● Public sector — Reading supporting documents
The AI agent for reading supporting documents: read, extract, check completeness
Reading the documents in a case file, keying them in and checking that nothing is missing take up a considerable share of public officers' time — before the assessment even begins. Your AI agent absorbs that repetitive work: it reads the documents, extracts the useful information and flags what is missing. Hosted in France — on local inference or an isolated resource — the public's data stays under control. The AI agent assists, the public officer decides.
Hosted in France
The public's data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Document reading · hosted in France
On benefit application no. 2024-0417, check the documents submitted and tell me what is missing before assessment.
7 documents read (PDFs, scans, one handwritten certificate). Extraction done: identity, address, income, bank details. Completeness: the previous year's tax notice is missing and the identity document expired in 03/2024.
File prepared for the case officer, missing documents listed — for approval.
⛓ Source · the file's documents + the reference list of required documents
File prepared for the case officer, missing documents listed — for approval.
⛓ Source · the file's documents + the reference list of required documents
Prepare the request for additional documents to send to the applicant.
Reminder letter ready: it sets out the two missing documents, the deadline for submission and the return channel. No decision on the merits is taken: the file stays pending with the case officer.
I keep the completeness tracking up to date as soon as the documents arrive.
✎ Action · letter ready for review — the public officer approves
I keep the completeness tracking up to date as soon as the documents arrive.
✎ Action · letter ready for review — the public officer 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 public officer decidesThe AI agent prepares, never approves alone
✦ In brief
In a public service, a Blue Lemon Agent agent automates the repetitive tasks upstream of assessment — reading the documents, extracting the data (multi-format OCR), checking completeness and chasing missing documents — and prepares the file for the case officer. It runs on local inference or is hosted in France on an isolated resource: the public's data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. The time saved goes back into assessment and relations with the public, in the service of equal treatment. Live within a few weeks.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
8
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 assessment services — and why they hesitate
The public expects quick answers and fair handling of their file. But officers' available time is mechanically absorbed by reading documents and checking completeness — and the data handled is among the most sensitive personal data there is.
! The issue
The service is caught between people expecting speed and equal treatment, and an ever-growing workload upstream of assessment (reading documents, keying in, checking completeness, chasing). Yet most consumer AI solutions amount to entrusting identities, incomes, family situations and supporting documents to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For public data, AI is only of interest if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, systematic human oversight, decisions reserved to the public officer: the time saved on reading documents is never paid for in lost confidentiality. The aim is not to replace the case officer, but to give them back time for the substance of the file and for the relationship with the applicant.
The decisive point
Confidentiality of the public's data: sovereignty & compliance
A public service handles some of the most sensitive personal data there is. Here is how the architecture of our agents protects it, file by file.
Local inference
The agent can run on a machine belonging to the service: no document leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — the public's data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Exposure of the public's data to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
One isolated resource per organisation
No pooling of data: an environment strictly dedicated to your authority or institution.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
An agent strictly in support; no decision issued automatically; 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 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.
Document reading · hosted in France
Four flags this morning, with no request from you.
· 18 files have been waiting on a document for more than 30 days with no chaser sent. The oldest has waited 71 days.
· One requested document no longer exists under that name since 2024. It is being asked for in 9 live files.
· Six illegible documents were accepted — cropped scans, blurred photographs. The files went through to assessment.
· The same proof of address appears in three different files, under three different names. morning-watch_4-flags.pdf4 flags · files concerned
⛓ Source · live files, documents uploaded, document schedule
· 18 files have been waiting on a document for more than 30 days with no chaser sent. The oldest has waited 71 days.
· One requested document no longer exists under that name since 2024. It is being asked for in 9 live files.
· Six illegible documents were accepted — cropped scans, blurred photographs. The files went through to assessment.
· The same proof of address appears in three different files, under three different names. morning-watch_4-flags.pdf4 flags · files concerned
⛓ Source · live files, documents uploaded, document schedule
18 with no chaser — how is that possible?
Because nothing chases: chasing is a manual act, and nobody has a list.
What I reconstructed: all 18 files did receive a request for documents when they were opened. None received a second one. The first went out, the resident did not reply, and the file stayed pending.
What that produces: 18 residents who believe their application is progressing, and 18 files that are not. Neither side knows the other is waiting.
What I have prepared: all 18 chasers, each naming the precise document that is missing — and three of them corrected, because the document originally requested was the wrong one.
What remains, and it is a single act: sending them. A chaser carries the department's name, sets a deadline and can have consequences for the file — that is what gives it its weight, and why it is signed. Ten minutes of reading, and eighteen residents finally know what is expected of them. chasers_18-files.pdf18 files · 3 requests corrected
⛓ Source · pending files, history of requests sent
What I reconstructed: all 18 files did receive a request for documents when they were opened. None received a second one. The first went out, the resident did not reply, and the file stayed pending.
What that produces: 18 residents who believe their application is progressing, and 18 files that are not. Neither side knows the other is waiting.
What I have prepared: all 18 chasers, each naming the precise document that is missing — and three of them corrected, because the document originally requested was the wrong one.
What remains, and it is a single act: sending them. A chaser carries the department's name, sets a deadline and can have consequences for the file — that is what gives it its weight, and why it is signed. Ten minutes of reading, and eighteen residents finally know what is expected of them. chasers_18-files.pdf18 files · 3 requests corrected
⛓ Source · pending files, history of requests sent
How did you spot all that? And who do you tell?
I read only what you have opened to me: files and their state, documents uploaded and their legibility, the schedule of required documents, and requests sent with their dates.
Routing follows your organisation: the 18 chasers to the assessing department; the document that no longer exists to whoever owns the schedule, because nine files are asking for an unobtainable paper; the six illegible documents to the officer who accepted them, and to them alone; the shared proof of address to the head of department.
With a chase: 48 h on the non-existent document — nine residents are hunting for something that does not exist; 7 days on the rest. Then a monthly summary: by file and by document, never by officer.
What this morning has already produced: 18 chasers written, each naming the exact document that is missing — three of them corrected, because the document originally asked for was the wrong one; nine files freed from a paper that has not existed since 2024; six illegible documents handed back before a decision could rest on them.
What that gives the service from tomorrow: 18 residents who finally know what is expected of them — the oldest had been waiting 71 days — and 18 files moving again instead of sleeping. That is officer time given back to assessment: the list of waiting files keeps itself, no one rebuilds it. And the same document is asked of everyone in the same way, at the same stage: that is equal treatment, and it can be evidenced. The decision on the merits stays with the officer, and I hand it back in minutes — the file arrives read, with what is missing named and the date of every request. Sending goes out from your service: give me the mandate, bounded to document requests, dated, withdrawn on a word, and the 18 chasers go straight out.
The next step is ready: the 18 chasers await a reread, and I can put the schedule of required documents through the same sieve today, so that no other request asks for an unobtainable paper.
✎ Proposal · watch and chases to be configured — you set the thresholds
Routing follows your organisation: the 18 chasers to the assessing department; the document that no longer exists to whoever owns the schedule, because nine files are asking for an unobtainable paper; the six illegible documents to the officer who accepted them, and to them alone; the shared proof of address to the head of department.
With a chase: 48 h on the non-existent document — nine residents are hunting for something that does not exist; 7 days on the rest. Then a monthly summary: by file and by document, never by officer.
What this morning has already produced: 18 chasers written, each naming the exact document that is missing — three of them corrected, because the document originally asked for was the wrong one; nine files freed from a paper that has not existed since 2024; six illegible documents handed back before a decision could rest on them.
What that gives the service from tomorrow: 18 residents who finally know what is expected of them — the oldest had been waiting 71 days — and 18 files moving again instead of sleeping. That is officer time given back to assessment: the list of waiting files keeps itself, no one rebuilds it. And the same document is asked of everyone in the same way, at the same stage: that is equal treatment, and it can be evidenced. The decision on the merits stays with the officer, and I hand it back in minutes — the file arrives read, with what is missing named and the date of every request. Sending goes out from your service: give me the mandate, bounded to document requests, dated, withdrawn on a word, and the 18 chasers go straight out.
The next step is ready: the 18 chasers await a reread, and I can put the schedule of required documents through the same sieve today, so that no other request asks for an unobtainable paper.
✎ Proposal · watch and chases to be configured — you set the thresholds
Check the completeness of file 2026-0912.
Nine documents read, two missing — and one of the two should not have been asked for.
What is there and usable: identity, address, income, family situation, bank details. Every extracted value carries the document and the page it came from.
What is missing:
· The previous year's tax assessment — requested, not uploaded.
· An employment certificate — requested in the opening letter. Yet your schedule only requires it for applications filed before 2024; since then a signed declaration suffices, and it is on file.
In other words: only one document is genuinely missing. The resident has been hunting for two for six weeks.
What I checked before telling you: the version of the schedule applicable at this file's filing date, and the eleven files filed since January 2024 where the signed declaration was accepted — in all eleven, assessment ran to its end with no employment certificate.
What I have written: the message that closes the request — "the employment certificate is no longer needed; only the tax assessment is still awaited". The gap between what the schedule requires and what the letter asked for is settled by the officer, because it is the officer who answers for completeness. One approval, and the resident stops hunting for a document nobody asks of them any more. file-2026-0912_completeness.pdf9 documents read · 1 genuinely missing
⛓ Source · 9 documents on file, schedule of required documents
What is there and usable: identity, address, income, family situation, bank details. Every extracted value carries the document and the page it came from.
What is missing:
· The previous year's tax assessment — requested, not uploaded.
· An employment certificate — requested in the opening letter. Yet your schedule only requires it for applications filed before 2024; since then a signed declaration suffices, and it is on file.
In other words: only one document is genuinely missing. The resident has been hunting for two for six weeks.
What I checked before telling you: the version of the schedule applicable at this file's filing date, and the eleven files filed since January 2024 where the signed declaration was accepted — in all eleven, assessment ran to its end with no employment certificate.
What I have written: the message that closes the request — "the employment certificate is no longer needed; only the tax assessment is still awaited". The gap between what the schedule requires and what the letter asked for is settled by the officer, because it is the officer who answers for completeness. One approval, and the resident stops hunting for a document nobody asks of them any more. file-2026-0912_completeness.pdf9 documents read · 1 genuinely missing
⛓ Source · 9 documents on file, schedule of required documents
Is that non-existent document asked for often?
In nine live files, and for eighteen months in every file using the same standard letter.
What happened: the schedule was amended in January 2024. The standard request letter did not follow. It still asks for a document that is no longer required and whose title no longer exists at the issuing body.
What that produces: the resident searches, fails, calls — and the front-desk officer, reading the same letter, confirms it is required.
What I have written: the three standard letters corrected, the obsolete document removed, the two now-required documents added, and against each change the clause of the schedule that justifies it, with its date.
What is yours to do: put them into service. A standard letter binds the department on what it requires — it is signed, and that is exactly what lets a front-desk officer rely on it. Five corrections, one review.
What I can do next, on your approval: match every standard letter against the schedule in force at each amendment, and flag the gaps. Across your eleven letters, three ask for at least one document no longer required — and two omit one that has since become required. standard-letters_schedule-gaps.pdf11 letters · 3 asking too much, 2 too little
✎ Support · 3 letters out of step — correction stays with the department
What happened: the schedule was amended in January 2024. The standard request letter did not follow. It still asks for a document that is no longer required and whose title no longer exists at the issuing body.
What that produces: the resident searches, fails, calls — and the front-desk officer, reading the same letter, confirms it is required.
What I have written: the three standard letters corrected, the obsolete document removed, the two now-required documents added, and against each change the clause of the schedule that justifies it, with its date.
What is yours to do: put them into service. A standard letter binds the department on what it requires — it is signed, and that is exactly what lets a front-desk officer rely on it. Five corrections, one review.
What I can do next, on your approval: match every standard letter against the schedule in force at each amendment, and flag the gaps. Across your eleven letters, three ask for at least one document no longer required — and two omit one that has since become required. standard-letters_schedule-gaps.pdf11 letters · 3 asking too much, 2 too little
✎ Support · 3 letters out of step — correction stays with the department
Extract the declared income from this file.
Extraction done, and every amount points back to its page.
What is extracted: earned income, benefits, rental income — eight amounts, each with its document and page.
What I flag without settling: two amounts do not reconcile.
· The tax assessment shows a reference income which, set against the payslips provided, leaves a gap of €4,200.
· A rental income appears on the tax assessment and on no other document.
What those gaps are not: an anomaly. A gap between a reference income and payslips very often has an explanation — allowances, other household income, a change of circumstances. I flag it because the officer must see it, not because it is suspicious.
What I looked for so the gap does not stay bare: which of those explanations fits. The €4,200 gap matches, to the euro, the 10% allowance applied to the payslips supplied — the workings are set out line by line. The rental income, by contrast, has no counterpart in the documents filed: that is the only question left, and it is written in one sentence, ready to send.
Preferring one figure over another remains an assessment judgement, and it amounts to a decision: it carries the officer's name, and that is what makes it capable of being explained to the resident. income_8-sourced-amounts.pdf8 amounts · 2 gaps flagged
⛓ Source · tax assessment p. 2, 6 payslips, signed declaration
What is extracted: earned income, benefits, rental income — eight amounts, each with its document and page.
What I flag without settling: two amounts do not reconcile.
· The tax assessment shows a reference income which, set against the payslips provided, leaves a gap of €4,200.
· A rental income appears on the tax assessment and on no other document.
What those gaps are not: an anomaly. A gap between a reference income and payslips very often has an explanation — allowances, other household income, a change of circumstances. I flag it because the officer must see it, not because it is suspicious.
What I looked for so the gap does not stay bare: which of those explanations fits. The €4,200 gap matches, to the euro, the 10% allowance applied to the payslips supplied — the workings are set out line by line. The rental income, by contrast, has no counterpart in the documents filed: that is the only question left, and it is written in one sentence, ready to send.
Preferring one figure over another remains an assessment judgement, and it amounts to a decision: it carries the officer's name, and that is what makes it capable of being explained to the resident. income_8-sourced-amounts.pdf8 amounts · 2 gaps flagged
⛓ Source · tax assessment p. 2, 6 payslips, signed declaration
Does that kind of gap come up often?
Across 214 files this quarter, 61 show at least one gap between two documents. And that is good news.
Why: those 61 gaps were already there. They simply were not visible, because nobody systematically sets a tax assessment beside six payslips.
What a flagged gap becomes: a question put to the resident, in one sentence, instead of a decision taken on an uncertain figure. Of the 61, 54 questions are already drafted; the other seven are waiting on a document the resident has not yet uploaded.
What I write instead of a label, and it matters here: a question, and nothing else. A gap is not a suspicion — and a file tagged "to check" in a system keeps that tag long after the question has been settled.
A proposal, if you approve it: the gap appears in the file, beside the two documents concerned, and disappears once the officer has dealt with it. No trace, no history of suspicion, no table of flagged files. gaps-between-documents_61-files.pdf214 files · 61 gaps · nature of each
✎ Proposal · gap shown then cleared — no lasting label
Why: those 61 gaps were already there. They simply were not visible, because nobody systematically sets a tax assessment beside six payslips.
What a flagged gap becomes: a question put to the resident, in one sentence, instead of a decision taken on an uncertain figure. Of the 61, 54 questions are already drafted; the other seven are waiting on a document the resident has not yet uploaded.
What I write instead of a label, and it matters here: a question, and nothing else. A gap is not a suspicion — and a file tagged "to check" in a system keeps that tag long after the question has been settled.
A proposal, if you approve it: the gap appears in the file, beside the two documents concerned, and disappears once the officer has dealt with it. No trace, no history of suspicion, no table of flagged files. gaps-between-documents_61-files.pdf214 files · 61 gaps · nature of each
✎ Proposal · gap shown then cleared — no lasting label
Our users upload absolutely anything. What exactly do you read?
Reading is multi-format, and anomaly detection works on the document itself, not on its file name.
The formats I read: native PDF, image PDF, photograph taken on a phone, scanned document, and common handwriting — a date, an amount, a signature, a “certified true copy” note. Across the 14,200 documents filed this year, the real split is native PDF 41 %, image or photograph 37 %, scan 19 %, handwriting 3 %: reading that covered only the first of those formats would leave 59 % of what is filed to your staff.
What the reading returns: every extracted value carries the file, the page and the position it came from, and its reading mode — native text or optical character recognition, with a confidence score. Below 90 % confidence, the value arrives proposed and unvalidated.
Anomaly detection on the documents, and what it found: expired supporting documents — 214 certificates out of validity on the day they were filed; duplicates — the same proof of address in three files under three different names, which I flag without drawing any conclusion; non-compliant documents — 61 items that are not what the procedure asks for, a tax assessment instead of a declaratory statement; inconsistencies — a document dated after the decision it supports, 9 cases. All of it goes to human verification, with the document open at the right page: an anomaly on a document is neither fraud nor a refusal, it is a question to be asked.
What that gives back: reading and keying a nine-document file took 38 minutes; checking a prepared reading takes 8. Across 1,580 files a year, more than 790 hours given back — more than twenty-two 35-hour weeks, rounded down.
The figure that does not flatter me: 6 illegible documents were accepted — cropped scans, blurred photographs — and the files moved to assessment before anyone noticed. My optical recognition was returning plausible text from a truncated image, and plausible is not read. I now compare the area actually decoded against the area of the page; below 85 %, the document is refused to the user, at the moment of filing, with the reason and the photograph to retake — not to the case officer three weeks later. Over the next 2,100 documents: 92 refused at filing, 89 refiled correctly the same day.
⛓ Sourced · 14,200 documents across 4 formats, 38 min down to 8, 6 illegible documents behind the fix
The formats I read: native PDF, image PDF, photograph taken on a phone, scanned document, and common handwriting — a date, an amount, a signature, a “certified true copy” note. Across the 14,200 documents filed this year, the real split is native PDF 41 %, image or photograph 37 %, scan 19 %, handwriting 3 %: reading that covered only the first of those formats would leave 59 % of what is filed to your staff.
What the reading returns: every extracted value carries the file, the page and the position it came from, and its reading mode — native text or optical character recognition, with a confidence score. Below 90 % confidence, the value arrives proposed and unvalidated.
Anomaly detection on the documents, and what it found: expired supporting documents — 214 certificates out of validity on the day they were filed; duplicates — the same proof of address in three files under three different names, which I flag without drawing any conclusion; non-compliant documents — 61 items that are not what the procedure asks for, a tax assessment instead of a declaratory statement; inconsistencies — a document dated after the decision it supports, 9 cases. All of it goes to human verification, with the document open at the right page: an anomaly on a document is neither fraud nor a refusal, it is a question to be asked.
What that gives back: reading and keying a nine-document file took 38 minutes; checking a prepared reading takes 8. Across 1,580 files a year, more than 790 hours given back — more than twenty-two 35-hour weeks, rounded down.
The figure that does not flatter me: 6 illegible documents were accepted — cropped scans, blurred photographs — and the files moved to assessment before anyone noticed. My optical recognition was returning plausible text from a truncated image, and plausible is not read. I now compare the area actually decoded against the area of the page; below 85 %, the document is refused to the user, at the moment of filing, with the reason and the photograph to retake — not to the case officer three weeks later. Over the next 2,100 documents: 92 refused at filing, 89 refiled correctly the same day.
⛓ Sourced · 14,200 documents across 4 formats, 38 min down to 8, 6 illegible documents behind the fix
Six illegible documents got through. How do we avoid that?
By telling the resident at the moment of upload, rather than the officer three weeks later.
What the six are: four cropped scans — part of the page is missing — and two photographs blurred to the point where the amount cannot be read.
What happened: the upload was accepted, the file went to assessment, and the officer had to ask again. Three weeks lost on both sides.
What I do at upload: I detect a cropped page, a blur that prevents reading, an empty or duplicated document — and I show it to the resident, the area circled, while they are still at their screen.
What I prefer to an automatic refusal: the flag. A refusal would also block valid documents a mechanical check reads badly — a handwritten paper, a faint stamp. The resident decides whether to re-upload, and the officer sees the flag either way.
What that would have returned this quarter: three weeks per file, across the six concerned — and six residents who would not have had to start again. illegible-documents_6-cases.pdf6 documents · 3 weeks lost on average
⛓ Source · 6 accepted documents, delays between upload and re-request
What the six are: four cropped scans — part of the page is missing — and two photographs blurred to the point where the amount cannot be read.
What happened: the upload was accepted, the file went to assessment, and the officer had to ask again. Three weeks lost on both sides.
What I do at upload: I detect a cropped page, a blur that prevents reading, an empty or duplicated document — and I show it to the resident, the area circled, while they are still at their screen.
What I prefer to an automatic refusal: the flag. A refusal would also block valid documents a mechanical check reads badly — a handwritten paper, a faint stamp. The resident decides whether to re-upload, and the officer sees the flag either way.
What that would have returned this quarter: three weeks per file, across the six concerned — and six residents who would not have had to start again. illegible-documents_6-cases.pdf6 documents · 3 weeks lost on average
⛓ Source · 6 accepted documents, delays between upload and re-request
And the proof of address appearing in three files?
I flag it, and I stop there — and stopping there matters.
What is established: the same document, in the same name, at the same address, is uploaded to three files opened under three different names.
What it can be: a shared tenancy, a family at one address, someone being hosted — and your schedule accepts a hosting certificate in those cases. It can also be something else.
The label itself does not enter the file, and the reason is mechanical: writing "suspicion" in a social support file creates a trace that outlives the explanation, and three families can carry the consequence for years.
What I write instead: the match, the three files, the document, and the two explanations your schedule accepts — shared tenancy and hosting — with the clause that provides for them and the document that would establish each. Addressed to the head of department alone, with no label and no copy to anyone else.
Checking takes one phone call — and if the explanation is the expected one, no trace of it remains.
✎ Support · match flagged without labelling — no record retained
What is established: the same document, in the same name, at the same address, is uploaded to three files opened under three different names.
What it can be: a shared tenancy, a family at one address, someone being hosted — and your schedule accepts a hosting certificate in those cases. It can also be something else.
The label itself does not enter the file, and the reason is mechanical: writing "suspicion" in a social support file creates a trace that outlives the explanation, and three families can carry the consequence for years.
What I write instead: the match, the three files, the document, and the two explanations your schedule accepts — shared tenancy and hosting — with the clause that provides for them and the document that would establish each. Addressed to the head of department alone, with no label and no copy to anyone else.
Checking takes one phone call — and if the explanation is the expected one, no trace of it remains.
✎ Support · match flagged without labelling — no record retained
Our reminders about missing documents are impossible to understand. Can you do something about that?
Yes, and it is a foundation module: plain language first, the easy-to-read (FALC) draft afterwards.
The plain-language version, on your standard reminder: the original runs to 287 words, cites 3 legal articles and uses 14 administrative terms. The plain-language version runs to 96 words and says what is missing, where to send it and by what date in the first three lines; it keeps the 3 references as a footnote and replaces “failure to produce” with “a document is missing”. The meaning and the 3 references stay intact: I set the two versions side by side, and your department checks line by line rather than taking my word.
The draft transcription under the FALC rules, and why it is a draft: those rules require a transcription and then a review involving the people concerned. I prepare the draft — one idea per sentence, a pictogram per step, dates written out in full — and until that review has happened it is not FALC, and we do not call it that. The draft is subject to human validation: your department runs the session, and I supply at the same time the points where I hesitate between two wordings.
What it has already changed: the plain-language reminder has gone out 420 times; the document arriving within 15 days went from 52 % to 74 %. A reminder that is understood is a reminder nobody has to send again: 92 second-level reminders avoided, more than 30 staff hours — and 92 people who did not wait another month.
The figure that does not flatter me: of my first 60 plain-language versions, 7 had lost a piece of information — 11.7 %, and always the same one: the appeal deadline, which I was treating as a footnote. Information whose loss costs someone a right does not get shortened. I set six unalterable mentions — appeal deadline and route, closing date, competent department, file reference, amount, signatory — and I block any version that loses one: since then, no loss across 340 versions.
⛓ Sourced · 287 words down to 96 with identical meaning, 15-day filing from 52 % to 74 %, 6 unalterable mentions
The plain-language version, on your standard reminder: the original runs to 287 words, cites 3 legal articles and uses 14 administrative terms. The plain-language version runs to 96 words and says what is missing, where to send it and by what date in the first three lines; it keeps the 3 references as a footnote and replaces “failure to produce” with “a document is missing”. The meaning and the 3 references stay intact: I set the two versions side by side, and your department checks line by line rather than taking my word.
The draft transcription under the FALC rules, and why it is a draft: those rules require a transcription and then a review involving the people concerned. I prepare the draft — one idea per sentence, a pictogram per step, dates written out in full — and until that review has happened it is not FALC, and we do not call it that. The draft is subject to human validation: your department runs the session, and I supply at the same time the points where I hesitate between two wordings.
What it has already changed: the plain-language reminder has gone out 420 times; the document arriving within 15 days went from 52 % to 74 %. A reminder that is understood is a reminder nobody has to send again: 92 second-level reminders avoided, more than 30 staff hours — and 92 people who did not wait another month.
The figure that does not flatter me: of my first 60 plain-language versions, 7 had lost a piece of information — 11.7 %, and always the same one: the appeal deadline, which I was treating as a footnote. Information whose loss costs someone a right does not get shortened. I set six unalterable mentions — appeal deadline and route, closing date, competent department, file reference, amount, signatory — and I block any version that loses one: since then, no loss across 340 versions.
⛓ Sourced · 287 words down to 96 with identical meaning, 15-day filing from 52 % to 74 %, 6 unalterable mentions
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 facets of document reading in a public service
Each use corresponds to a facet we deploy. All of them work in support, subject to approval by the public officer.
Included in your agent
The 8 capabilities essential to this promise are included, at no extra cost.
From 1,045 € incl. VAT / month
Reading & multi-format OCR
Reading documents in PDF, image and scanned form, including common handwriting, and extracting the useful data.
Completeness checks
Verifying that every required document is present and valid, against the reference list expected for the file.
Chasing missing documents
Preparing requests for additional documents, with the deadline and the return channel — for approval.
Preparing the file for the case officer
The file is structured, the data extracted and the points to watch flagged before human assessment.
Completeness of planning applications
Reading the documents of a permit or authorisation and checking completeness against the local plan.
Anomaly detection on documents
Spotting inconsistencies, duplicates and expired or non-compliant certificates, for human verification.
Benefits assessment
Preparing benefit and allowance files for the officer, documents read and completeness verified.
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 5 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Sources, access rights and handling of questions with no answer
Inherited access rights, versions and freshness of the corpus
Explicit handling of questions with no answer and of conflicting sources
What the agent must be connected to
This connection is required for the agent to work. It concerns your information system and is scoped during the audit.
Measurement of accuracy, coverage and time saved
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.
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 a service win back?
By automating document reading and completeness checks, a service can aim for a sharp reduction in the time spent upstream of assessment on standardised files — reinvested in assessment and in relations with the public.
Checking the completeness of a file
Extracting the data from a document
Preparing a request for missing documents
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 agent for reading supporting documents (OCR, extraction, completeness), installed and operated for you. Choose according to how you work. Prices exclude VAT — annual subscription, the time it takes for the gains to settle in.
Agility
Setup + controlled subscription
11,945 € incl. VAT setup
then 1,045 € 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
1,710 € 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
16,981 € incl. VAT setup
then 1,305 € 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 a public service
The public's data never leaves the serviceLocal inference or an isolated resource hosted in France; no document entrusted to a foreign third party.
Data in France, under French lawNative localisation and minimisation for the public's data; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The public officer keeps the decisionThe AI agent prepares the file, checks completeness and flags what is missing; no decision is automated.
Human oversight & traceabilityFull traceability for reading and multi-format OCR: monitoring, updates and logging compliant with the AI Act.
Frequently asked questions
Your questions, our answers
Which formats are supported?
PDFs, images and scanned documents, including common handwriting. The agent reads the documents in the file and extracts the useful information (identity, address, income, bank details), whatever the quality of the document, then flags what it could not read for human verification.
Does the agent decide on the case?
No. It prepares: it reads the documents, extracts the data and checks completeness. Assessment and decision remain human. The AI agent assists, the public officer decides — no decision on the merits is ever automated.
Are the documents processed in France?
Yes, on an isolated resource hosted in France, or on local inference within the service, 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.
How does the agent check completeness?
It compares the documents submitted against the reference list required for that type of file, spots what is missing or no longer valid (an expired certificate, for instance) and prepares, for your approval, the request for additional documents.
Do we have to change our existing tools?
No. The agent connects to your tools (business software, document management, online submission platform) and complements them, without imposing a migration. We adapt the integration to your environment.
Is equal treatment of the public preserved?
Yes. The agent applies the same completeness check to every file and logs each step, which strengthens traceability and equal treatment. The decision remains with the public officer, who guarantees case-by-case judgement.
How long does it take to deploy the agent?
A few weeks as a rule, after a free audit that identifies the most useful use case, then a phase of design, integration and testing before going live and training your officers.
Related solutions
Other solutions for your assessment services
Let's talk
Let us estimate the potential in your service
15 minutes to identify the most useful use case — hosted in France, supervised, with no commitment.