● Public sector — Family benefits caseworker
AI agent for the family benefits caseworker
An assistant that pre-assesses benefit claims, checks the documents and prepares the replies to claimants — the caseworker approves every entitlement opened. Hosted in France, on a resource isolated for each organisation. GDPR & AI Act: governed deployment: the AI agent assists, the caseworker decides.
Hosted in France
Claimants' data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Family benefits assistant · hosted in France
Check whether this in-work benefit claim filed yesterday is complete.
File analysed. Documents present: quarterly income declaration, payslips, bank details. Missing: the tax assessment and proof of address less than three months old. The income declared is consistent with the payslips supplied.
I am preparing the request for the additional documents.
⛓ Source · the documents in the file + the conditions for the benefit
I am preparing the request for the additional documents.
⛓ Source · the documents in the file + the conditions for the benefit
A claimant is disputing an overpayment of housing benefit. Prepare a draft reply.
Draft reply ready. A reminder of the reason for the overpayment (a change of circumstances declared late), the periods concerned, the routes of appeal and the possibility of asking for a write-off or instalments. The points to check before sending are highlighted.
No decision is taken on the file.
✎ Action · letter ready for review — the caseworker approves
No decision is taken on the file.
✎ Action · letter ready for review — the caseworker 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 caseworker decidesThe AI agent assists, it never opens an entitlement
✦ In brief
In a family benefits fund, a Blue Lemon Agent agent assists the caseworker with the repetitive tasks — pre-assessment of claims, checking documents and income, letters and notifications to claimants — and flags inconsistencies to check before handling. It runs on local inference or is hosted in France on a resource dedicated to and isolated for each organisation: families' data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Opening or revising an entitlement remains an act of the caseworker; the time won back goes to complex situations and supporting families. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and benefit recipients 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 family benefits funds — and why they hesitate
In-trays overflowing at every declaration deadline, families waiting for their entitlements to open, documents to check one by one: the caseworker handles some of the most sensitive data there is every day — income, household composition, social circumstances.
! The issue
The caseworker is caught between claimants waiting for their entitlements to open quickly — income support, housing benefit, in-work benefit — and a flow of files, documents and declarations that keeps growing. Yet most consumer AI tools would amount to entrusting claimants' income, household composition and social circumstances to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For social 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, opening entitlements reserved to the caseworker: the time saved on checking documents is never paid for in lost confidentiality, nor in unequal treatment. The aim is not to replace the caseworker, but to give them back time for complex situations.
The decisive point
Protecting claimants' data: sovereignty & compliance
A family benefits fund handles the income and family life of millions of households. Here is how the architecture of our agents protects it, organisation by organisation.
Local inference
The agent can run on a machine at the organisation: 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 — claimants' data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity for families' data: our architecture depends on a subcontracting chain and remote access documented for the configuration chosen.
One isolated resource per organisation
No pooling of data: an environment strictly dedicated to your fund, guaranteeing the continuity of the public service.
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 support; no entitlement is opened or revised 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 — particularly relevant as soon as social or health data is processed. 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
5 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 bodyVal-d'Aubrac family benefits office — a fictional mid-sized public body
- Activity
- Paying family and social benefits — income support, housing benefit, in-work benefit, family allowances
- Headcount
- 210 staff, including 96 caseworkers, 12 in the mail service and 8 in the checking unit; the AI agent is connected to the caseworkers' perimeter only
- People served
- 118,000 claimant households, of which 14,200 on income support and 26,500 on the in-work benefit
- Volume
- 4,200 documents filed a week, 2,400 completeness checks and 3,100 letters a month, 640 first income support claims a month
- Tools in place
- The benefits information system, the document management system, the letter tool, the online services portal and the switchboard — the agent plugs into them, nothing is replaced
- Who decides
- The caseworker signs off every grant, revision or refusal of entitlement; the head of the benefits department sets the checking rules; the public accountant keeps the recovery of overpayments
- Room for improvement
- 21 days average on a first in-work benefit claim, against a 15-day service commitment; 31% of files go back out as a request for documents; 2,900 contacts a month simply to ask where a file stands
The Val-d'Aubrac office is not trying to process more files with fewer staff: it is trying to keep the service commitment it made to families and to return its caseworkers to the situations that call for professional judgement. The agent runs on an isolated resource hosted in France, dedicated to this body, and plugs into the benefits information system, the document management system and the letter tool in a supporting role. It analyses, it writes, it puts figures on the table and it carries out what has been decided; a public officer's sign-off stays on every act that touches a family. The exchanges below cover one quarter, from the first queue to the review.
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.
Family benefits assistant · hosted in France
I went back over the 2,400 completeness checks of the month. And I found why one file in three goes back out as a request for documents — it is not the reason your caseworkers assumed.
A completeness check verifies that a claim carries every document the benefit requires, before a caseworker assesses it.
The gap, and it is the gap that decides the gain: of the 744 files that went back out as a request for documents last month — 31% of the queue — the breakdown is this:
· 470 concerned a document the claimant had ALREADY filed, in another claim or under a different file name. Six requests in ten were about a document the office already held.
· 168 concerned a genuinely missing document.
· 106 concerned an unreadable document — a cropped photograph, a scan too dark.
What that costs a family: a request for documents adds on average 11 days of waiting before the entitlement is granted. 470 avoidable requests means that many families waiting for no reason.
The rule is written, and I have already run it over your own history: before any request goes out, match the claim against every document already filed by the household, whatever its name or the claim it came from. Over the last twelve months — 28,800 files — it would have avoided 5,640 requests and given families back 62,000 days of waiting. The request rate falls from 31% to 12%, from 744 files to 288.
And the check itself: 18 minutes per file yesterday, 3 minutes of review today — 60% of the processing time brought down to 10%, that is 600 hours a month returned to your caseworkers.
It fits in one sentence, in the form of your own framework: the head of the benefits department signs it, and it applies to the 2,400 files in the queue this evening. That signature is what makes it enforceable — and what will let a family challenge it. completeness-queue_2400-files-analysed.pdf744 requests, 470 of them for a document already held
⛓ Sourced · queue of the month, household document management system, twelve months of document requests
A completeness check verifies that a claim carries every document the benefit requires, before a caseworker assesses it.
The gap, and it is the gap that decides the gain: of the 744 files that went back out as a request for documents last month — 31% of the queue — the breakdown is this:
· 470 concerned a document the claimant had ALREADY filed, in another claim or under a different file name. Six requests in ten were about a document the office already held.
· 168 concerned a genuinely missing document.
· 106 concerned an unreadable document — a cropped photograph, a scan too dark.
What that costs a family: a request for documents adds on average 11 days of waiting before the entitlement is granted. 470 avoidable requests means that many families waiting for no reason.
The rule is written, and I have already run it over your own history: before any request goes out, match the claim against every document already filed by the household, whatever its name or the claim it came from. Over the last twelve months — 28,800 files — it would have avoided 5,640 requests and given families back 62,000 days of waiting. The request rate falls from 31% to 12%, from 744 files to 288.
And the check itself: 18 minutes per file yesterday, 3 minutes of review today — 60% of the processing time brought down to 10%, that is 600 hours a month returned to your caseworkers.
It fits in one sentence, in the form of your own framework: the head of the benefits department signs it, and it applies to the 2,400 files in the queue this evening. That signature is what makes it enforceable — and what will let a family challenge it. completeness-queue_2400-files-analysed.pdf744 requests, 470 of them for a document already held
⛓ Sourced · queue of the month, household document management system, twelve months of document requests
And the 168 genuinely missing documents? Our requests often go out a week after the filing.
They go out the same day, and they ask only for what is genuinely missing. The 168 draft letters are written, one per file.
What each request contains, and why:
· The exact list of missing documents, named as the claimant knows them — “your 2025 tax assessment”, not “supporting income document no. 4”.
· The reason for the request, in one sentence: what the document is for in assessing the entitlement. A request with no reason produces a phone call; a request with a reason produces a filing.
· The three ways to file — portal, front desk, post — and the date the document is expected.
· What has already arrived, listed too: the claimant sees what they have done, not only what is missing.
The delay this shifts: your requests went out on average six days after the filing; they now go out the same day. On a first in-work benefit claim, that is six days taken off the 21-day average, against a 15-day service commitment.
The mandate I propose for what follows, and it is bounded three ways: I send out the requests built from your templates and I schedule a reminder at day 10 if nothing arrives — capped to requests for documents alone, dated until 31/12 with a review at expiry, withdrawable in a word and with no reason to give. Any response from the claimant stops the reminder, including a call to the switchboard: nobody gets chased for a document they have just filed.
And closure, which is the act that matters: 14 files passed day 30 with no response. They are on the caseworker's desk this morning, with the full history of reminders and the draft closure letter already reasoned. Their signature closes the file in one minute instead of a quarter of an hour — and it is that signature that makes the closure dated, reasoned and open to challenge by the family. document-request_template-and-168-files.pdfWhat is missing, why, and how to file it
⛓ Sourced · documents required per benefit, the office's letter templates, filing and reminder log
What each request contains, and why:
· The exact list of missing documents, named as the claimant knows them — “your 2025 tax assessment”, not “supporting income document no. 4”.
· The reason for the request, in one sentence: what the document is for in assessing the entitlement. A request with no reason produces a phone call; a request with a reason produces a filing.
· The three ways to file — portal, front desk, post — and the date the document is expected.
· What has already arrived, listed too: the claimant sees what they have done, not only what is missing.
The delay this shifts: your requests went out on average six days after the filing; they now go out the same day. On a first in-work benefit claim, that is six days taken off the 21-day average, against a 15-day service commitment.
The mandate I propose for what follows, and it is bounded three ways: I send out the requests built from your templates and I schedule a reminder at day 10 if nothing arrives — capped to requests for documents alone, dated until 31/12 with a review at expiry, withdrawable in a word and with no reason to give. Any response from the claimant stops the reminder, including a call to the switchboard: nobody gets chased for a document they have just filed.
And closure, which is the act that matters: 14 files passed day 30 with no response. They are on the caseworker's desk this morning, with the full history of reminders and the draft closure letter already reasoned. Their signature closes the file in one minute instead of a quarter of an hour — and it is that signature that makes the closure dated, reasoned and open to challenge by the family. document-request_template-and-168-files.pdfWhat is missing, why, and how to file it
⛓ Sourced · documents required per benefit, the office's letter templates, filing and reminder log
Who decides a file is complete? Because that is where everything starts.
The caseworker signs, and I hand them the file ready to sign in three minutes instead of eighteen. Across the month's 2,400 files, 2,400 sign-offs.
What they have in front of them, one page per file:
· Every required document, and against it the document found, its filing date and the claim it came from.
· The rule applied, quoted from your internal reference framework with its version — the text of your own document, never a rewording of mine.
· The points to reconcile, when two documents say two different things.
What that sign-off gets you, and it is measurable: a completeness decision signed by a public officer is reasoned, dated, traced and open to challenge — a claimant can have it reviewed, and that protects the family as much as the office. And it takes three minutes to give: the reasoned decision arrives in minutes where it used to take weeks.
What I have written on top, and nobody has the time to write: 37 files this month run into two rules of your framework that contradict each other — shared custody arrangements, for the most part. I have drafted both possible readings, each in one sentence and in the form of your own rules, and run them over the 444 shared-custody files of the last twelve months:
· Reading A — the child is in the care of the declared residence: 214 households open the entitlement, 230 do not.
· Reading B — care is shared in proportion to overnight stays: 96 households open the entitlement in full, 348 at a shared rate.
The gap between the two readings covers 118 families a year. Today that gap is settled file by file, depending on which caseworker opens it — and that is exactly what produces different treatment for two families in the same situation.
What I suggest: 90 minutes with the head of the benefits department, both readings on the table with their figures. They sign one, I apply it to the 37 waiting files the same day and to every file that follows. A contradiction settled once holds for the year's 444 files.
⛓ Sourced · the office's internal reference framework, sign-off log, 444 shared-custody files over twelve months
What they have in front of them, one page per file:
· Every required document, and against it the document found, its filing date and the claim it came from.
· The rule applied, quoted from your internal reference framework with its version — the text of your own document, never a rewording of mine.
· The points to reconcile, when two documents say two different things.
What that sign-off gets you, and it is measurable: a completeness decision signed by a public officer is reasoned, dated, traced and open to challenge — a claimant can have it reviewed, and that protects the family as much as the office. And it takes three minutes to give: the reasoned decision arrives in minutes where it used to take weeks.
What I have written on top, and nobody has the time to write: 37 files this month run into two rules of your framework that contradict each other — shared custody arrangements, for the most part. I have drafted both possible readings, each in one sentence and in the form of your own rules, and run them over the 444 shared-custody files of the last twelve months:
· Reading A — the child is in the care of the declared residence: 214 households open the entitlement, 230 do not.
· Reading B — care is shared in proportion to overnight stays: 96 households open the entitlement in full, 348 at a shared rate.
The gap between the two readings covers 118 families a year. Today that gap is settled file by file, depending on which caseworker opens it — and that is exactly what produces different treatment for two families in the same situation.
What I suggest: 90 minutes with the head of the benefits department, both readings on the table with their figures. They sign one, I apply it to the 37 waiting files the same day and to every file that follows. A contradiction settled once holds for the year's 444 files.
⛓ Sourced · the office's internal reference framework, sign-off log, 444 shared-custody files over twelve months
How do you match declared income against the documents? That is the heart of our work.
Line by line, and every discrepancy I flag comes with the document that establishes it and the field where it reads. Never a flag without its proof.
The quarterly income declaration is the form on which a claimant declares what they received over the last three months; it determines the amount paid.
What I compare: the declared amounts, field by field, against the payslips, certificates and supporting documents in the file — and against the same household's declarations for previous quarters.
What one month produces: 214 discrepancies flagged across 2,400 files examined.
· 96 amount discrepancies between the declaration and the payslips provided.
· 61 changes of circumstance visible in a document and absent from the declaration — an address, an employer, a household composition.
· 57 fields entered as zero while a document in the file shows income over the period.
What a flag contains, and this is where it all turns: the declared amount, the amount read, the image of the document at the exact place where it reads, the period concerned and the field. The caseworker does not have to take my word for it: they look.
And what I add to the proof: the three possible readings of the discrepancy, already weighed against your own history — a misunderstood box 61%, a payslip that arrived after the declaration 26%, income genuinely left out 13%. Nine discrepancies in ten are explained without calling a family into question — and it is that sorting, done before you, that makes the tenth worth having. Characterising it belongs to the caseworker, and to the checking unit when it is called in: I hand them the three hypotheses already ranked, and examination time halves.
The figure that counts for families: 88 of those 214 discrepancies concern a payment that has not yet gone out. An overpayment caught before payment is a family that will not have to pay anything back — and one overpayment notice fewer to write, to contest and to recover.
What I suggest: handling the 214 by their effect on the family's entitlement, not in order of arrival. The sorting is done: the 88 caught before payment come first, then the 29 that change the amount paid by more than €50 a month, and the rest waits for the next review. declaration-discrepancies_214-flags-and-4-rules.pdfEvery discrepancy with its document, and 4 rules costed over 12 months
⛓ Sourced · quarterly income declarations, payslips and certificates in the file, previous quarters' history
The quarterly income declaration is the form on which a claimant declares what they received over the last three months; it determines the amount paid.
What I compare: the declared amounts, field by field, against the payslips, certificates and supporting documents in the file — and against the same household's declarations for previous quarters.
What one month produces: 214 discrepancies flagged across 2,400 files examined.
· 96 amount discrepancies between the declaration and the payslips provided.
· 61 changes of circumstance visible in a document and absent from the declaration — an address, an employer, a household composition.
· 57 fields entered as zero while a document in the file shows income over the period.
What a flag contains, and this is where it all turns: the declared amount, the amount read, the image of the document at the exact place where it reads, the period concerned and the field. The caseworker does not have to take my word for it: they look.
And what I add to the proof: the three possible readings of the discrepancy, already weighed against your own history — a misunderstood box 61%, a payslip that arrived after the declaration 26%, income genuinely left out 13%. Nine discrepancies in ten are explained without calling a family into question — and it is that sorting, done before you, that makes the tenth worth having. Characterising it belongs to the caseworker, and to the checking unit when it is called in: I hand them the three hypotheses already ranked, and examination time halves.
The figure that counts for families: 88 of those 214 discrepancies concern a payment that has not yet gone out. An overpayment caught before payment is a family that will not have to pay anything back — and one overpayment notice fewer to write, to contest and to recover.
What I suggest: handling the 214 by their effect on the family's entitlement, not in order of arrival. The sorting is done: the 88 caught before payment come first, then the 29 that change the amount paid by more than €50 a month, and the rest waits for the next review. declaration-discrepancies_214-flags-and-4-rules.pdfEvery discrepancy with its document, and 4 rules costed over 12 months
⛓ Sourced · quarterly income declarations, payslips and certificates in the file, previous quarters' history
And the false alarms? We are not going to call families in for nothing.
There were some, and here is the exact figure: of the 214 discrepancies flagged, 63 turned out to be groundless. 29% of my flags served no purpose.
The cause, and it is a single one in 80% of cases: 50 of the 63 concerned maintenance payments received, entered by families under “other income” rather than in the dedicated field. The amount was right, the box was not — and I was flagging a discrepancy where there was only a misread form.
What I did about it, in two steps: I refined the rule — maintenance found in another field, for an identical amount, becomes an entry note addressed to the caseworker — and I reported the wording of the field to the head of the benefits department. Fifty families in one month do not get it wrong by chance: it is the form that leads them there.
Result measured over the following two months: groundless flags fell from 29% to 11%, that is 39 families a month who stop being looked at for nothing.
What I do on top, and nobody has the time to do: I re-read your nine checking rules in the light of the 28,800 files actually examined over twelve months, and I have written you four new ones, in the form of your own, already costed. For each: the sentence in plain language, the number of flags it would have produced over twelve months, the share confirmed after examination, and the files it would have let through.
· “Field entered as zero for three consecutive quarters while a document in the file shows a payment over the period” — 340 flags, 214 confirmed after examination (63%), 11 files let through.
· “Return to work visible on a payslip and absent from the following quarter's declaration” — 268 flags, 191 confirmed (71%), 24 let through. Those 191 are overpayments caught before payment: 191 families with nothing to pay back.
· “Address differing between two documents filed in the same quarter” — 1,190 flags, 106 confirmed (9%). That one I advise against as it stands: it would have 1,084 households looked at for nothing.
· Its tightened version, which I wrote and ran over the same twelve months — “address differing and a change of household composition declared in the same quarter”: 148 flags instead of 1,190, all 106 confirmed cases kept, 42 households looked at instead of 1,084. 1,042 families spared, and not one confirmed case lost.
And putting them in force takes one signature: the head of the checking unit signs the rules they keep, and they apply that same evening. A written, published and dated rule can be argued with in front of a claimant; a departmental habit cannot. I save you the writing and the measuring; the decision takes ten minutes instead of a committee.
And for every rule put in force, I hand you two lists: the files it selects, and the files it excludes. It is the second that proves the rule is fair, and I hand it over as a matter of course with every rule put in force.
⛓ Sourced · follow-up on the 214 flags and their outcomes, 28,800 files examined over twelve months, wording of the declaration form
The cause, and it is a single one in 80% of cases: 50 of the 63 concerned maintenance payments received, entered by families under “other income” rather than in the dedicated field. The amount was right, the box was not — and I was flagging a discrepancy where there was only a misread form.
What I did about it, in two steps: I refined the rule — maintenance found in another field, for an identical amount, becomes an entry note addressed to the caseworker — and I reported the wording of the field to the head of the benefits department. Fifty families in one month do not get it wrong by chance: it is the form that leads them there.
Result measured over the following two months: groundless flags fell from 29% to 11%, that is 39 families a month who stop being looked at for nothing.
What I do on top, and nobody has the time to do: I re-read your nine checking rules in the light of the 28,800 files actually examined over twelve months, and I have written you four new ones, in the form of your own, already costed. For each: the sentence in plain language, the number of flags it would have produced over twelve months, the share confirmed after examination, and the files it would have let through.
· “Field entered as zero for three consecutive quarters while a document in the file shows a payment over the period” — 340 flags, 214 confirmed after examination (63%), 11 files let through.
· “Return to work visible on a payslip and absent from the following quarter's declaration” — 268 flags, 191 confirmed (71%), 24 let through. Those 191 are overpayments caught before payment: 191 families with nothing to pay back.
· “Address differing between two documents filed in the same quarter” — 1,190 flags, 106 confirmed (9%). That one I advise against as it stands: it would have 1,084 households looked at for nothing.
· Its tightened version, which I wrote and ran over the same twelve months — “address differing and a change of household composition declared in the same quarter”: 148 flags instead of 1,190, all 106 confirmed cases kept, 42 households looked at instead of 1,084. 1,042 families spared, and not one confirmed case lost.
And putting them in force takes one signature: the head of the checking unit signs the rules they keep, and they apply that same evening. A written, published and dated rule can be argued with in front of a claimant; a departmental habit cannot. I save you the writing and the measuring; the decision takes ten minutes instead of a committee.
And for every rule put in force, I hand you two lists: the files it selects, and the files it excludes. It is the second that proves the rule is fair, and I hand it over as a matter of course with every rule put in force.
⛓ Sourced · follow-up on the 214 flags and their outcomes, 28,800 files examined over twelve months, wording of the declaration form
And a full pre-assessment of an income support claim — what does it actually look like?
A file ready to decide on in ten minutes instead of fifty, with every amount tied to the document that establishes it.
Pre-assessing means gathering and checking everything the caseworker needs in order to decide — without taking the decision.
What the pre-assessed file contains, in this order:
· The household composition at the date of the claim, with the documents that establish it.
· The income base rebuilt over the required period, amount by amount, each one tied to its payslip or certificate. No amount without its document: an amount with no attachable supporting document is displayed as pending, with the exact document to request and the letter already written.
· The verifiable conditions, ticked with the rule quoted from your framework and its version.
· The points to reconcile with the claimant — two here: a period of employment overlapping two declarations, and an address that differs between two documents.
· The draft decision, written and reasoned both ways — grant and refusal, each with its grounds and its effective date. The caseworker appraises, chooses and signs; they no longer have to write.
The time this shifts: a full income support pre-assessment took 100% of the processing time, that is 50 minutes. It now takes 20% — 10 minutes of checking. Across 640 first claims a month, that is 427 hours returned every month.
And what I carry out the moment the sign-off is given, under a written mandate: notifying the claimant from your template, entering the effective date in the system, updating the portal and the payment schedule. Four acts that took the service two days and now take four minutes — capped to files already signed off, dated until 31/12, withdrawable in a word.
What I suggest next: applying the same pre-assessed file to reviews, which run at three times the volume of first claims. It is the same work on a flow three times the size — and that is where your 15-day service commitment is won or lost. income-support-pre-assessment_file-ready-to-decide.pdfEvery amount tied to its document, the decision left to the caseworker
⛓ Sourced · documents in the file, the office's internal framework, timings measured over one month
Pre-assessing means gathering and checking everything the caseworker needs in order to decide — without taking the decision.
What the pre-assessed file contains, in this order:
· The household composition at the date of the claim, with the documents that establish it.
· The income base rebuilt over the required period, amount by amount, each one tied to its payslip or certificate. No amount without its document: an amount with no attachable supporting document is displayed as pending, with the exact document to request and the letter already written.
· The verifiable conditions, ticked with the rule quoted from your framework and its version.
· The points to reconcile with the claimant — two here: a period of employment overlapping two declarations, and an address that differs between two documents.
· The draft decision, written and reasoned both ways — grant and refusal, each with its grounds and its effective date. The caseworker appraises, chooses and signs; they no longer have to write.
The time this shifts: a full income support pre-assessment took 100% of the processing time, that is 50 minutes. It now takes 20% — 10 minutes of checking. Across 640 first claims a month, that is 427 hours returned every month.
And what I carry out the moment the sign-off is given, under a written mandate: notifying the claimant from your template, entering the effective date in the system, updating the portal and the payment schedule. Four acts that took the service two days and now take four minutes — capped to files already signed off, dated until 31/12, withdrawable in a word.
What I suggest next: applying the same pre-assessed file to reviews, which run at three times the volume of first claims. It is the same work on a flow three times the size — and that is where your 15-day service commitment is won or lost. income-support-pre-assessment_file-ready-to-decide.pdfEvery amount tied to its document, the decision left to the caseworker
⛓ Sourced · documents in the file, the office's internal framework, timings measured over one month
A claimant is disputing a housing benefit overpayment. Draft a reply.
The draft is ready, with the six points to check highlighted, and it fits on one page the family will understand.
An overpayment is a sum paid in error that the office must recover; it most often arises from a change of circumstance learned about after the fact.
What the reply says, in this order:
· The reason for the overpayment — here, a return to work declared two months late — written in one sentence, not three lines of references.
· The periods concerned and the amount for each, month by month, because a total broken down is a total understood, and an understood total is disputed ten times less.
· What was paid and what was due, side by side.
· The routes and time limits for appeal, taken word for word from your templates — a mandatory notice is never rewritten.
· The option of asking for a write-off or an instalment plan, with how to do it and the documents to attach. It is the paragraph families miss most often, and it is set out plainly.
The six points highlighted for your review: the return-to-work date, recorded differently in two documents · the third month's amount, rounded in the system · the notification address · the signatory's authority · the limitation period, calculated and dated, to approve in one click · the earlier letter of September, for which I have asked the postal provider for proof of posting — it arrives within 48 hours.
The time this shifts: a notification letter took 40% of the processing time, that is 12 minutes. It now takes 10% — 3 minutes of review. Across 3,100 letters a month, that is 465 hours returned.
What goes out under mandate, as soon as the caseworker has signed off: I produce it, send it, file the copy and record the notification date that starts the appeal clock. Four acts that took the mail service two days and now take four minutes — capped to letters built from your templates, dated until 31/12, withdrawable in a word. Recovery is the public accountant's act, and that is what makes it sound: they receive it from me complete on the same day.
What I suggest taking on next: write-off requests, which today take three weeks. It is the paragraph I have just written into this letter, and the one the family cares about most. draft-reply_overpayment-dispute.pdfReason, periods, appeals, write-off — 6 points to check
⛓ Sourced · claimant's file, the office's letter templates, payment history over the period
An overpayment is a sum paid in error that the office must recover; it most often arises from a change of circumstance learned about after the fact.
What the reply says, in this order:
· The reason for the overpayment — here, a return to work declared two months late — written in one sentence, not three lines of references.
· The periods concerned and the amount for each, month by month, because a total broken down is a total understood, and an understood total is disputed ten times less.
· What was paid and what was due, side by side.
· The routes and time limits for appeal, taken word for word from your templates — a mandatory notice is never rewritten.
· The option of asking for a write-off or an instalment plan, with how to do it and the documents to attach. It is the paragraph families miss most often, and it is set out plainly.
The six points highlighted for your review: the return-to-work date, recorded differently in two documents · the third month's amount, rounded in the system · the notification address · the signatory's authority · the limitation period, calculated and dated, to approve in one click · the earlier letter of September, for which I have asked the postal provider for proof of posting — it arrives within 48 hours.
The time this shifts: a notification letter took 40% of the processing time, that is 12 minutes. It now takes 10% — 3 minutes of review. Across 3,100 letters a month, that is 465 hours returned.
What goes out under mandate, as soon as the caseworker has signed off: I produce it, send it, file the copy and record the notification date that starts the appeal clock. Four acts that took the mail service two days and now take four minutes — capped to letters built from your templates, dated until 31/12, withdrawable in a word. Recovery is the public accountant's act, and that is what makes it sound: they receive it from me complete on the same day.
What I suggest taking on next: write-off requests, which today take three weeks. It is the paragraph I have just written into this letter, and the one the family cares about most. draft-reply_overpayment-dispute.pdfReason, periods, appeals, write-off — 6 points to check
⛓ Sourced · claimant's file, the office's letter templates, payment history over the period
And the write-off, can you handle it? Today a write-off request takes three weeks.
I build the complete file in an hour, and I have written the route that gets simple cases decided in 48 hours instead of three weeks. The committee keeps its signature and gets its time back: it concentrates that time on the files that deserve it.
A write-off is the full or partial cancellation of an overpayment in view of the household's situation; an instalment plan spreads the repayment.
What I prepare for each request: the household's situation at the date of the request, disposable income calculated from the documents in the file rather than from a declaration, the history of earlier overpayments, the origin of this one — office error, late declaration, or deliberate concealment, and the distinction changes everything — and the reasoned proposal, with your committee's scale quoted and dated.
The route I have written for you, and it is ready to sign: a delegation of signature to the head of the benefits department for requests meeting three cumulative conditions — overpayment below €150, origin with no concealment, no earlier overpayment in 24 months.
· Capped to those three conditions, and nothing else.
· Dated — until 31/12, with a review at expiry.
· Withdrawable in a word, with no reason to give, effective immediately.
· Fully traced: every decision taken under delegation appears in the committee's monthly record.
What this changes, costed over your last twelve months: 61% of your write-off requests meet the three conditions. They would go from three weeks to 48 hours, and the committee would get its time back for the 39% that call for a real examination — those where a family is in lasting difficulty.
What the public accountant gets from me on the day of the decision, and did not have before: the origin of the overpayment established from documents, the capacity to repay calculated from the file, the proposed schedule and the letter ready. The movement of funds remains their act, and that is what makes it sound — what I give them back is the eleven days they spent rebuilding the file.
And the best overpayment is still the one that never happens: 88 discrepancies caught each month before the payment goes out means 88 families with nothing to pay back and 88 notices that never have to be written. delegation-of-signature_simple-write-offs.pdfThree conditions, capped, dated, withdrawable
✎ Framework · appeals committee scale, write-off files of the last twelve months, measured turnaround times
A write-off is the full or partial cancellation of an overpayment in view of the household's situation; an instalment plan spreads the repayment.
What I prepare for each request: the household's situation at the date of the request, disposable income calculated from the documents in the file rather than from a declaration, the history of earlier overpayments, the origin of this one — office error, late declaration, or deliberate concealment, and the distinction changes everything — and the reasoned proposal, with your committee's scale quoted and dated.
The route I have written for you, and it is ready to sign: a delegation of signature to the head of the benefits department for requests meeting three cumulative conditions — overpayment below €150, origin with no concealment, no earlier overpayment in 24 months.
· Capped to those three conditions, and nothing else.
· Dated — until 31/12, with a review at expiry.
· Withdrawable in a word, with no reason to give, effective immediately.
· Fully traced: every decision taken under delegation appears in the committee's monthly record.
What this changes, costed over your last twelve months: 61% of your write-off requests meet the three conditions. They would go from three weeks to 48 hours, and the committee would get its time back for the 39% that call for a real examination — those where a family is in lasting difficulty.
What the public accountant gets from me on the day of the decision, and did not have before: the origin of the overpayment established from documents, the capacity to repay calculated from the file, the proposed schedule and the letter ready. The movement of funds remains their act, and that is what makes it sound — what I give them back is the eleven days they spent rebuilding the file.
And the best overpayment is still the one that never happens: 88 discrepancies caught each month before the payment goes out means 88 families with nothing to pay back and 88 notices that never have to be written. delegation-of-signature_simple-write-offs.pdfThree conditions, capped, dated, withdrawable
✎ Framework · appeals committee scale, write-off files of the last twelve months, measured turnaround times
One problem remains: our letters are unreadable for many families. The switchboard pays for it.
I rewrote 400 letters following your templates and your plain-language charter, and the switchboard saw it before you did: calls asking what the letter meant more than halved.
What the rewriting changes, and it comes down to four simple moves:
· The amount and the date at the top, not in the third paragraph. A family is looking for two things: how much, and by when.
· One idea per paragraph, short sentences, active voice. “You must send us” rather than “it will be appropriate for transmission to be effected”.
· What to do, at the end of the letter, as a list, with the date expected.
· Abbreviations spelled out on first use. An abbreviation is obvious to whoever writes it and a wall to whoever reads it.
What is taken word for word, because it is what protects the claimant: mandatory notices, routes and time limits for appeal and your template references. Making a letter readable means moving the information around — never thinning out what the right of appeal rests on.
The measurement, across the 400 rewritten letters: calls received within ten days to understand the letter fell from 27% to 11%. These are not only hours returned to the switchboard: they are 64 families who understood first time what was being asked of them, so 64 documents filed on time and as many entitlements granted on time.
What I suggest: extending the rewrite to the eleven most-sent templates, which carry 74% of your letters. They are written and set side by side with the current ones; you approve them one at a time, and each stays editable in a word. letter-templates_rewriting-and-measurement.pdfCalls asking for clarification: 27% → 11%
⛓ Sourced · the office's letter templates, internal plain-language charter, log of calls received after sending
What the rewriting changes, and it comes down to four simple moves:
· The amount and the date at the top, not in the third paragraph. A family is looking for two things: how much, and by when.
· One idea per paragraph, short sentences, active voice. “You must send us” rather than “it will be appropriate for transmission to be effected”.
· What to do, at the end of the letter, as a list, with the date expected.
· Abbreviations spelled out on first use. An abbreviation is obvious to whoever writes it and a wall to whoever reads it.
What is taken word for word, because it is what protects the claimant: mandatory notices, routes and time limits for appeal and your template references. Making a letter readable means moving the information around — never thinning out what the right of appeal rests on.
The measurement, across the 400 rewritten letters: calls received within ten days to understand the letter fell from 27% to 11%. These are not only hours returned to the switchboard: they are 64 families who understood first time what was being asked of them, so 64 documents filed on time and as many entitlements granted on time.
What I suggest: extending the rewrite to the eleven most-sent templates, which carry 74% of your letters. They are written and set side by side with the current ones; you approve them one at a time, and each stays editable in a word. letter-templates_rewriting-and-measurement.pdfCalls asking for clarification: 27% → 11%
⛓ Sourced · the office's letter templates, internal plain-language charter, log of calls received after sending
2,900 contacts a month just to ask where a file stands. Our caseworkers do nothing else.
I answer 2,460 of them in under a minute, at any hour, with the real state of the file. And I pass 440 to a member of staff, because those deserve professional judgement.
What I answer, and where it comes from: the current stage, the document awaited if one is missing, the date the file will be examined — read from the system, so a date the office keeps. If a file has slipped, the claimant learns it from us before discovering it.
The 440 I pass on: an emergency — eviction, a utility cut-off, no income at all —, a disagreement with a decision, a social difficulty expressed, a request where the honest answer is “we need to talk about this”. An emergency gets a human voice within the minute, at any hour — that is what the other 2,460 answers make possible.
What I announce in the first sentence, and it is not a setting: I say that I am an artificial intelligence. Since 2 August 2026, article 50 §1 of the European AI regulation requires that anyone interacting with an AI system be informed, except where it is obvious. The announcement is in the welcome message, in the language of the person, and anyone can ask for a member of staff at any time — without having to insist and without giving a reason.
The time returned: your caseworkers spent the equivalent of two days a week per team on these status enquiries. They are left with the 440 real ones, and those call for an appraisal.
What I have already written and suggest you switch on: a message at each stage reached — filing recorded, file complete, decision notified. People ask again when they do not know: over the first three weeks of trial, “where does my file stand” contacts fell by 34%. The mandate is capped to those three stages, dated until 31/12, and withdrawable in a word. answering-claimants_2900-contacts-a-month.pdf2,460 immediate answers, 440 for a member of staff
⛓ Sourced · switchboard and online channel log, real file states, three-week trial
What I answer, and where it comes from: the current stage, the document awaited if one is missing, the date the file will be examined — read from the system, so a date the office keeps. If a file has slipped, the claimant learns it from us before discovering it.
The 440 I pass on: an emergency — eviction, a utility cut-off, no income at all —, a disagreement with a decision, a social difficulty expressed, a request where the honest answer is “we need to talk about this”. An emergency gets a human voice within the minute, at any hour — that is what the other 2,460 answers make possible.
What I announce in the first sentence, and it is not a setting: I say that I am an artificial intelligence. Since 2 August 2026, article 50 §1 of the European AI regulation requires that anyone interacting with an AI system be informed, except where it is obvious. The announcement is in the welcome message, in the language of the person, and anyone can ask for a member of staff at any time — without having to insist and without giving a reason.
The time returned: your caseworkers spent the equivalent of two days a week per team on these status enquiries. They are left with the 440 real ones, and those call for an appraisal.
What I have already written and suggest you switch on: a message at each stage reached — filing recorded, file complete, decision notified. People ask again when they do not know: over the first three weeks of trial, “where does my file stand” contacts fell by 34%. The mandate is capped to those three stages, dated until 31/12, and withdrawable in a word. answering-claimants_2900-contacts-a-month.pdf2,460 immediate answers, 440 for a member of staff
⛓ Sourced · switchboard and online channel log, real file states, three-week trial
And the families who cannot manage the portal? They are often the ones who need it most.
I guide them screen by screen, field by field, and not a character is entered without their say-so. 1,180 procedures guided this quarter — 232 were about to be abandoned.
How a guided session goes: I read the screen the person is on, I explain what the question is asking in everyday words, I say which document is expected and in what form, and I offer a front-desk appointment as soon as that is the shortest route for the family.
The 232 abandonments avoided, and what caused them: the income screen, at 71% — 165 of the 232. A family reaches the income section, does not know where to put a maintenance payment, an allowance or a bonus — and closes the page. It is the same screen that produced my 50 groundless maintenance flags: one single defect, seen from two places.
What I have written, and what is waiting for a signature: the eight screens that concentrate 84% of abandonments — 195 of the 232 — are rewritten, labels, entry help and question order included, in the form of your own portal and without changing a single item of data requested. For each: today's screen, the proposed screen, and the measurement run over the 4,720 procedures guided in the past year:
· Income screen: 660 abandonments over twelve months. The rewritten version, run over the same procedures: 168. 492 families who see it through.
· The seven other screens: 120 abandonments, 31 with the rewrite.
Your portal team puts them live when it decides to: it settles the matter on measured figures rather than on intuition, and on screens that are already written. I save you the writing and the measuring; the decision takes one meeting.
And for those who do not want the digital route: guidance also works by telephone, and it ends with an appointment at the front desk if the person asks, without having to explain themselves. Public service is not an online queue. online-procedure-guidance_1180-sessions.pdf232 abandonments avoided, 8 screens rewritten and costed
⛓ Sourced · twelve months of guidance logs, portal drop-off points, reasons given by users
How a guided session goes: I read the screen the person is on, I explain what the question is asking in everyday words, I say which document is expected and in what form, and I offer a front-desk appointment as soon as that is the shortest route for the family.
The 232 abandonments avoided, and what caused them: the income screen, at 71% — 165 of the 232. A family reaches the income section, does not know where to put a maintenance payment, an allowance or a bonus — and closes the page. It is the same screen that produced my 50 groundless maintenance flags: one single defect, seen from two places.
What I have written, and what is waiting for a signature: the eight screens that concentrate 84% of abandonments — 195 of the 232 — are rewritten, labels, entry help and question order included, in the form of your own portal and without changing a single item of data requested. For each: today's screen, the proposed screen, and the measurement run over the 4,720 procedures guided in the past year:
· Income screen: 660 abandonments over twelve months. The rewritten version, run over the same procedures: 168. 492 families who see it through.
· The seven other screens: 120 abandonments, 31 with the rewrite.
Your portal team puts them live when it decides to: it settles the matter on measured figures rather than on intuition, and on screens that are already written. I save you the writing and the measuring; the decision takes one meeting.
And for those who do not want the digital route: guidance also works by telephone, and it ends with an appointment at the front desk if the person asks, without having to explain themselves. Public service is not an online queue. online-procedure-guidance_1180-sessions.pdf232 abandonments avoided, 8 screens rewritten and costed
⛓ Sourced · twelve months of guidance logs, portal drop-off points, reasons given by users
We are under a lot of pressure to act on non-take-up. Can you do anything about that?
Yes, and without guessing at anything: I flagged 96 situations where a benefit condition is met on your own data and no claim exists, from the information families have already declared.
Non-take-up describes someone not claiming a benefit they might be entitled to — through not knowing, through discouragement, or because nobody told them.
How a situation is spotted, and what I do with it: I bring together the household composition, the declared income and the benefits in payment. Where the condition is met on the documents you already hold and no claim exists, I build the file and put it on the caseworker's desk, pre-assessed like any other. Assessing the entitlement stays theirs — and I hand it to them in ten minutes, documents already matched.
What goes out to the family: an invitation to check, written to make people act: “based on the information you have given us, you may be entitled to …, here is how to check in ten minutes”. That wording produced 41 claims out of 96 invitations — the best response rate I have measured on a non-mandatory letter.
What the quarter produced: 96 situations flagged, 41 claims filed, 33 entitlements granted by your caseworkers. Thirty-three households were receiving less than they were entitled to, and did not know it.
Equal treatment, because it is the first question you will be asked: the same rule is applied to all 118,000 claimants, with no targeting by neighbourhood, by length of claim or by any other population criterion. Geographic targeting would produce an unequal take-up rate between two identical families, and that inequality would show up in your own figures the following year.
What I suggest: putting the three most affected benefits through a systematic review each quarter. The rule is written and already run over twelve months; the spotting costs nothing extra, it works on information you already hold. non-take-up_96-situations-flagged.pdf41 claims filed, 33 entitlements granted
⛓ Sourced · information declared by households, benefits in payment, outcomes of the 96 flags
Non-take-up describes someone not claiming a benefit they might be entitled to — through not knowing, through discouragement, or because nobody told them.
How a situation is spotted, and what I do with it: I bring together the household composition, the declared income and the benefits in payment. Where the condition is met on the documents you already hold and no claim exists, I build the file and put it on the caseworker's desk, pre-assessed like any other. Assessing the entitlement stays theirs — and I hand it to them in ten minutes, documents already matched.
What goes out to the family: an invitation to check, written to make people act: “based on the information you have given us, you may be entitled to …, here is how to check in ten minutes”. That wording produced 41 claims out of 96 invitations — the best response rate I have measured on a non-mandatory letter.
What the quarter produced: 96 situations flagged, 41 claims filed, 33 entitlements granted by your caseworkers. Thirty-three households were receiving less than they were entitled to, and did not know it.
Equal treatment, because it is the first question you will be asked: the same rule is applied to all 118,000 claimants, with no targeting by neighbourhood, by length of claim or by any other population criterion. Geographic targeting would produce an unequal take-up rate between two identical families, and that inequality would show up in your own figures the following year.
What I suggest: putting the three most affected benefits through a systematic review each quarter. The rule is written and already run over twelve months; the spotting costs nothing extra, it works on information you already hold. non-take-up_96-situations-flagged.pdf41 claims filed, 33 entitlements granted
⛓ Sourced · information declared by households, benefits in payment, outcomes of the 96 flags
We handle incomes, household compositions, social circumstances. Where does that data go?
Nowhere. I run on a resource dedicated strictly to your office, hosted in France, under French law — or on a machine of your own if you prefer.
The two modes, for the office to choose:
· Local inference — the model computes on a machine inside your network: no claimant data leaves the body, nothing passes through a cloud.
· A resource isolated per body, hosted in France. No pooling with another office: your environment is yours alone, and that is what guarantees continuity of your service.
What protects families, point by point: encryption in transit and at rest · role-based access, meaning rights that follow the function and not the person · strong authentication · logging · hosting whose architecture is designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including against a US player hosting in Europe. Every access is traced and withdrawable in a word.
Two points your data protection officer will look at first:
· No claimant data trains a model. What is learned here stays here.
· No grant, revision or refusal of entitlement is automated — and that is what makes every decision reasoned, dated, traced and open to challenge by the family: a claimant can have it reviewed. What I change is the delay: the reasoned decision arrives in minutes where it used to take weeks.
For the most sensitive data, reinforced hosting options exist — SecNumCloud, and health-data hosting as soon as health information enters a file. A single architecture is designed to answer both the GDPR and extraterritorial exposure.
What I suggest: that I keep up to date the table of processing operations, retention periods and access by role, in the form your data protection impact assessment expects. It is the document that gets asked for once a year and takes three weeks to rebuild; it is ready and dated this evening. technical-framework_where-family-data-lives.pdfFrance, resource isolated per body, 0 transfer outside the EU in the target architecture
✎ Framework · deployment architecture, access log, quarterly transfer record
The two modes, for the office to choose:
· Local inference — the model computes on a machine inside your network: no claimant data leaves the body, nothing passes through a cloud.
· A resource isolated per body, hosted in France. No pooling with another office: your environment is yours alone, and that is what guarantees continuity of your service.
What protects families, point by point: encryption in transit and at rest · role-based access, meaning rights that follow the function and not the person · strong authentication · logging · hosting whose architecture is designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including against a US player hosting in Europe. Every access is traced and withdrawable in a word.
Two points your data protection officer will look at first:
· No claimant data trains a model. What is learned here stays here.
· No grant, revision or refusal of entitlement is automated — and that is what makes every decision reasoned, dated, traced and open to challenge by the family: a claimant can have it reviewed. What I change is the delay: the reasoned decision arrives in minutes where it used to take weeks.
For the most sensitive data, reinforced hosting options exist — SecNumCloud, and health-data hosting as soon as health information enters a file. A single architecture is designed to answer both the GDPR and extraterritorial exposure.
What I suggest: that I keep up to date the table of processing operations, retention periods and access by role, in the form your data protection impact assessment expects. It is the document that gets asked for once a year and takes three weeks to rebuild; it is ready and dated this evening. technical-framework_where-family-data-lives.pdfFrance, resource isolated per body, 0 transfer outside the EU in the target architecture
✎ Framework · deployment architecture, access log, quarterly transfer record
After a quarter, what has it given back to the office?
1,492 hours a month — 9.8 full-time equivalents — and not one post cut: ten caseworkers returned to the situations that call for professional judgement.
The three items you were measuring:
· Completeness check on a file: 60% → 10% of the processing time. 18 minutes become 3, that is 600 hours a month.
· Notification letter: 40% → 10%. 12 minutes become 3, that is 465 hours a month.
· Full pre-assessment of an income support claim: 100% → 20%. 50 minutes become 10, that is 427 hours a month.
What families saw, and it is the only figure that counts for them: the average time on a first in-work benefit claim went from 21 days to 12. Your 15-day service commitment is being kept, and kept across all files, not on average.
The quarter's other measurements: requests for documents 31% → 12% · groundless flags 29% → 11% · calls for clarification after a letter 27% → 11% · 33 entitlements granted through non-take-up work · 264 overpayments caught before payment, that is 264 families with nothing to pay back.
What those hours became, from your own records: +38% of time on guidance appointments, and two rural drop-in offices reopened, closed for two years for want of available staff. It is the only line of this review that can be seen from the front desk.
And the framework measures: 0 entitlements granted, revised or refused without a caseworker's sign-off, out of 41,000 files processed · 0 data leaving the European Union · 0 automated individual decisions.
The next step I propose, with its figures: extend the pre-assessed file to reviews. Three times the volume of first claims, the same mechanics already proven — it is the lever that takes your turnaround from 12 days to under 10 across the whole flow. quarterly-review_what-is-returned-to-the-service.pdf1,492 hours a month, turnaround 21 → 12 days
⛓ Sourced · quarterly activity records, sign-off log, measured turnaround on first claims
The three items you were measuring:
· Completeness check on a file: 60% → 10% of the processing time. 18 minutes become 3, that is 600 hours a month.
· Notification letter: 40% → 10%. 12 minutes become 3, that is 465 hours a month.
· Full pre-assessment of an income support claim: 100% → 20%. 50 minutes become 10, that is 427 hours a month.
What families saw, and it is the only figure that counts for them: the average time on a first in-work benefit claim went from 21 days to 12. Your 15-day service commitment is being kept, and kept across all files, not on average.
The quarter's other measurements: requests for documents 31% → 12% · groundless flags 29% → 11% · calls for clarification after a letter 27% → 11% · 33 entitlements granted through non-take-up work · 264 overpayments caught before payment, that is 264 families with nothing to pay back.
What those hours became, from your own records: +38% of time on guidance appointments, and two rural drop-in offices reopened, closed for two years for want of available staff. It is the only line of this review that can be seen from the front desk.
And the framework measures: 0 entitlements granted, revised or refused without a caseworker's sign-off, out of 41,000 files processed · 0 data leaving the European Union · 0 automated individual decisions.
The next step I propose, with its figures: extend the pre-assessed file to reviews. Three times the volume of first claims, the same mechanics already proven — it is the lever that takes your turnaround from 12 days to under 10 across the whole flow. quarterly-review_what-is-returned-to-the-service.pdf1,492 hours a month, turnaround 21 → 12 days
⛓ Sourced · quarterly activity records, sign-off log, measured turnaround on first claims
Last question, and it is the one staff representatives will ask: what does the agent do without being asked?
Three actions on its own initiative, three mandates you gave me in writing, and five acts that wait for a sign-off. The list is on display, and every line is withdrawable in a word.
The three actions I take alone:
· Attach a document the household has already filed. And the reverse is true too: a caseworker detaches it in one click if the match is wrong, and the rejected match does not come back. It is the action that took document requests from 31% to 12%.
· Acknowledge a filing within five minutes, with the list of what is still awaited. The acknowledgement lists what has arrived and what is still awaited: it is what removes tomorrow's phone call.
· Schedule a reminder at day 10 on a requested document. Any response from the claimant stops it, including a call to the switchboard.
The three written mandates, all capped, dated to 31/12 and withdrawable with no reason given: sending out the letters built from your templates once signed off · carrying out the signed decision — notification, effective date, portal, payment schedule, in four minutes instead of two days · sending the stage message at the three milestones you chose.
The five acts that wait for a sign-off: granting, revising or refusing an entitlement · notifying an overpayment · granting a write-off or an instalment plan · opening a check · closing an incomplete file. They are exactly the five that affect a family, and the system makes them decidable in minutes: a complete file, every amount tied to its document, the grounds written out — all that is left is the act of signing.
What activity monitoring measures, and why that is a design choice: flows — queues, turnaround times and reasons for return —, never people. A per-caseworker file counter would become a target; the target would pick the quickest file out of the queue; and complex situations, the very ones the returned time is meant for, would come last. Measuring flows protects exactly what the system gains you — and it is the answer your staff representatives are waiting for.
What I suggest for the coming quarter: extend the pre-assessed file to reviews, three times more numerous than first claims · put live the eight portal screens I have rewritten, which carry 84% of abandonments · put three benefits through a non-take-up review. The first weighs most on your service commitment, and it is ready to start on Monday. who-decides-what_3-actions-and-5-acts-to-sign-off.pdfWhat runs on its own, what runs under mandate, what waits for a sign-off
✎ Framework · list of automatic actions and mandates, sign-off log, activity monitoring settings
The three actions I take alone:
· Attach a document the household has already filed. And the reverse is true too: a caseworker detaches it in one click if the match is wrong, and the rejected match does not come back. It is the action that took document requests from 31% to 12%.
· Acknowledge a filing within five minutes, with the list of what is still awaited. The acknowledgement lists what has arrived and what is still awaited: it is what removes tomorrow's phone call.
· Schedule a reminder at day 10 on a requested document. Any response from the claimant stops it, including a call to the switchboard.
The three written mandates, all capped, dated to 31/12 and withdrawable with no reason given: sending out the letters built from your templates once signed off · carrying out the signed decision — notification, effective date, portal, payment schedule, in four minutes instead of two days · sending the stage message at the three milestones you chose.
The five acts that wait for a sign-off: granting, revising or refusing an entitlement · notifying an overpayment · granting a write-off or an instalment plan · opening a check · closing an incomplete file. They are exactly the five that affect a family, and the system makes them decidable in minutes: a complete file, every amount tied to its document, the grounds written out — all that is left is the act of signing.
What activity monitoring measures, and why that is a design choice: flows — queues, turnaround times and reasons for return —, never people. A per-caseworker file counter would become a target; the target would pick the quickest file out of the queue; and complex situations, the very ones the returned time is meant for, would come last. Measuring flows protects exactly what the system gains you — and it is the answer your staff representatives are waiting for.
What I suggest for the coming quarter: extend the pre-assessed file to reviews, three times more numerous than first claims · put live the eight portal screens I have rewritten, which carry 84% of abandonments · put three benefits through a non-take-up review. The first weighs most on your service commitment, and it is ready to start on Monday. who-decides-what_3-actions-and-5-acts-to-sign-off.pdfWhat runs on its own, what runs under mandate, what waits for a sign-off
✎ Framework · list of automatic actions and mandates, sign-off log, activity monitoring settings
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 for the family benefits caseworker
Each use corresponds to an agent we deploy. All work in support, subject to the officer's approval.
Included in your agent
The 6 capabilities essential to this promise are included, at no extra cost.
From 1,020 € incl. VAT / month
Pre-assessment of applications
Checking documents and income conditions, flagging discrepancies before the caseworker's decision.
Letters & notifications
Writing the letters and notifications to claimants from your templates, ready for review.
Spotting anomalies
Detection of anomalies and inconsistencies to check before handling, with no automated decision.
Completeness checks
Checking income support, housing benefit and in-work benefit files, with a list of the missing documents to request. The request for a document can be announced by text message, without ever naming the benefit.
Declaration inconsistencies
Reconciling declarations and documents to flag what needs to be compared before a decision.
Information for claimants
Answer everyday questions (entitlements, procedures, documents) and direct people to the right service, 24/7.
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
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
Step-by-step support for the claimant through the online procedure is handed to the dedicated agent, which guides the form filling and gathers the documents. The scope of this page stops at pre-assessment and information; human validation stays with the public officer.
Support with online procedures from 780 € incl. VAT / month Help with procedures →Pre-assessment of social assistance (local welfare centre)
Strictly in support (administrative). Award decision by a human.
Pre-assessment of social assistance (local welfare centre) from 1,050 € 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 a family benefits caseworker win back?
By automating the checking of documents, the letters and the spotting of anomalies, a fund can aim for a clear reduction in time spent on repetitive tasks — reinvested in handling complex situations.
Checking the completeness of a file
A notification letter to a claimant
Full pre-assessment of an income support file
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
A benefits agent (pre-assessment, checking documents, letters to claimants), installed and operated for you. Choose according to how you are organised — available by direct award below the public procurement thresholds.
Agility
Setup + controlled subscription
11,605 € incl. VAT setup
then 1,020 € 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
1,665 € 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
16,586 € incl. VAT setup
then 1,280 € 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 family benefits fund
Families' data never leaves the organisationLocal inference or an isolated resource hosted in France; no data entrusted to a foreign third party.
Data in France, under French lawExposure of families' data to the Cloud Act and FISA 702 is reduced by design, that data being located and minimised in France; location alone does not guarantee immunity.
The caseworker keeps the decisionThe AI agent prepares checks, letters and summaries that can be verified; no entitlement is opened automatically.
Human oversight & traceabilityPre-assessment of claims: monitoring, updates and logging, compliant with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent open entitlements automatically?
No: it prepares and checks; opening or revising an entitlement remains a caseworker's decision, with no automation of decisions affecting claimants.
How is families' data protected?
Hosting in France, a resource isolated for each organisation, encryption and logging. GDPR: governed deployment.
Does the agent decide whether an entitlement is opened?
No: it pre-assesses and prepares the decision; granting or refusing a benefit remains an act of the caseworker, in line with the AI Act.
Does it handle overpayments?
It spots the situations to be examined and prepares the notification letters, but recovery and write-offs are a matter for the caseworker.
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, then a phase of design, integration and testing before going live and handing over to your teams.
Do we need a technical team in-house?
No: the agent is designed, installed and operated for you — integration, supervision and updates included. Your teams concentrate on their work, with guided familiarisation.
Do we have to change software?
No. The agent relies on your templates and complements your existing tools, with no migration imposed. We adapt the integration to your environment.
Which tools can benefit recipients use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: a benefit recipient has no account to create and no application to install. Because family data is sensitive, the agent announces without detailing: it signals that a document is missing from the file without ever naming the benefit or the situation. 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, and the gateway bills the text messages sent — are passed on at actual cost, with no margin, outside the subscription.
Can the agent notify claimants by text message?
Yes, as an option — and it is the channel that reaches the people the others never do. A text message arrives on an ordinary phone: no email address, no account, no app, no data plan and no smartphone. Telling claimants by text message that a document is missing from their file, before an entitlement lapses means fewer entitlements going unclaimed and more equal access to the service. It is an outbound notification channel, not a conversation channel: users do not write to the agent there — they reach you on WhatsApp Business, the website chat or email. The message announces without detailing — never the reason, never the detail of the situation — and never contains a payment link, a login or a password: users reach the service the way they usually do. Delivery goes through a gateway established in the European Union; the messages sent are passed on at actual cost, with no margin, and the channel itself is an optional module, set out on the pricing page.
Let's talk
Let's size up the potential in your fund
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.