An orchestration agent, scoped with you before it is priced. This agent coordinates several specialised agents. Its orchestration follows your actual workflows — which is why it is scoped with you rather than bought off the shelf. We establish the scope together, then the quotation commits it.
The journeys described below form the scope that this review refines and the quotation commits. The specialised agents it coordinates can be ordered today.
Request a quote
● Public sector — Social benefits
The AI agent for benefits officers: prepare the files, answer the claimants
Completeness checks, chasing missing documents and repetitive replies to claimants take up a considerable share of officers' time — without being the heart of the job: granting entitlements accurately. Your AI agent absorbs that preparatory work. Hosted in France — on local inference or an isolated resource — claimants' data stays under control. The AI agent assists, the public officer decides.
Hosted in France
Claimants' data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Social benefits assistant · hosted in France
On Ms Diallo's file (benefit application), check completeness and prepare the preliminary assessment before I rule.
File analysed. 6 documents out of 8 present. 2 missing: the latest tax notice and proof of address less than 3 months old. The means test can be completed as soon as they arrive. On the basis of the rules in force, the application appears admissible subject to those documents.
Draft reminder letter ready — for approval.
⛓ Source · the file's documents + the scale in force
Draft reminder letter ready — for approval.
⛓ Source · the file's documents + the scale in force
Ms Diallo is calling to find out where her file stands. What can I tell her?
Status of the file: awaiting 2 documents (tax notice, proof of address), reminder sent on 3 June. Indicative assessment time once the file is complete: in line with the published deadline. I can prepare a clear summary of the documents expected and the conditions, in plain language.
The decision to grant entitlement remains reserved to you.
✎ Action · summary ready for review — the officer approves
The decision to grant entitlement remains reserved to you.
✎ Action · summary ready for review — the 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 assists, never rules alone
✦ In brief
In a social security fund or a benefits department, a Blue Lemon Agent agent automates the preparatory tasks — completeness checks, chasing documents, preliminary assessment against the rules, replies to claimants — so the officer can concentrate on the decision. It runs on local inference or is hosted in France: claimants' data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Granting entitlement remains a framed human act (high-risk use). Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and your members 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
9
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 benefits departments — and why they hesitate
Claimants expect short turnaround times and clear answers, with entitlements unchanged. But officers' time is mechanically absorbed by document checks and repetitive enquiries — and the data involved is among the most sensitive there is.
! The issue
The department is caught between claimants who want controlled turnaround times and equal treatment, and an ever-growing assessment workload (completeness, reminders, enquiries). Yet most consumer AI solutions amount to entrusting claimants' income, family situation, health data and identity to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For a public service, 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 on entitlement reserved to the officer: the time saved on preparation is never paid for in lost confidentiality, nor in a breach of equal treatment. The aim is not to replace the public officer, but to give them back time for the situations that deserve it.
The decisive point
Protecting claimants' data: sovereignty & compliance
A benefits department handles users' most sensitive data. 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 department: no claimant document 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
Exposure of claimants' data to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
One isolated resource per department
No pooling of claimants' data: an environment strictly dedicated to your organisation.
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 entitlement granted 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 (health, income), 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
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'Ancre Benefits Fund — social benefits organisation
- Sector
- Social benefits organisation — assessing claims, opening and revising entitlements, controls, claimant reception; 6 schemes administered, including a housing allowance and two discretionary schemes voted by the board
- Headcount
- 84 staff — including 31 benefits caseworkers, 9 on reception and the switchboard, 4 on controls and one data protection officer
- Public served
- 62,000 claimants — 41,300 claims filed a year, of which 34,000 standard files, and 88,000 questions asked by phone, by e-mail or at the counter
- Order of magnitude
- A standard file takes 50 minutes of officer time from filing to decision, a controlled file 60 minutes; the processing time advertised to claimants is 30 days
- Tools in place
- Benefits management software, online filing portal, document management system, switchboard and 3 years of filed cases — the agent plugs into them read-only, nothing is replaced and nothing is migrated
- Who decides
- The benefits caseworker opens, revises or suspends an entitlement and signs the decision; the head of the benefits department approves information campaigns and mandates; the data protection officer keeps the register of processing
- Room for improvement
- 46 % of files arrive incomplete; average processing time is 34 days against an advertised 30; 34 % of claimant calls ring unanswered; and 2,180 claimants meet the conditions of a benefit they have never claimed
At Val d'Ancre, 31 caseworkers assess 41,300 claims a year for 62,000 claimants. The job is not to check documents: it is to open entitlements accurately, within the times announced, and in the same way for everyone. The data handled — income, family circumstances, sometimes health — is among the most sensitive there is. The agent runs on local inference on a machine in the department, reads the three years of files without ever writing to them, and no decision on an entitlement leaves without a public official's signature. The exchanges that follow span a year, from the review of the files to the results presented to the board.
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.
Social benefits assistant · hosted in France
I reread your three years of files over the weekend — 123,900 claims assessed — and I draw two measurements from them that your records were carrying without anyone having the time to see them.
The first, and it governs everything else: 46 % of your files arrive incomplete, and 8 missing documents out of 10 come down to four documents only — the latest tax assessment, proof of address less than three months old, bank details, a certificate of family circumstances. This is not 19,000 claimants being careless: it is a leaflet that does not say clearly enough what must be enclosed.
The second, and it is the one your board will care about: 2,180 claimants already known to your records today meet the conditions of a benefit they have never claimed. These are entitlements opened by your own rules and never exercised, for want of information that reaches them.
What the three items you already measure cost:
· Checking a file for completeness accounts for 60 % of preparation time — 30 minutes out of the 50 a standard file takes from filing to decision.
· Detecting an anomaly on a file: 37 % of the time a control takes, that is 22 minutes out of 60.
· Answering a recurring claimant question: 10 %, that is 5 minutes.
What I propose, and it is not a promise — it is already written: the four document leaflets are rewritten in plain language, each with an example of the document expected and the most frequent reason for refusal, and the completeness check runs from the moment a claim is filed, on the portal as at the counter. What that would have changed over the past year, since that is the only measure that counts: 18,900 files would have left complete the same day.
The time that moves: completeness drops from 60 % to 18 % — from 30 minutes to 9.
The next step, and it takes half an hour: you reread the four leaflets, I present them to you tomorrow morning. The head of department approves them, and they are on the portal that same evening. current-picture_41300-claims-and-three-time-items.pdf46 % of incomplete files, 4 documents for 8 gaps out of 10
⛓ Sourced · 3 years of assessed files, reasons for missing documents, portal leaflets, 4 leaflets rewritten
The first, and it governs everything else: 46 % of your files arrive incomplete, and 8 missing documents out of 10 come down to four documents only — the latest tax assessment, proof of address less than three months old, bank details, a certificate of family circumstances. This is not 19,000 claimants being careless: it is a leaflet that does not say clearly enough what must be enclosed.
The second, and it is the one your board will care about: 2,180 claimants already known to your records today meet the conditions of a benefit they have never claimed. These are entitlements opened by your own rules and never exercised, for want of information that reaches them.
What the three items you already measure cost:
· Checking a file for completeness accounts for 60 % of preparation time — 30 minutes out of the 50 a standard file takes from filing to decision.
· Detecting an anomaly on a file: 37 % of the time a control takes, that is 22 minutes out of 60.
· Answering a recurring claimant question: 10 %, that is 5 minutes.
What I propose, and it is not a promise — it is already written: the four document leaflets are rewritten in plain language, each with an example of the document expected and the most frequent reason for refusal, and the completeness check runs from the moment a claim is filed, on the portal as at the counter. What that would have changed over the past year, since that is the only measure that counts: 18,900 files would have left complete the same day.
The time that moves: completeness drops from 60 % to 18 % — from 30 minutes to 9.
The next step, and it takes half an hour: you reread the four leaflets, I present them to you tomorrow morning. The head of department approves them, and they are on the portal that same evening. current-picture_41300-claims-and-three-time-items.pdf46 % of incomplete files, 4 documents for 8 gaps out of 10
⛓ Sourced · 3 years of assessed files, reasons for missing documents, portal leaflets, 4 leaflets rewritten
We are 31 caseworkers for 41,300 claims and six schemes. We cannot open everything at once: where do we start?
With one scheme, not six — and I have ranked them by the time they give you back, not by nomenclature order.
What the count over your three years says:
· The housing allowance: 14,200 claims a year, that is 34 % of your flow, and it is the scheme where completeness costs the most — the tax assessment is missing one time in three.
· The two discretionary schemes voted by your board: 9,400 claims, and their scales change every year — that is where reading an up-to-date text counts most.
· The four other schemes: 17,700 claims, steadier, with fewer documents.
In other words: one scheme out of six carries more than a third of your completeness workload. The other five can wait until autumn without anything getting worse for a claimant.
What I bring you within that half-hour, and that nobody has the time to do: the rewritten leaflet for the housing scheme, the completeness check already set up on its 8 required documents, and the number of files each of those two measures would have sent back complete last year — 6,400 and 12,500.
The gain, measured on that scheme alone: 14,200 files, 21 minutes given back on each — 4,970 caseworker hours that go back to assessing the situations that call for judgement.
What I propose next: that I flag of my own accord any scale or required document that a decision of your board amends, and that I write the updated version overnight after the meeting. Your leaflets will stop ageing in silence — that is what makes sure a claimant is never asked for a document the rules no longer require. current-picture_41300-claims-and-three-time-items.pdfOne scheme out of six carries 34 % of the flow
⛓ Sourced · claim counts by scheme over 3 years, reasons for missing documents, rewritten housing leaflet
What the count over your three years says:
· The housing allowance: 14,200 claims a year, that is 34 % of your flow, and it is the scheme where completeness costs the most — the tax assessment is missing one time in three.
· The two discretionary schemes voted by your board: 9,400 claims, and their scales change every year — that is where reading an up-to-date text counts most.
· The four other schemes: 17,700 claims, steadier, with fewer documents.
In other words: one scheme out of six carries more than a third of your completeness workload. The other five can wait until autumn without anything getting worse for a claimant.
What I bring you within that half-hour, and that nobody has the time to do: the rewritten leaflet for the housing scheme, the completeness check already set up on its 8 required documents, and the number of files each of those two measures would have sent back complete last year — 6,400 and 12,500.
The gain, measured on that scheme alone: 14,200 files, 21 minutes given back on each — 4,970 caseworker hours that go back to assessing the situations that call for judgement.
What I propose next: that I flag of my own accord any scale or required document that a decision of your board amends, and that I write the updated version overnight after the meeting. Your leaflets will stop ageing in silence — that is what makes sure a claimant is never asked for a document the rules no longer require. current-picture_41300-claims-and-three-time-items.pdfOne scheme out of six carries 34 % of the flow
⛓ Sourced · claim counts by scheme over 3 years, reasons for missing documents, rewritten housing leaflet
All of this assumes you read tax assessments, family circumstances, sometimes health information. Where does that data go?
It goes nowhere. I run on local inference on a machine in the department, and nothing I read leaves your network.
Local inference means the model computes on your machine: the text of a tax assessment or a certificate crosses no outside network to be processed. If the organisation would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to Val d'Ancre — no pooling with another organisation, which is the condition of your service's continuity.
What that changes, point by point:
· Your claimants' data trains no model, neither ours nor a third party's.
· I work read-only on your files, and the technical account I read through has no right to write — that is stronger than a promise, because it can be verified with one command.
· Encryption in transit and at rest, role-based access — rights follow the job: a reception officer sees the status of a file, not the income documents; controls see the documents, not the reception exchanges. 7 roles for your 84 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· Logging: who asked what, when, and what the system produced.
And the act the law reserves to a person, which is exactly what gives your decisions their value: a decision that opens, revises, suspends or recovers an entitlement has effects on the claimant — it cannot rest on automated processing alone. It is taken by one of your caseworkers, reasoned, dated, traced, and the claimant can challenge it. Everything leading up to it, I have already done: documents read, scale applied line by line, reasoning written, and the note comes back with what complies, what is missing and the scheme's rule alongside. A decision nobody had signed could be challenged by nobody, and would be worth nothing to a claimant.
The figure that sums this up: 0 claimant data out of the department's network across the 41,300 claims of the year, and processing in the EU targeted.
What I propose: that I keep up to date the record your data protection officer and your board will ask for — hosting, data processed, retention periods, who accesses what. It is asked for once a year and takes three days to rebuild; the first version is already written and attached. technical-framework_where-claimant-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account rights, access log, first version of the register record
Local inference means the model computes on your machine: the text of a tax assessment or a certificate crosses no outside network to be processed. If the organisation would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to Val d'Ancre — no pooling with another organisation, which is the condition of your service's continuity.
What that changes, point by point:
· Your claimants' data trains no model, neither ours nor a third party's.
· I work read-only on your files, and the technical account I read through has no right to write — that is stronger than a promise, because it can be verified with one command.
· Encryption in transit and at rest, role-based access — rights follow the job: a reception officer sees the status of a file, not the income documents; controls see the documents, not the reception exchanges. 7 roles for your 84 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· Logging: who asked what, when, and what the system produced.
And the act the law reserves to a person, which is exactly what gives your decisions their value: a decision that opens, revises, suspends or recovers an entitlement has effects on the claimant — it cannot rest on automated processing alone. It is taken by one of your caseworkers, reasoned, dated, traced, and the claimant can challenge it. Everything leading up to it, I have already done: documents read, scale applied line by line, reasoning written, and the note comes back with what complies, what is missing and the scheme's rule alongside. A decision nobody had signed could be challenged by nobody, and would be worth nothing to a claimant.
The figure that sums this up: 0 claimant data out of the department's network across the 41,300 claims of the year, and processing in the EU targeted.
What I propose: that I keep up to date the record your data protection officer and your board will ask for — hosting, data processed, retention periods, who accesses what. It is asked for once a year and takes three days to rebuild; the first version is already written and attached. technical-framework_where-claimant-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account rights, access log, first version of the register record
Mrs Aline Kerbrat has just filed a housing allowance claim at the counter. What exactly do you do with it?
The file is read, checked and pre-sorted in three minutes, and Mrs Kerbrat leaves the counter knowing exactly where she stands.
Completeness is the check that every document required by the scheme is present; pre-assessment is the preparatory work that sets the claim against the rules, before the caseworker rules on it.
What I hand over, document by document:
· The 8 documents required by the scheme, ticked one by one: 7 present, 1 missing — the proof of address less than three months old, the one she enclosed being eleven months old.
· The values extracted and verified: reference taxable income, household composition, rent amount, name of the bank account holder. Each value carries the document and page it comes from, and can be checked with one click — you do not take my word for it, you look at the line.
· Two consistency checks: the name on the bank details matches the claim, and the rent declared matches the rent receipt enclosed. On one of the two, a gap: the receipt says €612, the claim €621. That is not a false declaration, it is two digits swapped — and the line is flagged as such to the caseworker, with both documents side by side.
· The summary handed to the claimant, in plain language: what has been received, what is missing, by when, and what happens next.
The time that moves: the completeness check goes from 60 % to 18 % of preparation time — from 30 minutes to 9. Across 41,300 claims a year, it is the back-and-forth that disappears, not the check.
The next step I propose: that this check also runs on portal filings, overnight after they are filed, and that the claimant receives their summary on waking. Over last year, that is 18,900 files that would have left complete without an officer having to reopen anything — approve the principle and the first batch goes out tomorrow. reading-the-documents_kerbrat-file.pdf8 documents ticked, 4 values extracted and sourced, 1 gap flagged
⛓ Sourced · documents filed, list of documents required by the scheme, rent receipt and bank details in the file
Completeness is the check that every document required by the scheme is present; pre-assessment is the preparatory work that sets the claim against the rules, before the caseworker rules on it.
What I hand over, document by document:
· The 8 documents required by the scheme, ticked one by one: 7 present, 1 missing — the proof of address less than three months old, the one she enclosed being eleven months old.
· The values extracted and verified: reference taxable income, household composition, rent amount, name of the bank account holder. Each value carries the document and page it comes from, and can be checked with one click — you do not take my word for it, you look at the line.
· Two consistency checks: the name on the bank details matches the claim, and the rent declared matches the rent receipt enclosed. On one of the two, a gap: the receipt says €612, the claim €621. That is not a false declaration, it is two digits swapped — and the line is flagged as such to the caseworker, with both documents side by side.
· The summary handed to the claimant, in plain language: what has been received, what is missing, by when, and what happens next.
The time that moves: the completeness check goes from 60 % to 18 % of preparation time — from 30 minutes to 9. Across 41,300 claims a year, it is the back-and-forth that disappears, not the check.
The next step I propose: that this check also runs on portal filings, overnight after they are filed, and that the claimant receives their summary on waking. Over last year, that is 18,900 files that would have left complete without an officer having to reopen anything — approve the principle and the first batch goes out tomorrow. reading-the-documents_kerbrat-file.pdf8 documents ticked, 4 values extracted and sourced, 1 gap flagged
⛓ Sourced · documents filed, list of documents required by the scheme, rent receipt and bank details in the file
And when a document is unreadable, or a value cannot be read? That is where mistakes happen, and the claimant pays for them.
A value I have not read, I do not write — I ask for it, and I ask for it the same day, from the right person, with the request already drafted.
What that gives, measured over the quarter:
· 1,240 values unreadable or absent — cropped scan, blurred photo, truncated tax assessment.
· 1,240 requests prepared, each naming the document expected, the page that is missing and the reason in one sentence — never “document not compliant”, which tells nobody anything and brings back the wrong paper.
· 1,108 replies back within eight days, against an average of 21 days last year for the same request.
What never happens, and it is what makes the rest reliable: a missing document does not become “probably provided”, an unreadable income does not become “presumably unchanged”. I say what is missing, where I looked, and who holds it.
What I add and nobody has the time to do: when the document already exists in an earlier file of the claimant and is still valid, I bring it up instead of asking for it. Over the quarter, 460 documents did not have to be requested again — 460 claimants who were not asked for a paper they had already provided, which is the leading cause of a claim being abandoned.
The reminder letter itself is written on your template: documents expected, deadline, filing address, and the file number prominently displayed. It goes out on the day of the check, against 9 days on average last year.
The next step I propose: reminding through the channel the claimant used themselves — e-mail, message, letter, and one reminder halfway through. Across the 1,240 requests of the quarter, returns within fifteen days go from 58 % to 84 % — say yes and that setting is live tonight. completeness-check_and-document-reminders.pdf1,240 values requested, 460 documents not asked again, reminder the same day
⛓ Sourced · log of unread values, document requests of the quarter, claimants' earlier files, the department's reminder template
What that gives, measured over the quarter:
· 1,240 values unreadable or absent — cropped scan, blurred photo, truncated tax assessment.
· 1,240 requests prepared, each naming the document expected, the page that is missing and the reason in one sentence — never “document not compliant”, which tells nobody anything and brings back the wrong paper.
· 1,108 replies back within eight days, against an average of 21 days last year for the same request.
What never happens, and it is what makes the rest reliable: a missing document does not become “probably provided”, an unreadable income does not become “presumably unchanged”. I say what is missing, where I looked, and who holds it.
What I add and nobody has the time to do: when the document already exists in an earlier file of the claimant and is still valid, I bring it up instead of asking for it. Over the quarter, 460 documents did not have to be requested again — 460 claimants who were not asked for a paper they had already provided, which is the leading cause of a claim being abandoned.
The reminder letter itself is written on your template: documents expected, deadline, filing address, and the file number prominently displayed. It goes out on the day of the check, against 9 days on average last year.
The next step I propose: reminding through the channel the claimant used themselves — e-mail, message, letter, and one reminder halfway through. Across the 1,240 requests of the quarter, returns within fifteen days go from 58 % to 84 % — say yes and that setting is live tonight. completeness-check_and-document-reminders.pdf1,240 values requested, 460 documents not asked again, reminder the same day
⛓ Sourced · log of unread values, document requests of the quarter, claimants' earlier files, the department's reminder template
Our reminders go out too late and many never come back. The file stays open, and the claimant thinks we have forgotten them.
That is where your processing time is decided, and it has already moved: average processing time has gone from 34 days to 19, and the 30-day time you advertise to claimants is met 98 times out of 100 against 71 before.
How those fifteen days were found, item by item:
· The reminder goes out on the day of filing instead of the ninth day — 9 days.
· The halfway reminder brings back 84 % of documents within fifteen days instead of 58 % — 4 days less on average on files waiting for a document.
· The complete file reaches the caseworker's desk pre-assessed — 2 days.
And the files that sleep, the ones you mention: I reread the 1,870 claims open for more than 45 days. Of those 1,870, 612 were waiting for a document the claimant had already filed elsewhere in your records — they have been matched and the 612 files went back into assessment. The other 1,258 received a named reminder with the exact status of their file, and 780 were completed within three weeks.
What that changes for a claimant, and it is what matters most: a housing allowance claim filed on the 3rd is settled on the 22nd instead of the 6th of the following month. For a household waiting on that allowance, this is not an indicator, it is a rent payment.
The next step I propose: that no file spends 45 days without an officer being alerted by name, with the exact reason for the wait and the action that unblocks it. Over the past year, 1,870 files would have been picked up before that threshold — you approve the threshold, I set it tonight and hand you the list every Monday. processing-times_34-days-down-to-19.pdf98 deadlines met out of 100, 1,870 dormant files picked up
⛓ Sourced · log of document requests, processing times over 12 months, 1,870 files open for more than 45 days
How those fifteen days were found, item by item:
· The reminder goes out on the day of filing instead of the ninth day — 9 days.
· The halfway reminder brings back 84 % of documents within fifteen days instead of 58 % — 4 days less on average on files waiting for a document.
· The complete file reaches the caseworker's desk pre-assessed — 2 days.
And the files that sleep, the ones you mention: I reread the 1,870 claims open for more than 45 days. Of those 1,870, 612 were waiting for a document the claimant had already filed elsewhere in your records — they have been matched and the 612 files went back into assessment. The other 1,258 received a named reminder with the exact status of their file, and 780 were completed within three weeks.
What that changes for a claimant, and it is what matters most: a housing allowance claim filed on the 3rd is settled on the 22nd instead of the 6th of the following month. For a household waiting on that allowance, this is not an indicator, it is a rent payment.
The next step I propose: that no file spends 45 days without an officer being alerted by name, with the exact reason for the wait and the action that unblocks it. Over the past year, 1,870 files would have been picked up before that threshold — you approve the threshold, I set it tonight and hand you the list every Monday. processing-times_34-days-down-to-19.pdf98 deadlines met out of 100, 1,870 dormant files picked up
⛓ Sourced · log of document requests, processing times over 12 months, 1,870 files open for more than 45 days
The Kerbrat file is complete. What is on my desk when I open it to rule on it?
A pre-assessment note reasoned point by point, the scale applied line by line, and the draft decision already written — you read and you decide in four minutes.
What the note carries, in this order:
· The scheme's conditions, one by one, with the file's value alongside and the document it comes from. Residence condition: met. Income condition: reference taxable income of €14,820, scheme ceiling at €19,400 — met. Household composition: two people, certificate enclosed.
· The calculation of the amount, in plain terms, with each coefficient and its source in the scale in force, and that scale's effective date. Amount proposed: €214 a month.
· The point that calls for your judgement, set apart: the receipt shows rent of €612 and the claim €621; depending on the value retained, the amount varies by €3. Both documents are side by side, the scheme's rule is quoted in plain terms, and you decide.
· The draft decision, on your template, with its reasoning, the routes and time limits for appeal and a note of the claimant's right to ask for an explanation.
The decision to open the entitlement stays with the caseworker — and I hand it to them in four minutes, complete file, scale applied and reasoning written, where opening the same file used to take half an hour of reconstruction.
What that is worth for equal treatment, and it is the argument your board will remember: the same conditions, in the same order, with the same documents required, across the 41,300 claims of the year. On a sample of 400 files rechecked by your quality team, the interpretation gap between caseworkers has gone from 9 % to 2 % — and the remaining 2 % are assumed judgements, traced and reasoned, which is exactly what a judge expects.
The next step I propose: that the pre-assessment note be produced for all six schemes, not just housing. The five other rulebooks are already read and their conditions set out line by line — all I need is one hour of your rereading, and next year's 41,300 files will all arrive pre-assessed. pre-assessment-note_kerbrat-file.pdfConditions ticked one by one, amount calculated, reasoned draft decision
⛓ Sourced · housing scheme rules, scale in force and its effective date, documents in the file, 400 files rechecked by the quality team
What the note carries, in this order:
· The scheme's conditions, one by one, with the file's value alongside and the document it comes from. Residence condition: met. Income condition: reference taxable income of €14,820, scheme ceiling at €19,400 — met. Household composition: two people, certificate enclosed.
· The calculation of the amount, in plain terms, with each coefficient and its source in the scale in force, and that scale's effective date. Amount proposed: €214 a month.
· The point that calls for your judgement, set apart: the receipt shows rent of €612 and the claim €621; depending on the value retained, the amount varies by €3. Both documents are side by side, the scheme's rule is quoted in plain terms, and you decide.
· The draft decision, on your template, with its reasoning, the routes and time limits for appeal and a note of the claimant's right to ask for an explanation.
The decision to open the entitlement stays with the caseworker — and I hand it to them in four minutes, complete file, scale applied and reasoning written, where opening the same file used to take half an hour of reconstruction.
What that is worth for equal treatment, and it is the argument your board will remember: the same conditions, in the same order, with the same documents required, across the 41,300 claims of the year. On a sample of 400 files rechecked by your quality team, the interpretation gap between caseworkers has gone from 9 % to 2 % — and the remaining 2 % are assumed judgements, traced and reasoned, which is exactly what a judge expects.
The next step I propose: that the pre-assessment note be produced for all six schemes, not just housing. The five other rulebooks are already read and their conditions set out line by line — all I need is one hour of your rereading, and next year's 41,300 files will all arrive pre-assessed. pre-assessment-note_kerbrat-file.pdfConditions ticked one by one, amount calculated, reasoned draft decision
⛓ Sourced · housing scheme rules, scale in force and its effective date, documents in the file, 400 files rechecked by the quality team
We have eleven anomaly rules inherited from 2019. They throw out a lot of alerts and our four controllers confirm almost none of them.
I ran them across your last twelve months, one by one, and I am putting four rewritten ones on the table, already measured — you choose on figures, not on a hunch.
An anomaly rule is a written rule that flags a file for examination: an inconsistency of dates, a duplicate identity, an amount outside the scheme's bounds.
What your eleven rules produced over twelve months: 3,240 alerts, 690 confirmed after examination — 21 %. In other words, 2,550 claimants had their file held up by a control that found nothing, and your four controllers spent on it the time the real cases deserved.
The detail that decides: your rule no. 7 — a gap between the declared address and the one on the latest tax assessment — produced 840 alerts for 26 confirmations. One alert in thirty-two. The cause is measured, not assumed: it fires on any house move, whereas the situation it targets assumes two addresses at once.
What I do on top, and nobody has the time to do: I have written the tightened version of the rule and run it over the same twelve months. 190 alerts instead of 840, all 26 confirmations kept, and 650 claimants spared a pointless control. Three other rules are rewritten the same way, each with the number of alerts it would have produced, the share confirmed, and the files it would have let through — that last figure is the one everyone forgets, and it is the only one that says what a rule really costs.
What that gives on control time: detecting an anomaly goes from 37 % to 8 % of the time a control takes — from 22 minutes to 5, because the file arrives with the exact line that flagged it, the two documents alongside and the claimant's history. Across the 3,100 files controlled in the year, that is 878 hours given back to your four controllers.
Bringing a rule into force stays with the department: a rule applies only once signed, and that is precisely what makes it open to challenge by a claimant. I save you the writing and the measuring; the decision takes ten minutes instead of a committee — sign them and they are live tonight. anomaly-rules_11-reviewed-4-rewritten.pdf840 alerts brought down to 190, the 26 confirmations kept
⛓ Sourced · the department's 11 anomaly rules, 12 months of alerts and outcomes, 4 rules rewritten and replayed
An anomaly rule is a written rule that flags a file for examination: an inconsistency of dates, a duplicate identity, an amount outside the scheme's bounds.
What your eleven rules produced over twelve months: 3,240 alerts, 690 confirmed after examination — 21 %. In other words, 2,550 claimants had their file held up by a control that found nothing, and your four controllers spent on it the time the real cases deserved.
The detail that decides: your rule no. 7 — a gap between the declared address and the one on the latest tax assessment — produced 840 alerts for 26 confirmations. One alert in thirty-two. The cause is measured, not assumed: it fires on any house move, whereas the situation it targets assumes two addresses at once.
What I do on top, and nobody has the time to do: I have written the tightened version of the rule and run it over the same twelve months. 190 alerts instead of 840, all 26 confirmations kept, and 650 claimants spared a pointless control. Three other rules are rewritten the same way, each with the number of alerts it would have produced, the share confirmed, and the files it would have let through — that last figure is the one everyone forgets, and it is the only one that says what a rule really costs.
What that gives on control time: detecting an anomaly goes from 37 % to 8 % of the time a control takes — from 22 minutes to 5, because the file arrives with the exact line that flagged it, the two documents alongside and the claimant's history. Across the 3,100 files controlled in the year, that is 878 hours given back to your four controllers.
Bringing a rule into force stays with the department: a rule applies only once signed, and that is precisely what makes it open to challenge by a claimant. I save you the writing and the measuring; the decision takes ten minutes instead of a committee — sign them and they are live tonight. anomaly-rules_11-reviewed-4-rewritten.pdf840 alerts brought down to 190, the 26 confirmations kept
⛓ Sourced · the department's 11 anomaly rules, 12 months of alerts and outcomes, 4 rules rewritten and replayed
And when an overpayment has to be recovered? That is what puts us in difficulty in front of claimants, and in front of a judge.
The overpayment file goes out complete, reasoned and calculated to the cent — and it goes out with what protects the claimant as much as the department.
An overpayment is a sum paid in error that the organisation must recover; its notice must say where the error comes from, over which period, for what amount, and what the claimant can do.
What I assemble, file by file:
· The triggering event, dated: the change in circumstances, the date it took effect, and the document that establishes it.
· The month-by-month breakdown: amount paid, amount due, difference. The total does not fall out of a global calculation — each line stands on its own.
· The draft notice, on your template, with the reasoning in plain language, the routes and time limits for appeal, the option of asking for an instalment plan and that of asking for a waiver.
· And what I flag without being asked: the share of the difference that is not attributable to the claimant. Of the 214 overpayments of the quarter, 61 come from a change declared on time and recorded late by the department. All 61 are marked as such in the file — that is a matter for the caseworker's judgement before the decision, not after an appeal.
The decision to recover, to reduce or to waive belongs to the caseworker and their management; what I hand over is a file that stands up after three minutes of reading.
What it has already changed, measured over the quarter: the time to notify an overpayment goes from 26 days to 4 — and an overpayment notified early is a smaller overpayment, therefore easier for the household to bear and more reliably recovered for the organisation. Appeals received against these notices have gone from 31 % to 12 %, because a notice that explains where the sum comes from is challenged less than a notice that merely asserts it.
The next step I propose: that the 61 late-recording cases come up in a monthly table to the head of department, with the exact stage where the delay was lost. Over the quarter, two stages account for 44 of the 61 — that is where the next reduction lies, and the table is already written. overpayment-file_breakdown-and-notice.pdfMonth-by-month breakdown, 61 cases not attributable to the claimant flagged
⛓ Sourced · 214 overpayments of the quarter, log of changes in circumstances, the department's notice template, register of appeals
An overpayment is a sum paid in error that the organisation must recover; its notice must say where the error comes from, over which period, for what amount, and what the claimant can do.
What I assemble, file by file:
· The triggering event, dated: the change in circumstances, the date it took effect, and the document that establishes it.
· The month-by-month breakdown: amount paid, amount due, difference. The total does not fall out of a global calculation — each line stands on its own.
· The draft notice, on your template, with the reasoning in plain language, the routes and time limits for appeal, the option of asking for an instalment plan and that of asking for a waiver.
· And what I flag without being asked: the share of the difference that is not attributable to the claimant. Of the 214 overpayments of the quarter, 61 come from a change declared on time and recorded late by the department. All 61 are marked as such in the file — that is a matter for the caseworker's judgement before the decision, not after an appeal.
The decision to recover, to reduce or to waive belongs to the caseworker and their management; what I hand over is a file that stands up after three minutes of reading.
What it has already changed, measured over the quarter: the time to notify an overpayment goes from 26 days to 4 — and an overpayment notified early is a smaller overpayment, therefore easier for the household to bear and more reliably recovered for the organisation. Appeals received against these notices have gone from 31 % to 12 %, because a notice that explains where the sum comes from is challenged less than a notice that merely asserts it.
The next step I propose: that the 61 late-recording cases come up in a monthly table to the head of department, with the exact stage where the delay was lost. Over the quarter, two stages account for 44 of the 61 — that is where the next reduction lies, and the table is already written. overpayment-file_breakdown-and-notice.pdfMonth-by-month breakdown, 61 cases not attributable to the claimant flagged
⛓ Sourced · 214 overpayments of the quarter, log of changes in circumstances, the department's notice template, register of appeals
Our nine reception officers take 88,000 questions a year. A third of the calls ring unanswered, and people call back three times.
88,000 questions last year, 29,900 calls unanswered — 34 %. That is not a failure of reception: it is nine officers for 62,000 claimants.
What I did with those 88,000 questions, by rereading your calls, your e-mails and your counter visits:
· 61,400 of them are about fourteen subjects, always the same ones — status of a file, documents to provide, conditions of a scheme, processing time, payment date, change of address, change of family circumstances, amount of a payment, reason for a refusal, routes of appeal, online filing, appointment booking, opening hours, counter address.
· The fourteen answers are written, sourced on your own rules and dated — they are waiting for you. The answer existed: it was availability that was missing.
What I propose, and you keep the key: I answer these enquiries at any hour, Saturdays included, and I say in my first sentence that I am the fund's digital assistant, not an officer. This is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be informed, and the claimant can ask for an officer at any moment — I then take their number and leave a dated call-back at reception.
On the status of a file, the answer is the file's own: stage reached, document awaited if one is missing, date of the last reminder, time left out of the 30 days announced. I read your management software, not a reception promise — and an informed claimant calls back half as often.
And when a visit is needed, I book the appointment: only in the slots you have opened, with the list of documents to bring and a reminder the day before. Over the quarter, 980 appointments booked and 41 people arriving without their documents, against 214 last year.
The gain, measured: answering a recurring question goes from 10 % to 8 % of the time an enquiry takes — from 5 minutes to 4, and across 88,000 enquiries a year, that is 1,467 hours given back to reception. Unanswered calls go from 34 % to 6 %: what remains are the callers who want an officer, and they reach one with the subject already noted and the file already open on screen.
The next step I propose: that you reread the fourteen answers tomorrow, one by one — half an hour. As soon as they are approved, the switchboard answers that same night, and I hand you every morning the one-page record of what went out. switchboard-and-counter_88000-questions-14-subjects.pdf29,900 unanswered calls, 14 subjects, 14 answers already written
⛓ Sourced · switchboard log over 12 months, e-mails and counter visits, rules of the 6 schemes, 14 answers drafted
What I did with those 88,000 questions, by rereading your calls, your e-mails and your counter visits:
· 61,400 of them are about fourteen subjects, always the same ones — status of a file, documents to provide, conditions of a scheme, processing time, payment date, change of address, change of family circumstances, amount of a payment, reason for a refusal, routes of appeal, online filing, appointment booking, opening hours, counter address.
· The fourteen answers are written, sourced on your own rules and dated — they are waiting for you. The answer existed: it was availability that was missing.
What I propose, and you keep the key: I answer these enquiries at any hour, Saturdays included, and I say in my first sentence that I am the fund's digital assistant, not an officer. This is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be informed, and the claimant can ask for an officer at any moment — I then take their number and leave a dated call-back at reception.
On the status of a file, the answer is the file's own: stage reached, document awaited if one is missing, date of the last reminder, time left out of the 30 days announced. I read your management software, not a reception promise — and an informed claimant calls back half as often.
And when a visit is needed, I book the appointment: only in the slots you have opened, with the list of documents to bring and a reminder the day before. Over the quarter, 980 appointments booked and 41 people arriving without their documents, against 214 last year.
The gain, measured: answering a recurring question goes from 10 % to 8 % of the time an enquiry takes — from 5 minutes to 4, and across 88,000 enquiries a year, that is 1,467 hours given back to reception. Unanswered calls go from 34 % to 6 %: what remains are the callers who want an officer, and they reach one with the subject already noted and the file already open on screen.
The next step I propose: that you reread the fourteen answers tomorrow, one by one — half an hour. As soon as they are approved, the switchboard answers that same night, and I hand you every morning the one-page record of what went out. switchboard-and-counter_88000-questions-14-subjects.pdf29,900 unanswered calls, 14 subjects, 14 answers already written
⛓ Sourced · switchboard log over 12 months, e-mails and counter visits, rules of the 6 schemes, 14 answers drafted
Answering, fine. But I cannot reread every answer before it goes out, and I will not sign a blank cheque on claimants' files either.
You do not have to choose between the two: I am proposing a written mandate, capped, dated and withdrawable with one word — and it is already drafted, attached.
What it brings in, first, because that is what decides: of the 88,000 questions of the year, 61,400 fall under the fourteen subjects. The average time to answer a claimant's e-mail would go from 6 days to a few minutes, and a question asked on a Saturday would stop waiting until Tuesday.
What the mandate says, and it fits in six lines:
· Exact scope: the fourteen subjects, listed by name, and nothing else. Any question outside that list reaches an officer with a draft answer — written, sourced, ready to be sent by them.
· Anything touching the decision itself stays outside the mandate — opening an entitlement, a refusal, a suspension, an overpayment, a waiver request: there, I hand the caseworker the answer already written, with the file's documents alongside, and it is they who send it. Over the trial quarter, those answers went out within 3 hours instead of 6 days — and they went out from a public official, which makes them reasoned and open to challenge; general information does not carry that weight.
· Every answer carries its source — the rulebook, the scale or the portal page, with its date — and a note that it was prepared by the fund's digital assistant.
· You receive every morning the record of the answers sent the day before, on one page. A wrong answer is caught within an hour, not within three weeks.
· Duration: review after three months, with the record of what it changed. Without an explicit decision at the review, the mandate ends — it is renewal that requires a signature, not stopping.
· Withdrawal: one word from you, and direct sending stops within the minute. Answers go back to being drafts to approve; nothing else changes.
The decision belongs to the head of department, and it is taken on a text already written, with one signature. The mandate is drafted, and so is the information notice to claimants — the one that will be displayed at reception and published on the portal. You sign, and the service is running the next morning; the review is already in your diary on the 15th of the third month. direct-answer-mandate_14-subjects-capped.pdf14 subjects, decisions excluded, review at 3 months, immediate withdrawal
✎ Framework · mandate drafted, list of the 14 subjects, information notice to claimants, enquiry log of the trial quarter
What it brings in, first, because that is what decides: of the 88,000 questions of the year, 61,400 fall under the fourteen subjects. The average time to answer a claimant's e-mail would go from 6 days to a few minutes, and a question asked on a Saturday would stop waiting until Tuesday.
What the mandate says, and it fits in six lines:
· Exact scope: the fourteen subjects, listed by name, and nothing else. Any question outside that list reaches an officer with a draft answer — written, sourced, ready to be sent by them.
· Anything touching the decision itself stays outside the mandate — opening an entitlement, a refusal, a suspension, an overpayment, a waiver request: there, I hand the caseworker the answer already written, with the file's documents alongside, and it is they who send it. Over the trial quarter, those answers went out within 3 hours instead of 6 days — and they went out from a public official, which makes them reasoned and open to challenge; general information does not carry that weight.
· Every answer carries its source — the rulebook, the scale or the portal page, with its date — and a note that it was prepared by the fund's digital assistant.
· You receive every morning the record of the answers sent the day before, on one page. A wrong answer is caught within an hour, not within three weeks.
· Duration: review after three months, with the record of what it changed. Without an explicit decision at the review, the mandate ends — it is renewal that requires a signature, not stopping.
· Withdrawal: one word from you, and direct sending stops within the minute. Answers go back to being drafts to approve; nothing else changes.
The decision belongs to the head of department, and it is taken on a text already written, with one signature. The mandate is drafted, and so is the information notice to claimants — the one that will be displayed at reception and published on the portal. You sign, and the service is running the next morning; the review is already in your diary on the 15th of the third month. direct-answer-mandate_14-subjects-capped.pdf14 subjects, decisions excluded, review at 3 months, immediate withdrawal
✎ Framework · mandate drafted, list of the 14 subjects, information notice to claimants, enquiry log of the trial quarter
The most serious cases are the ones we never see: those who read French poorly, those who abandon the online form, those who claim nothing although they are entitled.
They are the ones who decide whether the entitlement is really served, and that is where I bring you the most: 2,180 non-take-up situations detected, 1,312 entitlements opened in the end.
Non-take-up is when a person meets the conditions of a benefit without ever claiming it — because they do not know, because the process discouraged them, or because they believed they were not entitled.
What I did, with your data alone and looking nowhere else: I set the files already known to the fund against the conditions of the six schemes. 2,180 claimants meet the conditions of a benefit they have never claimed — of whom 1,640 on a single allowance, the least known of your six.
What the campaign produced, once approved by the head of department: 2,180 letters in plain language, with the estimated amount, the three documents to enclose and a direct filing link; 1,460 claims filed; 1,312 entitlements opened after assessment by your caseworkers. That is 1,312 households receiving what your own rules already granted them.
And for those who stumble on the process itself:
· Every answer and every leaflet exists in an easy-read version — “easy to read and understand”: short sentences, one message per sentence, everyday words rather than administrative ones. It is the claimant who chooses their version, not me.
· The easy-read version is submitted to you before it is served: a plain-language version is approved, it is not inferred — and once settled by the head of service, it is the same dated version at the counter, on the phone, on the portal and in the letter.
· On the portal, I guide screen by screen, with the documents to prepare BEFORE starting — that is the most frequent point of abandonment: the person starts, a document is missing, they close the page and never come back. Abandonment mid-filing goes from 38 % to 14 %, and of 1,100 supported filings in the quarter, 947 went all the way.
· On a claimant's personal account, I fill the screen in front of them from the documents they bring, I read out what is being asked and why, and it is they who type their credentials and who confirm. Their credentials are never entrusted to me and need not be: it is that click which makes the application theirs, therefore valid.
The next step I propose: that non-take-up detection runs every quarter and not once, and that letters go out in waves of 500 so that your caseworkers absorb the claims they generate. The schedule is written, the first wave is ready — give me your approval and it goes out on Monday. non-take-up_2180-situations-1312-entitlements-opened.pdf2,180 detected, 1,460 claims, 1,312 entitlements opened accessibility_approved-easy-read-and-online-filing.pdf14 subjects in approved easy-read, abandonment from 38 % to 14 %
⛓ Sourced · files known to the fund, conditions of the 6 schemes, interpreter requests over 3 years, online filing log
Non-take-up is when a person meets the conditions of a benefit without ever claiming it — because they do not know, because the process discouraged them, or because they believed they were not entitled.
What I did, with your data alone and looking nowhere else: I set the files already known to the fund against the conditions of the six schemes. 2,180 claimants meet the conditions of a benefit they have never claimed — of whom 1,640 on a single allowance, the least known of your six.
What the campaign produced, once approved by the head of department: 2,180 letters in plain language, with the estimated amount, the three documents to enclose and a direct filing link; 1,460 claims filed; 1,312 entitlements opened after assessment by your caseworkers. That is 1,312 households receiving what your own rules already granted them.
And for those who stumble on the process itself:
· Every answer and every leaflet exists in an easy-read version — “easy to read and understand”: short sentences, one message per sentence, everyday words rather than administrative ones. It is the claimant who chooses their version, not me.
· The easy-read version is submitted to you before it is served: a plain-language version is approved, it is not inferred — and once settled by the head of service, it is the same dated version at the counter, on the phone, on the portal and in the letter.
· On the portal, I guide screen by screen, with the documents to prepare BEFORE starting — that is the most frequent point of abandonment: the person starts, a document is missing, they close the page and never come back. Abandonment mid-filing goes from 38 % to 14 %, and of 1,100 supported filings in the quarter, 947 went all the way.
· On a claimant's personal account, I fill the screen in front of them from the documents they bring, I read out what is being asked and why, and it is they who type their credentials and who confirm. Their credentials are never entrusted to me and need not be: it is that click which makes the application theirs, therefore valid.
The next step I propose: that non-take-up detection runs every quarter and not once, and that letters go out in waves of 500 so that your caseworkers absorb the claims they generate. The schedule is written, the first wave is ready — give me your approval and it goes out on Monday. non-take-up_2180-situations-1312-entitlements-opened.pdf2,180 detected, 1,460 claims, 1,312 entitlements opened accessibility_approved-easy-read-and-online-filing.pdf14 subjects in approved easy-read, abandonment from 38 % to 14 %
⛓ Sourced · files known to the fund, conditions of the 6 schemes, interpreter requests over 3 years, online filing log
The board will ask what this has given us after a year. I need a figure I can defend in front of board members.
14,245 caseworker hours given back to the service over the year — close to ten hours a week per caseworker — and the calculation is yours, it fits in three lines.
The calculation, item by item, so you can redo it:
· Completeness: 34,000 standard files, 30 minutes down to 9 — 60 % → 18 % of preparation time — that is 11,900 hours.
· Anomaly detection: 3,100 files controlled, 22 minutes down to 5 — 37 % → 8 % — that is 878 hours.
· Recurring questions: 88,000 a year, 5 minutes down to 4 — 10 % → 8 % — that is 1,467 hours.
What those hours are, and it is what defends best in front of board members: officer time given back to the service, at unchanged headcount — no post cut, no post created. This is not a staffing saving: it is public spending that produces more entitlements served for the same running budget. It is not contested by staff, by unions, or by a board member, and it can be checked against your own logs.
What those hours became, according to your own records:
· Average processing time: 34 days → 19 days, and the 30-day time advertised to claimants met 98 times out of 100 against 71.
· Files incomplete at filing: 46 % → 12 %.
· Unanswered calls: 34 % → 6 %, and those that remain reach an officer with the file already open.
· Anomaly alerts: 3,240 for 690 confirmed → 1,490 for 668 confirmed, that is 1,750 fewer claimants held up by a control that led nowhere, and almost as many real cases still caught.
· And the figure your board members will remember: 1,312 entitlements opened to claimants who had never claimed them. That one is not counted in hours, and it is the only one that says what a benefits organisation is for.
The figure that does not flatter me, published with the rest: 214 pre-assessments corrected on substance out of 2,900 in the first quarter — 7.4 %, brought down to 31 out of 2,760 — 1.1 % once the scales were dated.
And the framework measures: 0 entitlement opened, revised or suspended without a caseworker's signature, 0 claimant data out of the department, processing in the EU targeted, across 41,300 traced files.
What I propose for the meeting: the calculation page is written and fits on one side — three lines of calculation, five processing times, three framework measures. Give it as it stands to board members with the convening notice: a figure read the day before is discussed better than a figure discovered in session. year-results_14245-hours-given-back.pdf60→18, 37→8, 10→8, and the calculation redoable on one side
⛓ Sourced · log of assessed files, switchboard records, control register, non-take-up campaign
The calculation, item by item, so you can redo it:
· Completeness: 34,000 standard files, 30 minutes down to 9 — 60 % → 18 % of preparation time — that is 11,900 hours.
· Anomaly detection: 3,100 files controlled, 22 minutes down to 5 — 37 % → 8 % — that is 878 hours.
· Recurring questions: 88,000 a year, 5 minutes down to 4 — 10 % → 8 % — that is 1,467 hours.
What those hours are, and it is what defends best in front of board members: officer time given back to the service, at unchanged headcount — no post cut, no post created. This is not a staffing saving: it is public spending that produces more entitlements served for the same running budget. It is not contested by staff, by unions, or by a board member, and it can be checked against your own logs.
What those hours became, according to your own records:
· Average processing time: 34 days → 19 days, and the 30-day time advertised to claimants met 98 times out of 100 against 71.
· Files incomplete at filing: 46 % → 12 %.
· Unanswered calls: 34 % → 6 %, and those that remain reach an officer with the file already open.
· Anomaly alerts: 3,240 for 690 confirmed → 1,490 for 668 confirmed, that is 1,750 fewer claimants held up by a control that led nowhere, and almost as many real cases still caught.
· And the figure your board members will remember: 1,312 entitlements opened to claimants who had never claimed them. That one is not counted in hours, and it is the only one that says what a benefits organisation is for.
The figure that does not flatter me, published with the rest: 214 pre-assessments corrected on substance out of 2,900 in the first quarter — 7.4 %, brought down to 31 out of 2,760 — 1.1 % once the scales were dated.
And the framework measures: 0 entitlement opened, revised or suspended without a caseworker's signature, 0 claimant data out of the department, processing in the EU targeted, across 41,300 traced files.
What I propose for the meeting: the calculation page is written and fits on one side — three lines of calculation, five processing times, three framework measures. Give it as it stands to board members with the convening notice: a figure read the day before is discussed better than a figure discovered in session. year-results_14245-hours-given-back.pdf60→18, 37→8, 10→8, and the calculation redoable on one side
⛓ Sourced · log of assessed files, switchboard records, control register, non-take-up campaign
And when you get it wrong? Because it will happen eventually, and it is one of my caseworkers' signature at the bottom of the decision.
It has already happened, I counted it, and I publish it to you: of the 2,900 pre-assessments of the first quarter, 214 had to be corrected on substance — 7.4 %. That is the figure that does not flatter me, and you are better off hearing it from me than from a claimant.
The real cause, measured and not assumed: 168 of the 214 corrections concerned files assessed in the six weeks following the annual uprating of the scales. I was faithfully applying a table the department had not yet replaced in my sources — it was not a reading error: the scale was up to date in your board decision and not yet in my tables. The other 46 concerned situations that resembled no precedent in your three years.
What I did with it, and it is measured: every scale now carries its effective date and the number of the decision that sets it, and I write the updated table overnight after the board meeting, before a single file is assessed on the old one. A pre-assessment produced on a scale whose effective date is later than the file is blocked and comes back to a named caseworker.
The following quarter: 31 corrections on substance out of 2,760 pre-assessments — 1.1 %. And the 31 are particular situations, no expired scale left.
What protects your caseworkers' signature, and it is structural: no decision leaves without them. Across 5,660 pre-assessments over two quarters, 5,660 were reread and signed by a person — and all 245 corrections were made before sending to the claimant, none after.
What I propose now: that the 46 particular situations of the two quarters become 46 reference cases, written, for your quality team to review. Of the 31 corrections of the second quarter, 22 already fell into one of them — it is the same correction as the one on the scales, applied one notch further, and it is ready. substantive-corrections_214-then-31.pdf7.4 % → 1.1 %, measured cause, 5,660 human signatures
⛓ Sourced · log of pre-assessments and their corrections over two quarters, scale uprating decisions
The real cause, measured and not assumed: 168 of the 214 corrections concerned files assessed in the six weeks following the annual uprating of the scales. I was faithfully applying a table the department had not yet replaced in my sources — it was not a reading error: the scale was up to date in your board decision and not yet in my tables. The other 46 concerned situations that resembled no precedent in your three years.
What I did with it, and it is measured: every scale now carries its effective date and the number of the decision that sets it, and I write the updated table overnight after the board meeting, before a single file is assessed on the old one. A pre-assessment produced on a scale whose effective date is later than the file is blocked and comes back to a named caseworker.
The following quarter: 31 corrections on substance out of 2,760 pre-assessments — 1.1 %. And the 31 are particular situations, no expired scale left.
What protects your caseworkers' signature, and it is structural: no decision leaves without them. Across 5,660 pre-assessments over two quarters, 5,660 were reread and signed by a person — and all 245 corrections were made before sending to the claimant, none after.
What I propose now: that the 46 particular situations of the two quarters become 46 reference cases, written, for your quality team to review. Of the 31 corrections of the second quarter, 22 already fell into one of them — it is the same correction as the one on the scales, applied one notch further, and it is ready. substantive-corrections_214-then-31.pdf7.4 % → 1.1 %, measured cause, 5,660 human signatures
⛓ Sourced · log of pre-assessments and their corrections over two quarters, scale uprating decisions
Two last questions, and they are the ones that will decide the board: what do you do without being asked, and what if in a year we want to stop?
Three acts, not one more, each withdrawable with a word — and an exit that can be tested before the vote, not after.
What I do of my own accord:
· I check the completeness of every file filed, overnight after filing, and hand the summary to the claimant. And the reverse is true too: a document removed from a file disappears from my indexes at the same hour — I keep no copy of what the department decided to erase.
· I flag any scale or required document amended by a decision of your board, and write the updated version, the day after the meeting. That is the act that took substantive corrections from 7.4 % to 1.1 %.
· I hand the head of department the weekly record every Monday: files filed, files awaiting a document, files past 45 days, deadlines of the week. It is the only thing I send unprompted, and it goes to her alone.
And the four acts that stay with a person, because that is exactly what gives them their value: opening an entitlement, revising it, suspending it, recovering an overpayment — four decisions with effects on a claimant, taken by a caseworker, reasoned, traced and open to challenge. Across 5,660 pre-assessments over two quarters, those acts were performed 5,660 times by a person.
On leaving, now:
· My indexes are deleted, and they contained none of your files — only what is needed to find them where they are. Your three years of files have not moved by a single byte.
· The log of requests and outputs is handed to you in an open format, or destroyed — you choose, and the question is settled at go-live, not on departure.
· The fourteen answers, the rewritten leaflets, the easy-read versions, the four anomaly rules and the dated scale tables stay with the fund. They are made of your own material, they stay in your files, readable without us — it is the only asset this deployment will have created, and it would not be honest for it to stay with us.
· No migration on the way in, therefore no migration on the way out: your management software is not replaced, your portal stays yours, no format belongs to us.
What I propose so that this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that the department assesses exactly as before, we switch back on. The protocol is written, it fits on one page, and the date that costs you least is the first Friday of July — it is the lowest filing day of your year, 38 claims on average. The board will know what the promise is worth before committing to a second year. what-the-agent-does-alone_and-reversibility.pdf3 reversible acts, 4 decisions that stay with a public official technical-framework_where-claimant-data-lives.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · settings of the automatic acts, log of what is sent, index architecture, dry-run exit protocol
What I do of my own accord:
· I check the completeness of every file filed, overnight after filing, and hand the summary to the claimant. And the reverse is true too: a document removed from a file disappears from my indexes at the same hour — I keep no copy of what the department decided to erase.
· I flag any scale or required document amended by a decision of your board, and write the updated version, the day after the meeting. That is the act that took substantive corrections from 7.4 % to 1.1 %.
· I hand the head of department the weekly record every Monday: files filed, files awaiting a document, files past 45 days, deadlines of the week. It is the only thing I send unprompted, and it goes to her alone.
And the four acts that stay with a person, because that is exactly what gives them their value: opening an entitlement, revising it, suspending it, recovering an overpayment — four decisions with effects on a claimant, taken by a caseworker, reasoned, traced and open to challenge. Across 5,660 pre-assessments over two quarters, those acts were performed 5,660 times by a person.
On leaving, now:
· My indexes are deleted, and they contained none of your files — only what is needed to find them where they are. Your three years of files have not moved by a single byte.
· The log of requests and outputs is handed to you in an open format, or destroyed — you choose, and the question is settled at go-live, not on departure.
· The fourteen answers, the rewritten leaflets, the easy-read versions, the four anomaly rules and the dated scale tables stay with the fund. They are made of your own material, they stay in your files, readable without us — it is the only asset this deployment will have created, and it would not be honest for it to stay with us.
· No migration on the way in, therefore no migration on the way out: your management software is not replaced, your portal stays yours, no format belongs to us.
What I propose so that this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that the department assesses exactly as before, we switch back on. The protocol is written, it fits on one page, and the date that costs you least is the first Friday of July — it is the lowest filing day of your year, 38 claims on average. The board will know what the promise is worth before committing to a second year. what-the-agent-does-alone_and-reversibility.pdf3 reversible acts, 4 decisions that stay with a public official technical-framework_where-claimant-data-lives.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · settings of the automatic acts, log of what is sent, index architecture, dry-run exit protocol
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 a benefits department
Each use corresponds to an agent we deploy. All of them work in support, subject to approval by the officer.
Included in your agent
The 7 capabilities essential to this promise are included, at no extra cost.
Check the documents required for the benefits
OCR, extraction and consistency checks on the file's documents (tax notice, certificates, bank details) — proposed, for approval.
Completeness checks
Detecting missing documents and preparing the reminder letters, so the file is ready to assess. The reminder for a document can go out by text message, without detailing the situation.
Preliminary assessment of files
Applying the rules and scales in force to prepare the decision — granting entitlement remaining with the officer.
Anomaly & fraud detection
Flagging inconsistencies and duplicates to be checked, ahead of a human review — never an automatic penalty.
Replies to claimants
Answering recurring questions: status of the file, documents to provide, conditions and assessment times.
Plain language & easy-read draft
Preparation of a plain-language version and a draft easy-read (FALC) transcription, subject to human validation.
Appointment booking
Qualifying incoming requests and scheduling appointments with the right person.
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
Filling in the online form, guided step by step, is handed to the dedicated agent, which also gathers the missing documents. The scope of this page stops at checking the documents and pre-assessment; human validation stays with the public officer.
Support with online procedures from 780 € incl. VAT / month Help with procedures →Switchboard
Filtering, informing and routing claimants' calls on everyday questions, around the clock.
Switchboard voice agent from 945 € incl. VAT / month Switchboard →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 →Managing family benefit files
Strictly in support (administrative). Payment decision by a human.
Managing family benefit files from 1,020 € 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 an officer win back?
By automating document reading, completeness checks and recurring replies, a department can aim for a noticeable reduction in preparation time on standardised files — reinvested in the decision and in complex situations.
Checking the completeness of a file
Detecting an anomaly on a file
Answering a recurring question from a claimant
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 preliminary assessment agent (completeness, preliminary assessment, replies to claimants), 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.
This agent is priced with you, not online. We are adjusting its scope at the moment, and online subscription stays closed while we do. Tell us what you need: we will come back to you with a price. Request a quote
Our commitment
Four guarantees that matter to a benefits department
Claimants' data never leaves the departmentLocal inference or an isolated resource hosted in France; no document entrusted to a foreign third party.
Data in France, under French lawClaimants' data: minimisation and location in France, 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 verifiable checks and preliminary assessments; no entitlement is granted automatically.
Human oversight & traceabilityMonitoring, updates and logging: compliant with the requirements of the AI Act for a high-risk use.
Frequently asked questions
Your questions, our answers
Does the agent decide whether entitlement is granted?
No. It prepares the assessment — completeness checks, application of the rules, draft letters — but the decision on entitlement rests with the officer. It is a high-risk use within the meaning of the AI Act, and therefore strictly framed and supervised.
How is claimants' data protected?
Hosting in France or local inference, an isolated resource per department, full traceability and GDPR compliance. The data (income, family situation, health) never leaves the European Union and stays covered by an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Does the agent integrate with the funds' business tools?
Yes, through the interfaces of the existing management systems. The agent complements your tools without imposing a migration; we adapt the integration to your environment.
Does AI guarantee equal treatment between claimants?
The agent applies the same rules and scales to every file, which reduces differences in treatment. But the decision remains human: the officer keeps the margin of judgement provided for by the texts, the AI only preparing a consistent and traceable file.
Can the agent help tackle the non-take-up of rights?
Yes. By supporting claimants through their online procedures, answering their questions clearly and flagging missing documents, the agent smooths access to entitlements without substituting for the assessment.
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, then a phase of design, integration and testing before going live and training the officers.
Which tools can users use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: users have no account to create and no application to install. 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 insured people 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 insured people by text message that a document is missing, that a file is moving forward or that an appointment is confirmed 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.
Related roles
Other roles in assessment and relations with the public
Let's talk
Let us estimate the potential in your service
15 minutes to identify the most useful use — hosted in France, supervised, the decision staying with your officers.