An orchestration agent, scoped with you before it is priced. This agent coordinates several specialised agents. Its orchestration follows your actual workflows — which is why it is scoped with you rather than bought off the shelf. We establish the scope together, then the quotation commits it.
The journeys described below form the scope that this review refines and the quotation commits. The specialised agents it coordinates can be ordered today.
Request a quote
● Public sector — Workplace injury caseworker
AI agent for the workplace injury and sick leave caseworker
An assistant that pre-assesses accident reports and sick leave, checks the documents and prepares the replies — the caseworker keeps the decision and the classification. Deciding whether an accident is work-related stays strictly human, on medical advice. HDS hosting in France for health data.
HDS hosting in France
Medical data isolated
GDPR & AI Act: governed deployment
Human classification, always
Updated on
Deployed in a few weeks
Workplace injury assistant · hosted in France
Pre-assess this workplace accident report received this morning: check the documents and the deadlines.
File pre-assessed. Report received within the 48-hour limit, initial medical certificate attached, employer's reservations: none raised at this stage. The assessment period runs to the 28th of the month. Document to request: the salary certificate for daily benefits.
File passed on for classification — which stays yours, on medical advice.
⛓ Source · the report + the documents received, statutory deadlines
File passed on for classification — which stays yours, on medical advice.
⛓ Source · the report + the documents received, statutory deadlines
Prepare the notifications and the deadline tracking for the files whose assessment period ends this week.
5 files reaching the deadline: draft notifications prepared from your templates, decision fields left blank — recognition or refusal is a matter for your classification. Schedule updated with the deadline for each file.
All submitted for your check before sending.
✎ Action · notifications prepared — the caseworker approves and classifies
All submitted for your check before sending.
✎ Action · notifications prepared — the caseworker approves and classifies
Local inference · no data outside the EU
Data hosted in France
Sovereign by designHDS hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
Classification stays humanThe agent prepares, it never classifies
✦ In brief
A Blue Lemon Agent agent assists the workplace injury caseworker with the repetitive tasks — pre-assessment of workplace accident reports and sick leave (documents, deadlines), letters and notifications to claimants and employers, tracking the classification deadlines — and spots inconsistencies to check before handling. It is hosted in France with HDS hosting for health data, on an isolated resource, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Deciding whether an injury is work-related stays a human decision, on medical advice — never the machine's. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and insured people and employers 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 workplace injury services — and why they demand guarantees
Accident reports, medical certificates, salary certificates, statutory assessment periods: every workplace injury file mixes administrative documents and health data, under the pressure of strict statutory deadlines.
! The issue
The caseworker is caught between statutory assessment periods that cannot be negotiated and files whose assembly — reports, certificates, salary statements, employers' reservations — is time-consuming. Yet most consumer AI tools amount to entrusting health data, employment situations and medical information to a third party, often hosted outside Europe and subject to the Cloud Act — unacceptable for a social security body.
✓ Our answer
For files mixing medical and social data, AI is only of interest if it is sovereign, certified and strictly bounded. HDS hosting for health data, an isolated resource, minimised access, systematic human oversight: the AI agent prepares the files and watches the deadlines, but deciding whether an accident is work-related remains the caseworker's decision, on medical advice. The time won back goes to handling the files within the deadlines.
The decisive point
Protecting occupational health data: sovereignty & compliance
A workplace injury service handles health data and sensitive employment situations. Here is how the architecture of our agents protects it.
HDS hosting
The health-data hosting level required as soon as certificates and medical information are involved, for the processing concerned: settled with you at scoping.
Hosting in France
A dedicated and isolated resource, hosted in France under French law — claimants' data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Exposure of the medical data in workplace injury files to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
Local inference possible
For the most sensitive information, the agent can run on local inference: nothing leaves the organisation's network.
Encryption & controlled access
Encryption in transit and at rest, minimisation, role-based access, strong authentication and logging.
AI Act: governed deployment
The agent is strictly in support; no classification 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.
HDS hosting (the French certification for health data hosts) for the processing concerned — essential as soon as a medical certificate is involved — and a SecNumCloud option according to how demanding your requirements are. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
5 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The public body in this demonstration
Fictional public bodyVal-d'Aurance health insurance fund — occupational risks service
- Sector
- Occupational risks — handling reports of workplace accidents, commuting accidents and occupational illnesses, monitoring sick leave and daily allowances
- Headcount
- 62 staff — including 24 workplace injury caseworkers, 8 on the telephone service, 5 investigators, and a separate medical service, alone authorised on medical material
- Public served
- 18,700 employers and 214,000 employees covered, from the three-worker building site to the eight-hundred-strong industrial plant
- Order of magnitude
- 10,000 files a year — 9,400 accident reports and 600 occupational illness claims —, 31,000 sick leave spells monitored, 41,000 documents received, 12,000 letters and notifications, and 74,000 enquiries: 54,000 calls, 14,000 e-mails, 6,000 portal messages
- Tools in place
- National occupational risks management software, employer online reporting portal, telephone platform and 7 years of filed cases — the agent plugs into them read-only, nothing is replaced and nothing is migrated
- Who decides
- The caseworker prepares the file; classifying the accident as work-related is decided and signed by the authorised officer, on the medical service's opinion; the head of service rules on extensions; the medical service alone accesses medical material
- Points for improvement
- 1,400 files a year — 14 % — reached the decision deadline without the adversarial stage having been opened; 780 overran a deadline; nearly one report in three arrived incomplete from the employer portal; and 29 % of calls rang unanswered at accident peaks
At Val-d'Aurance, handling a workplace accident turns on deadlines that cannot be negotiated: when the deadline passes without a decision, the accident counts as work-related by default, and nobody has decided anything. The agent is hosted in France, on a resource dedicated to health data, with local inference available for the most sensitive material; it reads the seven years of files without ever writing to them, opens no medical item and releases nothing without a caseworker's approval. The exchanges below cover a year, from rebuilding the timetables to the review presented to management.
This public body, its figures and the exchanges that follow were invented for the demonstration. They illustrate a common situation; they describe no real service.
Workplace injury assistant · hosted in France
I reread your seven years of files over the weekend — 10,000 files a year, 41,000 documents, every instruction timetable — and I have drawn out the figure that decides everything else.
The gap measured: 1,400 files a year — 14 % — reached the decision deadline without the adversarial stage having been opened. The adversarial stage is the point at which the employer and the insured person may consult the material in the file and make their observations before the decision is taken; it is not a formality, it is what makes the decision defensible.
And the figure that follows, the costliest one: 780 files overran an instruction deadline. When the deadline passes without a decision, the accident counts as work-related by default — 780 times, nobody decided anything, one way or the other. These are neither refusals nor recognitions: they are files the calendar settled instead of your officers.
The cause, measured and not assumed: each file's timetable lived in the head of the caseworker holding it. Of the 780, 611 were files taken over by another caseworker mid-handling — leave, moves, year-end reinforcement. The date was not lost through negligence: it was lost at the hand-over.
What I have already done, and this is not a promise: I have rebuilt the timetable of all 10,000 files — date received, starting point of each deadline, date the adversarial stage was opened, decision deadline — and rerun it over the last twelve months. With alerts at fifteen days, seven days and two days before expiry: all 1,400 adversarial stages would have been opened in time, and the 780 overruns would have fallen to 41.
The next step I propose, and it takes one meeting: you read the standard timetable for each type of file — it is written in everyday wording, one line per stage. The service settles it, and from the next day every file carries its three alerts: your 24 caseworkers stop holding dates in their heads, and not one file decides itself again. instruction-deadlines_1400-adversarial-stages-and-780-overruns.pdf14 % of files with no adversarial stage, 611 overruns after hand-over
⛓ Sourced · 7 years of files, rebuilt instruction timetables, hand-over log, rerun over 12 months
The gap measured: 1,400 files a year — 14 % — reached the decision deadline without the adversarial stage having been opened. The adversarial stage is the point at which the employer and the insured person may consult the material in the file and make their observations before the decision is taken; it is not a formality, it is what makes the decision defensible.
And the figure that follows, the costliest one: 780 files overran an instruction deadline. When the deadline passes without a decision, the accident counts as work-related by default — 780 times, nobody decided anything, one way or the other. These are neither refusals nor recognitions: they are files the calendar settled instead of your officers.
The cause, measured and not assumed: each file's timetable lived in the head of the caseworker holding it. Of the 780, 611 were files taken over by another caseworker mid-handling — leave, moves, year-end reinforcement. The date was not lost through negligence: it was lost at the hand-over.
What I have already done, and this is not a promise: I have rebuilt the timetable of all 10,000 files — date received, starting point of each deadline, date the adversarial stage was opened, decision deadline — and rerun it over the last twelve months. With alerts at fifteen days, seven days and two days before expiry: all 1,400 adversarial stages would have been opened in time, and the 780 overruns would have fallen to 41.
The next step I propose, and it takes one meeting: you read the standard timetable for each type of file — it is written in everyday wording, one line per stage. The service settles it, and from the next day every file carries its three alerts: your 24 caseworkers stop holding dates in their heads, and not one file decides itself again. instruction-deadlines_1400-adversarial-stages-and-780-overruns.pdf14 % of files with no adversarial stage, 611 overruns after hand-over
⛓ Sourced · 7 years of files, rebuilt instruction timetables, hand-over log, rerun over 12 months
Rebuilding timetables means opening files that contain medical certificates. That has been our sticking point for two years.
I open no certificate, and it can be checked with one command: I read dates and document types, never medical content.
What I see of a medical certificate, and nothing else: that a document of that type arrived, on what date, and for which file. The content — the site of the injuries, the diagnosis, the medically justified duration — stays in the medical service, and I do not access it. What the caseworker sees is what the medical service has decided to pass on: an opinion, a date, a duration.
The architecture that guarantees it, point by point:
· The 9,400 medical documents received this year are hosted on the resource dedicated in France to hosting health data, separate from the rest, and the medical service alone accesses it.
· For the most sensitive material, local inference is available — the model then computes on a machine belonging to the body, and nothing crosses any external network.
· Role-based access — rights follow the job: a caseworker sees the administrative documents of their files, an investigator the circumstances, the medical service the medical material, and nobody sees all three. Over the year, the log shows 0 access to a medical item outside the medical service.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity; a dedicated resource, with no pooling with another body; encryption in transit and at rest; a read-only technical account, with no write permission on your systems; and no insured person's data trains any model.
What this partitioning lets you say, and it is an argument, not a constraint: to an insured person asking who has seen their file, you answer with a named and dated list; to an employer asking to consult the material, you show exactly what they are entitled to consult, with no risk of showing one document too many. The log can be inspected by your data protection officer whenever she decides, without telling us.
The next step I propose: that I keep up to date the record your officer and your supervising authority will ask for — hosting, data processed, retention periods, who has access to what, and the special place of health material. The first version is written and you have it in front of you. technical-framework_partitioning-of-health-data.pdf0 medical access outside the medical service, local inference available
✎ Framework · deployment architecture, the HDS hosting requirement, role-based access log
What I see of a medical certificate, and nothing else: that a document of that type arrived, on what date, and for which file. The content — the site of the injuries, the diagnosis, the medically justified duration — stays in the medical service, and I do not access it. What the caseworker sees is what the medical service has decided to pass on: an opinion, a date, a duration.
The architecture that guarantees it, point by point:
· The 9,400 medical documents received this year are hosted on the resource dedicated in France to hosting health data, separate from the rest, and the medical service alone accesses it.
· For the most sensitive material, local inference is available — the model then computes on a machine belonging to the body, and nothing crosses any external network.
· Role-based access — rights follow the job: a caseworker sees the administrative documents of their files, an investigator the circumstances, the medical service the medical material, and nobody sees all three. Over the year, the log shows 0 access to a medical item outside the medical service.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity; a dedicated resource, with no pooling with another body; encryption in transit and at rest; a read-only technical account, with no write permission on your systems; and no insured person's data trains any model.
What this partitioning lets you say, and it is an argument, not a constraint: to an insured person asking who has seen their file, you answer with a named and dated list; to an employer asking to consult the material, you show exactly what they are entitled to consult, with no risk of showing one document too many. The log can be inspected by your data protection officer whenever she decides, without telling us.
The next step I propose: that I keep up to date the record your officer and your supervising authority will ask for — hosting, data processed, retention periods, who has access to what, and the special place of health material. The first version is written and you have it in front of you. technical-framework_partitioning-of-health-data.pdf0 medical access outside the medical service, local inference available
✎ Framework · deployment architecture, the HDS hosting requirement, role-based access log
We already have alerts. There are so many of them that nobody looks at them any more.
That is exactly what I produced in the first twelve weeks, and I publish the figure: 3,200 alerts, of which 2,100 were pointless — 66 %. Only one alert in three deserved a look, and that is the surest way to make all three useless.
The cause, measured: I alerted on every file whose deadline was approaching, without looking at its state. 1,380 of the 2,100 pointless alerts concerned files already decided and 720 files whose handling was suspended pending an external element.
What I did about it, and it is measured: an alert now fires only if three conditions are met: the file is open, no decision is recorded, and the stage due on that date has not been reached. Over the following twelve weeks: 1,240 alerts, 1,180 useful — 95 %. And the remaining 60 are suspended files whose suspension had not been recorded: I flag those separately, because they are the ones that go back into overrun.
What each alert carries, and it is what makes it usable in ten seconds: the file, the missing stage, the deadline, the days remaining, and the document already drafted that unblocks the stage — letter opening the adversarial stage, document request, or draft decision when the file is complete.
The time this shifts: reviewing the caseload's deadlines goes from 100 % to 20 % of the task — from 2 hours a week per caseworker to 24 minutes. Across 24 caseworkers and 45 working weeks, that is 1,728 hours given back to handling itself.
The next step I propose: that the Monday record carry the only three figures that matter to your head of service — files due within fifteen days, adversarial stages not opened, files suspended for more than thirty days. One page, three columns — that is what allows reinforcements to be arranged on Monday morning rather than overruns to be noted at quarter end. deadline-alerts_3200-then-1240.pdf66 % pointless down to 5 %, 1,728 hours given back
⛓ Sourced · alert log over two twelve-week periods, file states, outcomes
The cause, measured: I alerted on every file whose deadline was approaching, without looking at its state. 1,380 of the 2,100 pointless alerts concerned files already decided and 720 files whose handling was suspended pending an external element.
What I did about it, and it is measured: an alert now fires only if three conditions are met: the file is open, no decision is recorded, and the stage due on that date has not been reached. Over the following twelve weeks: 1,240 alerts, 1,180 useful — 95 %. And the remaining 60 are suspended files whose suspension had not been recorded: I flag those separately, because they are the ones that go back into overrun.
What each alert carries, and it is what makes it usable in ten seconds: the file, the missing stage, the deadline, the days remaining, and the document already drafted that unblocks the stage — letter opening the adversarial stage, document request, or draft decision when the file is complete.
The time this shifts: reviewing the caseload's deadlines goes from 100 % to 20 % of the task — from 2 hours a week per caseworker to 24 minutes. Across 24 caseworkers and 45 working weeks, that is 1,728 hours given back to handling itself.
The next step I propose: that the Monday record carry the only three figures that matter to your head of service — files due within fifteen days, adversarial stages not opened, files suspended for more than thirty days. One page, three columns — that is what allows reinforcements to be arranged on Monday morning rather than overruns to be noted at quarter end. deadline-alerts_3200-then-1240.pdf66 % pointless down to 5 %, 1,728 hours given back
⛓ Sourced · alert log over two twelve-week periods, file states, outcomes
An accident report came in this morning from a roofing site. I need to know where I stand before midday.
The file is pre-assessed, complete but for one document, and its timetable runs through to the decision deadline.
Pre-assessment is everything that precedes the decision: checking the deadlines, gathering and checking the documents, cross-matching what must agree and drafting the letters. The decision itself is the classification of the accident as work-related.
What I hand over, on one page:
· The employer's report received within the 48-hour deadline that applies to them, with the portal timestamp.
· The initial medical certificate: present, dated, attached to the file — its type and its date, not its content: the content is with the medical service.
· The employer's reservations — the observations by which an employer disputes that the accident happened at work or in the circumstances reported — none has been made at this stage, and I say so because the period for making them is still running.
· The missing document, named: the wage certificate, without which daily allowances cannot be calculated. The request to the employer is written, it names the form and attaches the example.
· The file's full timetable: deadline for opening the adversarial stage, decision deadline, and the three alerts already set.
What remains with the officer, and it is the heart of the job: classifying the accident as work-related is decided and signed by the authorised officer, on the medical service's opinion, because it commits the body and is open to appeal — and I hand it to them in minutes, complete file, documents checked, deadlines met and reasoning already drafted both ways. It is that signature that makes the decision reasoned and challengeable, therefore defensible before the employer and before the insured person.
The time this shifts: pre-assessing an accident report goes from 60 % to 10 % of the file's time — from 27 minutes to 4 minutes 30. Across 10,000 files a year, that is 3,750 caseworker hours given back to the files that need real handling — those with reservations, an investigation, or a dispute in the making.
The next step I propose: that every report arriving complete from the portal be pre-assessed on receipt, without being asked for. Of the year's 9,400 reports, 6,400 are in that case — six thousand four hundred files waiting for you ready to decide the same morning. pre-assessment_workplace-accident-report.pdf27 minutes down to 4 min 30, timetable set through to the decision
⛓ Sourced · report received from the portal, documents in the file, instruction timetable, the establishment's history
Pre-assessment is everything that precedes the decision: checking the deadlines, gathering and checking the documents, cross-matching what must agree and drafting the letters. The decision itself is the classification of the accident as work-related.
What I hand over, on one page:
· The employer's report received within the 48-hour deadline that applies to them, with the portal timestamp.
· The initial medical certificate: present, dated, attached to the file — its type and its date, not its content: the content is with the medical service.
· The employer's reservations — the observations by which an employer disputes that the accident happened at work or in the circumstances reported — none has been made at this stage, and I say so because the period for making them is still running.
· The missing document, named: the wage certificate, without which daily allowances cannot be calculated. The request to the employer is written, it names the form and attaches the example.
· The file's full timetable: deadline for opening the adversarial stage, decision deadline, and the three alerts already set.
What remains with the officer, and it is the heart of the job: classifying the accident as work-related is decided and signed by the authorised officer, on the medical service's opinion, because it commits the body and is open to appeal — and I hand it to them in minutes, complete file, documents checked, deadlines met and reasoning already drafted both ways. It is that signature that makes the decision reasoned and challengeable, therefore defensible before the employer and before the insured person.
The time this shifts: pre-assessing an accident report goes from 60 % to 10 % of the file's time — from 27 minutes to 4 minutes 30. Across 10,000 files a year, that is 3,750 caseworker hours given back to the files that need real handling — those with reservations, an investigation, or a dispute in the making.
The next step I propose: that every report arriving complete from the portal be pre-assessed on receipt, without being asked for. Of the year's 9,400 reports, 6,400 are in that case — six thousand four hundred files waiting for you ready to decide the same morning. pre-assessment_workplace-accident-report.pdf27 minutes down to 4 min 30, timetable set through to the decision
⛓ Sourced · report received from the portal, documents in the file, instruction timetable, the establishment's history
Documents are our daily bread: certificates, wage certificates, employer letters. They arrive by every channel and we are the ones sorting them.
41,000 documents received this year, 37,720 checked and filed on their own in the right file — 92 %. The remaining 3,280 reach you with their reason, not in a pile.
What I check on each document, before a caseworker opens it: that it matches the document expected, that it is complete, that it carries the right insured person and the right employer, and that it arrives within the deadline that concerns it — that last point is the one nobody has time to check, and the one that counts on the day of an appeal. A crooked scan or a photo taken on a building site changes nothing: I read the administrative content, not the layout.
The 3,280 that do not pass, and what I have already done with them:
· 1,940 wage certificates incomplete or unreadable — most often a missing reference period. The request to redo it is written, it points at the single line to complete and attaches the example. That is what halves second submissions: an employer does not send a wrong document out of ill will, he sends it because he was asked for “the certificate”.
· 860 documents with no file reference. I attached 790 of them by cross-checking the insured person's name, the establishment and the date of the accident, and they reach you with the two candidate files side by side.
· 480 medical documents sent to the administrative service: they go straight to the medical service without anyone else opening them, and the sender receives the acknowledgement saying so.
What that gives, in figures: sorting and matching documents occupied two people half-time; half a day a week remains, and a document is on file the day it arrives — therefore before the deadline, not after.
The next step I propose: an automatic acknowledgement of every document, within the hour, saying what it completes, what is still missing and the file's deadline. Over a one-month trial, “did you get my certificate?” calls fell from 480 to 61 — it is your platform's second reason for calling, and it falls away on its own. document-checks_41000-received-37720-filed.pdf92 % filed on their own, 790 attached out of 860
⛓ Sourced · 41,000 documents received over 12 months, case files, log of second submissions and follow-up calls
What I check on each document, before a caseworker opens it: that it matches the document expected, that it is complete, that it carries the right insured person and the right employer, and that it arrives within the deadline that concerns it — that last point is the one nobody has time to check, and the one that counts on the day of an appeal. A crooked scan or a photo taken on a building site changes nothing: I read the administrative content, not the layout.
The 3,280 that do not pass, and what I have already done with them:
· 1,940 wage certificates incomplete or unreadable — most often a missing reference period. The request to redo it is written, it points at the single line to complete and attaches the example. That is what halves second submissions: an employer does not send a wrong document out of ill will, he sends it because he was asked for “the certificate”.
· 860 documents with no file reference. I attached 790 of them by cross-checking the insured person's name, the establishment and the date of the accident, and they reach you with the two candidate files side by side.
· 480 medical documents sent to the administrative service: they go straight to the medical service without anyone else opening them, and the sender receives the acknowledgement saying so.
What that gives, in figures: sorting and matching documents occupied two people half-time; half a day a week remains, and a document is on file the day it arrives — therefore before the deadline, not after.
The next step I propose: an automatic acknowledgement of every document, within the hour, saying what it completes, what is still missing and the file's deadline. Over a one-month trial, “did you get my certificate?” calls fell from 480 to 61 — it is your platform's second reason for calling, and it falls away on its own. document-checks_41000-received-37720-filed.pdf92 % filed on their own, 790 attached out of 860
⛓ Sourced · 41,000 documents received over 12 months, case files, log of second submissions and follow-up calls
And when the report, the certificate and the wage certificate do not say the same thing? That is when files get difficult.
I cross-match them systematically, and I publish the figure that does not flatter me: my first rule produced 1,460 flags over twelve months, of which 254 were confirmed after review — 17 %.
The cause, measured: my first rule flagged every discrepancy, including one-day differences between the reported date of the accident and the date of the certificate — 1,010 of the 1,460 flags were of that kind, and an accident happening at the end of the day where the employee sees a doctor the next morning is no anomaly.
What I did about it, and it is measured: the tightened rule flags only three discrepancies, the ones that change the handling: a difference in circumstances between the report and the insured person's account · a gap of more than forty-eight hours between the reported accident and the first medical record · a leave period that does not match the period on the wage certificate. Rerun over the same twelve months: 402 flags, 241 confirmed — 60 %. The 13 confirmed cases the tightened version lets through surface in the monthly sample check, and they came out of it every time.
What that changes on both sides: 1,058 fewer files to review a year for your caseworkers, 241 of the 254 discrepancies still caught, and 1,058 employers or insured people who no longer receive a request for explanation over a one-day difference.
What each flag carries, and it is what makes it usable: the two documents side by side, the exact line where they diverge, the size of the discrepancy, and the three most frequent explanations for that kind of discrepancy in your own files. In two thirds of cases the explanation is already the right one, and the review takes three minutes. A discrepancy is never a presumption against anyone: it is a question to be asked, and the letter asking it is already written.
The next step I propose: that the tightened rule take effect through your signature, and that I hand you its quarterly record — flags produced, share confirmed, discrepancies caught by the sample. A rule that is not measured drifts in silence; this one is corrected on figures, and the service alone puts it into service. discrepancies-between-documents_1460-then-402.pdf17 % confirmation raised to 60 %, 241 discrepancies out of 254
⛓ Sourced · both versions of the rule rerun over 12 months, files reviewed and outcomes, monthly sample check
The cause, measured: my first rule flagged every discrepancy, including one-day differences between the reported date of the accident and the date of the certificate — 1,010 of the 1,460 flags were of that kind, and an accident happening at the end of the day where the employee sees a doctor the next morning is no anomaly.
What I did about it, and it is measured: the tightened rule flags only three discrepancies, the ones that change the handling: a difference in circumstances between the report and the insured person's account · a gap of more than forty-eight hours between the reported accident and the first medical record · a leave period that does not match the period on the wage certificate. Rerun over the same twelve months: 402 flags, 241 confirmed — 60 %. The 13 confirmed cases the tightened version lets through surface in the monthly sample check, and they came out of it every time.
What that changes on both sides: 1,058 fewer files to review a year for your caseworkers, 241 of the 254 discrepancies still caught, and 1,058 employers or insured people who no longer receive a request for explanation over a one-day difference.
What each flag carries, and it is what makes it usable: the two documents side by side, the exact line where they diverge, the size of the discrepancy, and the three most frequent explanations for that kind of discrepancy in your own files. In two thirds of cases the explanation is already the right one, and the review takes three minutes. A discrepancy is never a presumption against anyone: it is a question to be asked, and the letter asking it is already written.
The next step I propose: that the tightened rule take effect through your signature, and that I hand you its quarterly record — flags produced, share confirmed, discrepancies caught by the sample. A rule that is not measured drifts in silence; this one is corrected on figures, and the service alone puts it into service. discrepancies-between-documents_1460-then-402.pdf17 % confirmation raised to 60 %, 241 discrepancies out of 254
⛓ Sourced · both versions of the rule rerun over 12 months, files reviewed and outcomes, monthly sample check
Five files fall due this week. The adversarial stage has to be opened and the notifications prepared, and I do not have the day.
The five letters opening the adversarial stage are written, and so are the five draft notifications — both ways, recognition and refusal.
What each opening letter carries: the exact list of documents its addressee may consult, the period during which they may make observations, and how to make them — in writing, through the portal or at an appointment.
And the list is not the same for everyone, which is the whole point: the employer consults the administrative documents and the circumstances; they see no medical item, and the system cannot show them any — this is not an instruction, it is role-based partitioning, and it spares you the gravest mistake an occupational risks service can make. The insured person consults everything that concerns them. Across the year's 1,240 openings, 0 document was made consultable outside its addressee's scope.
The draft notifications now: each is written both ways, reasoned point by point on the material in the file, and the decision fields are left blank. It is the authorised officer who decides, on the medical service's opinion — and they decide on a file whose reasoning is already drafted, which is exactly the work nobody has time to do twice.
The time this shifts: a standard notification to an insured person or an employer goes from 40 % to 10 % of the task's time — from 8 minutes to 2. Across 12,000 letters and notifications a year, that is 1,200 hours given back to the service.
And the figure that matters most, last year's: 1,400 files reached the deadline with no adversarial stage opened. Over the twelve months rerun with the alerts, none remains — and 739 files out of 780 stop being settled by the calendar.
The next step I propose: that the letter opening the adversarial stage be prepared as soon as the file is complete, rather than as the deadline nears. Opened early, the adversarial stage gives the employer and the insured person time to answer properly — over a one-quarter trial, observations received rose from 19 % to 41 % of files, and that is as many decisions taken in full knowledge. opening-the-adversarial-stage_who-consults-what.pdf1,240 openings, 0 document out of scope, observations 19 % → 41 %
⛓ Sourced · timetables of files falling due, consultation scopes by role, the service's notification templates
What each opening letter carries: the exact list of documents its addressee may consult, the period during which they may make observations, and how to make them — in writing, through the portal or at an appointment.
And the list is not the same for everyone, which is the whole point: the employer consults the administrative documents and the circumstances; they see no medical item, and the system cannot show them any — this is not an instruction, it is role-based partitioning, and it spares you the gravest mistake an occupational risks service can make. The insured person consults everything that concerns them. Across the year's 1,240 openings, 0 document was made consultable outside its addressee's scope.
The draft notifications now: each is written both ways, reasoned point by point on the material in the file, and the decision fields are left blank. It is the authorised officer who decides, on the medical service's opinion — and they decide on a file whose reasoning is already drafted, which is exactly the work nobody has time to do twice.
The time this shifts: a standard notification to an insured person or an employer goes from 40 % to 10 % of the task's time — from 8 minutes to 2. Across 12,000 letters and notifications a year, that is 1,200 hours given back to the service.
And the figure that matters most, last year's: 1,400 files reached the deadline with no adversarial stage opened. Over the twelve months rerun with the alerts, none remains — and 739 files out of 780 stop being settled by the calendar.
The next step I propose: that the letter opening the adversarial stage be prepared as soon as the file is complete, rather than as the deadline nears. Opened early, the adversarial stage gives the employer and the insured person time to answer properly — over a one-quarter trial, observations received rose from 19 % to 41 % of files, and that is as many decisions taken in full knowledge. opening-the-adversarial-stage_who-consults-what.pdf1,240 openings, 0 document out of scope, observations 19 % → 41 %
⛓ Sourced · timetables of files falling due, consultation scopes by role, the service's notification templates
Our notifications often come back: employers dispute them, insured people do not understand the reasoning.
I reread your 12,000 letters of the year and counted the ones that come back: 1,640 triggered a call, a request for explanation or a dispute — nearly one in seven.
What the rereading shows, and it is nobody's fault: three templates account for 1,190 of those 1,640 returns. They are the three that announce a decision without saying what it is based on or how it can be challenged. An insured person who does not understand the reasoning does not give up: they dispute, and they dispute without knowing what about.
What I have rewritten, and it is ready for your reading:
· The reasoning spelled out, element by element: what was taken into account, what was not, and the sentence that is always missing — “here is what would have led to a different decision”.
· The route and time limit for appeal at the top of the letter, not in the footer, because a notification that does not say how to challenge it is challenged badly, therefore at greater length.
· For a plain-language version, pass this content to the Accessibility, Easy Read and Translation agent — a dedicated agent, and it is not part of what this agent includes: I prepare the facts, the reasoning and the time limits it needs, and it is that agent which produces the adapted version.
· And for the employer, a reminder of what they were able to consult and what they replied — a dispute rarely starts from a disagreement on the merits, it starts from the feeling of not having been heard.
What that gives, measured over the trial quarter: returns on those three templates go from 1,190 to 340 a year, and requests for explanation from 512 to 96. These are not disputes stifled: they are addressees who understood first time — and the disputes that remain are about the merits, which is precisely their purpose.
The next step I propose: that I flag on my own any template a new instruction amends, and write the updated version within the 24 hours following its circulation. Your templates will stop ageing in silence — that is what avoids redoing this work in two years, and it will cost you one reading per instruction. notifications_3-templates-rewritten.pdf1,190 returns down to 340, reasoning spelled out
⛓ Sourced · 12,000 letters of the year, disputes and requests for explanation, notification templates, one-quarter trial
What the rereading shows, and it is nobody's fault: three templates account for 1,190 of those 1,640 returns. They are the three that announce a decision without saying what it is based on or how it can be challenged. An insured person who does not understand the reasoning does not give up: they dispute, and they dispute without knowing what about.
What I have rewritten, and it is ready for your reading:
· The reasoning spelled out, element by element: what was taken into account, what was not, and the sentence that is always missing — “here is what would have led to a different decision”.
· The route and time limit for appeal at the top of the letter, not in the footer, because a notification that does not say how to challenge it is challenged badly, therefore at greater length.
· For a plain-language version, pass this content to the Accessibility, Easy Read and Translation agent — a dedicated agent, and it is not part of what this agent includes: I prepare the facts, the reasoning and the time limits it needs, and it is that agent which produces the adapted version.
· And for the employer, a reminder of what they were able to consult and what they replied — a dispute rarely starts from a disagreement on the merits, it starts from the feeling of not having been heard.
What that gives, measured over the trial quarter: returns on those three templates go from 1,190 to 340 a year, and requests for explanation from 512 to 96. These are not disputes stifled: they are addressees who understood first time — and the disputes that remain are about the merits, which is precisely their purpose.
The next step I propose: that I flag on my own any template a new instruction amends, and write the updated version within the 24 hours following its circulation. Your templates will stop ageing in silence — that is what avoids redoing this work in two years, and it will cost you one reading per instruction. notifications_3-templates-rewritten.pdf1,190 returns down to 340, reasoning spelled out
⛓ Sourced · 12,000 letters of the year, disputes and requests for explanation, notification templates, one-quarter trial
That leaves the telephone. At accident peaks, 29 % of calls ring unanswered, and these are employees on sick leave waiting for their allowances.
54,000 calls last year, 15,660 unanswered — 29 %. This is not a failure of service, it is a peak eight people cannot absorb.
What I did with those calls, rereading your subject records, your e-mails and your portal messages:
· 44,000 of the year's 74,000 enquiries are about twelve subjects, always the same — where my file stands, which documents are missing, when my daily allowances will be paid, how to report an accident, how to make reservations, how to consult the file, what the decision deadline is, how to challenge, the wage certificate, returning to work, a relapse, and a change of contact details.
· All twelve have a written answer in your instructions or your standard letters, and all twelve answers are drafted, sourced and dated — they are waiting for you. The answer existed: what was missing was availability.
What I propose, and you keep the key: I answer those calls at any hour, and I say in the first sentence that I am a digital assistant of the body, not an officer. That is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be told so, and the caller can ask for an officer at any moment — I then take their number and their subject, and the call-back lands in your queue with the file already open.
The rule I hold most firmly, and it is the one that protects the body: on the progress of a file, I say where it stands, what it is waiting for and by what date the decision must be taken; on the outcome, nothing at all. And no medical item is ever given over the telephone, to anyone — not to the insured person, not to their employer, not to a relative: what is medical goes through the medical service, and the caller leaves with the exact path to obtain it.
The gain, in figures: unanswered calls go from 29 % to 4 %, and the reply time to an e-mail from 6 days to under 24 hours. What remains are the calls asking for an officer, and they arrive with the subject already noted and the file open.
The next step I propose: that you read the twelve answers tomorrow, one by one — twenty-five minutes. As soon as they are approved, the front line answers that same night, and I hand you each morning the page of what went out. front-line_74000-enquiries-12-subjects.pdf15,660 calls lost, 12 subjects, 0 medical item over the phone
⛓ Sourced · 12 months of platform logs, e-mails and portal messages, internal instructions, 12 drafted answers
What I did with those calls, rereading your subject records, your e-mails and your portal messages:
· 44,000 of the year's 74,000 enquiries are about twelve subjects, always the same — where my file stands, which documents are missing, when my daily allowances will be paid, how to report an accident, how to make reservations, how to consult the file, what the decision deadline is, how to challenge, the wage certificate, returning to work, a relapse, and a change of contact details.
· All twelve have a written answer in your instructions or your standard letters, and all twelve answers are drafted, sourced and dated — they are waiting for you. The answer existed: what was missing was availability.
What I propose, and you keep the key: I answer those calls at any hour, and I say in the first sentence that I am a digital assistant of the body, not an officer. That is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be told so, and the caller can ask for an officer at any moment — I then take their number and their subject, and the call-back lands in your queue with the file already open.
The rule I hold most firmly, and it is the one that protects the body: on the progress of a file, I say where it stands, what it is waiting for and by what date the decision must be taken; on the outcome, nothing at all. And no medical item is ever given over the telephone, to anyone — not to the insured person, not to their employer, not to a relative: what is medical goes through the medical service, and the caller leaves with the exact path to obtain it.
The gain, in figures: unanswered calls go from 29 % to 4 %, and the reply time to an e-mail from 6 days to under 24 hours. What remains are the calls asking for an officer, and they arrive with the subject already noted and the file open.
The next step I propose: that you read the twelve answers tomorrow, one by one — twenty-five minutes. As soon as they are approved, the front line answers that same night, and I hand you each morning the page of what went out. front-line_74000-enquiries-12-subjects.pdf15,660 calls lost, 12 subjects, 0 medical item over the phone
⛓ Sourced · 12 months of platform logs, e-mails and portal messages, internal instructions, 12 drafted answers
Nearly one report in three arrives incomplete from the portal. We spend our days chasing what is missing, and the deadline runs anyway.
3,000 incomplete reports out of 9,400 last year, and I have found where it plays out: five fields account for 78 % of the gaps.
The five, in order: the exact time of the accident · the exact place, when it is not the establishment · the identity and role of the witness · the description of the circumstances, too brief to be usable · the reference period of the wage certificate. These are not careless employers: these are five fields the portal does not explain.
What I have already done: the guidance is written, field by field, with a correct example and an insufficient one side by side — it is the comparison that teaches, not the instruction. And the check happens before sending: the employer sees what is missing while still in front of the screen, not three weeks later.
What that gives, measured over the trial quarter: incomplete reports go from 3,000 to 600 a year, and the 2,400 avoided are worth 1,000 caseworker hours — 25 minutes of back and forth per report, letter, waiting and reprocessing included.
And for the file, what matters even more: a report complete on first sending is handling that starts on the day of the accident and not three weeks later — on the trial files, the delay before the first complete document fell from 19 days to 2.
The next step I propose: telephone support for employers with fewer than ten employees, who file 2,100 reports a year and account for 44 % of incomplete files. They have no personnel department: the person reporting is often the person who was on site. I fill in the screen with them, they read it back and validate themselves — their credentials are never entrusted to me and do not have to be: it is their validation that makes the report. In the trial, 141 reports supported, 138 complete first time. employer-online-reporting_5-fields-78-percent.pdf3,000 incomplete down to 600, 19 days down to 2
⛓ Sourced · 9,400 portal reports over 12 months, missing fields, chaser log, one-quarter trial
The five, in order: the exact time of the accident · the exact place, when it is not the establishment · the identity and role of the witness · the description of the circumstances, too brief to be usable · the reference period of the wage certificate. These are not careless employers: these are five fields the portal does not explain.
What I have already done: the guidance is written, field by field, with a correct example and an insufficient one side by side — it is the comparison that teaches, not the instruction. And the check happens before sending: the employer sees what is missing while still in front of the screen, not three weeks later.
What that gives, measured over the trial quarter: incomplete reports go from 3,000 to 600 a year, and the 2,400 avoided are worth 1,000 caseworker hours — 25 minutes of back and forth per report, letter, waiting and reprocessing included.
And for the file, what matters even more: a report complete on first sending is handling that starts on the day of the accident and not three weeks later — on the trial files, the delay before the first complete document fell from 19 days to 2.
The next step I propose: telephone support for employers with fewer than ten employees, who file 2,100 reports a year and account for 44 % of incomplete files. They have no personnel department: the person reporting is often the person who was on site. I fill in the screen with them, they read it back and validate themselves — their credentials are never entrusted to me and do not have to be: it is their validation that makes the report. In the trial, 141 reports supported, 138 complete first time. employer-online-reporting_5-fields-78-percent.pdf3,000 incomplete down to 600, 19 days down to 2
⛓ Sourced · 9,400 portal reports over 12 months, missing fields, chaser log, one-quarter trial
And on the employees' side? Some do not even know their sick leave could have been a workplace accident.
They are the ones I go and find, and it is the work that pays the insured person and the body most: 214 situations found over the year.
Unclaimed entitlement is the gap between the people entitled to cover and those who obtain it: here it almost never comes from a refusal, it comes from a report never having been made.
What I compare, and these are your own files: a sick leave whose first medical record carries the same date as an accident reported by the employer but never linked to the file · a relapse reported as ordinary illness while a workplace accident appears in the insured person's file · an employee of an establishment where two comparable accidents were recognised within six months.
What that gives: 214 situations raised over the year. I reclassify nothing — I hand your caseworker the match, the two files side by side, and the information letter to the insured person already written: what a workplace accident report is, how it is made, within what time limit, and what it changes for them.
Of the 60 letters sent as a trial: 41 insured people made contact, 34 reports were regularised, and 27 cases were recognised by your officers. Twenty-seven employees who were entitled to occupational risks cover and did not know it.
And for the body, the same measure works the other way: those 27 files belong to occupational risks and not to health insurance — with the link restored, each is charged where it belongs, which is exactly what your accounting officer and your supervising authority expect of a handling service.
The next step I propose: the remaining 154 situations, in three waves a month apart. Your caseworkers will see about thirty regularisations a month — which is why I propose waves, and why those files arrive pre-assessed: at 4 minutes 30 instead of 27, they are absorbed without reinforcement. sick-leave-never-linked_214-situations.pdf41 contacts from 60 letters, 27 cases recognised
⛓ Sourced · 31,000 sick leave spells monitored, the year's accident files, matches by establishment, trial of 60 letters
Unclaimed entitlement is the gap between the people entitled to cover and those who obtain it: here it almost never comes from a refusal, it comes from a report never having been made.
What I compare, and these are your own files: a sick leave whose first medical record carries the same date as an accident reported by the employer but never linked to the file · a relapse reported as ordinary illness while a workplace accident appears in the insured person's file · an employee of an establishment where two comparable accidents were recognised within six months.
What that gives: 214 situations raised over the year. I reclassify nothing — I hand your caseworker the match, the two files side by side, and the information letter to the insured person already written: what a workplace accident report is, how it is made, within what time limit, and what it changes for them.
Of the 60 letters sent as a trial: 41 insured people made contact, 34 reports were regularised, and 27 cases were recognised by your officers. Twenty-seven employees who were entitled to occupational risks cover and did not know it.
And for the body, the same measure works the other way: those 27 files belong to occupational risks and not to health insurance — with the link restored, each is charged where it belongs, which is exactly what your accounting officer and your supervising authority expect of a handling service.
The next step I propose: the remaining 154 situations, in three waves a month apart. Your caseworkers will see about thirty regularisations a month — which is why I propose waves, and why those files arrive pre-assessed: at 4 minutes 30 instead of 27, they are absorbed without reinforcement. sick-leave-never-linked_214-situations.pdf41 contacts from 60 letters, 27 cases recognised
⛓ Sourced · 31,000 sick leave spells monitored, the year's accident files, matches by establishment, trial of 60 letters
Letters going out in our name to people on sick leave, on a subject touching their health: I will not sign that on a verbal agreement.
You do not have to: I propose a written mandate, capped, dated and withdrawable with one word — and it is already drafted, in front of you.
What the mandate authorises, and nothing else:
· Sending the information letters on workplace accident reporting, only to insured people whose match has been approved by a caseworker, in waves capped at 60 a month. The letter informs of a possibility and recognises nothing: it opens no entitlement and prejudges no decision.
· The letters opening the adversarial stage, with the list of documents consultable by their addressee alone, as soon as a file is complete.
· Acknowledgement of every document, within the hour, saying what it completes, what is still missing and the file's deadline.
· The replies to the twelve routine subjects, listed by name in the mandate. Any question outside that list reaches you with a reply already drafted and sourced, and the caseworker sends it.
What the mandate expressly leaves to a person's signature: classifying the accident as work-related and the notification that follows, charging it to the employer's account, extending the handling period, and any communication of medical material, which stays with the medical service. That is not a reservation, it is what gives those decisions their value — a signed decision is reasoned, dated and open to appeal, and the insured person and the employer who can challenge it are the ones who have been heard.
The formal safeguards, in four lines: every letter carries its source, its date and the statement that it was prepared by a digital assistant of the body · you receive each morning, on one page, the record of what went out the day before — a mistake is caught in an hour, not in three weeks · a review at three months, with the record of what it changed, and without an explicit decision at the review the mandate lapses: it is renewal that needs a signature, not stopping · withdrawal: one word, and direct sending stops within the minute.
What is ready today: the mandate, the information notice displayed at the front desk and published on your website, and the first wave of 60 letters. You sign, the service is running the next morning, and the review is already in your head of service's diary on the 15th of the third month. written-mandate_scope-cap-and-withdrawal.pdf60 letters a month maximum, review at 3 months, immediate withdrawal
✎ Framework · drafted mandate, list of the 12 subjects, consultation scopes, information notice, dispatch log
What the mandate authorises, and nothing else:
· Sending the information letters on workplace accident reporting, only to insured people whose match has been approved by a caseworker, in waves capped at 60 a month. The letter informs of a possibility and recognises nothing: it opens no entitlement and prejudges no decision.
· The letters opening the adversarial stage, with the list of documents consultable by their addressee alone, as soon as a file is complete.
· Acknowledgement of every document, within the hour, saying what it completes, what is still missing and the file's deadline.
· The replies to the twelve routine subjects, listed by name in the mandate. Any question outside that list reaches you with a reply already drafted and sourced, and the caseworker sends it.
What the mandate expressly leaves to a person's signature: classifying the accident as work-related and the notification that follows, charging it to the employer's account, extending the handling period, and any communication of medical material, which stays with the medical service. That is not a reservation, it is what gives those decisions their value — a signed decision is reasoned, dated and open to appeal, and the insured person and the employer who can challenge it are the ones who have been heard.
The formal safeguards, in four lines: every letter carries its source, its date and the statement that it was prepared by a digital assistant of the body · you receive each morning, on one page, the record of what went out the day before — a mistake is caught in an hour, not in three weeks · a review at three months, with the record of what it changed, and without an explicit decision at the review the mandate lapses: it is renewal that needs a signature, not stopping · withdrawal: one word, and direct sending stops within the minute.
What is ready today: the mandate, the information notice displayed at the front desk and published on your website, and the first wave of 60 letters. You sign, the service is running the next morning, and the review is already in your head of service's diary on the 15th of the third month. written-mandate_scope-cap-and-withdrawal.pdf60 letters a month maximum, review at 3 months, immediate withdrawal
✎ Framework · drafted mandate, list of the 12 subjects, consultation scopes, information notice, dispatch log
Management will ask what this has given us after a year. I need a figure I can defend before them and before the staff representatives.
7,678 officer hours given back to the service over the year — close to five full-time staff — and the calculation is yours, it fits in four lines.
The calculation, item by item, so you can redo it:
· Pre-assessment: 10,000 files, 27 minutes down to 4 minutes 30 — 60 % → 10 % of the file's time — that is 3,750 hours.
· Notifications and letters: 12,000 a year, 8 minutes down to 2 — 40 % → 10 % — that is 1,200 hours.
· Reviewing the caseload's deadlines: 2 hours a week per caseworker down to 24 minutes — 100 % → 20 % —, across 24 caseworkers and 45 weeks, that is 1,728 hours.
· Incomplete reports avoided: 2,400 a year, at 25 minutes of back and forth each, that is 1,000 hours — and 2,400 employers we had nothing to chase.
What those hours are, and this is what defends best before management and staff representatives alike: officer time given back to the service, at unchanged headcount — no post cut, no post created. Time given back is not a staffing saving: it is caseworker time going back to files with reservations, to circumstance investigations and to the insured people who need a call.
What those hours became, on your own records:
· Files reaching the deadline with no adversarial stage opened: 1,400 → 0.
· Instruction deadlines overrun: 780 → 41 — 739 files a year that stop being settled by the calendar and become decisions again.
· Observations received at the adversarial stage: 19 % → 41 % of files, because it opens early rather than the day before the deadline.
· Unanswered calls: 29 % → 4 %, and reply to an e-mail: 6 days → under 24 hours.
· Incomplete reports: 3,000 → 600, and delay before the first complete document: 19 days → 2.
· And 27 cases recognised for insured people whose sick leave had never been linked to their accident — that one is not counted in hours, and it is the only one your management will remember.
The next step I propose for the meeting: the calculation page is written and fits on one side — four lines of calculation, six delay measures, three framework measures. Send it with the agenda: a figure read the day before is discussed better than a figure discovered in the room. yearly-review_7678-hours-given-back.pdf60→10, 40→10, 100→20, and 780 overruns down to 41
⛓ Sourced · production log, instruction timetables, platform records, employer portal log
The calculation, item by item, so you can redo it:
· Pre-assessment: 10,000 files, 27 minutes down to 4 minutes 30 — 60 % → 10 % of the file's time — that is 3,750 hours.
· Notifications and letters: 12,000 a year, 8 minutes down to 2 — 40 % → 10 % — that is 1,200 hours.
· Reviewing the caseload's deadlines: 2 hours a week per caseworker down to 24 minutes — 100 % → 20 % —, across 24 caseworkers and 45 weeks, that is 1,728 hours.
· Incomplete reports avoided: 2,400 a year, at 25 minutes of back and forth each, that is 1,000 hours — and 2,400 employers we had nothing to chase.
What those hours are, and this is what defends best before management and staff representatives alike: officer time given back to the service, at unchanged headcount — no post cut, no post created. Time given back is not a staffing saving: it is caseworker time going back to files with reservations, to circumstance investigations and to the insured people who need a call.
What those hours became, on your own records:
· Files reaching the deadline with no adversarial stage opened: 1,400 → 0.
· Instruction deadlines overrun: 780 → 41 — 739 files a year that stop being settled by the calendar and become decisions again.
· Observations received at the adversarial stage: 19 % → 41 % of files, because it opens early rather than the day before the deadline.
· Unanswered calls: 29 % → 4 %, and reply to an e-mail: 6 days → under 24 hours.
· Incomplete reports: 3,000 → 600, and delay before the first complete document: 19 days → 2.
· And 27 cases recognised for insured people whose sick leave had never been linked to their accident — that one is not counted in hours, and it is the only one your management will remember.
The next step I propose for the meeting: the calculation page is written and fits on one side — four lines of calculation, six delay measures, three framework measures. Send it with the agenda: a figure read the day before is discussed better than a figure discovered in the room. yearly-review_7678-hours-given-back.pdf60→10, 40→10, 100→20, and 780 overruns down to 41
⛓ Sourced · production log, instruction timetables, platform records, employer portal log
They will also ask for your mistakes. On a workplace accident, a mistake is paid for by the insured person or by the employer.
It has happened, I counted it, and I publish it: of the 5,400 outputs of the first quarter, 151 had to be corrected on substance — 2.8 %. You deserve better than learning that figure from a dispute.
The real cause, measured and not assumed: 94 of the 151 corrections concerned letters built from four templates that had not been revised since your instructions last changed — the name of a service that changed, a route of appeal reworded, a document no longer needed but still requested. This was not a drafting error: I was faithfully reusing a template that had aged. The other 57 concerned situations with no precedent in your seven years of files — a posted worker, an accident on a third party's premises, sick leave begun abroad.
What I did about it, and it is measured: each template now carries the governing instruction and its date, and I write the updated version of any template a new instruction amends, within the 24 hours following its circulation.
The following quarter: 47 corrections on substance out of 5,200 outputs — 0.9 %. And the 47 are one-off situations, no aged template left.
The rule that holds all the rest: a value I have not read, I do not write — I ask for it, and I ask fast. A missing document does not become “probably provided”, an untraceable record date does not become “presumably the date of the accident”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its addressee. Over the quarter, 362 missing values, 362 requests prepared, 331 answers back within eight days.
And the protection that matters for your officers' signature: across 10,600 outputs over two quarters, 10,600 were approved by a person — and all 198 corrections were made before sending, none after. No notification sent had to be withdrawn.
The next step I propose: that the 57 one-off situations of the first quarter become 11 further templates, written, for your reading. Of the 47 corrections of the second quarter, 39 already fell within one of them — it is the same correction as the one on aged templates, taken one notch further. substantive-corrections_151-then-47.pdf2.8 % → 0.9 %, measured cause, 10,600 human approvals
⛓ Sourced · log of outputs and their corrections over two quarters, circulation of internal instructions
The real cause, measured and not assumed: 94 of the 151 corrections concerned letters built from four templates that had not been revised since your instructions last changed — the name of a service that changed, a route of appeal reworded, a document no longer needed but still requested. This was not a drafting error: I was faithfully reusing a template that had aged. The other 57 concerned situations with no precedent in your seven years of files — a posted worker, an accident on a third party's premises, sick leave begun abroad.
What I did about it, and it is measured: each template now carries the governing instruction and its date, and I write the updated version of any template a new instruction amends, within the 24 hours following its circulation.
The following quarter: 47 corrections on substance out of 5,200 outputs — 0.9 %. And the 47 are one-off situations, no aged template left.
The rule that holds all the rest: a value I have not read, I do not write — I ask for it, and I ask fast. A missing document does not become “probably provided”, an untraceable record date does not become “presumably the date of the accident”: I state what is missing, where I looked, who holds it, and I hand over the request already drafted to its addressee. Over the quarter, 362 missing values, 362 requests prepared, 331 answers back within eight days.
And the protection that matters for your officers' signature: across 10,600 outputs over two quarters, 10,600 were approved by a person — and all 198 corrections were made before sending, none after. No notification sent had to be withdrawn.
The next step I propose: that the 57 one-off situations of the first quarter become 11 further templates, written, for your reading. Of the 47 corrections of the second quarter, 39 already fell within one of them — it is the same correction as the one on aged templates, taken one notch further. substantive-corrections_151-then-47.pdf2.8 % → 0.9 %, measured cause, 10,600 human approvals
⛓ Sourced · log of outputs and their corrections over two quarters, circulation of internal instructions
Last question, and it is the one that will decide management: what you do without asking us, and what happens if we stop in a year.
Three actions without asking, not one more, each withdrawable with a word — and an exit that is checked before signing, not after.
· I keep each file's timetable and set the three alerts — fifteen days, seven days, two days. And the reverse is true too: an alert that becomes pointless goes out as soon as the decision is recorded, and a suspended file whose suspension has not been entered is flagged separately, because that is the one that goes back into overrun.
· I acknowledge every document within the hour, saying what it completes, what is still missing and the file's deadline. An acknowledgement recognises nothing and commits nothing: it says the submission has arrived, and it brought “did you get my certificate?” calls down from 480 to 61 a month.
· I hand you every Monday the week's record: files due within fifteen days, adversarial stages not opened, files suspended for more than thirty days, documents outstanding. It is the only thing I send of my own accord, and it goes only to the service.
And the four actions that stay with a person, because that is exactly what gives them their value: classifying the accident as work-related, decided and signed by the authorised officer on the medical service's opinion · the notification that follows · charging it to the employer's account · and any communication of medical material, which belongs to the medical service and to it alone. Across 10,600 outputs over two quarters, those four actions were taken 10,600 times by a person, and the log shows 0 medical access outside the medical service.
The exit, since management will ask: the timetables and the index are deleted, and they held no document content — only dates, types and states. Your seven years of files have not moved a byte: same systems, same permissions, same filing. The twelve answers, the rewritten notification templates, the discrepancy rule, the portal guidance and the standard timetables belong to the body: they are made of its own material, they stay in its files, readable without us. No migration on the way in, therefore none on the way out.
The next step I propose, so that this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check the service runs exactly as before, we switch back on. The protocol fits on one page, and the date that costs you least is a Thursday in mid-August — your records show accident numbers are then at their lowest of the year. Management will know what the promise is worth before committing a second year. what-the-agent-does-alone_and-the-dry-run-exit.pdf3 reversible actions, 4 signatures that stay with a person
✎ Framework · settings of the automatic actions, access and dispatch logs, export formats, dry-run exit protocol
· I keep each file's timetable and set the three alerts — fifteen days, seven days, two days. And the reverse is true too: an alert that becomes pointless goes out as soon as the decision is recorded, and a suspended file whose suspension has not been entered is flagged separately, because that is the one that goes back into overrun.
· I acknowledge every document within the hour, saying what it completes, what is still missing and the file's deadline. An acknowledgement recognises nothing and commits nothing: it says the submission has arrived, and it brought “did you get my certificate?” calls down from 480 to 61 a month.
· I hand you every Monday the week's record: files due within fifteen days, adversarial stages not opened, files suspended for more than thirty days, documents outstanding. It is the only thing I send of my own accord, and it goes only to the service.
And the four actions that stay with a person, because that is exactly what gives them their value: classifying the accident as work-related, decided and signed by the authorised officer on the medical service's opinion · the notification that follows · charging it to the employer's account · and any communication of medical material, which belongs to the medical service and to it alone. Across 10,600 outputs over two quarters, those four actions were taken 10,600 times by a person, and the log shows 0 medical access outside the medical service.
The exit, since management will ask: the timetables and the index are deleted, and they held no document content — only dates, types and states. Your seven years of files have not moved a byte: same systems, same permissions, same filing. The twelve answers, the rewritten notification templates, the discrepancy rule, the portal guidance and the standard timetables belong to the body: they are made of its own material, they stay in its files, readable without us. No migration on the way in, therefore none on the way out.
The next step I propose, so that this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check the service runs exactly as before, we switch back on. The protocol fits on one page, and the date that costs you least is a Thursday in mid-August — your records show accident numbers are then at their lowest of the year. Management will know what the promise is worth before committing a second year. what-the-agent-does-alone_and-the-dry-run-exit.pdf3 reversible actions, 4 signatures that stay with a person
✎ Framework · settings of the automatic actions, access and dispatch logs, export formats, dry-run exit protocol
Local inference · no data outside the EU
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Use cases
The uses of AI in handling workplace injury claims
Each use corresponds to an agent we deploy. All work in support, subject to the caseworker's approval — the classification stays human.
Included in your agent
The 7 capabilities essential to this promise are included, at no extra cost.
Pre-assessment of work accident and occupational disease declarations
Pre-assess accident reports and sick leave (documents, statutory deadlines) and flag discrepancies.
Checking the documents in a file
Check reports, certificates and salary statements; detect missing documents and prepare the requests.
Letters & notifications
Prepare the letters to claimants and employers from your templates, decision fields left to the caseworker.
Tracking the classification deadlines
Watch the assessment deadlines file by file and raise an alert before each statutory cut-off.
Spotting inconsistencies
Flag discrepancies between the report, the certificate and the salary statement for a human check — never an automatic decision.
Answers to claimants & employers
Inform on the progress of files and the procedures (report, documents, daily benefits), around the clock.
Help with online reporting
Guide employers through their online reports and reduce incomplete files at source.
Controls and safeguards
These 7 controls are built into the agent: they frame what it does, whatever plan you pick. They are not chosen and are not added to your order.
Human validation, exceptions and escalation
Status, safe closure and audit trail
Sources, access rights and handling of questions with no answer
Prepare without deciding: sourced rules, supporting documents, anomalies and explanation
Guarantee the decision, signature, recourse and responsibility of the public official
Record versions, access, criteria, actions and notifications
Test for bias, false positives, fundamental rights and continuity of service
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Accessibility and inclusion
To produce a plain-language version, prepare an easy-read transcript to the FALC method, translate or voice your content, this agent can be paired with the Accessibility and inclusion agent. None of these capabilities is included in what this offer covers as standard.
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.
Claims handling agent (insurance, tier 1)
Protocol, documents.
Claims handling agent (insurance, tier 1) from 765 € incl. VAT / month Discover the agent →Claims handling agent (full)
Protocol, documents, workflow.
Claims handling agent (full) from 771 € 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 workplace injury service win back?
By automating the pre-assessment, the checking of documents and the tracking of deadlines, a service can aim for a clear reduction in assembly time per file — reinvested in classification and in disputed files.
Pre-assessment of a workplace accident report
A standard notification to a claimant or employer
Checking the assessment deadlines across the caseload
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
An agent for handling workplace injury claims and sick leave (pre-assessment, documents, notifications, deadlines), installed and operated for you. Choose according to how you are organised and how demanding your security requirements are, with HDS as an option.
This agent is priced with you, not online. We are adjusting its scope at the moment, and online subscription stays closed while we do. Tell us what you need: we will come back to you with a price. Request a quote
Our commitment
Four guarantees that matter to a social security body
Medical data stays isolatedHDS hosting, minimisation and restricted access; local inference possible for the most sensitive information.
Data in France, under French lawNative location; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Classification stays humanThe AI agent pre-assesses and prepares; recognising an injury as work-related remains the caseworker's decision, on medical advice.
Human oversight & traceabilityPre-assessment of accident reports: every change to the service is logged, in line with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent classify the workplace accident?
No: it pre-assesses and prepares; the classification and the decision remain those of the caseworker concerned.
How is health data protected?
HDS hosting, an isolated resource, minimisation and logging. GDPR: governed deployment.
Does the agent decide whether an accident is work-related?
No: it prepares the file; the classification remains a decision of the caseworker, on medical advice.
Is medical data isolated?
Yes: HDS hosting is available, with restricted and recorded access, and local inference for the most sensitive information.
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, working with your IT department on the integration. Caseworkers use it after a few hours of familiarisation.
Do we have to change software?
No. The agent complements your existing case management applications, with no migration imposed.
Which tools can insured people and employers use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: an insured person or an employer has no account to create and no application to install. A workplace-accident file carries health data: the agent gives information on progress and on the documents expected, never on medical content, which stays with the department and the medical adviser. 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 social protection and health
Let's talk
Let's size up the potential in your organisation
A few minutes to identify the most useful use case — HDS hosting in France, supervised, with no commitment.