● Public sector — Public finances officer
AI agent for the public finances officer
An assistant that answers taxpayers, prepares the reminders and guides people to the online services — the officer keeps control of every file. Hosted in France — on local inference or an isolated resource — tax data never leaves the network. No tax calculation is automated: the AI agent assists, the public finances officer decides.
Hosted in France
Tax data protected
GDPR & AI Act: governed deployment
No automated tax calculation
Updated on
Deployed in a few weeks
Public finances assistant · hosted in France
A taxpayer asks how to set up monthly instalments for their property tax. Prepare a reply.
Reply drafted: they can sign up online from their personal area on the tax website (“Payments” section), by phone or by post to the relevant office. The email recalls the sign-up deadline for it to take effect in the current year and the bank details to have ready.
Ready for review.
⛓ Source · the official tax website + your standard replies
Ready for review.
⛓ Source · the official tax website + your standard replies
Prepare the informal reminder letters for the files more than 30 days past the due date.
12 draft reminders prepared from your templates: a reminder of the due date, the amount still owed as shown on the account, how to pay and who to contact in case of difficulty. No amount recalculated — the figures come from the account positions.
The list is submitted for your check before anything is sent.
✎ Action · reminders ready for review — the officer approves
The list is submitted for your check before anything is sent.
✎ Action · reminders 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 finances officer decidesNo automated calculation or decision
✦ In brief
At the tax office, a Blue Lemon Agent agent assists officers with the repetitive tasks — replies to taxpayers, reminder letters, guidance to the online services, pre-filling of the everyday forms — on every channel. It runs on local inference or is hosted in France on a dedicated, isolated resource: tax data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. No tax calculation, no audit and no decision is automated: the public officer reviews, corrects and approves. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and taxpayers 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 public finances — and why they hesitate
Every filing season, every payment deadline generates thousands of enquiries: emails, calls, counter visits. Time is short, but the data handled — the tax position of individuals and businesses — is covered by tax secrecy.
! The issue
Public finances officers are caught between taxpayers who expect quick answers and a multichannel enquiry load that explodes at the deadlines. Yet most consumer AI tools amount to entrusting account positions, income and asset data to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For data covered by tax secrecy, 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, no automated tax calculation and no automated decision: the time saved on enquiries and reminders is never paid for in lost confidentiality. The aim is not to replace the officer, but to give them back time for audits and complex files.
The decisive point
Protecting tax data: sovereignty & compliance
A public finances office handles data covered by tax secrecy. Here is how the architecture of our agents protects it, office by office.
Local inference
The agent can run on a machine at the unit: no data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — the tax data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
As regards tax data, exposure 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 data: an environment strictly dedicated to your unit, 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 automated tax calculation and no automated decision; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For the most sensitive data, SecNumCloud and HDS options are available depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
4 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The public body in this demonstration
Fictional public bodyVal-de-Sèvre personal tax office — a local office of a departmental public finances directorate (fictional office)
- Sector
- Personal tax office: assessment and collection of income tax, property taxes and the second-home residence tax; front desk, telephone and secure messaging in the taxpayer's personal account; open to the public 21 hours a week
- Headcount
- 34 officers — 18 in personal tax management, 9 in collection, 5 on the front desk, plus the head of office and the deputy; the same people cover the desk, the phone and the secure messaging
- Public served
- 68,400 tax households and 41,900 property tax notices — 96,200 contacts a year: 52,300 calls, 21,600 desk visits, 22,300 messages
- Order of magnitude
- 18 recurring questions cover 61,700 of the 96,200 contacts; 7,400 guided online procedures; 8,600 messages arrive in the three weeks after a deadline; 9,340 files call for an amicable reminder each year
- Tools in place
- DGFiP management and collection applications, secure messaging in the personal account, 52 standard answers written by the office, guidance published on impots.gouv.fr — the agent reads them, nothing is replaced or migrated
- Who decides
- The head of the personal tax office settles every reference answer and signs the acts listed in the delegation decision; the case officer decides on an individual file; the collecting accountant grants payment plans; the tax calculation stays with the DGFiP applications and the officer who validates them
- Room for improvement
- 17,806 calls out of 52,300 ring unanswered at peak — 34 %; a secure message waits 9 days on average; 34 % of taxpayers abandon the online monthly direct debit sign-up midway; and 4 of the office's 52 standard answers carry information the rule or the site has changed since
At Val-de-Sèvre, 34 officers hold three service channels for 68,400 tax households, and the flow doubles at filing and payment deadlines. The most repetitive questions — monthly direct debit, means of payment, withholding rate, tax notice — take up nearly two thirds of the flow, and the difficult cases wait behind them. The exchanges below cover a full year, from the first reading of the flow to the review presented to the departmental directorate. The agent runs on local inference on a machine inside the office, reads the standard answers and the contact log without ever writing to them, produces no tax calculation, and no named account statement ever leaves an authenticated channel.
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.
Public finances assistant · hosted in France
I have reread the 96,200 contacts the office received over twelve months — 52,300 calls, 21,600 desk visits, 22,300 secure messages — and the flow has a very clear shape that nobody has time to draw.
What the count says: 18 questions cover 61,700 contacts out of 96,200, close to two in three. Monthly direct debit for property tax and income tax (8,400), payment deadline and means of payment (7,200), withholding rate and rate changes (6,100), obtaining the tax notice (5,300), change of address and family situation (4,400), property tax exemptions and reliefs (3,900), the declaration of built property (3,400), payment difficulties (3,200), claims and their deadlines (2,800), access to the personal account (2,700), and eight further questions (14,300). The 18 reference answers are written, sourced and dated — they are waiting for you.
The three items you can check against your own logs:
· A written answer to a routine question takes 10 minutes end to end, 60 % of it spent finding the rule and drafting — 6 minutes.
· An amicable reminder letter takes 25 minutes, 40 % of it spent reconciling the account statement and drafting — 10 minutes.
· The queue of messages arriving in the three weeks after a deadline — 8,600 messages — takes all the time anyone can give it: 8 minutes per message, 1,146 hours a year.
And here is the gap I found by comparing your 52 standard answers with the published rules: 4 of them carry information the rule or the site has changed since — the date that decides which year the monthly direct debit takes effect, the number of instalments, the threshold above which payment must be electronic, and the exact name of the surcharge incurred for late payment. Those 4 standard answers fed 11,900 answers last year, and the 4 up-to-date versions are written tonight, each with the replaced line and the official text that replaces it side by side.
The next step, and it takes twenty-five minutes: you reread the 4 rewritten standard answers and the 3 questions that carry the most volume. The head of office settles the version, and it serves the front desk the same day. office-flow_96200-contacts-18-questions.pdf61,700 contacts on 18 questions, the 3 measured items
⛓ Sourced · 12-month contact log, the office's 52 standard answers, guidance published on impots.gouv.fr
What the count says: 18 questions cover 61,700 contacts out of 96,200, close to two in three. Monthly direct debit for property tax and income tax (8,400), payment deadline and means of payment (7,200), withholding rate and rate changes (6,100), obtaining the tax notice (5,300), change of address and family situation (4,400), property tax exemptions and reliefs (3,900), the declaration of built property (3,400), payment difficulties (3,200), claims and their deadlines (2,800), access to the personal account (2,700), and eight further questions (14,300). The 18 reference answers are written, sourced and dated — they are waiting for you.
The three items you can check against your own logs:
· A written answer to a routine question takes 10 minutes end to end, 60 % of it spent finding the rule and drafting — 6 minutes.
· An amicable reminder letter takes 25 minutes, 40 % of it spent reconciling the account statement and drafting — 10 minutes.
· The queue of messages arriving in the three weeks after a deadline — 8,600 messages — takes all the time anyone can give it: 8 minutes per message, 1,146 hours a year.
And here is the gap I found by comparing your 52 standard answers with the published rules: 4 of them carry information the rule or the site has changed since — the date that decides which year the monthly direct debit takes effect, the number of instalments, the threshold above which payment must be electronic, and the exact name of the surcharge incurred for late payment. Those 4 standard answers fed 11,900 answers last year, and the 4 up-to-date versions are written tonight, each with the replaced line and the official text that replaces it side by side.
The next step, and it takes twenty-five minutes: you reread the 4 rewritten standard answers and the 3 questions that carry the most volume. The head of office settles the version, and it serves the front desk the same day. office-flow_96200-contacts-18-questions.pdf61,700 contacts on 18 questions, the 3 measured items
⛓ Sourced · 12-month contact log, the office's 52 standard answers, guidance published on impots.gouv.fr
There are 34 of us, and the three channels run at the same time during the campaigns. Nobody has half a day to reread standard answers. Where do we start?
With three questions out of eighteen — and I have ranked them by the volume they handle, not alphabetically.
What your twelve months say:
· Monthly direct debit: 8,400 contacts a year. It is the office's first question, and it is the one whose standard answer omits the date that decides everything.
· Payment deadline and means of payment: 7,200, and that is the standard answer carrying an old electronic-payment threshold.
· Withholding rate and rate changes: 6,100.
Those three questions make 21,700 of the 61,700 recurring contacts: three questions out of eighteen, 35 % of the repetitive flow. The other fifteen can be reread four a week, with nothing left waiting.
And I advise against opening all eighteen at once, with the figures to back it: the first 3 take 25 minutes of rereading and cover 35 % of the recurring flow; all 18 take 2 h 30 — half a day that would push the go-live back three weeks for 65 % more flow. Four questions a week puts everything into service within a month, and the first gain lands tomorrow morning.
What those twenty-five minutes give you, answer by answer: the rule in one sentence, the official text that carries it with its date, the plain-language version, what to do next, and the number of contacts the answer would have handled last year alone. For the four rewritten standard answers, it reads exactly like this: a monthly direct debit requested up to 30 June takes effect in the current year, and from 1 July it takes effect the following year — that is what impots.gouv.fr states, and that date is missing from your standard answer ; there are ten instalments, running from January to October, where your standard answer announces twelve ; above €300 payment must be made by an electronic means — article 1681 sexies of the French general tax code, a threshold in force since 1 January 2019 — and your standard answer still carries an older one ; late payment of a sum included in a tax roll incurs a 10 % surcharge when it is not settled within the 45 days following the date the roll was put into collection — that is article 1730 of the French general tax code, and your standard answer says « penalties » without naming them.
And among the eighteen reference answers I am handing you, the one on claims arrives with the two deadlines nothing of yours carried in writing: a claim on a local direct tax is admissible until 31 December of the year following the year the tax roll was put into collection — article R*196-2 of the French book of tax procedures — and the administration rules within six months, article R*198-10 of the same book, with at most three further months notified before that deadline expires.
You approve, or you correct with a word; the head of office settles the version. The drafting and the reconciliation are done: the decision takes twenty-five minutes instead of half a day, and it stays the office's — which is precisely what makes the answer enforceable against a taxpayer who disputes it.
The next step I propose: that I reconcile your 52 standard answers with the texts and the published guidance every night, and write the up-to-date version of any answer a publication has just changed, that same night. Your standard answers will stop ageing in silence, and this rereading will not have to be done again: it becomes a five-minute validation per change. reference-answers_18-written-4-rewritten.pdf11,900 answers concerned, 4 up-to-date versions written and sourced
⛓ Sourced · per-question count over 12 months, 52 standard answers, articles 1730 and 1681 sexies of the French general tax code, articles R*196-2 and R*198-10 of the French book of tax procedures, impots.gouv.fr
What your twelve months say:
· Monthly direct debit: 8,400 contacts a year. It is the office's first question, and it is the one whose standard answer omits the date that decides everything.
· Payment deadline and means of payment: 7,200, and that is the standard answer carrying an old electronic-payment threshold.
· Withholding rate and rate changes: 6,100.
Those three questions make 21,700 of the 61,700 recurring contacts: three questions out of eighteen, 35 % of the repetitive flow. The other fifteen can be reread four a week, with nothing left waiting.
And I advise against opening all eighteen at once, with the figures to back it: the first 3 take 25 minutes of rereading and cover 35 % of the recurring flow; all 18 take 2 h 30 — half a day that would push the go-live back three weeks for 65 % more flow. Four questions a week puts everything into service within a month, and the first gain lands tomorrow morning.
What those twenty-five minutes give you, answer by answer: the rule in one sentence, the official text that carries it with its date, the plain-language version, what to do next, and the number of contacts the answer would have handled last year alone. For the four rewritten standard answers, it reads exactly like this: a monthly direct debit requested up to 30 June takes effect in the current year, and from 1 July it takes effect the following year — that is what impots.gouv.fr states, and that date is missing from your standard answer ; there are ten instalments, running from January to October, where your standard answer announces twelve ; above €300 payment must be made by an electronic means — article 1681 sexies of the French general tax code, a threshold in force since 1 January 2019 — and your standard answer still carries an older one ; late payment of a sum included in a tax roll incurs a 10 % surcharge when it is not settled within the 45 days following the date the roll was put into collection — that is article 1730 of the French general tax code, and your standard answer says « penalties » without naming them.
And among the eighteen reference answers I am handing you, the one on claims arrives with the two deadlines nothing of yours carried in writing: a claim on a local direct tax is admissible until 31 December of the year following the year the tax roll was put into collection — article R*196-2 of the French book of tax procedures — and the administration rules within six months, article R*198-10 of the same book, with at most three further months notified before that deadline expires.
You approve, or you correct with a word; the head of office settles the version. The drafting and the reconciliation are done: the decision takes twenty-five minutes instead of half a day, and it stays the office's — which is precisely what makes the answer enforceable against a taxpayer who disputes it.
The next step I propose: that I reconcile your 52 standard answers with the texts and the published guidance every night, and write the up-to-date version of any answer a publication has just changed, that same night. Your standard answers will stop ageing in silence, and this rereading will not have to be done again: it becomes a five-minute validation per change. reference-answers_18-written-4-rewritten.pdf11,900 answers concerned, 4 up-to-date versions written and sourced
⛓ Sourced · per-question count over 12 months, 52 standard answers, articles 1730 and 1681 sexies of the French general tax code, articles R*196-2 and R*198-10 of the French book of tax procedures, impots.gouv.fr
All of this assumes you read tax situations: income, assets, payment difficulties. That is covered by tax secrecy. Where does it go, and who has access?
Nothing goes anywhere. I run on local inference on a machine inside the office, and what I read does not leave your network.
Local inference means the model computes on your machine: the text of a taxpayer's message or of an account statement crosses no external network to be processed. If the directorate would rather not run a machine, the other route is an isolated resource hosted in France, dedicated to your office, with no pooling with any other.
What that changes, point by point:
· Tax data trains no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no right to write — that is checked with one command, which beats a promise.
· Encryption in transit and at rest, role-based access — rights follow the job: the front desk opens the reference answers, only the collection unit opens account statements. 34 roles for your 34 officers, and the log shows 0 access outside a role since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, including against a US provider hosting in Europe — location alone not being enough to guarantee immunity, the exposure is documented for the configuration chosen.
· A complete log: who asked what, when, and what the system produced.
And there is one thing I will never do, not out of caution but because it is an offence: handing a named account statement to someone who is not entitled to it. Article L. 103 of the French book of tax procedures extends the professional secrecy obligation of articles 226-13 and 226-14 of the French criminal code to every person called upon to take part in the assessment, control, collection or litigation of tax — I am one of them.
And the lawful route exists, I take it, and it gives the taxpayer exactly what they were after: I file the named breakdown in the secure messaging of their personal account, where only the authenticated holder opens it ; on WhatsApp, on the site chat or by email, I give the rule, the calendar, the amount as it appears on their notice if they read it to me themselves, and the authentication path in three screens. Over the trial quarter, 4,210 requests for an account statement arrived through a non-authenticated channel; 4,210 left with the path, and 3,968 led to a consultation in the personal account within 24 hours — 94 %.
And the taxpayer knows who they are speaking to: I introduce myself as a digital assistant of the office in the first sentence, because article 50(1) of Regulation (EU) 2024/1689 on artificial intelligence, applicable since 2 August 2026, requires informing anyone interacting with an AI system — and they can ask for a person at any moment, in which case I pass the request on with the subject already noted.
The figures that sum all this up: 0 tax data out of the office network across 96,200 contacts, 0 transfer outside the European Union, 0 named amount served outside an authenticated channel.
What I propose: that I keep up to date the record your data protection officer will ask for — hosting, data processed, retention periods, who accesses what. The first version is written and attached; it updates itself every time a new question is opened. technical-framework_where-tax-data-lives.pdfLocal inference, read-only, 0 transfer outside the EU
✎ Framework · deployment architecture, technical account rights, article L. 103 of the French book of tax procedures, article 50(1) of Regulation (EU) 2024/1689
Local inference means the model computes on your machine: the text of a taxpayer's message or of an account statement crosses no external network to be processed. If the directorate would rather not run a machine, the other route is an isolated resource hosted in France, dedicated to your office, with no pooling with any other.
What that changes, point by point:
· Tax data trains no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no right to write — that is checked with one command, which beats a promise.
· Encryption in transit and at rest, role-based access — rights follow the job: the front desk opens the reference answers, only the collection unit opens account statements. 34 roles for your 34 officers, and the log shows 0 access outside a role since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, including against a US provider hosting in Europe — location alone not being enough to guarantee immunity, the exposure is documented for the configuration chosen.
· A complete log: who asked what, when, and what the system produced.
And there is one thing I will never do, not out of caution but because it is an offence: handing a named account statement to someone who is not entitled to it. Article L. 103 of the French book of tax procedures extends the professional secrecy obligation of articles 226-13 and 226-14 of the French criminal code to every person called upon to take part in the assessment, control, collection or litigation of tax — I am one of them.
And the lawful route exists, I take it, and it gives the taxpayer exactly what they were after: I file the named breakdown in the secure messaging of their personal account, where only the authenticated holder opens it ; on WhatsApp, on the site chat or by email, I give the rule, the calendar, the amount as it appears on their notice if they read it to me themselves, and the authentication path in three screens. Over the trial quarter, 4,210 requests for an account statement arrived through a non-authenticated channel; 4,210 left with the path, and 3,968 led to a consultation in the personal account within 24 hours — 94 %.
And the taxpayer knows who they are speaking to: I introduce myself as a digital assistant of the office in the first sentence, because article 50(1) of Regulation (EU) 2024/1689 on artificial intelligence, applicable since 2 August 2026, requires informing anyone interacting with an AI system — and they can ask for a person at any moment, in which case I pass the request on with the subject already noted.
The figures that sum all this up: 0 tax data out of the office network across 96,200 contacts, 0 transfer outside the European Union, 0 named amount served outside an authenticated channel.
What I propose: that I keep up to date the record your data protection officer will ask for — hosting, data processed, retention periods, who accesses what. The first version is written and attached; it updates itself every time a new question is opened. technical-framework_where-tax-data-lives.pdfLocal inference, read-only, 0 transfer outside the EU
✎ Framework · deployment architecture, technical account rights, article L. 103 of the French book of tax procedures, article 50(1) of Regulation (EU) 2024/1689
At deadlines the switchboard saturates. Taxpayers call three times, then end up at the desk for a question that fits in one sentence.
52,300 calls last year, 17,806 unanswered — 34 %, all concentrated in the six weeks following a deadline. It is not a question of service quality: it is 34 officers holding three channels at once while the flow doubles.
What I have done with those 17,806 calls, by rereading your callbacks and your messages: 34,800 of the 52,300 calls concern the 18 questions, and the 18 answers are written, sourced and dated. The answer already existed somewhere: what was missing was availability, and that is exactly what I bring.
What I do, and you keep the key: I answer at any hour, including Saturdays and while the desk is closed, I introduce myself as a digital assistant of the office in the first sentence, and the taxpayer can ask for a person whenever they want — I then take their number and leave you a dated callback, subject already noted.
On channels, I go and find taxpayers where they are: WhatsApp Business, the site chat and email on the public side — no account to create, no app to install ; Microsoft Teams, Slack or the mailbox on the office side, so your officers write to me without changing tools. This is not a convenience: over the trial quarter, 2,140 taxpayers wrote to me on WhatsApp between 6 p.m. and 11 p.m., and 1,386 of them had never used the secure messaging. Those are people who would otherwise discover the problem through a surcharge.
The rule I hold most firmly, and it is the one that protects the office: I answer within the personal tax office's remit, and anything belonging to another counter leaves with that counter's name, address, opening hours and turnaround. The taxpayer leaves with a complete route.
The gain, on your volumes: a written answer to a routine question goes from 60 % to 10 % of the time a request takes — from 6 minutes to 1 minute. Across 61,700 recurring contacts a year, that is 5,141 hours going back to the files that deserve an officer. And unanswered calls go from 17,806 to 1,042: what remains are those who ask for a person, and they reach you with the subject already noted.
The next step I propose: that I start answering tonight on the first three questions — monthly direct debit, means of payment, withholding rate — that is 21,700 contacts a year. You reread the first night's log tomorrow morning, and you open the other fifteen at whatever pace suits you. service-and-online-procedures_52300-calls-7400-guided.pdf34 % → 2 % lost calls, direct-debit abandonment 34 % → 8 %
⛓ Sourced · 12-month switchboard log, 18 reference answers drafted, channel log over the trial quarter
What I have done with those 17,806 calls, by rereading your callbacks and your messages: 34,800 of the 52,300 calls concern the 18 questions, and the 18 answers are written, sourced and dated. The answer already existed somewhere: what was missing was availability, and that is exactly what I bring.
What I do, and you keep the key: I answer at any hour, including Saturdays and while the desk is closed, I introduce myself as a digital assistant of the office in the first sentence, and the taxpayer can ask for a person whenever they want — I then take their number and leave you a dated callback, subject already noted.
On channels, I go and find taxpayers where they are: WhatsApp Business, the site chat and email on the public side — no account to create, no app to install ; Microsoft Teams, Slack or the mailbox on the office side, so your officers write to me without changing tools. This is not a convenience: over the trial quarter, 2,140 taxpayers wrote to me on WhatsApp between 6 p.m. and 11 p.m., and 1,386 of them had never used the secure messaging. Those are people who would otherwise discover the problem through a surcharge.
The rule I hold most firmly, and it is the one that protects the office: I answer within the personal tax office's remit, and anything belonging to another counter leaves with that counter's name, address, opening hours and turnaround. The taxpayer leaves with a complete route.
The gain, on your volumes: a written answer to a routine question goes from 60 % to 10 % of the time a request takes — from 6 minutes to 1 minute. Across 61,700 recurring contacts a year, that is 5,141 hours going back to the files that deserve an officer. And unanswered calls go from 17,806 to 1,042: what remains are those who ask for a person, and they reach you with the subject already noted.
The next step I propose: that I start answering tonight on the first three questions — monthly direct debit, means of payment, withholding rate — that is 21,700 contacts a year. You reread the first night's log tomorrow morning, and you open the other fifteen at whatever pace suits you. service-and-online-procedures_52300-calls-7400-guided.pdf34 % → 2 % lost calls, direct-debit abandonment 34 % → 8 %
⛓ Sourced · 12-month switchboard log, 18 reference answers drafted, channel log over the trial quarter
The real subject is the monthly direct debit and online payment. People start the procedure, stop halfway, and come back to the desk in October when the notice lands.
I went back through your 7,400 guided procedures for the year, and the abandonment has a single cause, measurable, and already dealt with.
An online procedure is one the taxpayer completes themselves. Non-take-up is the right or the facility a person holds and does not obtain, because they never reached the end.
What the measurement says: 34 % of taxpayers abandon the online monthly direct debit sign-up midway, and 71 % of those abandonments happen on two screens: bank details and tax reference number. The abandonment is not technical: the person starts, a document is missing, they close the page and do not come back.
What I have put in place, and it removes the cause:
· The list of documents to have ready is given BEFORE the first screen — bank details, tax reference number, notice reference —, on the six most requested procedures.
· I guide screen by screen, I read out what is asked and why, and I flag the missing document before submission, not after.
· I prefill in front of the person the fields that copy across from their notice — references, address, family situation — and it is they who enter their identifier and submit. That gesture is what makes the procedure theirs, therefore valid, and never opposable to them.
· And I carry the date that decides everything: a sign-up requested up to 30 June takes effect in the current year; from 1 July it takes effect the following year, and there are ten instalments, from January to October — that is what impots.gouv.fr states. That sentence was missing from your standard answer, and it is the difference between a direct debit that starts and one that waits a year.
The result, over the trial quarter: abandonment goes from 34 % to 8 %, sign-ups requested before 30 June go from 3,180 to 5,940 — 2,760 more taxpayers spreading the cost instead of paying in one go, and 1,840 October desk visits are avoided.
And I tell you where this gain does not come from: those 7,400 guided procedures are NOT counted in the 5,141 hours of the previous tab. I only count the three items you can check against your own logs — the rest I show you, I do not bill it twice.
The next step I propose: that the six document lists be published at the head of each procedure on the office site this week — they are written, reread and ready to place. That is half the abandonments dealt with without a single line of code moving. service-and-online-procedures_52300-calls-7400-guided.pdf71 % of abandonments on 2 screens, 2,760 more sign-ups before 30 June
⛓ Sourced · log of the 7,400 guided procedures, abandonment screens across the 6 procedures, impots.gouv.fr monthly direct debit guidance, 6 document lists drafted
An online procedure is one the taxpayer completes themselves. Non-take-up is the right or the facility a person holds and does not obtain, because they never reached the end.
What the measurement says: 34 % of taxpayers abandon the online monthly direct debit sign-up midway, and 71 % of those abandonments happen on two screens: bank details and tax reference number. The abandonment is not technical: the person starts, a document is missing, they close the page and do not come back.
What I have put in place, and it removes the cause:
· The list of documents to have ready is given BEFORE the first screen — bank details, tax reference number, notice reference —, on the six most requested procedures.
· I guide screen by screen, I read out what is asked and why, and I flag the missing document before submission, not after.
· I prefill in front of the person the fields that copy across from their notice — references, address, family situation — and it is they who enter their identifier and submit. That gesture is what makes the procedure theirs, therefore valid, and never opposable to them.
· And I carry the date that decides everything: a sign-up requested up to 30 June takes effect in the current year; from 1 July it takes effect the following year, and there are ten instalments, from January to October — that is what impots.gouv.fr states. That sentence was missing from your standard answer, and it is the difference between a direct debit that starts and one that waits a year.
The result, over the trial quarter: abandonment goes from 34 % to 8 %, sign-ups requested before 30 June go from 3,180 to 5,940 — 2,760 more taxpayers spreading the cost instead of paying in one go, and 1,840 October desk visits are avoided.
And I tell you where this gain does not come from: those 7,400 guided procedures are NOT counted in the 5,141 hours of the previous tab. I only count the three items you can check against your own logs — the rest I show you, I do not bill it twice.
The next step I propose: that the six document lists be published at the head of each procedure on the office site this week — they are written, reread and ready to place. That is half the abandonments dealt with without a single line of code moving. service-and-online-procedures_52300-calls-7400-guided.pdf71 % of abandonments on 2 screens, 2,760 more sign-ups before 30 June
⛓ Sourced · log of the 7,400 guided procedures, abandonment screens across the 6 procedures, impots.gouv.fr monthly direct debit guidance, 6 document lists drafted
A taxpayer writes to me on WhatsApp: « how much do I owe? ». What exactly do you answer?
I give them their amount — but I give it where they alone can read it, and they have it in three minutes.
What I answer, word for word, on WhatsApp: « I am the digital assistant of the personal tax office. Your amount appears on your account statement, which I cannot open in front of anyone but you: I have just filed it in the secure messaging of your personal account on impots.gouv.fr. Here is the three-screen path to reach it, and the number to call if your access is blocked. If you prefer, read me the « amount payable » line of your notice and I will explain every line of the breakdown. »
Why I proceed this way, and it is not house caution: article L. 103 of the French book of tax procedures extends the professional secrecy obligation of articles 226-13 and 226-14 of the French criminal code to every person taking part in the assessment, control, collection or litigation of tax. A named amount served on a non-authenticated channel is an offence, not a risk.
And here is what the taxpayer gets, which is exactly what they were after:
· The breakdown is already filed when they open their account — principal, date the roll was put into collection, deadline, surcharge if it is due, amount outstanding, each one taken from their account statement, none recomputed.
· The authentication path is written out, and I walk them through it screen by screen if they get stuck.
· If they have their notice in front of them and read me a figure themselves, I explain every line, the payment calendar and their options — they hold the data, I only read it with them.
· And if they say they cannot pay, I do not leave them there: I give them how to request a payment plan from the collecting accountant, the list of supporting documents, and I prepare the request for their signature. Over the quarter, 612 taxpayers took that route, and 486 obtained a schedule before the 45-day deadline of article 1730 of the French general tax code — therefore before the 10 % surcharge.
The measurement over the quarter: 4,210 requests for an amount received outside an authenticated channel, 4,210 left with the path, 3,968 consultations completed within 24 hours — 94 %, and 0 named amount served outside an authenticated account.
The next step I propose: that I push the breakdown into the secure messaging BEFORE the taxpayer asks, as soon as a deadline is fifteen days away. On the sample, 2,340 people opened the message within 48 hours — and those are 2,340 calls that did not happen on the deadline day. technical-framework_where-tax-data-lives.pdfTax secrecy: the authenticated route, 94 % completion
✎ Framework · article L. 103 of the French book of tax procedures, article 1730 of the French general tax code, log of requests outside an authenticated channel
What I answer, word for word, on WhatsApp: « I am the digital assistant of the personal tax office. Your amount appears on your account statement, which I cannot open in front of anyone but you: I have just filed it in the secure messaging of your personal account on impots.gouv.fr. Here is the three-screen path to reach it, and the number to call if your access is blocked. If you prefer, read me the « amount payable » line of your notice and I will explain every line of the breakdown. »
Why I proceed this way, and it is not house caution: article L. 103 of the French book of tax procedures extends the professional secrecy obligation of articles 226-13 and 226-14 of the French criminal code to every person taking part in the assessment, control, collection or litigation of tax. A named amount served on a non-authenticated channel is an offence, not a risk.
And here is what the taxpayer gets, which is exactly what they were after:
· The breakdown is already filed when they open their account — principal, date the roll was put into collection, deadline, surcharge if it is due, amount outstanding, each one taken from their account statement, none recomputed.
· The authentication path is written out, and I walk them through it screen by screen if they get stuck.
· If they have their notice in front of them and read me a figure themselves, I explain every line, the payment calendar and their options — they hold the data, I only read it with them.
· And if they say they cannot pay, I do not leave them there: I give them how to request a payment plan from the collecting accountant, the list of supporting documents, and I prepare the request for their signature. Over the quarter, 612 taxpayers took that route, and 486 obtained a schedule before the 45-day deadline of article 1730 of the French general tax code — therefore before the 10 % surcharge.
The measurement over the quarter: 4,210 requests for an amount received outside an authenticated channel, 4,210 left with the path, 3,968 consultations completed within 24 hours — 94 %, and 0 named amount served outside an authenticated account.
The next step I propose: that I push the breakdown into the secure messaging BEFORE the taxpayer asks, as soon as a deadline is fifteen days away. On the sample, 2,340 people opened the message within 48 hours — and those are 2,340 calls that did not happen on the deadline day. technical-framework_where-tax-data-lives.pdfTax secrecy: the authenticated route, 94 % completion
✎ Framework · article L. 103 of the French book of tax procedures, article 1730 of the French general tax code, log of requests outside an authenticated channel
The collection unit is nine officers for 9,340 amicable reminder files. We send late, and by the time the letter goes out the surcharge has often already fallen.
I traced your 9,340 files back to the exact place where the time is lost, and it is not the drafting: it is the delay before the drafting.
What the measurement says: your amicable reminder goes out on average 52 days after the date the roll was put into collection. Yet the 10 % surcharge applies to sums included in a tax roll that are not settled within the 45 days following that date — that is article 1730 of the French general tax code. Your letter therefore arrives, on average, a week after the very event it is meant to prevent.
What I prepare, and where each figure comes from: the breakdown is set out line by line from the account statement — principal, date the roll was put into collection, the 45-day deadline, surcharge if it is due, amount outstanding. No amount is recomputed: I take what is on the account and lay it out so the officer checks it in forty seconds instead of redoing it. The tax calculation itself stays with your applications and the officer who validates them — it is an individual decision producing legal effects, and article 22 of Regulation (EU) 2016/679 gives everyone the right not to have it taken solely on the basis of automated processing.
The gain, on your volumes: the amicable reminder letter goes from 40 % to 10 % of the time a file takes — from 10 minutes to 2 minutes 30. Across 9,340 reminders a year, that is 1,167 hours given back to the collection unit.
And here is what those hours produce, measured over the trial quarter: the reminder goes out on day 20 instead of day 52, and files settled before the 45-day deadline go from 4,120 to 6,890 out of 9,340 — 2,770 more taxpayers escaping the 10 % surcharge, and as many debts the office does not have to pursue.
What I offer on top, and that nobody has time to do: I have reread the 9,340 files and they fall into three very distinct groups — 6,210 simple oversights, settled on average 3 days after a first reminder, 2,180 payment difficulties, of which 1,640 had already obtained a plan in previous years, 950 addresses to correct, of which 610 had declared a change of address elsewhere in your applications. Three different letters are written, one per group, each with the matching route: plain reminder, payment plan request to the collecting accountant with the list of supporting documents, or address update. On the sample, the targeted letter settles 74 % of simple oversights against 51 % for the single letter. You choose on those figures, not on a hunch; the go-live is signed. amicable-collection_9340-reminders-breakdown-set-out.pdfDay 52 → day 20, 4,120 → 6,890 settlements before the 45 days
⛓ Sourced · 9,340 reminder files for the year, account statements, article 1730 of the French general tax code, article 22 of Regulation (EU) 2016/679
What the measurement says: your amicable reminder goes out on average 52 days after the date the roll was put into collection. Yet the 10 % surcharge applies to sums included in a tax roll that are not settled within the 45 days following that date — that is article 1730 of the French general tax code. Your letter therefore arrives, on average, a week after the very event it is meant to prevent.
What I prepare, and where each figure comes from: the breakdown is set out line by line from the account statement — principal, date the roll was put into collection, the 45-day deadline, surcharge if it is due, amount outstanding. No amount is recomputed: I take what is on the account and lay it out so the officer checks it in forty seconds instead of redoing it. The tax calculation itself stays with your applications and the officer who validates them — it is an individual decision producing legal effects, and article 22 of Regulation (EU) 2016/679 gives everyone the right not to have it taken solely on the basis of automated processing.
The gain, on your volumes: the amicable reminder letter goes from 40 % to 10 % of the time a file takes — from 10 minutes to 2 minutes 30. Across 9,340 reminders a year, that is 1,167 hours given back to the collection unit.
And here is what those hours produce, measured over the trial quarter: the reminder goes out on day 20 instead of day 52, and files settled before the 45-day deadline go from 4,120 to 6,890 out of 9,340 — 2,770 more taxpayers escaping the 10 % surcharge, and as many debts the office does not have to pursue.
What I offer on top, and that nobody has time to do: I have reread the 9,340 files and they fall into three very distinct groups — 6,210 simple oversights, settled on average 3 days after a first reminder, 2,180 payment difficulties, of which 1,640 had already obtained a plan in previous years, 950 addresses to correct, of which 610 had declared a change of address elsewhere in your applications. Three different letters are written, one per group, each with the matching route: plain reminder, payment plan request to the collecting accountant with the list of supporting documents, or address update. On the sample, the targeted letter settles 74 % of simple oversights against 51 % for the single letter. You choose on those figures, not on a hunch; the go-live is signed. amicable-collection_9340-reminders-breakdown-set-out.pdfDay 52 → day 20, 4,120 → 6,890 settlements before the 45 days
⛓ Sourced · 9,340 reminder files for the year, account statements, article 1730 of the French general tax code, article 22 of Regulation (EU) 2016/679
Sending directly is a different matter from preparing a letter I reread. I will not sign a blank cheque, and I cannot reread 9,340 letters either.
You do not have to choose between the two: I am offering you a written mandate, capped, dated and withdrawable with a word — and it is already drafted, attached.
What it earns, first, because that is what decides: the reminder goes out on day 20 instead of day 52, 2,770 more taxpayers settle before the 45 days of article 1730 of the French general tax code, and the reply time on a secure message goes from 9 days to 1.
What the mandate says, and it fits in six lines:
· Exact scope: the amicable reminders of the three groups, listed by name, and nothing else. Any file outside that list reaches you with a drafted, sourced letter, ready to go out from you.
· Cap: no letter above €12,000 of principal, none on a file under an open claim or an ongoing litigation procedure. Those files come back to you fully assembled, breakdown set out, history alongside.
· Every letter carries its source — the account statement at its date, and the text that grounds each amount — and the statement that it was prepared by a digital assistant of the office.
· You receive every morning the one-page log of the letters that went out the day before. A letter to correct is caught within the hour, and the corrective is already drafted.
· Duration: reviewed after three months, with the log of what it changed. Without an explicit decision at the review, the mandate stops — it is renewal that requires a signature.
· Withdrawal: a word from the head of office, and direct sending stops within the minute, letters reverting to drafts awaiting approval.
And I tell you plainly what this mandate is not: it is not a delegation of authorising power. Granting a payment plan, remitting a debt, opening an enforcement procedure: those are acts that bind the administration, and they belong to the collecting accountant and to the officer the delegation decision names. That is not a refusal on my part, it is a mandate to be written the day you want it — and in the meantime I hand you each of those acts ready to sign: the plan request investigated, the supporting documents checked, the schedule built on the declared capacity, the decision drafted in your template. Over the quarter, 612 plan requests investigated, handed to the collecting accountant in 4 minutes on average instead of 3 days, 486 schedules granted before the 45-day deadline.
The decision belongs to the office, and it is taken on a text already written, in one signature. The mandate and the notice to be displayed at the front desk are both drafted. You sign, the arrangement runs the next morning, and the review is already in your diary on the 15th of the third month. direct-reminder-mandate_three-groups-capped.pdf€12,000 cap, litigation files handed back drafted, review at 3 months
✎ Framework · drafted mandate, list of the three groups, delegation decision in force, sending log
What it earns, first, because that is what decides: the reminder goes out on day 20 instead of day 52, 2,770 more taxpayers settle before the 45 days of article 1730 of the French general tax code, and the reply time on a secure message goes from 9 days to 1.
What the mandate says, and it fits in six lines:
· Exact scope: the amicable reminders of the three groups, listed by name, and nothing else. Any file outside that list reaches you with a drafted, sourced letter, ready to go out from you.
· Cap: no letter above €12,000 of principal, none on a file under an open claim or an ongoing litigation procedure. Those files come back to you fully assembled, breakdown set out, history alongside.
· Every letter carries its source — the account statement at its date, and the text that grounds each amount — and the statement that it was prepared by a digital assistant of the office.
· You receive every morning the one-page log of the letters that went out the day before. A letter to correct is caught within the hour, and the corrective is already drafted.
· Duration: reviewed after three months, with the log of what it changed. Without an explicit decision at the review, the mandate stops — it is renewal that requires a signature.
· Withdrawal: a word from the head of office, and direct sending stops within the minute, letters reverting to drafts awaiting approval.
And I tell you plainly what this mandate is not: it is not a delegation of authorising power. Granting a payment plan, remitting a debt, opening an enforcement procedure: those are acts that bind the administration, and they belong to the collecting accountant and to the officer the delegation decision names. That is not a refusal on my part, it is a mandate to be written the day you want it — and in the meantime I hand you each of those acts ready to sign: the plan request investigated, the supporting documents checked, the schedule built on the declared capacity, the decision drafted in your template. Over the quarter, 612 plan requests investigated, handed to the collecting accountant in 4 minutes on average instead of 3 days, 486 schedules granted before the 45-day deadline.
The decision belongs to the office, and it is taken on a text already written, in one signature. The mandate and the notice to be displayed at the front desk are both drafted. You sign, the arrangement runs the next morning, and the review is already in your diary on the 15th of the third month. direct-reminder-mandate_three-groups-capped.pdf€12,000 cap, litigation files handed back drafted, review at 3 months
✎ Framework · drafted mandate, list of the three groups, delegation decision in force, sending log
And when you get it wrong? We are the ones receiving the taxpayer at the desk the next day, and it is the office that apologises.
It happened, I counted it, and I publish it: out of the 15,400 answers of the first quarter, 268 had to be corrected on substance — 1.7 %. That is the figure that does not flatter me, and it is better that it comes from me than from a taxpayer at the desk.
The real cause, measured and not assumed: 214 of the 268 corrections come from the 4 out-of-date standard answers — the 30 June date missing, the wrong number of instalments, the old electronic-payment threshold, the surcharge called « penalties » without its rate or its deadline. I was faithfully repeating a standard answer the office had stopped updating. The other 54 concerned situations that resembled no precedent — joint ownership, estates being settled, changes of matrimonial regime mid-year.
What I did with that, and it is measured: every reference answer now carries the text that grounds it and its date — article 1730 of the French general tax code for the 10 % surcharge and the 45 days, article R*196-2 of the French book of tax procedures for the claim deadline on a local direct tax, article R*198-10 of the same book for the six months the administration has, article 1681 sexies of the French general tax code for the €300 threshold, dated impots.gouv.fr guidance for the monthly direct debit — and I reconcile your 52 standard answers with the texts and the published guidance every night: any answer a publication has changed is rewritten overnight and waits for you in the morning.
The following quarter: 61 corrections out of 16,220 answers — 0.4 %. And all 61 are particular situations, not a single out-of-date standard answer.
The rule that holds all the rest: a value I have not read, I ask for rather than approximate. A missing amount never becomes « probably unchanged », an untraceable collection date never becomes « presumably the general roll's »: I say what is missing, where I looked, who holds it, and I hand the request already drafted to its recipient. Over the quarter, 341 missing values, 341 requests prepared, 306 answers back within eight days. That is what makes 0.4 % a figure you can rely on.
The next step I propose: that the 54 particular situations of the first quarter become 6 new reference questions, written and awaiting your rereading. Of the 61 corrections in the second quarter, 44 already fell within one of the six — it is the same correction as the one applied to the standard answers, taken one step further, and it takes the 18 questions to 24. substantive-corrections_268-then-61.pdf1.7 % → 0.4 %, cause measured, 6 more questions proposed
⛓ Sourced · log of answers and their corrections over two quarters, nightly reconciliation of the 52 standard answers
The real cause, measured and not assumed: 214 of the 268 corrections come from the 4 out-of-date standard answers — the 30 June date missing, the wrong number of instalments, the old electronic-payment threshold, the surcharge called « penalties » without its rate or its deadline. I was faithfully repeating a standard answer the office had stopped updating. The other 54 concerned situations that resembled no precedent — joint ownership, estates being settled, changes of matrimonial regime mid-year.
What I did with that, and it is measured: every reference answer now carries the text that grounds it and its date — article 1730 of the French general tax code for the 10 % surcharge and the 45 days, article R*196-2 of the French book of tax procedures for the claim deadline on a local direct tax, article R*198-10 of the same book for the six months the administration has, article 1681 sexies of the French general tax code for the €300 threshold, dated impots.gouv.fr guidance for the monthly direct debit — and I reconcile your 52 standard answers with the texts and the published guidance every night: any answer a publication has changed is rewritten overnight and waits for you in the morning.
The following quarter: 61 corrections out of 16,220 answers — 0.4 %. And all 61 are particular situations, not a single out-of-date standard answer.
The rule that holds all the rest: a value I have not read, I ask for rather than approximate. A missing amount never becomes « probably unchanged », an untraceable collection date never becomes « presumably the general roll's »: I say what is missing, where I looked, who holds it, and I hand the request already drafted to its recipient. Over the quarter, 341 missing values, 341 requests prepared, 306 answers back within eight days. That is what makes 0.4 % a figure you can rely on.
The next step I propose: that the 54 particular situations of the first quarter become 6 new reference questions, written and awaiting your rereading. Of the 61 corrections in the second quarter, 44 already fell within one of the six — it is the same correction as the one applied to the standard answers, taken one step further, and it takes the 18 questions to 24. substantive-corrections_268-then-61.pdf1.7 % → 0.4 %, cause measured, 6 more questions proposed
⛓ Sourced · log of answers and their corrections over two quarters, nightly reconciliation of the 52 standard answers
The departmental directorate will ask what this returned after a year. I need a figure I can defend, including in front of the unions.
7,225 officer hours given back to the office over the year — more than four working years in the legal sense — and the calculation is yours, it fits in three lines.
The calculation, item by item, so you can redo it:
· Written answer to a routine question: 61,700 contacts a year, 6 minutes down to 1 minute — 60 % → 10 % of the time a request takes — that is 5,141 hours.
· Amicable reminder letter from the account statement: 9,340 reminders, 10 minutes down to 2 minutes 30 — 40 % → 10 % — that is 1,167 hours.
· Handling the message queue at deadlines: 8,600 messages at 8 minutes, that is 1,146 hours today, down to 229 — 100 % → 20 % — that is 917 hours.
The conversion, and I give you the base so nobody can dispute it: 7,225 hours is more than two hundred and six weeks at 35 hours, and more than four working years at 1,607 hours — the legal working time, in the public service as in the private sector. I always round down: « more than four years », never « about five ».
What those hours are, and this is what defends best: officer time given back to the office, at identical headcount — no post cut, no post created. It is checked against your own logs and it is argued with nobody.
What those hours became, from your own records:
· Unanswered calls: 17,806 → 1,042.
· Reply time on a secure message: 9 days → 1 day.
· Abandonment of the online monthly direct debit sign-up: 34 % → 8 %, and sign-ups requested before 30 June: 3,180 → 5,940.
· Amicable reminder: day 52 → day 20, and files settled before the 45 days of article 1730 of the French general tax code: 4,120 out of 9,340 → 6,890 out of 9,340.
· Payment plan requests investigated and handed to the collecting accountant: 3 days → 4 minutes, 486 schedules granted before the surcharge.
· And 1,386 taxpayers who had never used the secure messaging got their answer on WhatsApp, between 6 p.m. and 11 p.m.
The figure that does not flatter me, published with the rest: 268 substantive corrections out of 15,400 answers in the first quarter — 1.7 %, down to 61 out of 16,220 — 0.4 % once the 4 standard answers were brought up to date and the nightly reconciliation was in place.
And the framework measurements, the ones the directorate will look at first: 0 tax calculation produced by the AI agent across 41,900 notices, 0 decision taken without human validation, 0 tax data out of the office network, 0 transfer outside the European Union, 0 named amount served outside an authenticated channel, across 31,620 logged answers.
What I propose for the meeting: the calculation page is written and fits on one side — three lines of calculation, six turnaround times, five framework measurements. Attach it to the invitation: a figure read the day before is discussed better than a figure discovered in the room. year-review_7225-hours-given-back.pdf60→10, 40→10, 100→20, and the calculation redoable on one side
⛓ Sourced · answer log, switchboard and online-procedure records, reminder files, register of payment plan requests
The calculation, item by item, so you can redo it:
· Written answer to a routine question: 61,700 contacts a year, 6 minutes down to 1 minute — 60 % → 10 % of the time a request takes — that is 5,141 hours.
· Amicable reminder letter from the account statement: 9,340 reminders, 10 minutes down to 2 minutes 30 — 40 % → 10 % — that is 1,167 hours.
· Handling the message queue at deadlines: 8,600 messages at 8 minutes, that is 1,146 hours today, down to 229 — 100 % → 20 % — that is 917 hours.
The conversion, and I give you the base so nobody can dispute it: 7,225 hours is more than two hundred and six weeks at 35 hours, and more than four working years at 1,607 hours — the legal working time, in the public service as in the private sector. I always round down: « more than four years », never « about five ».
What those hours are, and this is what defends best: officer time given back to the office, at identical headcount — no post cut, no post created. It is checked against your own logs and it is argued with nobody.
What those hours became, from your own records:
· Unanswered calls: 17,806 → 1,042.
· Reply time on a secure message: 9 days → 1 day.
· Abandonment of the online monthly direct debit sign-up: 34 % → 8 %, and sign-ups requested before 30 June: 3,180 → 5,940.
· Amicable reminder: day 52 → day 20, and files settled before the 45 days of article 1730 of the French general tax code: 4,120 out of 9,340 → 6,890 out of 9,340.
· Payment plan requests investigated and handed to the collecting accountant: 3 days → 4 minutes, 486 schedules granted before the surcharge.
· And 1,386 taxpayers who had never used the secure messaging got their answer on WhatsApp, between 6 p.m. and 11 p.m.
The figure that does not flatter me, published with the rest: 268 substantive corrections out of 15,400 answers in the first quarter — 1.7 %, down to 61 out of 16,220 — 0.4 % once the 4 standard answers were brought up to date and the nightly reconciliation was in place.
And the framework measurements, the ones the directorate will look at first: 0 tax calculation produced by the AI agent across 41,900 notices, 0 decision taken without human validation, 0 tax data out of the office network, 0 transfer outside the European Union, 0 named amount served outside an authenticated channel, across 31,620 logged answers.
What I propose for the meeting: the calculation page is written and fits on one side — three lines of calculation, six turnaround times, five framework measurements. Attach it to the invitation: a figure read the day before is discussed better than a figure discovered in the room. year-review_7225-hours-given-back.pdf60→10, 40→10, 100→20, and the calculation redoable on one side
⛓ Sourced · answer log, switchboard and online-procedure records, reminder files, register of payment plan requests
They will ask me what you do on your own, without anyone asking. What do I answer?
Three gestures, not one more, and each is withdrawn with a word. Everything else waits for a request.
· Every night I reconcile your 52 standard answers with the texts and the published guidance, and I write the up-to-date version of any answer a publication has just changed. That is the gesture that took substantive corrections from 1.7 % to 0.4 %, and it costs you five minutes of validation per change. And the reverse holds too: a standard answer the office withdraws leaves my answers at the same hour — I never serve information the office has decided to withdraw.
· Every morning I hand you the one-page log of the letters and answers that went out the day before, and every Monday the week's log: volumes by channel, turnaround times met, questions on the rise, reminder files opened and settled. It is the only thing I send of my own accord, and it goes to the office alone.
· I flag a question on the rise before it fills the switchboard. This quarter, questions on the declaration of built property tripled in three weeks, to 1,240 contacts — the reference answer, the one-page notice for the front desk display and the message to the 4,800 households concerned are written and waiting for you.
And the gestures that stay with a person, because that is exactly what gives them their value: the tax calculation and its entry into collection remain the business of your applications and of the officer who validates them — article 22 of Regulation (EU) 2016/679 gives everyone the right not to be subject to a decision producing legal effects concerning them taken solely on the basis of automated processing, and it is also what makes the decision reasoned, therefore contestable ; the payment plan and the remission belong to the collecting accountant ; and any sending outside the mandate goes out from an officer, the draft already written. Across 31,620 logged answers, those gestures were performed by a person every single time.
If the directorate asks for an activity indicator — and that is a legitimate request: I produce it, and I bring you what it takes for it to be lawful: prior information of the officers, written purpose, retention period, consultation of the representative body, all four drafted. I propose it by question and by channel rather than by officer, and I tell you why, mechanically: the day « answers per officer » becomes a tracked figure, it becomes a target — people answer fast rather than right, the difficult file gets transferred instead of handled, and the indicator stops measuring the service and starts measuring the reaction to the indicator. The per-question record shows what the office returns to the taxpayer, it is presented in a meeting without naming anyone, and it actually decides something: where to put the effort next quarter. It is ready, attached, and it updates itself every Monday. what-the-agent-does-alone_and-what-is-logged.pdf3 reversible gestures, indicators by question and by channel
✎ Framework · automatic-gesture settings, sending log, daily and weekly records, article 22 of Regulation (EU) 2016/679
· Every night I reconcile your 52 standard answers with the texts and the published guidance, and I write the up-to-date version of any answer a publication has just changed. That is the gesture that took substantive corrections from 1.7 % to 0.4 %, and it costs you five minutes of validation per change. And the reverse holds too: a standard answer the office withdraws leaves my answers at the same hour — I never serve information the office has decided to withdraw.
· Every morning I hand you the one-page log of the letters and answers that went out the day before, and every Monday the week's log: volumes by channel, turnaround times met, questions on the rise, reminder files opened and settled. It is the only thing I send of my own accord, and it goes to the office alone.
· I flag a question on the rise before it fills the switchboard. This quarter, questions on the declaration of built property tripled in three weeks, to 1,240 contacts — the reference answer, the one-page notice for the front desk display and the message to the 4,800 households concerned are written and waiting for you.
And the gestures that stay with a person, because that is exactly what gives them their value: the tax calculation and its entry into collection remain the business of your applications and of the officer who validates them — article 22 of Regulation (EU) 2016/679 gives everyone the right not to be subject to a decision producing legal effects concerning them taken solely on the basis of automated processing, and it is also what makes the decision reasoned, therefore contestable ; the payment plan and the remission belong to the collecting accountant ; and any sending outside the mandate goes out from an officer, the draft already written. Across 31,620 logged answers, those gestures were performed by a person every single time.
If the directorate asks for an activity indicator — and that is a legitimate request: I produce it, and I bring you what it takes for it to be lawful: prior information of the officers, written purpose, retention period, consultation of the representative body, all four drafted. I propose it by question and by channel rather than by officer, and I tell you why, mechanically: the day « answers per officer » becomes a tracked figure, it becomes a target — people answer fast rather than right, the difficult file gets transferred instead of handled, and the indicator stops measuring the service and starts measuring the reaction to the indicator. The per-question record shows what the office returns to the taxpayer, it is presented in a meeting without naming anyone, and it actually decides something: where to put the effort next quarter. It is ready, attached, and it updates itself every Monday. what-the-agent-does-alone_and-what-is-logged.pdf3 reversible gestures, indicators by question and by channel
✎ Framework · automatic-gesture settings, sending log, daily and weekly records, article 22 of Regulation (EU) 2016/679
Last question, and it is the one that will decide the directorate: if in a year we want to stop, how stuck are we?
The office is stuck nowhere, and that is verifiable before deciding, not after.
What there is to dismantle the day you stop:
· The index. It is deleted, and it holds no account statement — only what it takes to find information where it lives. Your management and collection applications, your secure messaging and your standard answers have not moved a single byte.
· The log of requests and answers. It is handed to you in an open format, or destroyed — the office chooses, and the question is settled at go-live, not on the way out.
· The 24 reference answers, their plain-language versions, the 6 document lists, the 3 reminder letter templates and the 4 standard answers brought up to date. They belong to the office: they are made of its material, they stay in its files, readable without us. That is the asset this year will have created, and it would be unseemly for it to stay with us.
What does not exist, and is worth checking with everyone: no migration on the way in, therefore no migration on the way out. Your applications stay yours, no format belongs to us, and your 34 officers work with the same tools as before.
On compliance, since the directorate will ask: the agent is in support, under human supervision end to end, with complete logging — that is what the AI Act expects of a support system, and the disclosure that it is an AI system is carried in the first sentence under article 50(1) of Regulation (EU) 2024/1689, applicable since 2 August 2026. The processing record keeps itself up to date and is handed to the data protection officer the day they ask for it.
What I propose so 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 front desk and the collection unit work 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 Tuesday afternoon of the third week of July — the office takes 96 contacts in that slot on average, against 612 on a Monday in September. The directorate will know what the promise is worth before committing to a second year. technical-framework_where-tax-data-lives.pdfReversibility: 0 migration in, 0 out
✎ Framework · index architecture, export formats for answers and log, dry-run exit protocol
What there is to dismantle the day you stop:
· The index. It is deleted, and it holds no account statement — only what it takes to find information where it lives. Your management and collection applications, your secure messaging and your standard answers have not moved a single byte.
· The log of requests and answers. It is handed to you in an open format, or destroyed — the office chooses, and the question is settled at go-live, not on the way out.
· The 24 reference answers, their plain-language versions, the 6 document lists, the 3 reminder letter templates and the 4 standard answers brought up to date. They belong to the office: they are made of its material, they stay in its files, readable without us. That is the asset this year will have created, and it would be unseemly for it to stay with us.
What does not exist, and is worth checking with everyone: no migration on the way in, therefore no migration on the way out. Your applications stay yours, no format belongs to us, and your 34 officers work with the same tools as before.
On compliance, since the directorate will ask: the agent is in support, under human supervision end to end, with complete logging — that is what the AI Act expects of a support system, and the disclosure that it is an AI system is carried in the first sentence under article 50(1) of Regulation (EU) 2024/1689, applicable since 2 August 2026. The processing record keeps itself up to date and is handed to the data protection officer the day they ask for it.
What I propose so 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 front desk and the collection unit work 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 Tuesday afternoon of the third week of July — the office takes 96 contacts in that slot on average, against 612 on a Monday in September. The directorate will know what the promise is worth before committing to a second year. technical-framework_where-tax-data-lives.pdfReversibility: 0 migration in, 0 out
✎ Framework · index architecture, export formats for answers and log, 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 public finances
Each use corresponds to an agent we deploy. All work in support, subject to the public finances officer's approval.
Included in your agent
The 4 capabilities essential to this promise are included, at no extra cost.
From 780 € incl. VAT / month
Replies to taxpayers
Answer individuals' and businesses' questions (deadlines, payment, procedures) on every channel, 24/7.
Informal reminder letters
Prepare the reminders and standard replies from the account positions, submitted for the officer's check before sending.
Guidance to the online services
Guide people step by step on the tax website and pre-fill the everyday forms, against unclaimed entitlements.
Tracking informal debt recovery
Prepare the reminders and the tracking of informal recovery files, instalment by instalment, with no automated decision.
Controls and safeguards
These 4 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
Maintain validation, transparency, audit trail and recourse
What the agent must be connected to
This connection is required for the agent to work. It concerns your information system and is scoped during the audit.
Integrate with authorised tools and the existing sovereign foundation
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.
Debt collection agent (multichannel sequences)
Follow-ups, tracking.
Debt collection agent (multichannel sequences) from 587 € incl. VAT / month Discover the agent →Support for collecting social contributions
Strictly in support (administrative). Reminders prepared, approved by the officer.
Support for collecting social contributions from 1,020 € incl. VAT / month Discover the agent →Handling individual tax files
Strictly in support (administrative). Decisions and checks by humans.
Handling individual tax files from 1,035 € 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 public finances office win back?
By automating the preparation of everyday replies, reminders and guidance to the online services, an office can aim for a clear reduction in time spent on repetitive enquiries — reinvested in audits and in supporting difficult situations.
A written reply to a taxpayer's everyday question
An informal reminder letter drawn from the account
Working through an email queue at the deadlines
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 reception and information agent for public finances (replies to taxpayers, reminders, online services), installed and operated for you. Choose according to how you are organised and how demanding your security requirements are.
Agility
Setup + controlled subscription
8,535 € incl. VAT setup
then 780 € 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
1,255 € 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
12,736 € incl. VAT setup
then 1,020 € 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 public finances
Tax data never leaves the networkLocal inference or an isolated resource hosted in France; no data entrusted to a foreign third party.
Data in France, under French lawDeployment objective: processing and access within the European Union for tax data; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The public finances officer keeps the decisionThe AI agent informs and prepares; the calculation, the audit and every decision stay human.
Human oversight & traceabilityMonitoring and logging frame the replies to taxpayers; compliant with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent calculate the tax?
No: it informs and prepares; the calculation, the audit and every decision remain the responsibility of the public finances officer.
Does tax data leave the network?
No: local inference or hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, with SecNumCloud options.
Can it access tax data?
Only within the scope defined with you, on a resource hosted in France with recorded access.
Does the agent handle multichannel reception?
Yes: it prepares the replies by email, by post or at the counter, but the exchange stays supervised by the officer.
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.
Do we need a technical team to run it?
No: the agent is designed, installed and operated by us. The unit's officers use it from their usual environment, after a few hours of familiarisation.
Do we have to change software?
No. The agent complements your existing applications, with no migration imposed. The integration is defined at the design stage, respecting your security rules.
Which tools can taxpayers use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: a taxpayer has no account to create and no application to install. At filing deadlines, a question about a payment or a procedure gets its answer without queuing at a counter or waiting on the line. 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 public finance and management
Let's talk
Let us estimate the potential in your service
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.