● Public sector — Collections caseworker
AI agent for the collections caseworker
An assistant that prepares the reminders, answers contributors and sorts the requests for time to pay — the officer keeps control of every collection decision. Hosted in France, on a resource dedicated to and isolated for each organisation, contributors' data stays under French law. The AI agent assists, the officer decides.
Hosted in France
Contributors' data protected
GDPR & AI Act: governed deployment
Human oversight
Updated on
Deployed in a few weeks
Collections assistant · hosted in France
Prepare an informal reminder for this account: two instalments unpaid since March.
Draft reminder ready. It sets out the account position: the March and April instalments, a reminder of the late-payment surcharges that apply, the department's contact details and the option of requesting time to pay online. An informal tone, in line with your first-level templates.
For review before sending.
⛓ Source · the account position + your reminder templates
For review before sending.
⛓ Source · the account position + your reminder templates
The contributor is asking to spread it over six months. Pre-qualify the request.
Request pre-qualified. Declarations up to date, first payment incident, amount consistent with a standard arrangement: the file falls within the usual criteria. A draft six-month payment plan has been prepared, with the covering letter.
No commitment is made until you have approved it.
✎ Action · draft payment plan — the officer decides and signs
No commitment is made until you have approved it.
✎ Action · draft payment plan — the officer decides and signs
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 officer decidesThe AI agent assists, it never pursues alone
✦ In brief
In a collections body, a Blue Lemon Agent agent assists the caseworker with the repetitive tasks — informal reminders, replies to contributors, pre-qualification of requests for time to pay, draft payment plans — and reconciles declarations against payments to flag discrepancies. It runs on local inference or is hosted in France on a resource dedicated to and isolated for each organisation: businesses' accounting data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Any move to enforced collection remains decided and signed by the officer; the time won back goes to supporting contributors in difficulty. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and contributors 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 collections bodies — and why they hesitate
Thousands of reminders, requests for time to pay, recurring questions on due dates and rates: the collections caseworker handles a considerable volume of enquiries, on accounting and financial data that is sensitive for every business.
! The issue
The collections caseworker is caught between contributors who expect quick answers — due dates, rates, time to pay — and account positions to be handled in bulk, where every error can weaken a business. Yet most consumer AI tools would amount to entrusting contributors' account positions, declarations and financial difficulties to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For financial 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, the collection decision reserved to the officer: the time saved on reminders is never paid for in lost confidentiality. The aim is not to replace the caseworker, but to give them back time for supporting contributors in difficulty.
The decisive point
Protecting contributors' data: sovereignty & compliance
A collections body handles the accounting and financial data of thousands of businesses. 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 — contributors' data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Contributors' data: architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
One isolated resource per organisation
No pooling of data: an environment strictly dedicated to your organisation, 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 enforced collection is automated; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For the most sensitive data, SecNumCloud and HDS options are available depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
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-Boréal social contributions collection body — county branch
- Sector
- Collection of social contributions — accounts of self-employed workers and employers, amicable collection, support for contributors in difficulty
- Headcount
- 88 staff — including 26 amicable collection caseworkers, 12 on the telephone platform, 9 in enforced collection and 4 supporting contributors in difficulty
- Public served
- 34,200 contributor accounts — 21,400 self-employed workers and 12,800 employers, from the two-person garage to the three-hundred-strong cooperative
- Order of magnitude
- 24,000 amicable reminders a year, 4,800 requests for time to pay, 2,300 requests to waive late-payment surcharges, and 128,000 enquiries: 96,000 calls, 18,000 e-mails, 14,000 messages filed in the online account
- Tools in place
- National contributor account management software, telephone platform, e-mail, letter templates and 9 years of filed correspondence — the agent plugs into them read-only, nothing is replaced and nothing is migrated
- Who decides
- The head of amicable collection signs the instalment plans; the amicable appeals committee rules on waivers; the caseworker approves every letter before it goes out; the director alone starts enforced collection
- Points for improvement
- 11 % of reminders went out on an account whose payment had already arrived; 41 % of requests for time to pay waited more than 21 days for a first reply; 38 % of calls rang unanswered at due-date peaks; and 620 contributors met the conditions for a waiver without ever having asked for one
At Val-Boréal, amicable collection is a struggling company's first contact with national solidarity: done badly it damages an already strained cash position; done well it prevents a dispute. The agent runs on local inference on a machine belonging to the body, reads the nine years of accounts and correspondence without ever writing to them, and releases nothing without a caseworker's approval. The exchanges below cover a year, from matching the accounts to the review presented to management.
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.
Collections assistant · hosted in France
I reread your nine years of contributor accounts over the weekend — 34,200 accounts, 216,000 amicable reminders, 9 years of correspondence — and I have drawn out a figure your reports did not produce.
The gap measured, and it decides everything else: 2,640 of your 24,000 reminders last year — 11 % — went out on an account whose payment had already arrived. These are not caseworker mistakes: they are reminders calculated on a snapshot of the account taken too early.
The cause, measured and not assumed:
· 1,940 reminders went out because the payment had been received but not yet allocated — allocation is the attaching of a payment to the instalment it settles; until it is done, the account looks in arrears while the money is there. The average delay between receipt and allocation is 6 working days in your own records; your reminder run goes out on the 4th.
· 700 reminders went out on accounts that had filed a corrective return — a return that amends an earlier one — after the balance was calculated and before the letters were sent.
What I have already done, and this is not a promise: I have written the matching rule that compares, instalment by instalment, the amount due to the amount actually allocated, and I have run it over the last twelve months. It would have held back 2,400 of the 2,640 pointless reminders — 2,400 companies that would not have received a demand for a sum they had already paid, and 600 caseworker hours that would not have gone into the disputing calls each of those letters brings.
The next step I propose, and it takes twenty minutes: you read the rule line by line — it is written in plain language, not in code. The service signs it, and next month's run goes out matched: it is that signature that makes it enforceable against a contributor who would challenge it, and it is the only action I leave to you. returns-payments-matching_2640-reminders.pdf11 % of reminders, the measured cause, 2,400 avoided over 12 months
⛓ Sourced · 9 years of contributor accounts, allocation log, 216,000 reminders sent, rule tested over 12 months
The gap measured, and it decides everything else: 2,640 of your 24,000 reminders last year — 11 % — went out on an account whose payment had already arrived. These are not caseworker mistakes: they are reminders calculated on a snapshot of the account taken too early.
The cause, measured and not assumed:
· 1,940 reminders went out because the payment had been received but not yet allocated — allocation is the attaching of a payment to the instalment it settles; until it is done, the account looks in arrears while the money is there. The average delay between receipt and allocation is 6 working days in your own records; your reminder run goes out on the 4th.
· 700 reminders went out on accounts that had filed a corrective return — a return that amends an earlier one — after the balance was calculated and before the letters were sent.
What I have already done, and this is not a promise: I have written the matching rule that compares, instalment by instalment, the amount due to the amount actually allocated, and I have run it over the last twelve months. It would have held back 2,400 of the 2,640 pointless reminders — 2,400 companies that would not have received a demand for a sum they had already paid, and 600 caseworker hours that would not have gone into the disputing calls each of those letters brings.
The next step I propose, and it takes twenty minutes: you read the rule line by line — it is written in plain language, not in code. The service signs it, and next month's run goes out matched: it is that signature that makes it enforceable against a contributor who would challenge it, and it is the only action I leave to you. returns-payments-matching_2640-reminders.pdf11 % of reminders, the measured cause, 2,400 avoided over 12 months
⛓ Sourced · 9 years of contributor accounts, allocation log, 216,000 reminders sent, rule tested over 12 months
Holding back a reminder means delaying a recovery. How many genuine debtors does your rule let through?
41 over twelve months, and I counted them before you asked — then I wrote the version that catches them.
The figure that does not flatter me: of the 2,400 reminders my first rule held back, 41 concerned accounts genuinely in arrears. Their recovery would have been delayed by 14 days on average — that is the gap until the following run.
The cause, and it is a single one: the 41 accounts had received a partial payment. My first rule looked at whether a payment had arrived; it did not look at whether it covered the whole instalment. A payment of €400 against an instalment of €1,900 was enough to hold the reminder back.
What I did about it, and it is measured: the rule now compares the amount allocated to the amount due, instalment by instalment, and holds a reminder back only where the gap is nil. Rerun over the same twelve months: 6 accounts instead of 41, and the 6 surface in the Monday weekly check, so they are caught within 8 days instead of 14.
What the tightened rule gives overall: 2,400 pointless reminders avoided, 6 recoveries delayed by 8 days. The ratio is 400 to 1, and it is verifiable on your own ledgers, not on my word.
The next step I propose: that I hand you every Monday the list of what the rule held back over the past week, with the reason for each hold and the accounts returned to collection. One page, four columns — it is what allows your head of service to defend the rule at a management meeting with current figures rather than a recollection. returns-payments-matching_2640-reminders.pdf41 accounts down to 6, Monday weekly check
⛓ Sourced · rule rerun over 12 months, allocation entries, accounts returned to collection
The figure that does not flatter me: of the 2,400 reminders my first rule held back, 41 concerned accounts genuinely in arrears. Their recovery would have been delayed by 14 days on average — that is the gap until the following run.
The cause, and it is a single one: the 41 accounts had received a partial payment. My first rule looked at whether a payment had arrived; it did not look at whether it covered the whole instalment. A payment of €400 against an instalment of €1,900 was enough to hold the reminder back.
What I did about it, and it is measured: the rule now compares the amount allocated to the amount due, instalment by instalment, and holds a reminder back only where the gap is nil. Rerun over the same twelve months: 6 accounts instead of 41, and the 6 surface in the Monday weekly check, so they are caught within 8 days instead of 14.
What the tightened rule gives overall: 2,400 pointless reminders avoided, 6 recoveries delayed by 8 days. The ratio is 400 to 1, and it is verifiable on your own ledgers, not on my word.
The next step I propose: that I hand you every Monday the list of what the rule held back over the past week, with the reason for each hold and the accounts returned to collection. One page, four columns — it is what allows your head of service to defend the rule at a management meeting with current figures rather than a recollection. returns-payments-matching_2640-reminders.pdf41 accounts down to 6, Monday weekly check
⛓ Sourced · rule rerun over 12 months, allocation entries, accounts returned to collection
All of this assumes you read the accounts, the returns and the financial difficulties of 34,200 companies. Where do they go?
They go nowhere. I run on local inference on a machine belonging to the body, and nothing I read leaves your network.
Local inference means the model computes on your machine: an account balance, a return or a contributor's letter crosses no external network to be processed. If the body would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to your branch — no pooling with another body, which is the condition of your service's continuity.
What that changes, point by point:
· Contributor data trains no model, neither ours nor a third party's. What I learn from your accounts serves your accounts.
· I work read-only on your management software, and the technical account I read through has no write permission — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, role-based access — rights follow the job: the telephone platform sees the state of an account, not the material behind a waiver request; enforced collection sees its own files, not those of amicable collection. Over the year, the log shows 0 out-of-role access.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A full log: who asked what, when, and what I produced. It is that log which lets you answer a contributor who asks how their file was handled.
The figure that sums this up: 0 contributor record out of the body's network across the 128,000 enquiries of the year, and processing in the EU targeted.
The next step I propose: that I keep up to date the record your data protection officer and your supervising authority will ask for — hosting, data processed, retention periods, who has access to what. It is asked for once a year and takes three days to reconstruct; the first version is written and you have it in front of you. technical-framework_where-contributor-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account permissions, access log, first version of the processing record
Local inference means the model computes on your machine: an account balance, a return or a contributor's letter crosses no external network to be processed. If the body would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to your branch — no pooling with another body, which is the condition of your service's continuity.
What that changes, point by point:
· Contributor data trains no model, neither ours nor a third party's. What I learn from your accounts serves your accounts.
· I work read-only on your management software, and the technical account I read through has no write permission — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, role-based access — rights follow the job: the telephone platform sees the state of an account, not the material behind a waiver request; enforced collection sees its own files, not those of amicable collection. Over the year, the log shows 0 out-of-role access.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· A full log: who asked what, when, and what I produced. It is that log which lets you answer a contributor who asks how their file was handled.
The figure that sums this up: 0 contributor record out of the body's network across the 128,000 enquiries of the year, and processing in the EU targeted.
The next step I propose: that I keep up to date the record your data protection officer and your supervising authority will ask for — hosting, data processed, retention periods, who has access to what. It is asked for once a year and takes three days to reconstruct; the first version is written and you have it in front of you. technical-framework_where-contributor-data-lives.pdfLocal inference, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account permissions, access log, first version of the processing record
The 5th of the month has passed. The amicable reminder run has to go out, and I do not have three days to spend on it.
The run is built, matched and sorted by level: 1,840 accounts, and all you have to do is read.
An amicable reminder is the first letter sent to a contributor whose instalment is unsettled; it precedes any procedure and has no binding effect — it is a reminder, and it is often the only contact a struggling director will have with you.
What the matching rule has already done: 2,106 accounts looked in arrears on the 4th day; 266 had paid, their payment simply not yet allocated. They will receive nothing. That leaves 1,840 accounts.
What I hand you, sorted by what the situation calls for:
· 1,510 first reminders — one instalment, a first incident. Amicable tone, amount due, instalment concerned, the late-payment surcharges that apply and how they are calculated, in plain language, the online link to request time to pay and the service's direct number.
· 246 second reminders — two instalments, or an incident already chased. The letter recalls the first reminder and its date, which nobody has time to check by hand.
· 84 situations I flag without being asked: accounts whose amount due has doubled over two instalments after three years fully up to date. These are not bad payers, these are companies slipping — and for each one I have drafted the letter referring them to your support unit, with the three free slots in its diary.
The time this shifts: preparing an amicable reminder goes from 60 % to 10 % of the time the task takes — from 9 minutes to 1 minute 30. Across the 21,600 reminders that will go out this year once the pointless ones are removed, that is 2,700 caseworker hours going back to the contributors who need a phone call.
The next step I propose: that the 84 slipping situations surface on their own each month, before the run rather than after. A director called at the first incident settles in one go; the same director, six months later, asks to spread it over twelve. You approve the principle once, and the list waits for you on the 6th of each month. amicable-reminder-run_1840-accounts.pdf1,510 first reminders, 246 second, 84 slipping accounts flagged
⛓ Sourced · balances for the 5th, allocation log, your reminder templates, three years of payment history
An amicable reminder is the first letter sent to a contributor whose instalment is unsettled; it precedes any procedure and has no binding effect — it is a reminder, and it is often the only contact a struggling director will have with you.
What the matching rule has already done: 2,106 accounts looked in arrears on the 4th day; 266 had paid, their payment simply not yet allocated. They will receive nothing. That leaves 1,840 accounts.
What I hand you, sorted by what the situation calls for:
· 1,510 first reminders — one instalment, a first incident. Amicable tone, amount due, instalment concerned, the late-payment surcharges that apply and how they are calculated, in plain language, the online link to request time to pay and the service's direct number.
· 246 second reminders — two instalments, or an incident already chased. The letter recalls the first reminder and its date, which nobody has time to check by hand.
· 84 situations I flag without being asked: accounts whose amount due has doubled over two instalments after three years fully up to date. These are not bad payers, these are companies slipping — and for each one I have drafted the letter referring them to your support unit, with the three free slots in its diary.
The time this shifts: preparing an amicable reminder goes from 60 % to 10 % of the time the task takes — from 9 minutes to 1 minute 30. Across the 21,600 reminders that will go out this year once the pointless ones are removed, that is 2,700 caseworker hours going back to the contributors who need a phone call.
The next step I propose: that the 84 slipping situations surface on their own each month, before the run rather than after. A director called at the first incident settles in one go; the same director, six months later, asks to spread it over twelve. You approve the principle once, and the list waits for you on the 6th of each month. amicable-reminder-run_1840-accounts.pdf1,510 first reminders, 246 second, 84 slipping accounts flagged
⛓ Sourced · balances for the 5th, allocation log, your reminder templates, three years of payment history
After every run, contributors send documents back in bulk: certificates, statements, letters scanned crooked. That is what clogs the service.
3,100 documents received last month, 2,852 matched on their own to the right account and the right instalment — 92 %. The remaining 248 reach you with their reason, not in a pile.
What I do with each document: I read it, pick out the account number, the period, the amount and the nature of the document, and file it in the contributor's record with its receipt timestamp — it is that timestamp which will count if the contributor disputes a deadline. A crooked scan, a photo taken on a phone or a submission with no reference change nothing: I read the content, not the layout.
The 248 that do not match on their own, and what I have already drawn from them:
· 191 carried an incomplete reference or a namesake. I found 176 of them by cross-checking name, address and amount — they reach you with the proposed attachment and the two candidate accounts side by side, so the choice takes five seconds.
· 57 are documents of a different nature from the one expected — a bank statement where a certificate was needed. The letter asking for the right document is written, it names the document precisely and it attaches the example. That is what halves second submissions: a contributor does not send the wrong document out of ill will, he sends it because he was asked for “supporting evidence”.
What that gives back to the service, in figures: filing and matching documents took 3 staff days a week across two people; half a day remains. And for the contributor: the document is on file the day he sends it, not the following week — that is what prevents the one reminder too many, the one that arrives after he has replied.
The next step I propose: an automatic acknowledgement of every document received, within the hour, saying what it is, what it settles and what is still missing. Over a one-month trial, “did you receive my documents?” chasers fell from 340 to 41. Approve it and it starts on Monday. documents-received_3100-read-2852-matched.pdf92 % matched on their own, 176 found out of 191
⛓ Sourced · documents received this month, contributor records, matching and second-submission logs
What I do with each document: I read it, pick out the account number, the period, the amount and the nature of the document, and file it in the contributor's record with its receipt timestamp — it is that timestamp which will count if the contributor disputes a deadline. A crooked scan, a photo taken on a phone or a submission with no reference change nothing: I read the content, not the layout.
The 248 that do not match on their own, and what I have already drawn from them:
· 191 carried an incomplete reference or a namesake. I found 176 of them by cross-checking name, address and amount — they reach you with the proposed attachment and the two candidate accounts side by side, so the choice takes five seconds.
· 57 are documents of a different nature from the one expected — a bank statement where a certificate was needed. The letter asking for the right document is written, it names the document precisely and it attaches the example. That is what halves second submissions: a contributor does not send the wrong document out of ill will, he sends it because he was asked for “supporting evidence”.
What that gives back to the service, in figures: filing and matching documents took 3 staff days a week across two people; half a day remains. And for the contributor: the document is on file the day he sends it, not the following week — that is what prevents the one reminder too many, the one that arrives after he has replied.
The next step I propose: an automatic acknowledgement of every document received, within the hour, saying what it is, what it settles and what is still missing. Over a one-month trial, “did you receive my documents?” chasers fell from 340 to 41. Approve it and it starts on Monday. documents-received_3100-read-2852-matched.pdf92 % matched on their own, 176 found out of 191
⛓ Sourced · documents received this month, contributor records, matching and second-submission logs
And the contributors who reply to the letter? Some dispute, some describe their situation, some ask for time without saying so.
412 replies last month, sorted the day they arrived, and each one goes back out with what it actually calls for.
· 168 disputes over the amount. I redid the calculation from the returns and the allocated payments, and I hand you the two columns side by side: what the contributor counts, what the account carries, and the exact line where the two diverge. Of the 168, 39 divergences came from us — a payment wrongly allocated, a corrective return not taken into account —, and the 39 corrections are prepared, ready to be entered by an authorised caseworker.
· 121 requests for time to pay that do not say their name — “I cannot pay this month”. They are reclassified as requests for time to pay and pre-assessed, instalment plan included: a contributor does not have to know the name of the procedure to be entitled to it. It is the biggest pocket of unclaimed entitlement in your service, and it costs nothing to recover.
· 84 accounts of a situation — cash difficulty, the director on sick leave, the loss of a main customer. Those go to your support unit with the account position, three years of payment history and a ten-line summary of what the contributor wrote.
· 39 replies promising payment, with the date announced. A follow-up is set for that date: if the payment does not arrive, you know the next day.
The delay, which is the whole point: a contributor's reply waited 11 days for first handling; it is sorted, reclassified and handed to the caseworker the day it arrives. Across 412 replies, 0 was left unattended for more than 48 hours.
The next step I propose: that the 121 unnamed requests trigger the pre-assessment on their own, as they arrive. The contributor then receives a proposed instalment plan in 3 days instead of 21, and you get back the only thing that matters: the decision, taken on a complete file. contributor-replies_412-sorted.pdf39 corrections prepared, 121 requests recovered
⛓ Sourced · letters and e-mails received this month, returns, allocated payments, payment history
· 168 disputes over the amount. I redid the calculation from the returns and the allocated payments, and I hand you the two columns side by side: what the contributor counts, what the account carries, and the exact line where the two diverge. Of the 168, 39 divergences came from us — a payment wrongly allocated, a corrective return not taken into account —, and the 39 corrections are prepared, ready to be entered by an authorised caseworker.
· 121 requests for time to pay that do not say their name — “I cannot pay this month”. They are reclassified as requests for time to pay and pre-assessed, instalment plan included: a contributor does not have to know the name of the procedure to be entitled to it. It is the biggest pocket of unclaimed entitlement in your service, and it costs nothing to recover.
· 84 accounts of a situation — cash difficulty, the director on sick leave, the loss of a main customer. Those go to your support unit with the account position, three years of payment history and a ten-line summary of what the contributor wrote.
· 39 replies promising payment, with the date announced. A follow-up is set for that date: if the payment does not arrive, you know the next day.
The delay, which is the whole point: a contributor's reply waited 11 days for first handling; it is sorted, reclassified and handed to the caseworker the day it arrives. Across 412 replies, 0 was left unattended for more than 48 hours.
The next step I propose: that the 121 unnamed requests trigger the pre-assessment on their own, as they arrive. The contributor then receives a proposed instalment plan in 3 days instead of 21, and you get back the only thing that matters: the decision, taken on a complete file. contributor-replies_412-sorted.pdf39 corrections prepared, 121 requests recovered
⛓ Sourced · letters and e-mails received this month, returns, allocated payments, payment history
Requests for time to pay are our black spot: 4,800 a year, and 41 % wait more than three weeks for a first reply. A company that waits three weeks is already elsewhere.
The year's 4,800 requests are pre-assessed, and each comes back with three instalment plans already costed — the service decides on figures, not on an impression.
Time to pay, or an instalment plan, spreads a contributions debt over several months; it assumes the returns are up to date, and it comes with a commitment from the contributor on future instalments.
What I hand over for each request, as a complete file:
· The state of the returns — up to date or not, and which are missing, with the regularisation letter already drafted where any are.
· Three years of payment history: incidents, earlier instalment plans, and whether they were kept or broken. This is what your caseworkers went looking for by hand, screen by screen, and it alone took half the time of the review.
· Three assumptions, costed on the contributor's declared income: 3 months — the heaviest monthly amount, debt cleared before the next annual adjustment · 6 months — the assumption closest to what your service usually grants · 12 months — the lightest monthly amount, but it straddles two financial years and your records show that is where plans break: 31 % breakage at 12 months against 9 % at 6.
· The draft letter granting and the draft letter refusing, both written and reasoned. The decision to grant time to pay stays with the public officer, because it commits the body — and I hand it to them in minutes, complete, costed and reasoned, rather than after three weeks of searching.
What that gives, measured over the trial quarter: pre-assessing a request for time to pay with an instalment plan goes from 100 % to 20 % of the file's time — from 40 minutes to 8. The first reply to the contributor goes from 21 days to 3, and across 4,800 requests that is 2,560 caseworker hours going back to the files that need a real review.
The next step I propose: that granted plans be followed on their own, instalment by instalment. A contributor who misses a monthly payment is called the next day, not three months later once the plan is lost — across the 312 live plans, that call would have saved 41 of last year's 47 breakages. request-for-time-to-pay_three-costed-plans.pdf40 minutes down to 8, 31 % breakage at 12 months against 9 % at 6
⛓ Sourced · 4,800 requests for time to pay over 12 months, history of plans granted and how they held, returns and declared income
Time to pay, or an instalment plan, spreads a contributions debt over several months; it assumes the returns are up to date, and it comes with a commitment from the contributor on future instalments.
What I hand over for each request, as a complete file:
· The state of the returns — up to date or not, and which are missing, with the regularisation letter already drafted where any are.
· Three years of payment history: incidents, earlier instalment plans, and whether they were kept or broken. This is what your caseworkers went looking for by hand, screen by screen, and it alone took half the time of the review.
· Three assumptions, costed on the contributor's declared income: 3 months — the heaviest monthly amount, debt cleared before the next annual adjustment · 6 months — the assumption closest to what your service usually grants · 12 months — the lightest monthly amount, but it straddles two financial years and your records show that is where plans break: 31 % breakage at 12 months against 9 % at 6.
· The draft letter granting and the draft letter refusing, both written and reasoned. The decision to grant time to pay stays with the public officer, because it commits the body — and I hand it to them in minutes, complete, costed and reasoned, rather than after three weeks of searching.
What that gives, measured over the trial quarter: pre-assessing a request for time to pay with an instalment plan goes from 100 % to 20 % of the file's time — from 40 minutes to 8. The first reply to the contributor goes from 21 days to 3, and across 4,800 requests that is 2,560 caseworker hours going back to the files that need a real review.
The next step I propose: that granted plans be followed on their own, instalment by instalment. A contributor who misses a monthly payment is called the next day, not three months later once the plan is lost — across the 312 live plans, that call would have saved 41 of last year's 47 breakages. request-for-time-to-pay_three-costed-plans.pdf40 minutes down to 8, 31 % breakage at 12 months against 9 % at 6
⛓ Sourced · 4,800 requests for time to pay over 12 months, history of plans granted and how they held, returns and declared income
And waivers of surcharges? That is the subject where we are most often accused of treating two contributors differently.
I reread your three years of waiver decisions, and I draw two things from them: the grid your decisions actually apply, and 620 contributors who met the conditions without ever having asked.
A waiver of surcharges is the writing off of all or part of the late-payment surcharges; it has to be asked for, it is not granted automatically — which is exactly why it passes by those who do not know it exists.
What your 1,940 decisions of the last three years say, once set side by side: four elements explain 91 % of the decisions handed down — returns up to date at the time of the request, debt cleared or plan kept, no incident over the two preceding years, and the amount of surcharges relative to the debt. This is not a grid I invent: it is yours, as it reads in your decisions, and nobody had had the time to write it down.
What I do with each new request: I hand over the four elements filled in, the comparable decisions your committee handed down on neighbouring situations, and a note of the situations where your grid produced two different decisions — there are 9 out of 1,940, and all nine are documented, which is precisely what allows the tenth to be decided either way in full knowledge.
Unclaimed entitlement now, and this is where the gain is largest: 620 contributors in the portfolio meet the four conditions and have never filed a request. I have written the 620 information letters — what a waiver is, the conditions, the form, the deadline. Of the 80 sent as a trial: 63 contributors filed a request, 58 waivers were granted by your committee, 5 files went to the support unit.
What that is worth, put plainly: 58 companies that were entitled to something and did not know it obtained it, and equal treatment stops depending on what a director happens to know about the administration.
The next step I propose: the remaining 540 letters, in two waves of 270, a month apart. The committee will see 200 extra requests over the quarter — which is why I propose taking them in waves, with the four elements already filled in: at 8 minutes a file instead of 40, your committee absorbs the wave without an extra sitting. surcharge-waivers_620-contributors-unclaimed.pdfGrid read from your decisions, 63 requests from 80 letters
⛓ Sourced · 1,940 waiver decisions over 3 years, portfolio accounts, trial of 80 information letters
A waiver of surcharges is the writing off of all or part of the late-payment surcharges; it has to be asked for, it is not granted automatically — which is exactly why it passes by those who do not know it exists.
What your 1,940 decisions of the last three years say, once set side by side: four elements explain 91 % of the decisions handed down — returns up to date at the time of the request, debt cleared or plan kept, no incident over the two preceding years, and the amount of surcharges relative to the debt. This is not a grid I invent: it is yours, as it reads in your decisions, and nobody had had the time to write it down.
What I do with each new request: I hand over the four elements filled in, the comparable decisions your committee handed down on neighbouring situations, and a note of the situations where your grid produced two different decisions — there are 9 out of 1,940, and all nine are documented, which is precisely what allows the tenth to be decided either way in full knowledge.
Unclaimed entitlement now, and this is where the gain is largest: 620 contributors in the portfolio meet the four conditions and have never filed a request. I have written the 620 information letters — what a waiver is, the conditions, the form, the deadline. Of the 80 sent as a trial: 63 contributors filed a request, 58 waivers were granted by your committee, 5 files went to the support unit.
What that is worth, put plainly: 58 companies that were entitled to something and did not know it obtained it, and equal treatment stops depending on what a director happens to know about the administration.
The next step I propose: the remaining 540 letters, in two waves of 270, a month apart. The committee will see 200 extra requests over the quarter — which is why I propose taking them in waves, with the four elements already filled in: at 8 minutes a file instead of 40, your committee absorbs the wave without an extra sitting. surcharge-waivers_620-contributors-unclaimed.pdfGrid read from your decisions, 63 requests from 80 letters
⛓ Sourced · 1,940 waiver decisions over 3 years, portfolio accounts, trial of 80 information letters
620 letters going out in our name on a financial matter — I will not sign that on a verbal agreement.
You do not have to: I propose a written mandate, capped, dated and withdrawable with one word — and it is already drafted, in front of you.
What the mandate authorises, and nothing else:
· Sending the information letters on waivers of surcharges, only to contributors meeting the four conditions, in waves capped at 270 a month. The letter informs of a right and grants nothing: it commits no sum and prejudges no decision of your committee.
· Acknowledgement of every document and every request, within the hour, saying what it is and what is still missing.
· The replies to the fourteen routine subjects, listed by name in the mandate. Any question outside that list reaches you with a reply already drafted and sourced, and you are the one who sends it.
What the mandate expressly leaves to an officer's signature: granting time to pay, granting a waiver, starting enforced collection, and any decision adverse to a contributor. That is not a reservation, it is what gives those decisions their value — a signed decision is a reasoned, traceable, challengeable decision, and a contributor who can challenge it is a contributor who has been heard.
The formal safeguards, in four lines: every letter carries its source and its date, and the statement that it was prepared by a digital assistant of the body · you receive each morning, on one page, the record of what went out the day before — a mistake is caught in an hour, not in three weeks · a review at three months, with the record of what it changed, and without an explicit decision at the review the mandate lapses: it is renewal that needs a signature, not stopping · withdrawal: one word, and direct sending stops within the minute, letters going back to drafts for approval.
What is ready today: the mandate, the information notice displayed at the front desk and published on your website, and the first wave of 270 letters. You sign, the service is running the next morning, and the review is already in your head of service's diary on the 15th of the third month. written-mandate_scope-cap-and-withdrawal.pdf270 letters a month maximum, review at 3 months, immediate withdrawal
✎ Framework · drafted mandate, list of the 14 subjects, information notice to contributors, dispatch log
What the mandate authorises, and nothing else:
· Sending the information letters on waivers of surcharges, only to contributors meeting the four conditions, in waves capped at 270 a month. The letter informs of a right and grants nothing: it commits no sum and prejudges no decision of your committee.
· Acknowledgement of every document and every request, within the hour, saying what it is and what is still missing.
· The replies to the fourteen routine subjects, listed by name in the mandate. Any question outside that list reaches you with a reply already drafted and sourced, and you are the one who sends it.
What the mandate expressly leaves to an officer's signature: granting time to pay, granting a waiver, starting enforced collection, and any decision adverse to a contributor. That is not a reservation, it is what gives those decisions their value — a signed decision is a reasoned, traceable, challengeable decision, and a contributor who can challenge it is a contributor who has been heard.
The formal safeguards, in four lines: every letter carries its source and its date, and the statement that it was prepared by a digital assistant of the body · you receive each morning, on one page, the record of what went out the day before — a mistake is caught in an hour, not in three weeks · a review at three months, with the record of what it changed, and without an explicit decision at the review the mandate lapses: it is renewal that needs a signature, not stopping · withdrawal: one word, and direct sending stops within the minute, letters going back to drafts for approval.
What is ready today: the mandate, the information notice displayed at the front desk and published on your website, and the first wave of 270 letters. You sign, the service is running the next morning, and the review is already in your head of service's diary on the 15th of the third month. written-mandate_scope-cap-and-withdrawal.pdf270 letters a month maximum, review at 3 months, immediate withdrawal
✎ Framework · drafted mandate, list of the 14 subjects, information notice to contributors, dispatch log
At due-date peaks, 38 % of calls ring unanswered. Contributors call back three times, and those who give up are the ones who were about to report a difficulty.
96,000 calls last year, 36,480 unanswered — 38 %, all concentrated in the five days after each due date. This is not a failure of service, it is a peak twelve people cannot absorb.
What I did with those calls, rereading your subject logs, your e-mails and your online account messages:
· 62,000 of the year's 128,000 enquiries are about fourteen subjects, always the same — amount and date of the next instalment, how surcharges are calculated, how to ask for time to pay, where my request stands, how to declare a change of income, compliance certificate, change of bank details, allocation of a payment, waiver of surcharges, disputing an amount, ceasing trading, first registration, live instalment plan, documents to provide.
· All fourteen have a written answer in your internal instructions or your standard letters, and all fourteen answers are drafted, sourced and dated — they are waiting for you. The answer existed: what was missing was availability.
What I propose, and you keep the key: I answer those calls at any hour, including through the five peak days, and I say in the first sentence that I am a digital assistant of the body, not an officer. That is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be told so, and the caller can ask for an officer at any moment — I then take their number and their subject, and the call-back lands in your queue with the file already open.
The rule I hold most firmly, and it is the one that protects the body: I answer on what the account carries, never on what a future decision will grant. A contributor asking whether their request will be granted receives the conditions, the list of documents, the review time and the date on which they will have their answer — a complete path, never a forecast made in the service's name.
The gain, in figures: answering a routine question goes from 40 % to 10 % of the enquiry's time — from 2 minutes to 30 seconds, and across 62,000 enquiries that is 1,550 hours given back to the platform. Unanswered calls go from 38 % to 5 %: what remains are those asking for an officer, and they arrive with the subject already noted.
The next step I propose: that you read the fourteen answers tomorrow, one by one — thirty minutes. As soon as they are approved, the switchboard answers that same night, and I hand you each morning the page of what went out. switchboard-and-front-desk_128000-enquiries-14-subjects.pdf36,480 calls lost, 14 subjects, 14 answers already written
⛓ Sourced · 12 months of switchboard logs, e-mails and online account messages, internal instructions, 14 drafted answers
What I did with those calls, rereading your subject logs, your e-mails and your online account messages:
· 62,000 of the year's 128,000 enquiries are about fourteen subjects, always the same — amount and date of the next instalment, how surcharges are calculated, how to ask for time to pay, where my request stands, how to declare a change of income, compliance certificate, change of bank details, allocation of a payment, waiver of surcharges, disputing an amount, ceasing trading, first registration, live instalment plan, documents to provide.
· All fourteen have a written answer in your internal instructions or your standard letters, and all fourteen answers are drafted, sourced and dated — they are waiting for you. The answer existed: what was missing was availability.
What I propose, and you keep the key: I answer those calls at any hour, including through the five peak days, and I say in the first sentence that I am a digital assistant of the body, not an officer. That is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be told so, and the caller can ask for an officer at any moment — I then take their number and their subject, and the call-back lands in your queue with the file already open.
The rule I hold most firmly, and it is the one that protects the body: I answer on what the account carries, never on what a future decision will grant. A contributor asking whether their request will be granted receives the conditions, the list of documents, the review time and the date on which they will have their answer — a complete path, never a forecast made in the service's name.
The gain, in figures: answering a routine question goes from 40 % to 10 % of the enquiry's time — from 2 minutes to 30 seconds, and across 62,000 enquiries that is 1,550 hours given back to the platform. Unanswered calls go from 38 % to 5 %: what remains are those asking for an officer, and they arrive with the subject already noted.
The next step I propose: that you read the fourteen answers tomorrow, one by one — thirty minutes. As soon as they are approved, the switchboard answers that same night, and I hand you each morning the page of what went out. switchboard-and-front-desk_128000-enquiries-14-subjects.pdf36,480 calls lost, 14 subjects, 14 answers already written
⛓ Sourced · 12 months of switchboard logs, e-mails and online account messages, internal instructions, 14 drafted answers
What we are most criticised for is that two contributors calling in the same week do not get the same answer.
That is true, I have measured it, and it is the first defect I correct: across the fourteen subjects, your letters carried thirty-one different wordings.
What rereading nine years of correspondence shows: on the order in which a partial payment is allocated, six wordings coexisted, two of which plainly contradicted each other. This is nobody's fault: it is nine years of successive caseworkers, each reusing a colleague's letter. For the contributor, though, it is the same question asked twice and two different answers.
What I have already done: one answer per subject, one only, each carrying its internal source and its date of last revision, and the name of the competent service where the subject is not yours. The thirty-one wordings come down to fourteen answers, and the seventeen set aside are kept with the reason for their withdrawal — which is what lets you answer a contributor producing an old letter: you know what he received, when, and why it changed.
What that gives, measured over the quarter:
· 62,000 enquiries handled, 0 divergent answer recorded on the same subject.
· Complaints about a contradictory answer: 74 in the previous quarter, 6 in this one, and the 6 concerned letters predating the rewrite.
· Every answer is dated and sourced: a contributor who disputes knows what he is disputing, and your caseworker knows what he was told.
And the same requirement applies to those who read French poorly or do not dare call: the fourteen answers exist in a plain-language version, short sentences and one message per sentence, and they can be obtained in writing at any hour — email, form or letter, without having to pick up the phone. Over the quarter, 214 contributors used a plain-language version, 61 of whom had never written to the body, and 138 asked for it outside opening hours.
The next step I propose: that I flag on my own any answer that a new instruction amends, and write the updated version within 24 hours. Your fourteen answers will stop ageing in silence — it is the only way this work is not to be done again in two years, and it will cost you one reading per instruction. equal-treatment_31-wordings-down-to-14.pdf0 divergence across 62,000, complaints 74 → 6
⛓ Sourced · 9 years of outgoing letters, internal instructions, complaints of the quarter, log of answers served
What rereading nine years of correspondence shows: on the order in which a partial payment is allocated, six wordings coexisted, two of which plainly contradicted each other. This is nobody's fault: it is nine years of successive caseworkers, each reusing a colleague's letter. For the contributor, though, it is the same question asked twice and two different answers.
What I have already done: one answer per subject, one only, each carrying its internal source and its date of last revision, and the name of the competent service where the subject is not yours. The thirty-one wordings come down to fourteen answers, and the seventeen set aside are kept with the reason for their withdrawal — which is what lets you answer a contributor producing an old letter: you know what he received, when, and why it changed.
What that gives, measured over the quarter:
· 62,000 enquiries handled, 0 divergent answer recorded on the same subject.
· Complaints about a contradictory answer: 74 in the previous quarter, 6 in this one, and the 6 concerned letters predating the rewrite.
· Every answer is dated and sourced: a contributor who disputes knows what he is disputing, and your caseworker knows what he was told.
And the same requirement applies to those who read French poorly or do not dare call: the fourteen answers exist in a plain-language version, short sentences and one message per sentence, and they can be obtained in writing at any hour — email, form or letter, without having to pick up the phone. Over the quarter, 214 contributors used a plain-language version, 61 of whom had never written to the body, and 138 asked for it outside opening hours.
The next step I propose: that I flag on my own any answer that a new instruction amends, and write the updated version within 24 hours. Your fourteen answers will stop ageing in silence — it is the only way this work is not to be done again in two years, and it will cost you one reading per instruction. equal-treatment_31-wordings-down-to-14.pdf0 divergence across 62,000, complaints 74 → 6
⛓ Sourced · 9 years of outgoing letters, internal instructions, complaints of the quarter, log of answers served
And those who never call at all? They are the ones who end up in enforced collection.
They are the ones I go and find, and it is the work that pays the service and the contributor most: 84 accounts a month spotted before the second missed payment.
What I look at, and these are your own ledgers: an account up to date for three years whose amount due doubles over two instalments · a contributor who still files returns but no longer pays · a plan kept for five months then interrupted · a return filed late for the first time in three years. Taken alone, none of these signals means much; together, they preceded 71 % of last year's entries into enforced collection.
What I do with the 84, every month: the file goes to your support unit with the account position, three years of payment history, the schemes the contributor may claim and the three free slots to call them. The call is made by an officer — it is a director in difficulty being called, and that conversation is not delegated to a machine.
What the trial quarter gave: 84 accounts called a month, 61 regularised with no plan or with short time to pay, 14 directed to a spread payment, 9 entered enforced collection. On the same profiles last year, with no call: 34 entries into enforced collection a month. Nine instead of thirty-four, and that is twenty-five companies a month that stay contributors instead of becoming files.
And the preventive reminder, which costs even less: three days before the due date, a message recalls the amount and the date to accounts that have had an incident in the last twelve months. In the trial, 1,100 accounts reminded, 214 payments arriving on time that would have gone into the reminder run — 214 fewer reminders to prepare, and 214 demand letters nobody received.
The next step I propose: extend the preventive reminder to the 2,400 accounts that have had an incident in the last twenty-four months. At the effect measured, that is close to 470 reminders avoided per due date — I hand you the list and the standard message, you approve the wording once, and the reminder goes out before the March due date. spotting-companies-slipping_84-accounts-a-month.pdf34 entries into enforced collection down to 9, 214 reminders avoided
⛓ Sourced · three years of payment history, the year's entries into enforced collection, preventive reminder trial on 1,100 accounts
What I look at, and these are your own ledgers: an account up to date for three years whose amount due doubles over two instalments · a contributor who still files returns but no longer pays · a plan kept for five months then interrupted · a return filed late for the first time in three years. Taken alone, none of these signals means much; together, they preceded 71 % of last year's entries into enforced collection.
What I do with the 84, every month: the file goes to your support unit with the account position, three years of payment history, the schemes the contributor may claim and the three free slots to call them. The call is made by an officer — it is a director in difficulty being called, and that conversation is not delegated to a machine.
What the trial quarter gave: 84 accounts called a month, 61 regularised with no plan or with short time to pay, 14 directed to a spread payment, 9 entered enforced collection. On the same profiles last year, with no call: 34 entries into enforced collection a month. Nine instead of thirty-four, and that is twenty-five companies a month that stay contributors instead of becoming files.
And the preventive reminder, which costs even less: three days before the due date, a message recalls the amount and the date to accounts that have had an incident in the last twelve months. In the trial, 1,100 accounts reminded, 214 payments arriving on time that would have gone into the reminder run — 214 fewer reminders to prepare, and 214 demand letters nobody received.
The next step I propose: extend the preventive reminder to the 2,400 accounts that have had an incident in the last twenty-four months. At the effect measured, that is close to 470 reminders avoided per due date — I hand you the list and the standard message, you approve the wording once, and the reminder goes out before the March due date. spotting-companies-slipping_84-accounts-a-month.pdf34 entries into enforced collection down to 9, 214 reminders avoided
⛓ Sourced · three years of payment history, the year's entries into enforced collection, preventive reminder trial on 1,100 accounts
Management will ask what this has given us after a year. I need a figure I can defend in front of them and in front of the staff representatives.
7,410 officer hours given back to the service over the year — close to five full-time staff — and the calculation is yours, it fits in four lines.
The calculation, item by item, so you can redo it:
· Amicable reminders: 21,600 reminders, preparation from 9 minutes down to 1 minute 30 — 60 % → 10 % of the task's time — that is 2,700 hours.
· Reminders now pointless: 2,400 letters that no longer go out, at 15 minutes from spotting to sending, that is 600 hours — and 2,400 companies that received no demand for a sum already paid.
· Routine questions: 62,000 enquiries, 2 minutes down to 30 seconds — 40 % → 10 % — that is 1,550 hours.
· Requests for time to pay: 4,800 files, 40 minutes down to 8 — 100 % → 20 % — that is 2,560 hours.
What those hours are, and this is what defends best before management and staff representatives alike: officer time given back to the service, at unchanged headcount — no post cut, no post created. Time given back is not a staffing saving: it is caseworker time going back to the call to the company that is slipping, to the review of a difficult situation and to the files nobody had time to open.
What those hours became, on your own records:
· First reply to a request for time to pay: 21 days → 3 days, and 100 % of requests handled within 5 days.
· Unanswered calls: 38 % → 5 %, and those that remain arrive with the subject already noted.
· Contributor replies left unattended beyond 48 hours: 0 out of 412.
· Monthly entries into enforced collection on the profiles followed: 34 → 9.
· Waivers granted to contributors who had never asked: 58 from the first 80 information letters — that one is not counted in hours, and it may be the only one your management remembers.
The next step I propose for the meeting: the calculation page is written and fits on one side — four lines of calculation, five delays, two framework measures. Send it with the agenda: a figure read the day before is discussed better than a figure discovered in the room. yearly-review_7410-hours-given-back.pdf60→10, 40→10, 100→20, and the calculation redoable on one side
⛓ Sourced · production log, switchboard records, register of requests for time to pay, entries into enforced collection
The calculation, item by item, so you can redo it:
· Amicable reminders: 21,600 reminders, preparation from 9 minutes down to 1 minute 30 — 60 % → 10 % of the task's time — that is 2,700 hours.
· Reminders now pointless: 2,400 letters that no longer go out, at 15 minutes from spotting to sending, that is 600 hours — and 2,400 companies that received no demand for a sum already paid.
· Routine questions: 62,000 enquiries, 2 minutes down to 30 seconds — 40 % → 10 % — that is 1,550 hours.
· Requests for time to pay: 4,800 files, 40 minutes down to 8 — 100 % → 20 % — that is 2,560 hours.
What those hours are, and this is what defends best before management and staff representatives alike: officer time given back to the service, at unchanged headcount — no post cut, no post created. Time given back is not a staffing saving: it is caseworker time going back to the call to the company that is slipping, to the review of a difficult situation and to the files nobody had time to open.
What those hours became, on your own records:
· First reply to a request for time to pay: 21 days → 3 days, and 100 % of requests handled within 5 days.
· Unanswered calls: 38 % → 5 %, and those that remain arrive with the subject already noted.
· Contributor replies left unattended beyond 48 hours: 0 out of 412.
· Monthly entries into enforced collection on the profiles followed: 34 → 9.
· Waivers granted to contributors who had never asked: 58 from the first 80 information letters — that one is not counted in hours, and it may be the only one your management remembers.
The next step I propose for the meeting: the calculation page is written and fits on one side — four lines of calculation, five delays, two framework measures. Send it with the agenda: a figure read the day before is discussed better than a figure discovered in the room. yearly-review_7410-hours-given-back.pdf60→10, 40→10, 100→20, and the calculation redoable on one side
⛓ Sourced · production log, switchboard records, register of requests for time to pay, entries into enforced collection
They will also ask where you got it wrong. I would rather have it before a contributor throws it at us.
It has happened, I counted it, and I publish it: of the 4,100 outputs of the first quarter, 118 had to be corrected on substance — 2.9 %. You deserve better than learning that figure from a complaint.
The real cause, measured and not assumed: 71 of the 118 corrections concerned letters built on four of your templates that had not been revised since your internal instructions were last rewritten — the address of a service that has moved, mention of a procedure replaced by its online version, the wording on allocating a partial payment. This was not a drafting error: I was faithfully reusing a template that had aged. The other 47 concerned situations resembling no precedent in your portfolio.
What I did about it, and it is measured: each template now carries the governing instruction and its date, and I write the updated version of any template a new instruction amends, within the 24 hours following its circulation.
The following quarter: 35 corrections on substance out of 3,900 outputs — 0.9 %. And the 35 are one-off situations, no aged template left.
The rule that holds all the rest: a value I have not read, I do not write — I ask for it, and I ask fast. A missing document does not become “probably provided”, an untraceable payment does not become “presumably arrived”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its addressee. Over the quarter, 302 missing values, 302 requests prepared, 271 answers back within eight days.
And the protection that matters for your officers' signature: across 8,000 outputs over two quarters, 8,000 were approved by a person — and all 153 corrections were made before sending, none after.
The next step I propose: that the 47 one-off situations of the first quarter become 12 further templates, written, for your reading. Of the 35 corrections of the second quarter, 29 already fell within one of them — it is the same correction as the one on aged templates, taken one notch further. substantive-corrections_118-then-35.pdf2.9 % → 0.9 %, measured cause, 8,000 human approvals
⛓ Sourced · log of outputs and their corrections over two quarters, circulation of internal instructions
The real cause, measured and not assumed: 71 of the 118 corrections concerned letters built on four of your templates that had not been revised since your internal instructions were last rewritten — the address of a service that has moved, mention of a procedure replaced by its online version, the wording on allocating a partial payment. This was not a drafting error: I was faithfully reusing a template that had aged. The other 47 concerned situations resembling no precedent in your portfolio.
What I did about it, and it is measured: each template now carries the governing instruction and its date, and I write the updated version of any template a new instruction amends, within the 24 hours following its circulation.
The following quarter: 35 corrections on substance out of 3,900 outputs — 0.9 %. And the 35 are one-off situations, no aged template left.
The rule that holds all the rest: a value I have not read, I do not write — I ask for it, and I ask fast. A missing document does not become “probably provided”, an untraceable payment does not become “presumably arrived”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its addressee. Over the quarter, 302 missing values, 302 requests prepared, 271 answers back within eight days.
And the protection that matters for your officers' signature: across 8,000 outputs over two quarters, 8,000 were approved by a person — and all 153 corrections were made before sending, none after.
The next step I propose: that the 47 one-off situations of the first quarter become 12 further templates, written, for your reading. Of the 35 corrections of the second quarter, 29 already fell within one of them — it is the same correction as the one on aged templates, taken one notch further. substantive-corrections_118-then-35.pdf2.9 % → 0.9 %, measured cause, 8,000 human approvals
⛓ Sourced · log of outputs and their corrections over two quarters, circulation of internal instructions
Last question, and it is the one that will decide management: what you do without asking us, and what happens if we stop in a year.
Three actions without asking, not one more, each withdrawable with a word — and an exit that is checked before signing, not after.
· I match payments and returns before each run, and I hold the reminder back when the instalment is settled. And the reverse is true too: a reminder held back wrongly surfaces in the Monday check and goes out in the next run — 6 cases over twelve months, all caught within 8 days.
· I acknowledge every document and every request within the hour, saying what it is, what it settles and what is still missing. An acknowledgement grants nothing and commits nothing: it tells the contributor his submission has arrived, and it brought “did you receive it?” chasers down from 340 to 41 a month.
· I hand you every Monday the week's record: reminders held back and why, documents outstanding, requests for time to pay falling due, instalments missed the day before. It is the only thing I send of my own accord, and it goes only to the service.
And the four actions that stay with an officer's signature, because that is exactly what gives them their value: granting time to pay · granting a waiver · starting enforced collection · any decision adverse to a contributor. A signed decision is reasoned, traceable and challengeable — and the contributor who challenges it comes to your committee with a complete file, which is the best guarantee that he will be heard. Across 8,000 outputs over two quarters, those four actions were taken 8,000 times by a person.
The exit, since management will ask: the index is deleted and it held none of your entries — only what is needed to find them where they are. Your nine years of accounts and correspondence have not moved a byte: same software, same files, same permissions. The fourteen answers, the updated templates, the matching rule written in plain language and the grid read from your waiver decisions belong to the body: they are made of its own material, they stay in its files, readable without us. No migration on the way in, therefore none on the way out.
The next step 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 the service runs exactly as before, we switch back on. The protocol fits on one page, and the date that costs you least is the third Wednesday of the month — your records show it is the quietest day between two due dates. Management will know what the promise is worth before committing a second year. what-the-agent-does-alone_and-the-dry-run-exit.pdf3 reversible actions, 4 signatures that stay with an officer
✎ Framework · settings of the automatic actions, dispatch log, export formats, dry-run exit protocol
· I match payments and returns before each run, and I hold the reminder back when the instalment is settled. And the reverse is true too: a reminder held back wrongly surfaces in the Monday check and goes out in the next run — 6 cases over twelve months, all caught within 8 days.
· I acknowledge every document and every request within the hour, saying what it is, what it settles and what is still missing. An acknowledgement grants nothing and commits nothing: it tells the contributor his submission has arrived, and it brought “did you receive it?” chasers down from 340 to 41 a month.
· I hand you every Monday the week's record: reminders held back and why, documents outstanding, requests for time to pay falling due, instalments missed the day before. It is the only thing I send of my own accord, and it goes only to the service.
And the four actions that stay with an officer's signature, because that is exactly what gives them their value: granting time to pay · granting a waiver · starting enforced collection · any decision adverse to a contributor. A signed decision is reasoned, traceable and challengeable — and the contributor who challenges it comes to your committee with a complete file, which is the best guarantee that he will be heard. Across 8,000 outputs over two quarters, those four actions were taken 8,000 times by a person.
The exit, since management will ask: the index is deleted and it held none of your entries — only what is needed to find them where they are. Your nine years of accounts and correspondence have not moved a byte: same software, same files, same permissions. The fourteen answers, the updated templates, the matching rule written in plain language and the grid read from your waiver decisions belong to the body: they are made of its own material, they stay in its files, readable without us. No migration on the way in, therefore none on the way out.
The next step 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 the service runs exactly as before, we switch back on. The protocol fits on one page, and the date that costs you least is the third Wednesday of the month — your records show it is the quietest day between two due dates. Management will know what the promise is worth before committing a second year. what-the-agent-does-alone_and-the-dry-run-exit.pdf3 reversible actions, 4 signatures that stay with an officer
✎ Framework · settings of the automatic actions, dispatch log, export formats, 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 in collections
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
Informal reminders
Reminder letters and emails prepared from the account positions, in line with your templates.
Replies to contributors
Answer everyday questions (due dates, rates, procedures) and direct people to the right department, 24/7.
Assessing payment extension and debt remission requests
Sorts requests for payment extensions and debt remission, and flags sensitive cases for review.
Draft payment plans
Preparing the payment plans and the replies to requests for time to pay, for the officer to approve.
Reconciling declarations and payments
Reconcile declarations against payments to flag the discrepancies and anomalies to deal with.
Document reading
Extraction and checking of the documents sent by contributors (certificates, evidence, letters).
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.
Accounting agent (summaries, anomalies)
Accounting connection, alerts.
Accounting agent (summaries, anomalies) from 781 € incl. VAT / month Discover the agent →Debt collection agent (multichannel sequences)
Follow-ups, tracking.
Debt collection agent (multichannel sequences) from 587 € 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 collections department win back?
By automating informal reminders, everyday replies and the reconciliation of accounts, an organisation can aim for a clear reduction in time spent on repetitive tasks — reinvested in supporting difficult situations.
Preparing an informal reminder
Answering a contributor's everyday question
Pre-qualifying a request for time to pay with a payment plan
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 collections agent (informal reminders, replies to contributors, pre-qualifying time to pay), 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 collections body
Contributors' 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 lawNative location and minimisation for contributors' data; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The officer keeps the decisionThe AI agent prepares reminders, replies and payment plans that can be checked; no proceedings are started automatically.
Human oversight & traceabilityFor the informal reminders: systematic updates and logging, compliant with the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent start proceedings?
No: it prepares informal reminders and replies; any move to enforced collection remains decided and signed by an officer.
Is businesses' accounting data protected?
Yes: hosting in France, a dedicated and isolated resource, role-based access and full traceability. GDPR: governed deployment.
Does the agent launch enforced collection procedures?
No: it prepares the informal reminders and the draft payment plans; any move to contentious collection remains the officer's decision.
Does it calculate the contributions owed?
It sets out what has been declared and flags anomalies, but the binding calculation remains that of the organisation's own systems.
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 contributors use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: a contributor has no account to create and no application to install. A tradesperson or the head of a very small business who wants to know where their request for extra time stands writes from the tool already in their hand, between two jobs. This is a lever for access to the service — WhatsApp and the telephone reach people an online form never does, which reduces the non-take-up of rights and serves equal access. Internally, your public-sector staff talk to the agent from Microsoft Teams, Slack or their email, without switching tools. Oversight runs from a web dashboard. These connectors rely on open standards, including the MCP protocol; they are included in every plan, at no extra cost, within the number of connections your level includes. Only the fees charged by the platforms themselves — WhatsApp Business bills per conversation — are passed on at actual cost, with no margin, outside the subscription.
Related roles
Other professions in the social and financial sphere
Let's talk
Let's size up the potential in your organisation
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.