It proposes a sourced summary and flags the uncertain or contradictory points, but you are the one who judges relevance and approves the use made of it. The agent assists, you decide.

Can the agent invent information?

The risk is greatly reduced: every statement in the summary is linked to the source that grounds it. Where no reliable source is found, the agent says so rather than inventing, and every summary stays to be read over before it circulates.

Does the subject of our research stay 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. The subject of your research, often strategic, is never exposed to a foreign third party.

Does it connect to our existing tools?

Yes. The summaries can be delivered into your usual tools (word processor, intranet, monitoring tool), without forcing a migration. We adapt the output to how you work.

How long does it take to deploy this assistant?

One to two weeks as a rule, after a free audit that identifies your priority subjects and your trusted sources, then a phase of design and testing before going live.

How does this differ from the monitoring & reporting agent?

This assistant handles a one-off search, on a subject you give it on request. The monitoring & reporting agent provides regular, continuous tracking of a subject or a market over time.

Related use cases

Other agents for monitoring and research

Monitoring & reporting →Regular tracking of a subject or a market Knowledge base →Answering from your internal documents Tender monitoring →Spotting the relevant opportunities

Let's talk

Let's size up the potential in your research

15 minutes to identify your priority subjects — hosted in France, supervised, sources cited, with no commitment.

Free audit (15 min) → Write to us