AI agent for the pensions adviser
An assistant that reconstructs career records, prepares the estimates and drafts the replies to members — the adviser checks and decides every award. Hosted in France, on a resource dedicated to and isolated for each organisation, members' career data never leaves the European Union. The AI agent assists, the adviser decides.
Updated on
I am preparing the correction request to the scheme concerned.
⛓ Source · the documents in the file + the cross-scheme career record
Reply letter to the member ready for review.
✎ Action · indicative estimate ready — the adviser approves
In a pension fund, a Blue Lemon Agent agent assists the adviser with the repetitive tasks — reconstructing career records, indicative estimates, pre-assessment of claims, letters and notifications — and flags missing or inconsistent periods. It runs on local inference or is hosted in France on a resource dedicated to and isolated for each organisation: members' career data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. The adviser checks and decides every award; the time won back goes to supporting members. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and your members reach it on WhatsApp Business, the website chat or email — with no account to create and nothing to install. Reaching the administration from the tool people already have means less non-take-up of rights and equal access to the service. These connections are included in every plan, at no extra cost, within the number of connections your level includes.
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.
Why AI matters to pension funds — and why they hesitate
Reconstructing a career, checking quarters, answering members anxious about their future pension: the adviser's work is meticulous and time-consuming, on data among the most sensitive there is — career, income, family situation.
! The issue
The pensions adviser is caught between members who expect quick answers about their future pension and increasingly complex files: fragmented careers, people with several schemes, multiple regimes. Yet most consumer AI tools would amount to entrusting members' career records, income and family situations to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For social data, AI is only of interest if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, systematic human oversight, calculation and award reserved to the adviser: the time saved on pre-assessment is never paid for in lost confidentiality. The aim is not to replace the pensions adviser, but to give them back time for supporting members.
Protecting claimants' data: sovereignty & compliance
A pension fund handles a whole lifetime's career and income. Here is how the architecture of our agents protects it, organisation by organisation.
Local inference
The agent can run on a machine at the organisation: no data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — claimants' data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
For members' data, the architecture aims to reduce exposure to the Cloud Act and FISA 702; being located in France or in the European Union does not, on its own, guarantee immunity.
One isolated resource per organisation
No pooling of data: an environment strictly dedicated to your fund, guaranteeing the continuity of the public service.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
The agent is strictly in support; no award 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.
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 bodyCaisse Ravenel — fictional pension fund, base and supplementary schemes of an occupational regime
- Activity
- Social protection — awarding and paying base and supplementary pensions for an occupational regime
- Headcount
- 310 staff, of whom 74 in the pensions department: 48 pensions advisers, 12 career-record officers, 8 on telephone reception, 4 managers, 2 regulatory specialists
- Public served
- 212,000 contributing members and 96,000 pensioners — 38% of them belong to several schemes
- Volume
- 9,400 pension claims a year, 41,000 calls from members, 1,260 incomplete career records to reconstruct, 6,400 reply letters
- Tools in place
- Career-record application, inter-scheme register, document management system with 12 years of archives, outgoing mail tool, appointment diary — the agent connects to them, nothing is replaced
- Who decides
- The pensions adviser checks and awards every file; the regulatory specialist settles the individual hard cases; the head of the department sets priorities and deadlines
- Room for improvement
- Reconstructing one incomplete career record takes a full day of an officer's time; a correction request to another scheme comes back in 47 days; 1 pension claim in 6 is put on hold for want of a document
Caisse Ravenel is an organisation invented for this demonstration. It lacks neither competence nor method: it lacks officer hours, and those hours go into hunting for documents and reformatting, not into advising members. The agent runs on local inference on a machine at the fund and connects to the career-record application, the inter-scheme register, the document management system, the mail tool and the diary: it reconstructs, it pre-assesses, it drafts, it flags what must be checked — the adviser checks and awards. The exchanges below cover one year of service, from taking up the backlog to the annual review.
This public body, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real service.
A career record is the list of periods worked and the quarters they open, scheme by scheme. One missing quarter can push back a retirement date by a whole year.
The gap I measured, and it settles everything else: of your 9,400 annual pension claims, those whose career record is complete on arrival are notified in 39 days; those with a single period to reconstruct take 141 days — 102 days of difference for one period. It is not the complexity of the calculation that lengthens a file, it is the search for a period — and 38% of your members belong to several schemes, meaning they contributed to more than one over their working life.
What I did without being asked: I cross-checked your career-record application, the inter-scheme register and your 12 years of document management, and I matched 41,200 documents — payslips, certificates, contracts — to the periods they support. I recognise them by their content, never by their file name: a payslip deposited in 2019 is often called scan_0043.pdf.
What the review gives, reason by reason: 891 files turned on a period to reconstruct, 263 on a document never received, 175 on a discrepancy between two schemes, 131 on an individual hard case. The 890 I can unblock are sorted into 9 families of situation — the largest, 412 files, is reviewed on a single criterion: whether the employer and the period match.
What that changes for an officer: reconstructing an incomplete career record took a full day, that is 7 hours; it takes 1 hr 25. Across 1,260 incomplete career records a year, that is 7,000 officer hours given back to the service — and not one post fewer, not one post more: those hours go back to the members waiting for an appointment.
What I propose, and it is half a day for 890 files unblocked: that your two regulatory specialists review the 9 families rather than the 890 files, and approve them family by family. I then put every new file through the same sieve as it arrives, and the 131 hard cases go straight to the specialist's desk, documents gathered and the precise question already written. files-on-hold_what-the-review-reveals.pdf890 files unblocked, sorted into 9 families of situation
⛓ Sourced · career-record application, inter-scheme register, 12 years of document management, 1,460 files on hold
What is established: 168 quarters across the base and supplementary schemes, 4 different schemes over the career. Every quarter points to its source — a scheme's record, a payslip in the file, an employer's certificate — and a quarter without a source is flagged as such and never counted.
The three anomalies, and what I did with them:
· 1999-2000 — activity declared, not recorded. The 18 payslips were already in the file, deposited in 2019 with an earlier claim and never used. The supporting document exists: this is a correction, not an investigation, and the request is written.
· 2001 — national service. The record of service is missing. The letter to the member is ready, with the exact name of the document, the address where to obtain it and the time to allow. It is the only document missing from 97 files in the same family — approve the template once and it goes out to the other 96 the same day.
· 1998, two quarters — a discrepancy between two schemes. One records 2 quarters, the other 1. I have settled it on the documents: the December 1998 payslip is in the file and it covers the disputed quarter, which proves the scheme recording two right. The correction line is written and quantified — one quarter more and €9 a month of pension — and it enters the record when the adviser signs.
What it is worth to the member, and it is the only measure that counts for her: those 8 quarters, once validated, bring forward her full-rate retirement date by two years and add €74 a month to her pension, for life. A badly reconstructed career record is not an administrative error: it is two years of someone's life.
What I have already written and what awaits a signature: the correction request to the scheme concerned, in your template, with the 18 payslips attached in chronological order and numbered — that order is what saves the other fund three weeks — and the letter telling the member where her file stands and what she does not have to do.
What I propose: that the adviser review the three anomalies — eight minutes — and approve the despatch. I chase the other scheme at 30 days without anyone having to remember, and I place the member in the queue for a support appointment the moment the quarters land: that is when she has a decision to make. career-record-reconstruction_1998-2001.pdf168 quarters, 3 anomalies dealt with, every line with its source
⛓ Sourced · inter-scheme career record, 18 payslips from the document system, records of the 4 schemes concerned
Local inference means the model computes on your machine: a member's name, income and career periods cross no outside network to be processed. And if you would rather not host a machine, the second route is just as watertight: an isolated resource hosted in France, dedicated to your organisation — no pooling with another fund, in memory included, and that is the condition your supervisory authorities look at first.
What that gives, line by line:
· Hosted in France, under French law, and zero data outside the European Union — architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, including against a US player hosting in Europe.
· Nothing I read trains any model, neither yours nor a third party's. What I learn from a member serves that member.
· Encryption in transit and at rest, strong authentication for the 74 officers of the department, and role-based access — rights follow the job: a reception officer sees how far a file has progressed, not the income figures; a career-record officer sees the periods in their portfolio; the regulatory specialist sees everything.
· A full log: who consulted what, when, and what was approved.
The figure that makes this concrete rather than theoretical: you receive 62 requests to exercise data rights a year, and each took half a day of officer time to find what the fund holds about a person across five systems. It now takes 20 minutes, and above all the answer is complete: the index knows everywhere the member appears, where a manual search depended on the memory of whoever was searching. An obligation you used to endure becomes an obligation you meet.
And you change no tool: I connect to your career-record application, your inter-scheme register, your document system, your mail tool and your diary. No migration, no extra software for 74 people to learn.
What I propose: that I keep up to date the processing record your data protection officer presents each year — sources, retention periods, who accesses what. It used to take three days to rebuild once a year; it will be accurate at all times, and the day a supervisory authority or a member asks for it, you print it. technical-framework_where-career-records-live.pdfLocal inference, role-based access, processing in the EU targeted
✎ Framework · deployment architecture, access log, register of data-rights requests
What it contains:
· The estimated amount from the base scheme and the one from the supplementary scheme, kept separate — a member who sees a single figure always takes it for the total.
· The conditions used: number of quarters taken into account, retirement age envisaged, application of a reduction or an increment — a decrease if the member retires without meeting the full-rate conditions, an increase if they continue beyond.
· The list of periods still to be consolidated, with their quantified effect. On this file: the 8 quarters from 1999-2001, correction requested on 14 March, chaser scheduled for 14 April, €74 a month at stake. These are not caveats, they are euros still to be claimed — and the request that claims them is already written.
· The statement I put at the top, in the same size as the amount, because it serves the member: indicative estimate — it is the fund's official statement that is authoritative, and on this file the only periods still open can only raise it. Written in the body of the letter and not in a footnote — that is the only way it gets read.
The time this shifts: writing the reply letter weighed 40% of handling a request; it weighs 10% — 35 minutes down to 9, and those 9 minutes are the adviser's review.
What it changes for the member: three weeks of waiting become 48 hours. A member who receives her estimate within two days does not call three times to find out where it is — and your 8 reception officers get back the calls they no longer have to take.
And the award carries the adviser's signature, which is a strength: a decision producing legal effects for a person is not taken on the sole basis of automated processing. That is what gives the member a dated, reasoned decision open to challenge — a pension is paid for life, it deserves a name at the foot of the page. What I do is make that name go down in ten minutes instead of two hours, on a file that is already complete.
What I propose: that this format go out across your 6,400 letters of the year as soon as your specialists have reviewed it. I run it first on 200 files, I count the member call-backs at 15 days, and I give you the figure before you commit the rest. indicative-estimate_line-by-line.pdfBase and supplementary kept separate, €74 a month still to be claimed
⛓ Sourced · consolidated career record, records of the 4 schemes, the fund's calculation parameters
The four, for this file:
· Retirement at 63, as things stand — without the 8 quarters under correction. Estimated amount, applicable reduction, first payment the month after the claim.
· Retirement at 63, correction obtained — +€74 a month, on the same date. That is the comparison that decides, and it existed nowhere.
· Retirement at the full rate — the date on which the conditions are met, the amount without reduction, and the number of months of waiting, written in plain figures.
· Continuing to work one year beyond — the effect of the increment, the estimated amount, and what one further year of contributions is worth.
What I add and what simulators forget: the ten-year cumulative for each scenario. Retiring later pays more each month and starts later: the two curves cross, and it is that crossing point that answers the question actually being asked. On this file, it falls at 11 years and 4 months — an arithmetical fact, set out and dated, that the member would obtain nowhere else.
What separates the scenarios, and what remains to be confirmed by document: a single document moves three scenarios out of four — the record of military service. The request that claims it is written, with the exact name of the document and where to obtain it, and until it comes back the four amounts can move in one direction only: upwards.
Who does what, and it is what makes the whole thing solid: the choice belongs to the member — her health, her partner's situation and her plans appear on no career record; the award belongs to the adviser, who makes it on a comparison already done. I give both of them back the time of the calculation, and I take no decision away from either.
What I propose: that the comparison page go with every estimate request, without anyone asking for it. Over the trial quarter it cut member call-backs after receiving an estimate by 31% — the second question was always the same, and it is now answered in advance. retirement-scenarios_4-assumptions.pdfMonthly amount, ten-year cumulative, crossing at 11 years and 4 months
⛓ Sourced · consolidated career record, the fund's calculation parameters, log of member call-backs
How I know: I matched your 6,400 outgoing letters against the calls received in the following 15 days. Five phrases come up systematically before a call, and I rank them by what they cost: « periods not validated to date » (1,340 calls), « it is for you to provide » (1,020), « your file is undergoing further examination » (780), « subject to the information in our possession » (610), « where applicable » (350).
What I rewrote, in your templates and without changing a word of their legal reach:
· « Periods not validated to date » becomes « Three periods are not yet validated: 1999, 2000 and 2001. We wrote to the scheme concerned on 14 March and we will chase on 14 April. There is nothing for you to do. » That last sentence is the one that removes the call.
· « It is for you to provide » becomes « One document is missing: your record of military service. You obtain it here, and this is how. »
· « Where applicable » disappears. It means nothing to the person reading and protects nothing for the fund writing.
What that produced over the trial quarter: 4,100 annual calls down to 1,500, that is 2,600 fewer calls. At 6 minutes a call, that is 260 hours given back to telephone reception — and above all 2,600 members who did not need to call to understand a letter.
What I leave untouched, and it is what makes the rewriting unassailable: the mandatory statements, the appeal channels and time limits, the wordings your regulatory specialists have settled. I rewrite what informs; what binds keeps the exact word they chose — that is what protects the member's right of appeal. The 14 templates went through their hands before go-live, and it is their approval that stands over the 6,400 letters of the year.
What I propose: an annual review of the templates based on the calls received, and in the meantime I bring you the next three wordings in order of cost — I already hold them, together they weigh 900 calls a year, and they are rewritten within the same hour of review. letter-templates_5-wordings-that-generate-calls.pdf4,100 annual calls down to 1,500
⛓ Sourced · 6,400 outgoing letters, log of calls received within 15 days, the fund's 14 templates
Why it happened: your advisers asked for documents from a general list, the same for everyone. I rebuilt the real list by situation from your 9,400 files of last year — not what the procedure provides for, but what actually made it possible to award without reopening.
What that gives, situation by situation:
· Single-scheme career, no unusual period — 2 documents instead of 8. That is 41% of your files, and they all went out with the list of eight.
· Members in several schemes, 38% of the total — 2 documents: the records of the other schemes, which I fetch myself from the inter-scheme register rather than asking for them. A document the fund can obtain on its own is not asked of a 62-year-old, and that rule produced most of the gain.
· Careers including assimilated periods — periods not worked but counted, such as compensated unemployment or sickness — 1 document, and it is nearly always already in your document system.
· Periods completed abroad, 5% of files — documents varying by country, and immediate referral to the regulatory specialist. It is the family I got wrong most often — 9% of files reworked — and it is fixed: I now refer to the specialist straight away, in 20 minutes, and the rate fell to 2.8% over the last quarter. I give you the full figures in the review tab.
What it produced: 1 file in 6 on hold became 1 in 28 — 1,567 files put on hold a year became 336. Every hold cost the member 24 days and the fund a second letter: 1,231 chasers avoided, and 1,231 members who did not have to wonder why their file was going nowhere.
The time this shifts: pre-assessing a file weighed 60% of the handling; it weighs 10% — 1 hr 40 down to 17 minutes. Across 9,400 files it took 15,600 hours a year; it takes 2,700.
What I propose: that I push the list by situation straight into your online claim form. The documents then arrive with the claim and the first hold no longer exists — of the 336 that remain, I put 210 in that category. The head of the department decides, and the configuration takes a day. documents-expected_by-member-situation.pdfWhat actually made it possible to award without reopening
⛓ Sourced · 9,400 files of the year, reasons for putting files on hold, inter-scheme register, document system
What I check and what reduces to one line when all is well: identity and civil status · consistency between declared periods and recorded periods · consistency of recorded earnings from one year to the next · documents required by the situation · quality and legibility of the documents supplied · conditions for entitlement · absence of double counting between schemes.
What I bring to the top, and it is the only point of the checklist: the 3 discrepancies in this file.
· An earnings figure recorded in 2007, ten times the neighbouring years. Probable keying error, one zero out — I have written the corrected figure, tested it against the eight neighbouring years and against the 2007 payslip in the file, and I set it beside the recorded figure. It enters the record when the adviser signs it: that signature is what makes the correction dated and open to challenge by the member, and that is what makes it solid.
· Two quarters counted by two different schemes for the same period. Double counting is the error that costs the most to repair once the pension has been awarded; both records are set side by side, with the one the document supports.
· An illegible document — a 1994 certificate scanned crooked. The request letter is written, stating what is illegible and why, so the member does not send the same document twice.
What the checklist says about itself, and it is what makes it readable: of the 22 points, 19 carry a certain result; 3 carry their degree of certainty, their source, the hypothesis I retain and what would settle it. A check that shows where it is sure is a check people keep reading — and that is what brings the adviser's decision from two hours down to ten minutes, on a file where the documents are to hand.
What I propose: that the checklist be reviewed with your two specialists at every change to your operating rules, and that the date of that review appear on every file. A check with no date is a check nobody knows is current — and I will flag of my own accord the day a point of the checklist has caught nothing in twelve months: it will have stopped earning its place. pre-assessment-checklist_22-points.pdf19 certain points, 3 set out with the hypothesis retained
⛓ Sourced · the fund's operating rules, inter-scheme records, document system, documents in the file
What I found in your 2,100 correction requests of last year: 38% came back with a request for further information, which added 21 days on average. The delay was not imposed on you: it was produced by our own incomplete despatches.
What I now do before any despatch:
· I build the request in the template the receiving scheme expects — your 7 corresponding schemes expect neither the same documents nor the same order, and that was written down nowhere until I rebuilt it from their own replies.
· I attach the documents in chronological order, numbered, with a schedule. On the test despatches, that point alone cut requests for further information from 38% to 9%.
· I chase at 30 days with the history of the despatch, without an officer having to remember. Your chasers used to go out when someone thought of it.
And here is what binds the fund, so here is what I do under mandate — this is not a refusal, it is a mandate, and it is short:
· Scope: the 3 correction request templates approved by your regulatory specialists, to the 7 corresponding schemes, to the exclusion of any other letter.
· Condition: a request goes out only if the supporting document is in the file and the pre-assessment checklist is green on the points concerned.
· No financial commitment — a correction request pays nothing and promises nothing.
· Duration: 6 months, renewable. Withdrawal: on a word, immediate effect, and the requests already sent stay on the record.
· What stays with the adviser: any letter to a member carrying a decision, any request outside the templates, any file flagged by the checklist.
What the mandate produced: average time down from 47 days to 12, and 0 requests sent without a supporting document over the year. For the member, 35 days of administrative life disappear — and they will never know those days existed.
What I propose next: extending the same mandate to the 2 record requests still made by hand with the two schemes that have no written template. I have drafted their form from their last 40 replies; it remains for you to put it to them, and the 7 schemes become 9. inter-scheme-corrections_47-days-to-12.pdfDespatch mandate, 3 templates, 7 schemes, chaser at 30 days
⛓ Sourced · 2,100 correction requests of the year, reasons for requests for further information, templates of the 7 corresponding schemes
The five, in order: « where is my file? » (9,100) · « which documents must I supply? » (6,200) · « at what age can I retire? » (4,300) · « did my career record reach you? » (3,400) · « how do I book an appointment? » (2,400).
What I answer, and where it comes from: the real status of the file in your application, the list of documents matching the member's exact situation, the general retirement conditions as your specialists have written them, and the date the record was sent with the means of retrieving it. Every answer cites its source, and a member who challenges it gets the origin of what they were told.
And, for any pension amount, the written route every time — the estimate goes out within 48 hours, as a document, reviewed by an adviser: a pension figure can be kept, re-read and challenged, which a spoken exchange never allows.
What it changes for the member: those 25,400 requests get an answer around the clock, including on Saturdays, when you were receiving 2,900 calls with nobody to take them. Average waiting time on those five questions falls from 6 minutes 40 to under a minute.
What it changes for your 8 officers: they take the remaining 15,600 calls, the ones that need a person — an unusual situation, a member in difficulty, a disagreement. They spend 11 minutes on them on average instead of 6, and that is the real gain: what counts is not the number of calls handled, it is the time you can give to the person who needs it.
What I do from the very first sentence, and it is not a configuration option: I state that I am an artificial intelligence. Since 2 August 2026, article 50 §1 of the European regulation on artificial intelligence requires that any person interacting with an AI system be informed, unless this is obvious. The member can ask for an officer at any moment, and that request holds for the rest of the file: they never have to repeat it.
What that produced: out of 25,400 exchanges, 640 members asked for an officer, that is 2.5%, and all were transferred with what had already been gathered. Announcing the machine does not drive people away: it is waiting that drives people away, and the waiting has gone.
What I propose next: opening the sixth question, « why has my amount changed? », which weighs 1,900 calls a year and which your officers answer today by recalling the file from memory. I have written the standard answer and run it over 300 real cases; it fits in five lines and cites the line of the record that moved. Your specialists review it, and it opens next month. member-information_5-questions-62-per-cent.pdf25,400 requests answered around the clock, 2.5% transferred
⛓ Sourced · 41,000 calls of the year, reasons logged at reception, career-record application, register of requests for a human officer
What I found in your diary: 5,800 appointment requests a year for 6,400 slots opened. You had more slots than requests and a 26-day wait — because 1,100 slots were reserved for situations that did not arise, and 870 were lost to members who did not attend.
What I do:
· I qualify the request on arrival — situation, schemes concerned, retirement date envisaged — and direct it to the right kind of appointment. 1,200 requests a year needed no more than a written answer: they no longer take a slot.
· I offer three slots in the diary of the competent adviser, and I book the appointment. It can be moved or cancelled on a word, by the member as much as by the reception officer, and the freed slot goes straight back into the offer.
· I prepare the file before the interview: career record reconstructed, points to check, retirement scenarios. The adviser no longer searches during the appointment — and that is what members notice first.
· I remind the member 48 hours ahead, with the list of what to bring. The 870 missed appointments fell to 210.
What that gives: waiting time from 26 days to 9, 660 slots recovered, and 2,900 additional support appointments over the year — made possible by the hours that reconstruction and pre-assessment freed up. That is where the 22,700 hours went: into appointments, not into a dashboard.
The order of the queue now, because that is where a waiting list is decided: it follows the date of the request, and the only priorities are the ones the fund has written down — a retirement date within six months, a situation flagged by a social worker. Neither the future pension amount, nor seniority, nor insistence enters the calculation: equality of treatment is the sorting rule, it is written into the configuration and it can be read line by line.
What I propose next: to open the 660 slots recovered first to the 580 members whose retirement date falls within six months and who have never asked for an appointment. I have identified them, the list is ready, the invitation is written in your templates — you approve, it goes out tomorrow. A right opened at the right amount often begins with an appointment nobody thought to offer. support-appointments_26-days-to-9.pdf660 slots recovered, 2,900 additional appointments
⛓ Sourced · appointment diary for the year, reasons for requests, attendance rate, the fund's priority rules
What I spot, and these are facts about a file: a file left unanswered after two letters · a document returned three times and still not compliant · a call whose question bears no relation to the letter received · a claim filed by a third party.
What I then do:
· I offer an appointment, by telephone or in person, and book it into a supported slot. Over the year, 340 members were directed this way and 296 files were completed within two months — where 41% of such files used to stay stuck for more than six months.
· I pass to the adviser what has already been gathered, so the member does not have to start their story again.
· I flag the letters at fault to your accessibility officer: across the 340 files, 3 templates come up 6 times out of 10. It was not the person who could not cope, it was the letter that was not readable — and that correction benefits everyone, including those who never call.
What it is worth to the fund: those 340 files represented 1,100 repeat contacts a year — calls, letters, back and forth. They now represent 380. The service improves and the public money goes elsewhere in the same movement, which does not happen so often.
What I go on, and nothing else: the facts of the file above. Verifiable, dated and challengeable facts — that is what makes the detection fair, and that is why I hold to them. A member who asks to have no further dealings with an automated system is transferred to an officer, and stays there without having to ask again.
What I propose: that the same detection run across the 1,260 incomplete career records of the year, where the sequence is identical — two letters, a non-compliant document, a file that stops. Of the 340 files already handled this way, 296 were completed within two months, that is 87%: it is the same move, applied where it is still missing. You set the trigger threshold, I bring you the list every Monday.
⛓ Sourced · files unanswered after two letters, log of non-compliant documents, diary of supported slots
For members first, because that is what you must be able to show:
· Average time to notify an award: 118 days → 61 days. On files with a career record to reconstruct, 141 days → 68.
· Correction with another scheme: 47 days → 12.
· Files put on hold for want of a document: 1,567 a year → 336.
· Support appointments obtained in 9 days instead of 26, and 2,900 more of them.
· Estimate received in 48 hours instead of three weeks.
For staff: reconstructing an incomplete career record 100% → 20% of the time, that is 7 hours down to 1 hr 25, 7,000 hours over the year · pre-assessing a file 60% → 10%, that is 1 hr 40 down to 17 minutes, 12,900 hours · writing a letter 40% → 10%, that is 35 minutes down to 9, 2,800 hours. End to end: 22,700 officer hours, and they went into appointments and difficult files — not one of them went into processing more claims, and not one post moved.
The measure that does not flatter me, and I would rather give it to you myself: of 9,400 files pre-assessed, 846 had to be reworked by an adviser because my pre-assessment was wrong or incomplete — 9%.
The cause, precisely: 516 of those 846 fall into just two families. Careers including periods completed abroad, where the rules differ according to the agreements applicable to the country concerned — I was applying them from an internal note of 2019 that was no longer current; and apprenticeship periods before 1990, for which your document system holds only a scan in which I misread two mentions out of three. It was not a reasoning problem, it was a source problem — and an out-of-date source produces a wrong answer with the same confidence as a current one.
What I did about it, and the measured result: on those two families I conclude differently. I gather the documents, I write out the precise question to be settled with both hypotheses quantified, and I refer the file to the regulatory specialist straight away, in 20 minutes instead of three days. Over the last quarter, the rework rate fell from 9% to 2.8%. And every case a specialist settles, I record and apply to the next: the second career completed in the same country no longer starts from nothing.
The scope now open, so your supervisory authority can read it at a glance: the seven uses ready to deploy for this role are all live — reconstructing career records · indicative estimates · retirement scenarios · pre-assessment of claims · notifications and document requests · information for members · appointment booking. None of them required changing a piece of software, and each went live once your specialists had reviewed what it produced, not before.
What I propose: that the internal note on periods completed abroad be dated and revised once a year, and that its date appear on every file where it applies. I will remind you twelve months to the day after the revision: it is the only reminder I ask of you all year. annual-review_22700-hours-and-9-per-cent.pdfWhat is given back to members, what was reworked, what was corrected
⛓ Sourced · 9,400 files of the year, notification times, rework log, appointment diary, department time records
The three automatic actions, and each undoes itself as fast as it is done:
· I acknowledge a claim within the hour, with what is already in the file and what is missing. And the reverse holds too: if the member replies that they want a person, the claim moves to human handling and stays there.
· I book an appointment in a slot the member chose from three. It can be moved or cancelled on a word, by them as much as by the reception officer, and the freed slot goes straight back into the offer.
· I send a correction request in the three approved templates, under the mandate you signed, only when the supporting document is in the file. The mandate is withdrawn on a word, with immediate effect.
The six acts that carry a public officer's signature, and that is a strength: award a pension · notify a decision to a member · reject a claim · correct recorded earnings · validate a quarter · settle an individual hard case. On all six I arrive with the file complete, quantified and reasoned: the act that remains is the signature, and it carries a name and a time.
The basis, and it fits in one sentence: a decision producing legal effects for a person is not taken on the sole basis of automated processing. This is not a setting someone could loosen one day: it is the rule, and it is what makes every right opened challengeable by the person it is opened for. Result over the year: 0 decisions taken without human validation across 9,400 files.
What that gives you facing a supervisory authority or a complaint: every file keeps what I read, the source of every quarter, what I flagged, what I proposed and what was decided, by whom and when. A complaint about a career record used to take three days of searching; it takes twenty minutes, with the documents to hand.
What I propose: a review of this list at every review of your operating rules, and I open it with the usage count of each action. An automatic action that has served no purpose in twelve months should come off the list — I will flag it before your supervisory authority asks. who-decides-what_three-actions-six-reserved-acts.pdfThree reversible actions, six signed acts, 0 automated awards
✎ Framework · list of automatic actions approved by the head of the department, despatch mandate, approval log for the year
The conditions to be met before any individual indicator: inform the staff concerned beforehand, consult the staff representatives before implementation, and ensure the indicator is proportionate to what is sought. With 310 staff, you are bound by this without discussion. The file is ready: definition of the indicator, information note to staff, material for the consultation. Management decides, and I carry out whichever decision it takes.
Why I advise against it, and this is not a moral objection but a mechanical one: a department assessed on the number of files awarded awards the simple files first. Members in several schemes, careers completed abroad, files with a period to reconstruct sink to the bottom of the pile — and those are precisely the members who need you most. The indicator would rise and equality of treatment would fall: you would lose the instrument at the same time as the result.
What I found instead, and it answers management's real question: I compared the 9,400 files and notification time varies from 41 to 96 days depending on the month the claim is filed, not on who handles it. Claims filed in January and September wait twice as long, because they arrive alongside the campaign peaks. Smoothing arrivals by staggering the invitations to file cuts 24 days off the average time, and it puts nobody in the dock.
The invitation calendar is written, month by month, set against your real handling capacity: it moves 1,400 claims out of the two peak months into the four quiet ones, without any member filing later than they wish.
And the indicators I keep in any event, because they bear on files and not on people: time per stage, hold rate, rework rate, share of files notified within the service commitment. Those are the ones your supervisory authority looks at, and they improve without anyone being counted.
What I propose: to put the two options side by side for management, with what each produces. I hand you both files on Friday; if smoothing is chosen, the first staggered invitation goes out on the 1st of the following month and the gain is measurable within the quarter.
✎ Framework · 9,400 files of the year, times by month of filing, operating rules, the fund's monitoring indicators
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
The uses of AI for the pensions adviser
Each use corresponds to an agent we deploy. All work in support, subject to the adviser's approval.
Reconstructing career records
Analysis of the career record from the documents, identification of missing periods and preparation of the corrections.
Indicative pension estimates
Preparation of estimates clearly presented as indicative, with the reply letters to members.
Pre-assessment of applications
Checking the documents and the conditions, alerting on the points to verify before the adviser's decision.
Retirement scenarios
Simulation of several assumptions (age, full rate, increment) on an indicative basis, to inform the member's choice.
Notifications & missing documents
Drafting the notifications and the requests for documents to members and to the schemes concerned. A short reminder can be sent by text message, with the detailed version staying in the letter.
Information for members
Answer everyday questions (retirement age, procedures, documents) and direct people to the right person, 24/7.
Appointment booking
Qualify incoming requests and schedule pension support appointments.
Multichannel administrative information
RAG on procedures; web, phone and front desk; 24/7; multilingual. Sourced answers.
On quote View the agent page →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.
Managing family benefit files
Strictly in support (administrative). Payment decision by a human.
Managing family benefit files from 1,020 € incl. VAT / month Discover the agent →Administrative management of social housing files
Strictly in support (administrative). Allocations decided by the panel.
Administrative management of social housing files 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 →In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
How much time can a pensions adviser win back?
By automating the reconstruction of career records, the estimates and the letters, a fund can aim for a clear reduction in time spent on repetitive tasks — reinvested in supporting members.
The stages of your AI agent project
Audit & scoping
15 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
Three options, one agent
A pensions agent (career reconstruction, indicative estimates, letters to members), installed and operated for you. Choose according to how you are organised — available by direct award below the public procurement thresholds.
Setup + controlled subscription
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
All inclusive, no setup fee
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
On site, you own it
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Four guarantees that matter to a pension fund
Your questions, our answers
Does the agent calculate the final pension?
Where does the data come from?
Does the agent handle the supplementary scheme?
Do the simulations bind the fund?
Does the agent state that it is an artificial intelligence?
How long does it take to deploy an agent?
Do we need a technical team in-house?
Do we have to change software?
Which tools can users use to reach the agent?
Can the agent notify insured people by text message?
Let's size up the potential in your fund
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.