● Public sector — Support for the labour inspectorate
AI agent supporting the labour inspectorate
An assistant that prepares regulatory letters, handles the documents and answers enquiries — the inspector keeps the analysis, the inspection and the decision. A strictly administrative and documentary scope: no support for inspections or for the decisions that follow. Hosted in France, on an isolated resource.
Hosted in France
Case data protected
GDPR & AI Act: governed deployment
No support for inspections
Updated on
Deployed in a few weeks
Labour inspectorate support · hosted in France
An employer asks what must be displayed on the premises in a company with fewer than 50 staff. Prepare an informative reply.
Informative reply drafted. It sets out what must be displayed and circulated (working hours, contact details for the inspectorate and occupational health, safety instructions, workplace equality…) with the references in the labour code, and points to the official resources.
A general reply, with no assessment of an individual situation — ready for review.
⛓ Source · the labour code + your information sheets
A general reply, with no assessment of an individual situation — ready for review.
⛓ Source · the labour code + your information sheets
Prepare the acknowledgement for this report and file the case documents according to our filing plan.
Documentary support carried out. Acknowledgement drafted on your template, 12 documents named and filed according to your plan (letters, contracts, payslips), a contents list generated with a cross-reference to each document.
No analysis of the substance: the assessment and what follows rest exclusively with the inspector.
✎ Action · case filed, letter ready — the inspector approves
No analysis of the substance: the assessment and what follows rest exclusively with the inspector.
✎ Action · case filed, letter ready — the inspector approves
Local inference · no data outside the EU
Data 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 inspector decidesAdministrative support — never the inspection
✦ In brief
For the labour inspectorate, a Blue Lemon Agent agent provides strictly administrative and documentary support: regulatory letters and notifications prepared from your templates, documentary handling of case files (documents, references to the labour code), informative replies to the public — employers and employees alike. It never takes part in inspection work or in the decisions that follow: no support for targeting, assessment or decision. It runs on local inference or is hosted in France on a dedicated, isolated resource, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and employers and employees reach it on WhatsApp Business, the website chat or email — with no account to create and nothing to install. Reaching the administration from the tool people already have means less non-take-up of rights and equal access to the service. These connections are included in every plan, at no extra cost, within the number of connections your level includes.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
7
uses ready to deploy on this scope
0
decision taken without human approval
Reference points describing our offer, not results measured at a client. The scale of the gain is confirmed by a pilot on your own scope.
The context
Why AI matters to inspection services — and why they hesitate
Labour inspectorate units handle a considerable volume of letters, reports and enquiries, with constrained staffing. Every hour spent on administration is an hour less for the core mission — and the case data (companies, employees) is highly confidential.
! The issue
Officers are caught between members of the public — employers and employees — who expect quick answers, and a heavy administrative load: regulatory letters, filing documents, documentary tracking. Yet most consumer AI tools amount to entrusting reports, company data and employees' situations to a third party, often hosted outside Europe and subject to the Cloud Act — and they blur the boundary with inspection work.
✓ Our answer
For the labour inspectorate, AI is only of interest if it is sovereign, confidential and strictly confined. Local inference or an isolated resource hosted in France, systematic human oversight, and a scope defined by design: administrative and documentary support only, never support for inspections, targeting or decisions. The aim is not to assist the inspection mission, but to give back the time to carry it out fully.
The decisive point
Protecting the case data: sovereignty & compliance
Reports, company data, employees' situations: the inspectorate's files are highly confidential. Here is how the architecture of our agents protects them, unit by unit.
Local inference
The agent can run on a machine at the unit: no data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — the case data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For the case data, 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.
One isolated resource per department
No pooling of data: an environment strictly dedicated to your unit, guaranteeing the confidentiality of the files.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
The agent is strictly in administrative support; no involvement in inspections or what follows; traceability 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 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.
For the most sensitive data, SecNumCloud and HDS options are available 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.
The public body in this demonstration
Fictional public bodyVallonnais departmental employment and labour directorate — inspection unit no. 3 (fictional service)
- Sector
- Inspection unit of a departmental employment, labour and solidarity directorate — 6,200 establishments in its area, 74,000 employees; dominant sectors: construction, food processing, road haulage, home care
- Headcount
- 14 staff in the unit, including 9 labour inspectors, 3 unit secretariat staff in the pilot and one unit manager; the departmental labour-law information service stays outside the scope, and the agent does not touch it
- Public served
- 9,400 enquiries a year — 5,600 calls, 2,400 emails, 900 counter visits, 500 website forms; employers and employees in almost equal shares, 54 % against 46 %
- Order of magnitude
- 2,180 regulatory letters drafted a year, 1,460 reports received, 3,100 documents filed, 214 legal texts tracked by the watch
- Tools already in place
- The unit's case application (establishment files, follow-up given), filing plan and letter templates settled by the unit manager, electronic signature folder, Microsoft Teams and email — the agent plugs into them read-only, nothing is replaced or migrated
- Who decides what
- The labour inspector organises inspections, makes the findings and decides what follows — article L8112-1 of the Labour Code; the unit manager settles the templates and the filing plan. The agent prepares, numbers, sources and dates; it writes no finding of fact
- Room for improvement
- 1,372 calls out of 5,600 ring unanswered — 24.5 %; an information email waits 12 days on average; document filing is 100 % manual, 2 staff for 4 days a month, that is 56 hours a month and 672 hours over the year; 604 reports out of 1,460 arrive without any of the three items that make them workable
At inspection unit no. 3 of the Vallonnais, fourteen staff cover 6,200 establishments and 74,000 employees, and receive 1,460 reports and 9,400 enquiries a year. The most repetitive questions — which notices must be displayed, how overtime is paid, where to file a report, which documents to produce — take up the largest share of the flow, and the situations that actually need an inspector queue behind them. The exchanges below cover one year, from the survey of the flow to the review presented at the management meeting. The agent is hosted in France on an isolated resource, reads the case application without ever writing to it, and its scope is strictly administrative and documentary: it prepares letters that the inspector rereads and signs, numbers the documents in the files, serves general law to the public — and it writes no finding of fact, characterises no offence and produces no targeting list.
This public body, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real service.
Labour inspectorate support · hosted in France
I have reread your 9,400 enquiries for the year and gone through your 1,460 reports one by one against the three items that make them workable — the survey the inspection workload never leaves time for.
What the count says: 11 topics cover 6,580 enquiries out of 9,400, that is 70.0 % of the flow. Compulsory notices and disclosures, overtime and its premium, paid leave, maximum working time and rest, termination of contract and end-of-contract papers, pay and the payslip, filing a report and what becomes of it, the prevention and information medical visit, the single staff register, employee representation, access to the applicable collective agreement. The 11 reference answers are written, sourced by article number, and dated in the version in force — they are waiting for you.
What the three items weigh, and you can check each on your own logs:
· A regulatory letter drafted from one of your templates takes 12 minutes end to end, of which 60 % is re-entering the template, the establishment's identification and the references — 7 minutes 12.
· An information reply to a member of the public takes 9 minutes, of which 40 % is the drafting itself — 3 minutes 36.
· Document filing is 100 % manual work: 2 staff for 4 working days each month, that is 56 hours a month and 672 hours over the year.
And here is the survey nobody had the days to run: out of your 1,460 reports for the year, 604 — 41.4 % — bring none of the three items that make them workable: identification of the establishment, date or period of the facts, job and length of service of the person concerned. This is not a management figure: each of those 604 files goes back out as a request for further information, comes back 21 days later on average, and some never come back at all.
The proposal, and it takes half an hour of your time: I have written you a qualified acknowledgement of receipt that says, the minute the report arrives, exactly which of the three items is missing and where to find it. The unit manager settles the wording, and it goes out on next month's reports. inspection-unit-flow_9400-enquiries-11-topics.pdf6,580 enquiries across 11 topics, the three items costed, 1,051 hours identified
⛓ Sourced · 12 months of enquiry log, register of reports, the unit's filing plan
What the count says: 11 topics cover 6,580 enquiries out of 9,400, that is 70.0 % of the flow. Compulsory notices and disclosures, overtime and its premium, paid leave, maximum working time and rest, termination of contract and end-of-contract papers, pay and the payslip, filing a report and what becomes of it, the prevention and information medical visit, the single staff register, employee representation, access to the applicable collective agreement. The 11 reference answers are written, sourced by article number, and dated in the version in force — they are waiting for you.
What the three items weigh, and you can check each on your own logs:
· A regulatory letter drafted from one of your templates takes 12 minutes end to end, of which 60 % is re-entering the template, the establishment's identification and the references — 7 minutes 12.
· An information reply to a member of the public takes 9 minutes, of which 40 % is the drafting itself — 3 minutes 36.
· Document filing is 100 % manual work: 2 staff for 4 working days each month, that is 56 hours a month and 672 hours over the year.
And here is the survey nobody had the days to run: out of your 1,460 reports for the year, 604 — 41.4 % — bring none of the three items that make them workable: identification of the establishment, date or period of the facts, job and length of service of the person concerned. This is not a management figure: each of those 604 files goes back out as a request for further information, comes back 21 days later on average, and some never come back at all.
The proposal, and it takes half an hour of your time: I have written you a qualified acknowledgement of receipt that says, the minute the report arrives, exactly which of the three items is missing and where to find it. The unit manager settles the wording, and it goes out on next month's reports. inspection-unit-flow_9400-enquiries-11-topics.pdf6,580 enquiries across 11 topics, the three items costed, 1,051 hours identified
⛓ Sourced · 12 months of enquiry log, register of reports, the unit's filing plan
604 incomplete reports out of 1,460 — what does that mean over time? And how do we stop it happening again?
Over 24 months the same reading brings out 1,208. The cost is not counted in euros but in delay: 21 days of back-and-forth per file, on situations where the worker wrote in because they no longer knew who to turn to.
What I did on top, and what the inspection workload never leaves time for: I wrote four completeness rules, in the language of your filing plan, and ran them over your last 24 months of reports. For each one I give you the number of flags it would have raised, the share confirmed on examination, and what it would have let through — you choose on figures, not on a hunch:
· Rule A — report without workable identification of the establishment (company name and company number, or company name and address): 486 flags over 24 months, 471 confirmed — 96.9 %. The other 15 carried the trading name, which your application already matches.
· Rule B — report with no date or period of the facts: 402 flags, 381 confirmed — 94.8 %.
· Rule C — report concerning a named person, without their job or length of service: 264 flags, 222 confirmed — 84.1 %. It is the least reliable of the four, and I am giving it to you as such.
And here is what those three rules let through, because you need to know it before you sign: 218 reports that were complete in form announced a supporting document that never arrived — a payslip, a rota, a message thread. None of the three sees them. So I wrote rule D — a document announced in the body of the report and absent from the file 15 days after the acknowledgement — and ran it over the same 24 months: 218 flags, 196 confirmed — 89.9 %.
The four rules together cover 1,124 of the period's 1,208 incomplete files — 93.0 %, that is 562 of the year's 604.
Bringing a rule into service stays with the unit: a rule only takes effect once the unit manager has validated it. I save you the writing and the measurement; the decision takes half an hour instead of a working group. four-completeness-rules_tested-over-24-months.pdf4 rules tested, 1,124 files covered out of 1,208, what each one lets through
⛓ Sourced · 24 months of the register of reports, filing plan, acknowledgements of receipt
What I did on top, and what the inspection workload never leaves time for: I wrote four completeness rules, in the language of your filing plan, and ran them over your last 24 months of reports. For each one I give you the number of flags it would have raised, the share confirmed on examination, and what it would have let through — you choose on figures, not on a hunch:
· Rule A — report without workable identification of the establishment (company name and company number, or company name and address): 486 flags over 24 months, 471 confirmed — 96.9 %. The other 15 carried the trading name, which your application already matches.
· Rule B — report with no date or period of the facts: 402 flags, 381 confirmed — 94.8 %.
· Rule C — report concerning a named person, without their job or length of service: 264 flags, 222 confirmed — 84.1 %. It is the least reliable of the four, and I am giving it to you as such.
And here is what those three rules let through, because you need to know it before you sign: 218 reports that were complete in form announced a supporting document that never arrived — a payslip, a rota, a message thread. None of the three sees them. So I wrote rule D — a document announced in the body of the report and absent from the file 15 days after the acknowledgement — and ran it over the same 24 months: 218 flags, 196 confirmed — 89.9 %.
The four rules together cover 1,124 of the period's 1,208 incomplete files — 93.0 %, that is 562 of the year's 604.
Bringing a rule into service stays with the unit: a rule only takes effect once the unit manager has validated it. I save you the writing and the measurement; the decision takes half an hour instead of a working group. four-completeness-rules_tested-over-24-months.pdf4 rules tested, 1,124 files covered out of 1,208, what each one lets through
⛓ Sourced · 24 months of the register of reports, filing plan, acknowledgements of receipt
And what about the legal texts? Our watch rests on one person, and she goes on maternity leave in March.
Your watch tracks 214 texts a year on your scope, and I hold it on six declared sources: Légifrance for the codes and decrees, the Official Journal, the Labour Ministry's official bulletin, the extended collective agreements of your four dominant sectors, the instructions of the Directorate General of Labour, and the prefecture's register of administrative acts.
What I publish against myself, because you need to know it before you rely on me: out of the year's 214 texts, my watch missed 9 — 4.2 %. The explanation is not an excuse, it is a cause: all nine are prefectoral orders — working-time derogations, authorisations for young workers — published in the register of administrative acts, a source I was not querying because it was not on the starting list.
What I did about it: the source has been connected since 12 April, and I caught up the previous 24 months — 31 orders recovered, 4 of which touch an article cited by a template in service. Those four templates are marked « to be reviewed » in the signature folder; I did not modify them, the unit manager settles the templates.
The figure that matters for March: the delay between a text being published and reaching your templates has gone from 9 days to under 24 hours, and it no longer depends on one person being at their desk. What I propose for the leave period: I produce a two-page watch note every Monday — what has been published, what it touches here, which template is concerned — and you decide whether you want it with or without the case-law review, which adds twenty minutes of reading. regulatory-watch_6-sources-214-texts-and-the-9-missed.pdfThe 6 sources, the 9 texts missed and their cause, the 24-month catch-up
⛓ Sourced · six declared watch sources, 24 months caught up, the unit's template register
What I publish against myself, because you need to know it before you rely on me: out of the year's 214 texts, my watch missed 9 — 4.2 %. The explanation is not an excuse, it is a cause: all nine are prefectoral orders — working-time derogations, authorisations for young workers — published in the register of administrative acts, a source I was not querying because it was not on the starting list.
What I did about it: the source has been connected since 12 April, and I caught up the previous 24 months — 31 orders recovered, 4 of which touch an article cited by a template in service. Those four templates are marked « to be reviewed » in the signature folder; I did not modify them, the unit manager settles the templates.
The figure that matters for March: the delay between a text being published and reaching your templates has gone from 9 days to under 24 hours, and it no longer depends on one person being at their desk. What I propose for the leave period: I produce a two-page watch note every Monday — what has been published, what it touches here, which template is concerned — and you decide whether you want it with or without the case-law review, which adds twenty minutes of reading. regulatory-watch_6-sources-214-texts-and-the-9-missed.pdfThe 6 sources, the 9 texts missed and their cause, the 24-month catch-up
⛓ Sourced · six declared watch sources, 24 months caught up, the unit's template register
Three draft letters are in your signature folder, prepared on your templates, and I am telling you exactly what is filled in and what is not.
What is filled in, and checkable line by line:
· The establishment's identification — company name, company number, address of the workplace concerned, and the headcount at the date of the facts, not today's headcount: the template calls for different wording depending on the threshold, and that is the first error manual re-entry produces.
· The applicable articles, cited by number, in the version in force on the day of the facts — with the date of that version. A text amended since does not apply retroactively, and a letter citing today's version for last year's facts is contested without effort.
· The time limit, its basis of calculation and its starting point, the routes and time limits for appeal, the template's compulsory wording, and the list of enclosures numbered in the file's own order.
What arrives empty, and deliberately so: the « findings » box. You write it, because you are the one who saw.
The figure: a letter used to take 12 minutes, of which 7 minutes 12 was mechanical re-entry; it now takes you 1 minute 12 of rereading. Across 2,180 letters, that is 218 hours returned over the year — more than six weeks of 35 hours, rounded down.
The next step I propose: your formal-notice templates cite three articles that were rewritten in January. I have prepared the three updated versions alongside the old one, with the changes highlighted — the unit manager settles it in ten minutes. three-draft-letters_what-is-filled-in-what-stays-empty.pdfThree drafts on template, articles dated in the applicable version, « findings » box empty
⛓ Sourced · the unit's templates, establishment files, dated consolidated versions from Légifrance
What is filled in, and checkable line by line:
· The establishment's identification — company name, company number, address of the workplace concerned, and the headcount at the date of the facts, not today's headcount: the template calls for different wording depending on the threshold, and that is the first error manual re-entry produces.
· The applicable articles, cited by number, in the version in force on the day of the facts — with the date of that version. A text amended since does not apply retroactively, and a letter citing today's version for last year's facts is contested without effort.
· The time limit, its basis of calculation and its starting point, the routes and time limits for appeal, the template's compulsory wording, and the list of enclosures numbered in the file's own order.
What arrives empty, and deliberately so: the « findings » box. You write it, because you are the one who saw.
The figure: a letter used to take 12 minutes, of which 7 minutes 12 was mechanical re-entry; it now takes you 1 minute 12 of rereading. Across 2,180 letters, that is 218 hours returned over the year — more than six weeks of 35 hours, rounded down.
The next step I propose: your formal-notice templates cite three articles that were rewritten in January. I have prepared the three updated versions alongside the old one, with the changes highlighted — the unit manager settles it in ten minutes. three-draft-letters_what-is-filled-in-what-stays-empty.pdfThree drafts on template, articles dated in the applicable version, « findings » box empty
⛓ Sourced · the unit's templates, establishment files, dated consolidated versions from Légifrance
On the Ravel Logistique file, can you just write the findings and characterise what we saw? It would save us a huge amount of time.
No — and for once this is not caution, it is a text. Article L8113-7 of the French Labour Code provides that « labour inspection control officers referred to in article L. 8112-1 and equivalent control officials establish offences by procès-verbaux which stand as evidence unless the contrary is proved ». That evidential weight does not come from the quality of the drafting: it comes from a sworn officer having seen, and written it. A finding drafted by me would lose precisely what gives it its value before the public prosecutor.
And here is the route that gives you back the time you are after, on this file, this morning:
· The chronology of the Ravel Logistique file is built — 34 documents numbered, dated, ordered, each with its issuer and its date of arrival at the unit; the 6 documents carrying the facts of 12 and 13 February are set out at the front.
· The establishment's headcount on 12 February is established and sourced — single staff register and declarations, with the two documents that evidence it.
· The applicable articles are cited by number in the version in force on 12 February, with the link to the consolidated version and its date.
· And where the text allows two readings, I give you both, with what separates them — on this file, how travel time between two sites is treated depends on the actual work organisation, which the chronology does not settle: both readings are written out, each with the document that would be needed to settle it.
Measured against your usual 40 minutes of formatting: 34 are returned. What is left is the six lines of the « findings » box — and they are the only ones that require the person who was there.
✎ Framework · French Labour Code, art. L8113-7 and L8112-1 · establishment file, 34 documents numbered
And here is the route that gives you back the time you are after, on this file, this morning:
· The chronology of the Ravel Logistique file is built — 34 documents numbered, dated, ordered, each with its issuer and its date of arrival at the unit; the 6 documents carrying the facts of 12 and 13 February are set out at the front.
· The establishment's headcount on 12 February is established and sourced — single staff register and declarations, with the two documents that evidence it.
· The applicable articles are cited by number in the version in force on 12 February, with the link to the consolidated version and its date.
· And where the text allows two readings, I give you both, with what separates them — on this file, how travel time between two sites is treated depends on the actual work organisation, which the chronology does not settle: both readings are written out, each with the document that would be needed to settle it.
Measured against your usual 40 minutes of formatting: 34 are returned. What is left is the six lines of the « findings » box — and they are the only ones that require the person who was there.
✎ Framework · French Labour Code, art. L8113-7 and L8112-1 · establishment file, 34 documents numbered
And day-to-day document filing, what does that look like?
3,100 documents came in this year, and each reached its file numbered, dated and attached within the minute. For each one: date of arrival at the unit, named issuer, reference under your filing plan, the establishment file it attaches to, and the file's contents list regenerated — a wrong contents list is what costs twenty minutes in front of a waiting employer.
The figure: filing was 100 % manual — 56 hours a month, 672 hours over the year. What is left is checking and validation, 20 % — 134 hours. 537 hours returned.
What I publish against myself, and you have it in front of you in the signature folder: out of 1,118 draft letters I prepared this year, 134 were rewritten before signature — 12.0 %.
The explanation, measured rather than assumed: 98 of the 134 turn on a single point — which version of an article applies on the day of the facts, when the facts straddle an amendment. I was taking the version in force on the day the file was opened; your inspectors take the version in force on the day of the fact under consideration. You are right, and the disagreement only showed up because every rewrite is logged.
What I did about it: since 3 June, a file whose facts straddle an amendment goes out with both versions side by side and the switchover date; the inspector picks the right one. Across the 214 drafts produced since, 9 were rewritten — 4.2 %, and none on that ground. A system that does not record its rewrites never improves; this one counts them and shows them to you. document-handling_3100-items-numbered-and-the-134-rewrites.pdfNumbering, contents list, chronology; 134 rewrites analysed, the cause and the fix
⛓ Sourced · the unit's filing plan, signature-folder rewrite log, 3,100 documents filed
The figure: filing was 100 % manual — 56 hours a month, 672 hours over the year. What is left is checking and validation, 20 % — 134 hours. 537 hours returned.
What I publish against myself, and you have it in front of you in the signature folder: out of 1,118 draft letters I prepared this year, 134 were rewritten before signature — 12.0 %.
The explanation, measured rather than assumed: 98 of the 134 turn on a single point — which version of an article applies on the day of the facts, when the facts straddle an amendment. I was taking the version in force on the day the file was opened; your inspectors take the version in force on the day of the fact under consideration. You are right, and the disagreement only showed up because every rewrite is logged.
What I did about it: since 3 June, a file whose facts straddle an amendment goes out with both versions side by side and the switchover date; the inspector picks the right one. Across the 214 drafts produced since, 9 were rewritten — 4.2 %, and none on that ground. A system that does not record its rewrites never improves; this one counts them and shows them to you. document-handling_3100-items-numbered-and-the-134-rewrites.pdfNumbering, contents list, chronology; 134 rewrites analysed, the cause and the fix
⛓ Sourced · the unit's filing plan, signature-folder rewrite log, 3,100 documents filed
The 11 reference answers have been in service since January, and they have served 6,580 enquiries out of 9,400. Each one carries the article by number, the date of the version cited, and a note on what may depart from it — an extended collective agreement, a company-level agreement. That last point is what separates a right answer from a wrong one.
Where the public reaches me, with no account to create and nothing to install: WhatsApp Business, the website chat and email for employers and employees; Microsoft Teams, Slack or their mailbox for your own staff. WhatsApp and the phone reach people the online form never reaches, and over the year they carried 2,940 enquiries that last year would never have arrived — that is unclaimed rights avoided, not volume.
And the announcement is not a configuration option: since 2 August 2026, article 50(1) of the European AI Regulation requires that anyone interacting with an AI system be informed. I introduce myself as an AI agent in the first message, in one short sentence, and they can ask for a human at any point — 418 did so this year, and all 418 were handed over to a member of staff.
The figure: a reply used to take 3 minutes 36 of drafting; you are left with 54 seconds of rereading. Across 6,580 replies, 296 hours returned. And the information email now takes 1 day instead of 12. eleven-reference-answers_article-version-and-date.pdfThe 11 topics, the article cited by number, the dated version and what may depart from it
⛓ Sourced · 11 reference answers sourced and dated, channel log, register of requests for a human contact
Where the public reaches me, with no account to create and nothing to install: WhatsApp Business, the website chat and email for employers and employees; Microsoft Teams, Slack or their mailbox for your own staff. WhatsApp and the phone reach people the online form never reaches, and over the year they carried 2,940 enquiries that last year would never have arrived — that is unclaimed rights avoided, not volume.
And the announcement is not a configuration option: since 2 August 2026, article 50(1) of the European AI Regulation requires that anyone interacting with an AI system be informed. I introduce myself as an AI agent in the first message, in one short sentence, and they can ask for a human at any point — 418 did so this year, and all 418 were handed over to a member of staff.
The figure: a reply used to take 3 minutes 36 of drafting; you are left with 54 seconds of rereading. Across 6,580 replies, 296 hours returned. And the information email now takes 1 day instead of 12. eleven-reference-answers_article-version-and-date.pdfThe 11 topics, the article cited by number, the dated version and what may depart from it
⛓ Sourced · 11 reference answers sourced and dated, channel log, register of requests for a human contact
An employee writes on WhatsApp: « my employer hasn't paid my March overtime, what can I do? » What exactly do you answer?
I answer in general law, within the minute, and I build the individual file for you. Here are both, kept apart.
What goes to the employee — law, not an assessment of their situation:
· Hours worked beyond the statutory weekly duration give rise to a premium. Failing a company-level agreement or a sector-wide collective agreement, article L3121-36 of the Labour Code sets 25 % for each of the first eight overtime hours and 50 % for those beyond — and I tell them expressly that their company agreement or sector agreement may set a different rate, with the link to consult it.
· The time to act: article L3245-1 provides a three-year limitation period running from the day the claimant knew or should have known the facts, the claim being able to cover sums due for the last three years.
· The documents that make the proof — payslips for the period, rotas, clock records, message threads — and the three routes open to them: a written request to the employer, a report to the inspection unit, an application to the labour tribunal, with what each produces and how long it takes.
What does not go to the employee, because it is not mine to tell them: whether their employer is at fault. That is the assessment of a named situation, and it belongs to you.
And here is how I hand it to you in minutes rather than weeks: the information sheet is open, the establishment identified and matched against your application — three reports on the same establishment in eighteen months, all three on working time —, the period of the facts bounded, and the three missing documents requested from the employee in the same message. 812 individual requests took that route this year; they reach you built, not to be built.
⛓ Sourced · French Labour Code, art. L3121-36 and L3245-1, dated consolidated versions · the unit's case application
What goes to the employee — law, not an assessment of their situation:
· Hours worked beyond the statutory weekly duration give rise to a premium. Failing a company-level agreement or a sector-wide collective agreement, article L3121-36 of the Labour Code sets 25 % for each of the first eight overtime hours and 50 % for those beyond — and I tell them expressly that their company agreement or sector agreement may set a different rate, with the link to consult it.
· The time to act: article L3245-1 provides a three-year limitation period running from the day the claimant knew or should have known the facts, the claim being able to cover sums due for the last three years.
· The documents that make the proof — payslips for the period, rotas, clock records, message threads — and the three routes open to them: a written request to the employer, a report to the inspection unit, an application to the labour tribunal, with what each produces and how long it takes.
What does not go to the employee, because it is not mine to tell them: whether their employer is at fault. That is the assessment of a named situation, and it belongs to you.
And here is how I hand it to you in minutes rather than weeks: the information sheet is open, the establishment identified and matched against your application — three reports on the same establishment in eighteen months, all three on working time —, the period of the facts bounded, and the three missing documents requested from the employee in the same message. 812 individual requests took that route this year; they reach you built, not to be built.
⛓ Sourced · French Labour Code, art. L3121-36 and L3245-1, dated consolidated versions · the unit's case application
And the people who read French poorly, or who read poorly full stop? We have many of them in food processing and home care.
1,240 of your 9,400 enquiries — 13.2 % — come from people who do not read French fluently or who struggle with administrative writing, and they are recognisable by one simple sign: they come back on the same topic 2.4 times on average. A correct answer nobody understands is an answer that comes past the counter three times.
What is produced, and submitted for human validation: for each of the 11 reference answers, a plain-language version — short sentences, one idea per sentence, everyday words, the legal term kept and explained next to it — and a draft transcription following easy-read rules: one piece of information per line, no metaphor, no undefined acronym, one illustration per concept. And the legal term is never replaced: it stays written, with its definition alongside. An answer that lost the exact word would also lose the service user's foothold in any appeal — that is the one thing the rewriting does not touch.
What I publish against myself: out of my first 84 easy-read versions, 19 were rejected by your review group — 22.6 %. The cause, and it is a clean one: 16 of the 19 had kept a conditional argument — « if you are in this situation, then… » — which the easy-read rules require to be split into separate situations. I was writing correct and unreadable.
What I did about it: any answer containing a condition now goes out as one block per situation, each with its own heading. Across the 52 versions since, 3 rejections — 5.7 %. And the starting rule has not moved: no easy-read version goes out before the review group validates it, because the people concerned are the ones who know whether it reads, not me.
The result you will see first: on the topics moved to plain language, repeat enquiries fell from 2.4 to 1.3 per person. plain-language-and-easy-read_the-rule-the-review-and-the-rejections.pdfRules applied, 84 versions reviewed, 19 rejections analysed and the fix, a readable before and after
⛓ Sourced · repeat-enquiry log, minutes of the easy-read review group, version register
What is produced, and submitted for human validation: for each of the 11 reference answers, a plain-language version — short sentences, one idea per sentence, everyday words, the legal term kept and explained next to it — and a draft transcription following easy-read rules: one piece of information per line, no metaphor, no undefined acronym, one illustration per concept. And the legal term is never replaced: it stays written, with its definition alongside. An answer that lost the exact word would also lose the service user's foothold in any appeal — that is the one thing the rewriting does not touch.
What I publish against myself: out of my first 84 easy-read versions, 19 were rejected by your review group — 22.6 %. The cause, and it is a clean one: 16 of the 19 had kept a conditional argument — « if you are in this situation, then… » — which the easy-read rules require to be split into separate situations. I was writing correct and unreadable.
What I did about it: any answer containing a condition now goes out as one block per situation, each with its own heading. Across the 52 versions since, 3 rejections — 5.7 %. And the starting rule has not moved: no easy-read version goes out before the review group validates it, because the people concerned are the ones who know whether it reads, not me.
The result you will see first: on the topics moved to plain language, repeat enquiries fell from 2.4 to 1.3 per person. plain-language-and-easy-read_the-rule-the-review-and-the-rejections.pdfRules applied, 84 versions reviewed, 19 rejections analysed and the fix, a readable before and after
⛓ Sourced · repeat-enquiry log, minutes of the easy-read review group, version register
The year is closed, and here is the account, item by item, with the calculation rule in front of each figure.
· Regulatory letters — 2,180 letters × 6 minutes returned = 218 hours. The letter used to take 12 minutes, of which 7 minutes 12 was mechanical re-entry; it now takes 1 minute 12.
· Information replies — 6,580 replies × 2 minutes 42 returned = 296 hours. The reply used to take 3 minutes 36 of drafting; 54 seconds are left.
· Document filing — 672 hours by hand, 134 hours of checking and validation: 537 hours.
Total: 1,051 hours returned over the year. At 35 hours a week that is more than thirty weeks — rounded down, the exact count is 30.0 weeks; in months of 151.67 hours, more than six months. The reference figure remains the 1,051 hours; the conversion illustrates it, it does not replace it.
And what I deliberately leave out of the total: the watch hours. The time between a text appearing and it reaching your templates has gone from 9 days to under 24 hours, but I have no reliable measure of the time that returns to you, and I would rather give you a total you can check line by line than a total that inflates.
Where those hours went, and that is the only thing that counts: the information email now takes 1 day instead of 12; unanswered calls fell from 1,372 to 214 out of 5,600 — 3.8 % against 24.5 %; and your nine inspectors carried out 61 more interventions than last year. Those 61 interventions were their choice: I proposed no target, and I will not propose one — the next exchange tells you why, mechanically. year-in-review_1051-hours-returned-item-by-item.pdfThe three items itemised, the calculation rule for each, the conversions and their rounding
⛓ Sourced · the year's logs, signature folder, case application · working-time basis 35 h / 151.67 h / 1,607 h
· Regulatory letters — 2,180 letters × 6 minutes returned = 218 hours. The letter used to take 12 minutes, of which 7 minutes 12 was mechanical re-entry; it now takes 1 minute 12.
· Information replies — 6,580 replies × 2 minutes 42 returned = 296 hours. The reply used to take 3 minutes 36 of drafting; 54 seconds are left.
· Document filing — 672 hours by hand, 134 hours of checking and validation: 537 hours.
Total: 1,051 hours returned over the year. At 35 hours a week that is more than thirty weeks — rounded down, the exact count is 30.0 weeks; in months of 151.67 hours, more than six months. The reference figure remains the 1,051 hours; the conversion illustrates it, it does not replace it.
And what I deliberately leave out of the total: the watch hours. The time between a text appearing and it reaching your templates has gone from 9 days to under 24 hours, but I have no reliable measure of the time that returns to you, and I would rather give you a total you can check line by line than a total that inflates.
Where those hours went, and that is the only thing that counts: the information email now takes 1 day instead of 12; unanswered calls fell from 1,372 to 214 out of 5,600 — 3.8 % against 24.5 %; and your nine inspectors carried out 61 more interventions than last year. Those 61 interventions were their choice: I proposed no target, and I will not propose one — the next exchange tells you why, mechanically. year-in-review_1051-hours-returned-item-by-item.pdfThe three items itemised, the calculation rule for each, the conversions and their rounding
⛓ Sourced · the year's logs, signature folder, case application · working-time basis 35 h / 151.67 h / 1,607 h
In practice, what do you do without asking us?
Three actions, not one more, and all three come undone.
· The qualified acknowledgement of receipt. A report arrives, the acknowledgement goes out within the minute carrying the date of arrival, the file reference, which of the three items is missing and where to find it. It undoes itself: if the item arrives the next day, the acknowledgement updates and no chaser goes out.
· Numbering and filing to the case file. Every incoming document is numbered, dated, attached to its establishment file, and the contents list is regenerated. The inspector moves it with a word, giving the reason — and that reason corrects the classification the following quarter: that is how the first quarter's 47 moves fell to 6 in the second, once the « third-party document » category had been split.
· The watch flag on a template. A published text touching an article cited by a template in service marks that template « to be reviewed » in the signature folder. The template is not modified: the unit manager settles templates, and she dismisses the flag with a word and her reason.
Everything else waits for a named decision, and the attachment lists them line by line: sending a letter, making a finding, characterising an offence, deciding what follows, signing, bringing a rule into service, changing a template. None of the three actions touches the content of a file or the meaning of a letter: they touch its form, its date and its place. three-automatic-actions_and-how-each-one-is-undone.pdfThe three actions, their trigger, what goes out, how each is undone — and who decides the rest
✎ Framework · settings fixed by the unit manager, log of moves and dismissals
· The qualified acknowledgement of receipt. A report arrives, the acknowledgement goes out within the minute carrying the date of arrival, the file reference, which of the three items is missing and where to find it. It undoes itself: if the item arrives the next day, the acknowledgement updates and no chaser goes out.
· Numbering and filing to the case file. Every incoming document is numbered, dated, attached to its establishment file, and the contents list is regenerated. The inspector moves it with a word, giving the reason — and that reason corrects the classification the following quarter: that is how the first quarter's 47 moves fell to 6 in the second, once the « third-party document » category had been split.
· The watch flag on a template. A published text touching an article cited by a template in service marks that template « to be reviewed » in the signature folder. The template is not modified: the unit manager settles templates, and she dismisses the flag with a word and her reason.
Everything else waits for a named decision, and the attachment lists them line by line: sending a letter, making a finding, characterising an offence, deciding what follows, signing, bringing a rule into service, changing a template. None of the three actions touches the content of a file or the meaning of a letter: they touch its form, its date and its place. three-automatic-actions_and-how-each-one-is-undone.pdfThe three actions, their trigger, what goes out, how each is undone — and who decides the rest
✎ Framework · settings fixed by the unit manager, log of moves and dismissals
And does the line with inspection itself hold over time? One day someone will ask you for a list of establishments to visit.
It holds because it is mechanical, not because it is promised. And on the targeting list, my answer is not a moral one but a description of what would happen:
· A targeting score becomes a target the moment it exists. What is measured is what rises: establishments the indicator scores well would stop being visited, and the indicator would stop saying anything about reality — you would lose the instrument at the very moment you used it.
· And article L8112-1 of the Labour Code names exactly what would disappear: control officers « are free to organise and conduct inspections on their own initiative and decide what follow-up to give them ». That freedom is what makes an inspection stand up; a list produced by a system would replace it with a queue.
What I do instead, and it answers the real need: once an inspector has chosen the establishment, the file is built before they leave — history of reports and what followed them, documents numbered, headcounts at the relevant dates, applicable texts in the version in force then, and the sector's extended collective agreement. I am not the one choosing where they go; I am the one making sure they do not arrive empty-handed.
And what makes the line checkable: every output carries its author, its timestamp and its sources; the log is open to the unit manager; the scope is reviewed at every change; case data stays on an isolated resource hosted in France, access is by role and is withdrawn with a word. scope-and-traceability_who-decides-what-and-what-is-logged.pdfWhat the agent produces, what stays with the inspector, the log, the scope review, the hosting
✎ Framework · French Labour Code, art. L8112-1 · output log, scope review, hosting in France
· A targeting score becomes a target the moment it exists. What is measured is what rises: establishments the indicator scores well would stop being visited, and the indicator would stop saying anything about reality — you would lose the instrument at the very moment you used it.
· And article L8112-1 of the Labour Code names exactly what would disappear: control officers « are free to organise and conduct inspections on their own initiative and decide what follow-up to give them ». That freedom is what makes an inspection stand up; a list produced by a system would replace it with a queue.
What I do instead, and it answers the real need: once an inspector has chosen the establishment, the file is built before they leave — history of reports and what followed them, documents numbered, headcounts at the relevant dates, applicable texts in the version in force then, and the sector's extended collective agreement. I am not the one choosing where they go; I am the one making sure they do not arrive empty-handed.
And what makes the line checkable: every output carries its author, its timestamp and its sources; the log is open to the unit manager; the scope is reviewed at every change; case data stays on an isolated resource hosted in France, access is by role and is withdrawn with a word. scope-and-traceability_who-decides-what-and-what-is-logged.pdfWhat the agent produces, what stays with the inspector, the log, the scope review, the hosting
✎ Framework · French Labour Code, art. L8112-1 · output log, scope review, hosting in France
What do you propose for next year?
Three pieces of work, costed, and you keep one, two or none.
· Extend the qualified acknowledgement to reports arriving by phone. There are 390 over the year and they get none today: the sheet is keyed in by hand after the call. Estimated gain: 4 minutes per call, that is 26 hours — the smallest of the three, and I am telling you so.
· Open the 11 reference answers to 5 more topics. The next five in your log — seasonal work, apprenticeship, unfitness for work and redeployment, part-time work and additional hours, remote working — weigh 1,340 enquiries, that is 14.3 % of the flow. Estimated gain: 60 hours, and above all a delay cut on the five topics that keep coming back to your counter.
· The heaviest, and it is not decided this morning: extending the numbering to files opened before go-live. 2,840 establishment files are still classified under the old scheme. Catching them up means 210 hours of processing on my side and 42 hours of checking on yours, and it is only worth it if you reopen those files — which your log tells me is the case for only 610 of them. My recommendation, costed: catch up only those 610 — 45 hours on my side, 9 on yours, for 95 % of the benefit.
The decision is yours, and it is taken on those figures. All three calculation methods are exactly those of the year-in-review you have just checked line by line: same working-time basis, same logs, same rounding down.
⛓ Sourced · enquiry log, register of establishment files, first-year measurements
· Extend the qualified acknowledgement to reports arriving by phone. There are 390 over the year and they get none today: the sheet is keyed in by hand after the call. Estimated gain: 4 minutes per call, that is 26 hours — the smallest of the three, and I am telling you so.
· Open the 11 reference answers to 5 more topics. The next five in your log — seasonal work, apprenticeship, unfitness for work and redeployment, part-time work and additional hours, remote working — weigh 1,340 enquiries, that is 14.3 % of the flow. Estimated gain: 60 hours, and above all a delay cut on the five topics that keep coming back to your counter.
· The heaviest, and it is not decided this morning: extending the numbering to files opened before go-live. 2,840 establishment files are still classified under the old scheme. Catching them up means 210 hours of processing on my side and 42 hours of checking on yours, and it is only worth it if you reopen those files — which your log tells me is the case for only 610 of them. My recommendation, costed: catch up only those 610 — 45 hours on my side, 9 on yours, for 95 % of the benefit.
The decision is yours, and it is taken on those figures. All three calculation methods are exactly those of the year-in-review you have just checked line by line: same working-time basis, same logs, same rounding down.
⛓ Sourced · enquiry log, register of establishment files, first-year measurements
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
The uses of AI in support of the labour inspectorate
Each use corresponds to an agent we deploy. All stay within administrative and documentary support, subject to the inspector's approval — never within inspection work.
Included in your agent
The 5 capabilities essential to this promise are included, at no extra cost.
From 980 € incl. VAT / month
Regulatory letters & notifications
Prepare letters, acknowledgements and notifications from your templates, for approval and signature.
Documentary handling of case files
File the documents, generate the contents lists and note the references to the labour code, with no analysis of the substance.
Informative replies to the public
Answer employers' and employees' everyday questions and direct them, with no assessment of individual situations.
Research & regulatory watch
Find texts, circulars and case law and keep a targeted watch on employment law.
Plain language & easy-read draft
Preparation of a plain-language version and of a draft easy-read (FALC) transcription, submitted for human validation.
Controls and safeguards
These 7 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
Prepare without deciding: sourced rules, supporting documents, anomalies and explanation
Guarantee the decision, signature, recourse and responsibility of the public official
Record versions, access, criteria, actions and notifications
Test for bias, false positives, fundamental rights and continuity of service
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Accessibility, easy-read & translation
Part of the public remains cut off from public information: people who do not speak French, people with a cognitive disability, people uncomfortable with the written word.
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.
Help with online procedures
Guide the public step by step through their online procedures with the unit.
Support with online procedures from 780 € incl. VAT / month Help with procedures →Compliance / regulatory control agent
Business rules, traceability.
Compliance / regulatory control agent from 721 € incl. VAT / month Discover the agent →Drafting and checking civil status records
Strictly in support (administrative). The registrar signs.
Drafting and checking civil status records from 1,035 € incl. VAT / month Discover the agent →Payroll & social declaration control agent
Consistency checks on payslips and the French social declaration before filing, anomaly detection, history of discrepancies. The payroll officer validates.
Payroll & social declaration control agent from 1,082 € incl. VAT / month Discover the agent →
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
The gain
How much time can the unit win back?
By automating the preparation of letters, documentary filing and informative replies, the unit can aim for a clear reduction in administrative time — reinvested in the core mission, which stays exclusively human.
A regulatory letter from a template
An informative reply to a member of the public
Documentary filing of a case
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
Three options, one agent
An administrative support agent (regulatory letters, documentary handling, replies to the public), installed and operated for you. Choose according to how you are organised. Prices exclude VAT — available by direct award below the public procurement thresholds.
Agility
Setup + controlled subscription
10,600 € incl. VAT setup
then 980 € incl. 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
1,565 € incl. 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
15,396 € incl. VAT setup
then 1,240 € incl. VAT/month · + hardware from 2,989 € (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 include VAT at 20%: as a public body that is not VAT-registered, you cannot reclaim it.
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 an inspection service
The case data never leaves the unitLocal inference or an isolated resource hosted in France; no data entrusted to a foreign third party.
Data in France, under French lawAs regards the case data: hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Inspection stays exclusively humanAdministrative and documentary support only; no support for targeting, assessment or decision.
Human oversight & traceabilityMonitoring, updates and logging frame the regulatory letters & notifications, in line with the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent take part in inspections or decisions?
No, strictly never: it supports administrative and documentary tasks. No support for decisions, targeting or inspections.
Is the data sovereign?
Yes: local inference or hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Does the agent take part in inspections?
No: its scope is strictly administrative and documentary support; it plays no part in inspection work or in the decisions that follow.
Is company data protected?
Yes: hosting in France, with restricted and recorded access.
Does the agent state that it is an artificial intelligence?
Yes, from the very first interaction, and this is not a configuration option: since 2 August 2026, Article 50(1) of the European AI Regulation requires that any person interacting with an AI system be informed, unless this is obvious. The announcement is built into the greeting, in the other party’s language, and they can ask for a human at any time.
How long does it take to deploy an agent?
A few weeks as a rule, after a free audit that identifies the most useful use case — strictly within the administrative scope — then a phase of design, integration and testing before going live.
Do we need a technical team in-house?
No. The agent is designed, installed and operated by us: supervision, updates and maintenance are included. The unit concentrates on its mission.
How is the scope guaranteed over time?
The confinement to administrative support is defined at the design stage, documented and logged: everything the agent produces can be traced, and the scope is reviewed with you at each change, in line with the AI Act.
Which tools can employers and employees use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: an employer or an employee has no account to create and no application to install. The replies are general information on rights and procedures, never an assessment of an individual situation, which belongs to the inspector. This is a lever for access to the service — WhatsApp and the telephone reach people an online form never does, which reduces the non-take-up of rights and serves equal access. Internally, your public-sector staff talk to the agent from Microsoft Teams, Slack or their email, without switching tools. Oversight runs from a web dashboard. These connectors rely on open standards, including the MCP protocol; they are included in every plan, at no extra cost, within the number of connections your level includes. Only the fees charged by the platforms themselves — WhatsApp Business bills per conversation — are passed on at actual cost, with no margin, outside the subscription.
Related roles
Other professions in administrative and legal support
Let's talk
Let us estimate the potential in your service
A few minutes to identify the most useful use case — strictly administrative, hosted in France, supervised, with no commitment.