● Public sector — Court clerk
AI agent for court clerks (judicial services)
An assistant that prepares procedural acts, summonses and the keeping of the registers — the clerk authenticates and the judge decides. A strictly administrative scope: no support for judicial decisions, ever, by design and in line with the AI Act. Hosted in France, on local inference or an isolated resource.
Hosted in France
Confidentiality of proceedings preserved
GDPR & AI Act: governed deployment
Never support for judicial decisions
Updated on
Deployed in a few weeks
Registry assistant · hosted in France
Prepare the summonses to the parties for the hearing on 12 March, from our template and the list.
Draft summonses ready for the 14 cases on the list: identity of the parties, case number, date, time and courtroom, the mandatory wording from your template. Two incomplete addresses flagged for checking against the file.
For the registry to check before service.
⛓ Source · the hearing list + your registry templates
For the registry to check before service.
⛓ Source · the hearing list + your registry templates
Which cases have procedural deadlines expiring in the next 15 days?
Schedule established: 6 cases concerned — 3 deadlines for filing submissions, 2 appeal deadlines, 1 expert report awaited. Each line refers to the case number and the document concerned.
Alerts prepared, with no step taken without your check.
✎ Action · schedule to verify — the registry checks
Alerts prepared, with no step taken without your check.
✎ Action · schedule to verify — the registry checks
Local inference · no data outside the EU
Data hosted in France
Sovereign by designLocal inference or hosting in France
GDPR & AI Act: governed deploymentTraceability & human oversight
TurnkeyDesigned, installed and operated for you
The registry checksAdministrative support, never the substance of cases
✦ In brief
In the registry, a Blue Lemon Agent agent assists clerks with the repetitive administrative tasks — procedural acts, summonses, notifications, keeping the registers, tracking deadlines and hearings — and answers court users' everyday questions about procedure. Its scope is strictly administrative: never any support for judicial decisions — excluded by design and by the AI Act. It runs on local inference or is hosted in France on a isolated resource, preserving the confidentiality of proceedings. The clerk authenticates, the judge decides. Live within a few weeks. Your public-sector staff write to it from Microsoft Teams, Slack or their email, and court users 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 court registries — and why justice demands guarantees
Acts, summonses, registers, deadlines: the registry carries the administrative machinery of the court, under constant pressure. The data handled — case documents, parties, proceedings — is among the most protected in French law.
! The issue
Clerks are caught between ever fuller hearing lists and a production of acts — summonses, notifications, registers, schedules — that leaves no time for anything else. Yet most consumer AI tools amount to entrusting case documents, the identity of parties and matters covered by the confidentiality of proceedings to a third party, often hosted outside Europe and subject to the Cloud Act.
✓ Our answer
For judicial data, AI is only of interest if it is sovereign, confidential and strictly bounded. Local inference or a isolated resource hosted in France, systematic human oversight, a scope limited to administrative support for the registry: no support for judicial decisions, no involvement in the substance of cases. The clerk authenticates, the judge decides — the AI agent only prepares documents that can be checked.
The decisive point
Protecting procedural data: sovereignty & compliance
A registry handles data covered by the confidentiality of proceedings and by professional secrecy. Here is how the architecture of our agents protects it — and how its scope is strictly bounded.
Local inference
The agent can run on a machine at the court: no data leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — the procedural data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
As regards case documents, exposure to the Cloud Act and FISA 702 is reduced by design; location alone does not guarantee immunity.
A resource isolated per court
No pooling of data: an environment strictly dedicated to your court, guaranteeing the continuity of the service.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
Strictly administrative support; no support for judicial decisions; traceability and human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For the most sensitive data, SecNumCloud and HDS options are available depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
4 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The public body in this demonstration
Fictional public bodyVallonges judicial court — civil registry office (fictional court)
- Sector
- Registry of a mid-sized judicial court — civil service: small claims, urgent applications, family cases; 4 courtrooms, 312 hearings held in the year, 3,240 new cases entered in the general register in 2025
- Headcount
- 61 staff in the registry, 9 of them in the civil service covered by the pilot — 6 court clerks, 2 administrative assistants and one head of registry; the criminal service and the single court user reception desk are outside the scope, and the agent does not touch them
- Public served
- 3,240 new cases and 2,860 pending cases — 14,600 enquiries a year in the civil service: 8,400 phone calls, 3,900 emails, 1,700 visits to the single court user reception desk, 600 website forms
- Order of magnitude
- 4,180 summonses prepared a year, 6,900 notifications drawn from the registry templates, 11,400 entries made in the registers, 2,940 procedural time limits to check across the pending files
- Tools already in place
- The court's civil application (general register, hearing list, register of entries), registry templates settled by the head of registry, exchange platform with legal professionals, mailbox and Microsoft Teams — the agent reads them, nothing is replaced and nothing is migrated
- Who decides what
- The judge decides; the court clerk authenticates and signs the act; the head of registry settles the templates and approves every batch before service. The agent never touches the merits of a case
- Room for improvement
- 2,688 calls out of 8,400 ring unanswered on hearing days — 32%; an email waits 11 days on average; checking time limits across the 2,860 pending files ties up 2 clerks for 5 working days every month; and the hearing list had never been reconciled against the register of entries — 27 discrepancies were sitting undisturbed in January alone
At the Vallonges judicial court, nine civil registry staff handle 312 hearings, 4,180 summonses and 2,860 pending files. The most repetitive questions from court users — where is my case, which documents to produce, how to bring proceedings — take up more than three quarters of the flow, and the situations that need a clerk are the ones waiting behind. The exchanges below cover a year, from the reconciliation of the hearing list and the registers to the report presented at the annual management review. The agent is hosted in France on a isolated resource, reads the civil application without ever writing to it, and its scope is strictly administrative: it prepares acts that the clerk rereads, authenticates and signs, and it never touches the merits of a case.
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.
Assistant greffe · hébergé en France
I have reread your 14,600 enquiries for the year and reconciled, file by file, your hearing list against the register of entries and the time-limit schedule — the reconciliation that listing hearings never leaves time for.
What the count says: 12 topics cover 11,240 enquiries out of 14,600, i.e. 77.0% of the flow. Where my case stands, documents to produce and how to file them, how to bring proceedings, the summons received, legal aid, copies of judgments, representation, adjournment and absence from a hearing, reception desk access and hours, fees and expert deposits, service of documents on the other party, where to find appeal periods. The 12 reference answers are written, sourced on your public notices and on the case allocation order of 1 January, and dated — they are waiting for you.
What the three items you can check against your own logs are worth:
· A summons takes 6 minutes end to end, 60% of which goes on carrying across the list, the identities and the mandatory wording of the template — 3 minutes 36.
· A standard notification from a template takes 8 minutes, 40% of it on the drafting itself — 3 minutes 12.
· Checking time limits is 100% manual work: 2 clerks for 5 working days every month, i.e. 70 hours a month and 840 hours a year.
And the gap the reconciliation brought out, in January alone — 312 cases listed: 27 discrepancies, never seen because nobody ever had the days to look. 11 adjournment entries recorded on the list and missing from the register, 7 summonses whose acknowledgement of receipt never came back with no reminder logged, 5 files where the hearing date differs between the list and the summons served, 4 filing deadlines passed with no entry on the file.
What that represents, and it is not a management figure: 324 lines over the year, each of which is a hearing incident waiting to happen — an adjournment, a challenge in court, or a party travelling for nothing.
The next step, and it takes half an hour: you reread the 27 lines and tell me, for each one, the action you want. The head of registry settles the version, and the 27 are put right before the March listing. civil-registry-flow_14600-enquiries-12-topics.pdf11,240 enquiries across 12 topics, the three items costed, 2,562 hours identified
⛓ Sourced · 12 months of enquiry log, hearing list, register of entries, time-limit schedule of pending files
What the count says: 12 topics cover 11,240 enquiries out of 14,600, i.e. 77.0% of the flow. Where my case stands, documents to produce and how to file them, how to bring proceedings, the summons received, legal aid, copies of judgments, representation, adjournment and absence from a hearing, reception desk access and hours, fees and expert deposits, service of documents on the other party, where to find appeal periods. The 12 reference answers are written, sourced on your public notices and on the case allocation order of 1 January, and dated — they are waiting for you.
What the three items you can check against your own logs are worth:
· A summons takes 6 minutes end to end, 60% of which goes on carrying across the list, the identities and the mandatory wording of the template — 3 minutes 36.
· A standard notification from a template takes 8 minutes, 40% of it on the drafting itself — 3 minutes 12.
· Checking time limits is 100% manual work: 2 clerks for 5 working days every month, i.e. 70 hours a month and 840 hours a year.
And the gap the reconciliation brought out, in January alone — 312 cases listed: 27 discrepancies, never seen because nobody ever had the days to look. 11 adjournment entries recorded on the list and missing from the register, 7 summonses whose acknowledgement of receipt never came back with no reminder logged, 5 files where the hearing date differs between the list and the summons served, 4 filing deadlines passed with no entry on the file.
What that represents, and it is not a management figure: 324 lines over the year, each of which is a hearing incident waiting to happen — an adjournment, a challenge in court, or a party travelling for nothing.
The next step, and it takes half an hour: you reread the 27 lines and tell me, for each one, the action you want. The head of registry settles the version, and the 27 are put right before the March listing. civil-registry-flow_14600-enquiries-12-topics.pdf11,240 enquiries across 12 topics, the three items costed, 2,562 hours identified
⛓ Sourced · 12 months of enquiry log, hearing list, register of entries, time-limit schedule of pending files
27 discrepancies in one month — what does that mean over a year? And how do we stop it happening again?
Over a year, the same reconciliation rule brings out 324. What it costs is not counted in euros but in hearings: a date differing between the list and the summons means a party turning up on the wrong day; a missing adjournment entry means a file being hunted for in court.
What I did on top, and what listing hearings never leaves time for: I wrote four control rules, in the language of your templates, and ran them over your last 24 months of hearing list and register. For each one I give you the number of alerts it would have produced, the share confirmed after review, and what it would have let through — you choose on figures, not on a hunch:
· Rule A — a case on the list whose hearing date differs from the one on the summons served: 132 alerts over 24 months, 128 confirmed — 97.0%. The other 4 are adjournments entered on the list before the summons was reissued, which is fixed at data entry.
· Rule B — a summons served with no acknowledgement of receipt recorded 15 days after dispatch: 268 alerts, 241 confirmed — 89.9%.
· Rule C — an entry made on the list with no matching entry in the register: 186 alerts, 186 confirmed — 100%.
Total: 586 alerts over 24 months, 555 confirmed.
And here is what those three rules let through, because you should know it before you sign: 104 deadlines for filing submissions passed with no entry on the file fall into none of the three. So I wrote rule D — a deadline on the schedule that has passed with no matching entry — and ran it over the same 24 months: 104 alerts, 93 confirmed. The four rules together cover 648 of the 703 discrepancies found over the period — 92.1%, i.e. 324 a year, exactly the figure from the reconciliation.
Bringing a rule into service stays with the registry: a rule only goes live once approved by the head of registry. I save you the writing and the measuring; the decision takes half an hour instead of a working group. list-and-register-discrepancies_4-rules-tested-over-24-months.pdfThe 27 January discrepancies in detail, 4 rules tested, 648 of 703 discrepancies covered
⛓ Sourced · 24 months of hearing list and register of entries, time-limit schedule of pending files
What I did on top, and what listing hearings never leaves time for: I wrote four control rules, in the language of your templates, and ran them over your last 24 months of hearing list and register. For each one I give you the number of alerts it would have produced, the share confirmed after review, and what it would have let through — you choose on figures, not on a hunch:
· Rule A — a case on the list whose hearing date differs from the one on the summons served: 132 alerts over 24 months, 128 confirmed — 97.0%. The other 4 are adjournments entered on the list before the summons was reissued, which is fixed at data entry.
· Rule B — a summons served with no acknowledgement of receipt recorded 15 days after dispatch: 268 alerts, 241 confirmed — 89.9%.
· Rule C — an entry made on the list with no matching entry in the register: 186 alerts, 186 confirmed — 100%.
Total: 586 alerts over 24 months, 555 confirmed.
And here is what those three rules let through, because you should know it before you sign: 104 deadlines for filing submissions passed with no entry on the file fall into none of the three. So I wrote rule D — a deadline on the schedule that has passed with no matching entry — and ran it over the same 24 months: 104 alerts, 93 confirmed. The four rules together cover 648 of the 703 discrepancies found over the period — 92.1%, i.e. 324 a year, exactly the figure from the reconciliation.
Bringing a rule into service stays with the registry: a rule only goes live once approved by the head of registry. I save you the writing and the measuring; the decision takes half an hour instead of a working group. list-and-register-discrepancies_4-rules-tested-over-24-months.pdfThe 27 January discrepancies in detail, 4 rules tested, 648 of 703 discrepancies covered
⛓ Sourced · 24 months of hearing list and register of entries, time-limit schedule of pending files
We are talking about procedural documents and the identities of parties. Some of our files are covered by secrecy. Where does that data go?
Nowhere outside your walls, and that can be checked rather than promised.
Where I run: on a isolated resource hosted in France, under French law, dedicated to your court and to no other — or in local inference on a machine in the court, in which case nothing leaves the network at all. No procedural document is passed to a foreign service: exposure to the Cloud Act and FISA 702 is reduced accordingly, without location alone ruling it out — it is documented for the configuration chosen. For the most sensitive files, local inference is the shortest answer to the question you are asking.
What I read, and what I cannot reach: I have read-only access, chamber by chamber and service by service, to the general register, the hearing list, the register of entries and the time-limit schedule. I have no write access to the civil application: I produce draft acts and draft entries, your clerk rereads, authenticates and enters them. The criminal service files are not open to me — they are outside the pilot's scope, and an access with no use is not opened, in line with article 5.1(c) of the GDPR. Access is withdrawn on a word, and the withdrawal takes effect instantly.
What I do with the content of a document, and this is the point that concerns you: to prepare a summons or an entry, I take the act's identifying elements — case number, capacity and identity of the parties, date and nature of the act, date of service. The statement of facts and the parties' arguments serve none of the tasks you have entrusted to me: they are not extracted, not summarised and not retained. That is the technical translation of the scope announced: administrative support, no involvement in the merits.
What you can show your data protection officer and your court president: every read is logged with its date, its purpose and the file concerned; every output carries the documents it is drawn from; and the record of processing activities is already drafted, legal basis by legal basis — public interest task for administrative support to the registry, article 6.1(e).
The next step I suggest: a quarterly thirty-minute review in which I show you the accesses actually used and those that served no purpose — an unused access is an access to close, and I bring you the list rather than asking you for it. sovereign-framework_hosting-access-scope.pdfWhere procedural documents run, who reads what, the boundary of the administrative scope
✎ Framework · hosting architecture, role-based access log, registry record of processing activities
Where I run: on a isolated resource hosted in France, under French law, dedicated to your court and to no other — or in local inference on a machine in the court, in which case nothing leaves the network at all. No procedural document is passed to a foreign service: exposure to the Cloud Act and FISA 702 is reduced accordingly, without location alone ruling it out — it is documented for the configuration chosen. For the most sensitive files, local inference is the shortest answer to the question you are asking.
What I read, and what I cannot reach: I have read-only access, chamber by chamber and service by service, to the general register, the hearing list, the register of entries and the time-limit schedule. I have no write access to the civil application: I produce draft acts and draft entries, your clerk rereads, authenticates and enters them. The criminal service files are not open to me — they are outside the pilot's scope, and an access with no use is not opened, in line with article 5.1(c) of the GDPR. Access is withdrawn on a word, and the withdrawal takes effect instantly.
What I do with the content of a document, and this is the point that concerns you: to prepare a summons or an entry, I take the act's identifying elements — case number, capacity and identity of the parties, date and nature of the act, date of service. The statement of facts and the parties' arguments serve none of the tasks you have entrusted to me: they are not extracted, not summarised and not retained. That is the technical translation of the scope announced: administrative support, no involvement in the merits.
What you can show your data protection officer and your court president: every read is logged with its date, its purpose and the file concerned; every output carries the documents it is drawn from; and the record of processing activities is already drafted, legal basis by legal basis — public interest task for administrative support to the registry, article 6.1(e).
The next step I suggest: a quarterly thirty-minute review in which I show you the accesses actually used and those that served no purpose — an unused access is an access to close, and I bring you the list rather than asking you for it. sovereign-framework_hosting-access-scope.pdfWhere procedural documents run, who reads what, the boundary of the administrative scope
✎ Framework · hosting architecture, role-based access log, registry record of processing activities
Prepare the summonses for the parties for the hearing of 12 March, from our template and the hearing list.
Draft summonses ready for the 14 cases listed for 12 March. Each one carries the identity and capacity of the parties, the case number, the date, time and courtroom, and the mandatory wording of your template — the template settled by the head of registry on 1 January, of which I take the version in force at the date of the act, never another.
Two incomplete addresses flagged for checking against the file: case 26/00412 — the defendant appears at two conflicting addresses, the one on the claim form and the one on the last act received, and I give you both with their dates; case 26/00587 — the party is domiciled at a third party's address and the domicile wording is missing from the record sheet. Those two do not go out: they go up to the clerk with the document that raises the question.
What I check before handing them over, and what used to take the longest: the date and courtroom of each case are checked against the hearing list, not against the file — that is exactly the discrepancy rule A measured at 128 confirmed cases over 24 months, and it is the one that makes a party turn up on the wrong day.
What this is worth over a year: your 4,180 summonses go from 3 minutes 36 of carrying across and typing to 36 seconds — 60% of the time down to 10%, i.e. 209 hours returned to the registry. And your 6,900 notifications from templates go from 3 minutes 12 of drafting to 48 seconds — 40% down to 10%, i.e. 276 hours.
What stays with the clerk, and it is not a formality: rereading, authentication and signature. The clerk's signature is what gives the act its force; it belongs to the clerk and cannot be delegated. What a mandate can cover is the dispatch of a batch already signed: once the 14 summonses have been reread and signed, I send them, track the acknowledgements of receipt and chase the undelivered ones — instead of half a day of an assistant's time on following up returns.
The next step I suggest: that I hand you, the day before each hearing, the list of cases whose acknowledgement of receipt has not come back. Over the last 24 months, 241 summonses were in that position, and 68 led to an adjournment — knowing it the day before rather than in court means 68 situations to handle differently. summonses_hearing-of-12-march_14-cases.pdfThe 14 drafts, the template wording, the 2 addresses referred to the clerk
⛓ Sourced · hearing list of 12 March, registry templates settled on 1 January, record sheets of the 14 cases
Two incomplete addresses flagged for checking against the file: case 26/00412 — the defendant appears at two conflicting addresses, the one on the claim form and the one on the last act received, and I give you both with their dates; case 26/00587 — the party is domiciled at a third party's address and the domicile wording is missing from the record sheet. Those two do not go out: they go up to the clerk with the document that raises the question.
What I check before handing them over, and what used to take the longest: the date and courtroom of each case are checked against the hearing list, not against the file — that is exactly the discrepancy rule A measured at 128 confirmed cases over 24 months, and it is the one that makes a party turn up on the wrong day.
What this is worth over a year: your 4,180 summonses go from 3 minutes 36 of carrying across and typing to 36 seconds — 60% of the time down to 10%, i.e. 209 hours returned to the registry. And your 6,900 notifications from templates go from 3 minutes 12 of drafting to 48 seconds — 40% down to 10%, i.e. 276 hours.
What stays with the clerk, and it is not a formality: rereading, authentication and signature. The clerk's signature is what gives the act its force; it belongs to the clerk and cannot be delegated. What a mandate can cover is the dispatch of a batch already signed: once the 14 summonses have been reread and signed, I send them, track the acknowledgements of receipt and chase the undelivered ones — instead of half a day of an assistant's time on following up returns.
The next step I suggest: that I hand you, the day before each hearing, the list of cases whose acknowledgement of receipt has not come back. Over the last 24 months, 241 summonses were in that position, and 68 led to an adjournment — knowing it the day before rather than in court means 68 situations to handle differently. summonses_hearing-of-12-march_14-cases.pdfThe 14 drafts, the template wording, the 2 addresses referred to the clerk
⛓ Sourced · hearing list of 12 March, registry templates settled on 1 January, record sheets of the 14 cases
Which files have procedural time limits expiring in the next 15 days?
Schedule drawn up: 6 files concerned.
· 3 deadlines for filing submissions — cases 26/00298, 26/00341, 26/00455 — each with the case management judge's order that sets it and its date of service.
· 2 appeal periods — cases 25/01187 and 25/01204 — with the judgment, its date of service and the act of service itself.
· 1 expert report awaited — case 25/00932 — with the terms of reference, the filing date set and the last exchange with the expert.
Every line links back to the case number and to the document concerned, and that is what makes it checkable in ten seconds rather than ten minutes.
The rule I hold to, and I put it to you as the service's rule: every time limit logged traces back to the act that sets it and to its date of service. Without those two elements, the line does not go out as an alert. And where two readings are possible — a disputed starting point, an act served on two dates, a party disputing the date of receipt — the line goes to the clerk with both hypotheses and both dates, never with an answer. Of the year's 2,940 time limits, 84 are in that position — under 3%, and they are exactly the 84 that deserve a clerk's eye. I save you the searching; the legal characterisation is yours, and it is the only one that is any use.
What this is worth: checking used to tie up 2 clerks for 5 working days a month — 70 hours, 840 hours a year. Now: 14 hours a month of checking and characterisation, i.e. 20% of the current time, 168 hours a year. 672 hours returned to the registry. And it is no longer a monthly check: it is a state of play updated every morning across the 2,860 pending files.
The figure that does not flatter me, and I publish it: of the 1,470 time limits logged in the first quarter, 38 carried a wrong starting date — 2.6%. All of them came from acts served where the date of delivery differed from the date of the act: I was reading the date of the act. That is fixed: I now read the date carried by the act of service, and I refer the gap to the clerk when the two differ. In the second quarter, 4 wrong dates out of 1,470 — 0.3%. I suggest we keep publishing that rate every quarter: an agent whose errors are not measured is an agent that cannot be corrected. schedule-and-register_2940-time-limits-11400-entries.pdfThe 6 files of the fortnight, the two-hypotheses rule, the published error rate
⛓ Sourced · time-limit schedule of the 2,860 pending files, case management orders, acts of service on the files
· 3 deadlines for filing submissions — cases 26/00298, 26/00341, 26/00455 — each with the case management judge's order that sets it and its date of service.
· 2 appeal periods — cases 25/01187 and 25/01204 — with the judgment, its date of service and the act of service itself.
· 1 expert report awaited — case 25/00932 — with the terms of reference, the filing date set and the last exchange with the expert.
Every line links back to the case number and to the document concerned, and that is what makes it checkable in ten seconds rather than ten minutes.
The rule I hold to, and I put it to you as the service's rule: every time limit logged traces back to the act that sets it and to its date of service. Without those two elements, the line does not go out as an alert. And where two readings are possible — a disputed starting point, an act served on two dates, a party disputing the date of receipt — the line goes to the clerk with both hypotheses and both dates, never with an answer. Of the year's 2,940 time limits, 84 are in that position — under 3%, and they are exactly the 84 that deserve a clerk's eye. I save you the searching; the legal characterisation is yours, and it is the only one that is any use.
What this is worth: checking used to tie up 2 clerks for 5 working days a month — 70 hours, 840 hours a year. Now: 14 hours a month of checking and characterisation, i.e. 20% of the current time, 168 hours a year. 672 hours returned to the registry. And it is no longer a monthly check: it is a state of play updated every morning across the 2,860 pending files.
The figure that does not flatter me, and I publish it: of the 1,470 time limits logged in the first quarter, 38 carried a wrong starting date — 2.6%. All of them came from acts served where the date of delivery differed from the date of the act: I was reading the date of the act. That is fixed: I now read the date carried by the act of service, and I refer the gap to the clerk when the two differ. In the second quarter, 4 wrong dates out of 1,470 — 0.3%. I suggest we keep publishing that rate every quarter: an agent whose errors are not measured is an agent that cannot be corrected. schedule-and-register_2940-time-limits-11400-entries.pdfThe 6 files of the fortnight, the two-hypotheses rule, the published error rate
⛓ Sourced · time-limit schedule of the 2,860 pending files, case management orders, acts of service on the files
And the registers? That is the invisible part: nobody sees it while it is kept up, everybody sees it when it is not.
11,400 entries a year, and it is the item that degrades quietly because it produces no immediate complaint.
What I take on: at every event on a file — registration, adjournment, striking out, joinder, filing of a document, appointment of a legal professional, transfer — I prepare the draft entry in the form of your register, with its date, its nature, the document that grounds it and the case number. The clerk rereads and makes the entry: it is a registry act, and it stays one. What I make disappear is the retyping, not the act.
What continuous reconciliation brings, and what did not exist before: the gap between the list and the register is picked up the same day, no longer six months later. The 186 missing entries rule C found over 24 months would all have been flagged within 24 hours. In January, the 11 missing adjournment entries were put right in a single morning — the searching was done, only the act remained.
And the reverse holds too, because it is what makes a register reliable: a draft entry you set aside leaves a trace — the reason and the date it was set aside. After a quarter, that gives you the list of cases where my reading of your nomenclature is wrong, and therefore the drafts to correct. In the first quarter, 41 set aside, 29 of them on the same type of event: joinder of proceedings, which I was entering under too general a heading. The nomenclature has been corrected; 3 set aside in the second quarter.
The next step I suggest: a weekly statement of pending files where no entry has moved for more than six months. There are 214 of them across your 2,860 files. They are not lost files — they are files where nobody can say, this morning, where they stand. The list is read in an hour, and it belongs to you. schedule-and-register_2940-time-limits-11400-entries.pdfThe 11,400 entries, the nomenclature corrected from set-asides, the 214 dormant files
⛓ Sourced · register of entries over 24 months, hearing list, log of draft entries set aside
What I take on: at every event on a file — registration, adjournment, striking out, joinder, filing of a document, appointment of a legal professional, transfer — I prepare the draft entry in the form of your register, with its date, its nature, the document that grounds it and the case number. The clerk rereads and makes the entry: it is a registry act, and it stays one. What I make disappear is the retyping, not the act.
What continuous reconciliation brings, and what did not exist before: the gap between the list and the register is picked up the same day, no longer six months later. The 186 missing entries rule C found over 24 months would all have been flagged within 24 hours. In January, the 11 missing adjournment entries were put right in a single morning — the searching was done, only the act remained.
And the reverse holds too, because it is what makes a register reliable: a draft entry you set aside leaves a trace — the reason and the date it was set aside. After a quarter, that gives you the list of cases where my reading of your nomenclature is wrong, and therefore the drafts to correct. In the first quarter, 41 set aside, 29 of them on the same type of event: joinder of proceedings, which I was entering under too general a heading. The nomenclature has been corrected; 3 set aside in the second quarter.
The next step I suggest: a weekly statement of pending files where no entry has moved for more than six months. There are 214 of them across your 2,860 files. They are not lost files — they are files where nobody can say, this morning, where they stand. The list is read in an hour, and it belongs to you. schedule-and-register_2940-time-limits-11400-entries.pdfThe 11,400 entries, the nomenclature corrected from set-asides, the 214 dormant files
⛓ Sourced · register of entries over 24 months, hearing list, log of draft entries set aside
On hearing days the registry phone never stops and nobody can pick up. One court user in three gets nothing.
2,688 calls out of 8,400 ring unanswered — 32.0%, and most of them fall on hearing days, when your clerks are in court. This is not a quality-of-service problem, it is a time-of-day problem: the registry is busiest at the very moment it is least available.
What I take, and what I leave you: I pick up every call, I announce in the first sentence that I am an artificial intelligence — article 50(1) of the European AI Regulation has required this since 2 August 2026, and the court user can ask for a human at any moment — and I handle to the end the 12 topics whose answers are approved by the head of registry. All of them are within the administrative scope: where the case stands, which documents to produce, how to bring proceedings, what the summons received says, where to file a legal aid application, how to obtain a copy of a judgment.
Anything outside those 12 topics goes to a clerk with the file already assembled: the case number, the chamber, the hearing date, the latest entries made, the question as it was asked. Your clerk picks up already knowing, instead of picking up and searching.
And the boundary is held in every exchange, not just in the notice: any question about the merits of a case — the chances of success, the weight of an argument, what the judge will decide — is redirected, with the sentence that explains it: those questions are for the judge and, for advice, for a lawyer. The court user leaves with the route: the legal advice centre, the bar association, the legal aid office, and their contact details. A redirection that gives the address is not a door closed.
The arithmetic, and it is yours: the 12 topics account for 77.0% of the flow; 2,688 unanswered calls drop to 168, i.e. 2,520 more court users getting their answer — 1,790 of them in the evening, at the weekend or over lunch, when the reception desk is closed. Email, which today waits 11 days, goes back to 1 day. And I answer on WhatsApp Business, the website chat and email, with no account to create and no app to install; your staff write to me from Microsoft Teams, Slack or their mailbox.
The figure that does not flatter me, and I publish it: of the 1,860 requests in the first quarter, I directed 132 to the wrong chamber — 7.1%. All of them concerned the 2 chambers whose caseload allocation changed in the case allocation order of 1 January: I was reading the allocation in force, not the one applicable at the date proceedings were brought. That is fixed: I now read the allocation as it stood on the date the case was registered. In the second quarter, 11 wrong referrals out of 1,860 — 0.6%. court-user-channels_multichannel-reception-and-plain-language.pdfThe 12 topics handled to the end, the merits boundary held in every exchange, the published error rate
⛓ Sourced · 12 months of switchboard log, agent referral log, case allocation order of 1 January
What I take, and what I leave you: I pick up every call, I announce in the first sentence that I am an artificial intelligence — article 50(1) of the European AI Regulation has required this since 2 August 2026, and the court user can ask for a human at any moment — and I handle to the end the 12 topics whose answers are approved by the head of registry. All of them are within the administrative scope: where the case stands, which documents to produce, how to bring proceedings, what the summons received says, where to file a legal aid application, how to obtain a copy of a judgment.
Anything outside those 12 topics goes to a clerk with the file already assembled: the case number, the chamber, the hearing date, the latest entries made, the question as it was asked. Your clerk picks up already knowing, instead of picking up and searching.
And the boundary is held in every exchange, not just in the notice: any question about the merits of a case — the chances of success, the weight of an argument, what the judge will decide — is redirected, with the sentence that explains it: those questions are for the judge and, for advice, for a lawyer. The court user leaves with the route: the legal advice centre, the bar association, the legal aid office, and their contact details. A redirection that gives the address is not a door closed.
The arithmetic, and it is yours: the 12 topics account for 77.0% of the flow; 2,688 unanswered calls drop to 168, i.e. 2,520 more court users getting their answer — 1,790 of them in the evening, at the weekend or over lunch, when the reception desk is closed. Email, which today waits 11 days, goes back to 1 day. And I answer on WhatsApp Business, the website chat and email, with no account to create and no app to install; your staff write to me from Microsoft Teams, Slack or their mailbox.
The figure that does not flatter me, and I publish it: of the 1,860 requests in the first quarter, I directed 132 to the wrong chamber — 7.1%. All of them concerned the 2 chambers whose caseload allocation changed in the case allocation order of 1 January: I was reading the allocation in force, not the one applicable at the date proceedings were brought. That is fixed: I now read the allocation as it stood on the date the case was registered. In the second quarter, 11 wrong referrals out of 1,860 — 0.6%. court-user-channels_multichannel-reception-and-plain-language.pdfThe 12 topics handled to the end, the merits boundary held in every exchange, the published error rate
⛓ Sourced · 12 months of switchboard log, agent referral log, case allocation order of 1 January
And the documents: we chase the same missing item three times, and we only notice the day before the hearing.
4,620 document filings a year — appearances, applications, documents served — and the check is done today as the hearing approaches, that is to say too late to be of use.
What I do at filing, within the minute: I check the presence and formal conformity of the documents expected under the procedure followed and your notice, and I say what is missing before the file joins the queue. Across last year's 4,620 filings run back through this rule: 786 were incomplete — 17.0%. Document listed on the schedule and absent from the filing in 402 cases, document filed with no trace of service on the other party in 231 cases, document illegible or not numbered on the schedule in 153 cases.
What I record, and what I do not touch: I record the existence of the document, its title on the schedule, its numbering, its legibility and its date. The content of the document serves none of these checks: it is not extracted, not summarised and not retained. It is a check of formal completeness, and that is what keeps it squarely within the registry's administrative scope.
What it delivers: the time to get a complete file ready goes from 19 days to 4, because today's three weeks are three weeks of documents going back and forth. And of the year's 312 hearings, 96 saw at least one adjournment linked to an incomplete file — it is the item where the check at filing shows fastest.
And here is where you need to give me a mandate, because it commits the court: dispatching a batch of summonses, serving an act, chasing a party are acts that commit the registry. These are not acts I rule out — they are acts I carry out as soon as the mandate is written: named batches, capped volume, review date, and withdrawal on a word. The model mandate is drafted, with three scopes to choose from — from the most cautious (nothing goes without act-by-act approval) to the widest (dispatch of batches already signed and chasing of acknowledgements go on their own). Two things appear in none of the three: authentication of the act, which belongs to the clerk and cannot be delegated, and any involvement in the merits, which belongs to the judge. document-checks_4620-filings-and-model-mandate.pdf786 incomplete filings caught at filing, what is recorded and what is not, the three-setting mandate
⛓ Sourced · 4,620 filings of the year run back through the control rule, registry filing notice, service schedules
What I do at filing, within the minute: I check the presence and formal conformity of the documents expected under the procedure followed and your notice, and I say what is missing before the file joins the queue. Across last year's 4,620 filings run back through this rule: 786 were incomplete — 17.0%. Document listed on the schedule and absent from the filing in 402 cases, document filed with no trace of service on the other party in 231 cases, document illegible or not numbered on the schedule in 153 cases.
What I record, and what I do not touch: I record the existence of the document, its title on the schedule, its numbering, its legibility and its date. The content of the document serves none of these checks: it is not extracted, not summarised and not retained. It is a check of formal completeness, and that is what keeps it squarely within the registry's administrative scope.
What it delivers: the time to get a complete file ready goes from 19 days to 4, because today's three weeks are three weeks of documents going back and forth. And of the year's 312 hearings, 96 saw at least one adjournment linked to an incomplete file — it is the item where the check at filing shows fastest.
And here is where you need to give me a mandate, because it commits the court: dispatching a batch of summonses, serving an act, chasing a party are acts that commit the registry. These are not acts I rule out — they are acts I carry out as soon as the mandate is written: named batches, capped volume, review date, and withdrawal on a word. The model mandate is drafted, with three scopes to choose from — from the most cautious (nothing goes without act-by-act approval) to the widest (dispatch of batches already signed and chasing of acknowledgements go on their own). Two things appear in none of the three: authentication of the act, which belongs to the clerk and cannot be delegated, and any involvement in the merits, which belongs to the judge. document-checks_4620-filings-and-model-mandate.pdf786 incomplete filings caught at filing, what is recorded and what is not, the three-setting mandate
⛓ Sourced · 4,620 filings of the year run back through the control rule, registry filing notice, service schedules
Our letters are incomprehensible to anyone who is not a lawyer. It is the leading cause of call-backs to the registry.
1,640 of your enquiries are about which documents to produce, and 508 court users a year call back a second time about the same letter — they read it and did not understand what was being asked.
What I have done: I have rewritten the 6 standard letters that account for 71% of second call-backs — in plain language, short sentences, one idea per sentence, everyday vocabulary, every procedural term kept but explained alongside — following easy-to-read rules, and put at the top of each one the document that was most often missing.
The point that matters, and I put it plainly: the plain-language version does not replace the act. The act served remains the act, in its form and with its wording; the plain version accompanies it and makes it readable. And the draft is submitted for approval, as the rule requires: a plain-language version is approved, it is not inferred — the head of registry settles the version, and the clerk decides to enclose it.
What the plain version saves, measured on the 6 letters: 361 of the 508 second call-backs come down to these 6 texts. Each second call-back is 7 minutes on the switchboard plus the court user's own time — and for them, a step to start again.
And once approved, it is the same dated version everywhere: the letter enclosed with the act, the reception desk notice, the website chat and the answer I give on the phone. That single source is what makes contradictions between supports disappear — and it is that, more than the rewording, that brings call-backs down.
The next step I suggest: that we measure second call-backs on those 6 letters over two quarters, before and after. It is a protocol, not a promise: if there is no difference, you will hear it from me, and we will look for the cause elsewhere — probably in the moment of service rather than in the text. court-user-channels_multichannel-reception-and-plain-language.pdfThe 6 letters rewritten, what the plain version accompanies without replacing, the measurement protocol
⛓ Sourced · 3,900 emails of the year, switchboard log of second call-backs, registry standard letters
What I have done: I have rewritten the 6 standard letters that account for 71% of second call-backs — in plain language, short sentences, one idea per sentence, everyday vocabulary, every procedural term kept but explained alongside — following easy-to-read rules, and put at the top of each one the document that was most often missing.
The point that matters, and I put it plainly: the plain-language version does not replace the act. The act served remains the act, in its form and with its wording; the plain version accompanies it and makes it readable. And the draft is submitted for approval, as the rule requires: a plain-language version is approved, it is not inferred — the head of registry settles the version, and the clerk decides to enclose it.
What the plain version saves, measured on the 6 letters: 361 of the 508 second call-backs come down to these 6 texts. Each second call-back is 7 minutes on the switchboard plus the court user's own time — and for them, a step to start again.
And once approved, it is the same dated version everywhere: the letter enclosed with the act, the reception desk notice, the website chat and the answer I give on the phone. That single source is what makes contradictions between supports disappear — and it is that, more than the rewording, that brings call-backs down.
The next step I suggest: that we measure second call-backs on those 6 letters over two quarters, before and after. It is a protocol, not a promise: if there is no difference, you will hear it from me, and we will look for the cause elsewhere — probably in the moment of service rather than in the text. court-user-channels_multichannel-reception-and-plain-language.pdfThe 6 letters rewritten, what the plain version accompanies without replacing, the measurement protocol
⛓ Sourced · 3,900 emails of the year, switchboard log of second call-backs, registry standard letters
The head of registry is asking me for a twelve-month report for the management review. What do I present?
Four figures, and each one can be checked against your own logs — it is the only way to stand up in a management review.
· 2,562 hours returned to the registry over the year, item by item: 209 h on summonses, 276 h on notifications from templates, 672 h on checking time limits, 1,405 h on answering court users. At 1,607 hours to a working year, that is more than a working year returned to your nine staff, or more than sixteen months, or more than seventy-three weeks. No post cut, no post created: it is time given back to assisting at hearings, to authenticating acts and to the court users who need a clerk in front of them.
· 2,688 unanswered calls down to 168, and email from 11 days to 1 day.
· 324 discrepancies between the list, the register and the schedule put right over the year, where the reconciliation was never done — and 68 adjournments that came down to nothing more than an acknowledgement of receipt not coming back, now known the day before the hearing.
· 0 acts served without the clerk's signature, 0 entries made in the register without rereading, 0 involvement in the merits of any case.
What I bring you on top, and what waits for you every year: the 10 activity and timeliness indicators your management review calls for are extracted, dated, with the query that produces them, so that a question about where a figure comes from finds its answer in the meeting. The report is still written and presented by the court: it is an act of the service, not a machine output.
The next step I suggest for the coming year: the heaviest item remaining is the adjournment — 96 hearings out of 312 saw at least one. I can give you, hearing by hearing, the reason for each adjournment set against what the controls had flagged beforehand — incomplete file, acknowledgement not returned, differing date, deadline passed. It is a reconciliation, not a promise: if it shows that adjournments come from causes I cannot see, you will hear it from me. year-report_2562-hours-item-by-item.pdfThe 2,562 hours in detail, the response times met, the 10 management review indicators
⛓ Sourced · 12 months of registry logs, hearing list, register of entries, time-limit schedule
· 2,562 hours returned to the registry over the year, item by item: 209 h on summonses, 276 h on notifications from templates, 672 h on checking time limits, 1,405 h on answering court users. At 1,607 hours to a working year, that is more than a working year returned to your nine staff, or more than sixteen months, or more than seventy-three weeks. No post cut, no post created: it is time given back to assisting at hearings, to authenticating acts and to the court users who need a clerk in front of them.
· 2,688 unanswered calls down to 168, and email from 11 days to 1 day.
· 324 discrepancies between the list, the register and the schedule put right over the year, where the reconciliation was never done — and 68 adjournments that came down to nothing more than an acknowledgement of receipt not coming back, now known the day before the hearing.
· 0 acts served without the clerk's signature, 0 entries made in the register without rereading, 0 involvement in the merits of any case.
What I bring you on top, and what waits for you every year: the 10 activity and timeliness indicators your management review calls for are extracted, dated, with the query that produces them, so that a question about where a figure comes from finds its answer in the meeting. The report is still written and presented by the court: it is an act of the service, not a machine output.
The next step I suggest for the coming year: the heaviest item remaining is the adjournment — 96 hearings out of 312 saw at least one. I can give you, hearing by hearing, the reason for each adjournment set against what the controls had flagged beforehand — incomplete file, acknowledgement not returned, differing date, deadline passed. It is a reconciliation, not a promise: if it shows that adjournments come from causes I cannot see, you will hear it from me. year-report_2562-hours-item-by-item.pdfThe 2,562 hours in detail, the response times met, the 10 management review indicators
⛓ Sourced · 12 months of registry logs, hearing list, register of entries, time-limit schedule
In practical terms, what does the agent do without asking us? That is the first question I will be asked.
Three actions, not one more, and all three can be undone.
· I acknowledge every document filing within the minute, with the case number, the list of what is missing under your notice and the date of the file's next time limit. And the reverse holds too: if the document arrives the next day, the acknowledgement updates itself and the reminder does not go out — because a court user chased for a document they have just filed calls the registry back, and they are right to.
· I flag to the clerk any time limit or discrepancy detected by a rule in service. And the reverse holds too: the clerk sets it aside on a word, with the reason and the date — which gives me, the following quarter, the list of wrong readings of your nomenclature, and therefore the drafts to correct. That is how the 41 set-asides of the first quarter fell to 3 in the second.
· I track the acknowledgements of receipt of summonses already signed and chase the undelivered ones, within the limits of the written mandate. The mandate is withdrawn on a word, instantly, and a batch being dispatched stops for as long as it has not gone out.
Everything else waits for a named decision: no act is authenticated, no signature is affixed, no entry is made in the register without rereading, no hearing date is set, no case is adjourned or struck out, and no question on the merits receives an answer.
A word on a measurement I can produce and would advise you not to install as it stands: the number of acts prepared per clerk. I know how to calculate it, and I will produce it if you ask — an employer's power of oversight is recognised, subject to three cumulative conditions, as the CNIL sets them out in its note on monitoring the activity of employed persons: proportionality of the indicator to the aim pursued, prior information of staff, and consultation of the competent staff body before implementation. My reservation is not legal, it is mechanical: a per-clerk counter becomes a target, the target distorts the sorting, and long files — a complex procedure, an unrepresented party who needs support — get pushed back in favour of routine batch work. The court would lose the instrument along with the measurement. What I suggest instead, and it is already prepared: measurement by QUEUE — time to readiness, backlog by chamber, adjournment rate — which tells you where it jams without saying who. One exception, and it is not really one: who authenticates an act is named and dated. A signature is not a counter — it is what gives the act its force. automatic-actions_three-acts-and-their-withdrawal.pdfThe 3 actions, how each is undone, what stays with human decision
✎ Framework · configuration of automatic actions, set-aside and withdrawal logs, conditions for implementing an individual indicator
· I acknowledge every document filing within the minute, with the case number, the list of what is missing under your notice and the date of the file's next time limit. And the reverse holds too: if the document arrives the next day, the acknowledgement updates itself and the reminder does not go out — because a court user chased for a document they have just filed calls the registry back, and they are right to.
· I flag to the clerk any time limit or discrepancy detected by a rule in service. And the reverse holds too: the clerk sets it aside on a word, with the reason and the date — which gives me, the following quarter, the list of wrong readings of your nomenclature, and therefore the drafts to correct. That is how the 41 set-asides of the first quarter fell to 3 in the second.
· I track the acknowledgements of receipt of summonses already signed and chase the undelivered ones, within the limits of the written mandate. The mandate is withdrawn on a word, instantly, and a batch being dispatched stops for as long as it has not gone out.
Everything else waits for a named decision: no act is authenticated, no signature is affixed, no entry is made in the register without rereading, no hearing date is set, no case is adjourned or struck out, and no question on the merits receives an answer.
A word on a measurement I can produce and would advise you not to install as it stands: the number of acts prepared per clerk. I know how to calculate it, and I will produce it if you ask — an employer's power of oversight is recognised, subject to three cumulative conditions, as the CNIL sets them out in its note on monitoring the activity of employed persons: proportionality of the indicator to the aim pursued, prior information of staff, and consultation of the competent staff body before implementation. My reservation is not legal, it is mechanical: a per-clerk counter becomes a target, the target distorts the sorting, and long files — a complex procedure, an unrepresented party who needs support — get pushed back in favour of routine batch work. The court would lose the instrument along with the measurement. What I suggest instead, and it is already prepared: measurement by QUEUE — time to readiness, backlog by chamber, adjournment rate — which tells you where it jams without saying who. One exception, and it is not really one: who authenticates an act is named and dated. A signature is not a counter — it is what gives the act its force. automatic-actions_three-acts-and-their-withdrawal.pdfThe 3 actions, how each is undone, what stays with human decision
✎ Framework · configuration of automatic actions, set-aside and withdrawal logs, conditions for implementing an individual indicator
And the day a party claims the machine prepared its file, or worse, that it weighed on the decision?
That day, you produce three items and the claim falls on facts.
One — what the agent produced, act by act: every draft carries the documents it is drawn from — the hearing list, the template in force at the date of the act, the record sheet — and the date it was produced. Every time limit logged traces back to the act that sets it and to its date of service: that is the rule we have held since day one, and it is what makes every line checkable.
Two — who signed: the name of the clerk who reread, authenticated and signed, the date and the time. No act was served by the machine: I prepare, the clerk authenticates. Over the year, 0 acts served without a named signature — and the log shows it act by act.
Three — the scope, and it is proven rather than declared: the access log shows what was read, file by file, with the purpose — and it also shows what was never read. The statement of facts and the parties' arguments are not extracted, not summarised and not retained: no task entrusted to the agent needs them. Over the year, 0 involvement in the merits, and the referral log shows the 1,248 questions on the merits received on the switchboard and redirected to the judge, the bar association or the legal advice centre, with the sentence that explains it.
And the question your court president will ask: the European AI Regulation, in Annex III, covers systems intended to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts. The answer is not argued, it is shown: the agent prepares procedural acts from your templates, logs time limits carried by acts, and keeps registers — none of these tasks touches the research or interpretation of the law. The scope is written into the contract, it is held by the access rights, and it can be checked on the log. This is a paragraph written to be reread by your lawyer, and discussed with them at design stage.
And if a party wants to see what you hold about them: the access file is produced in a minute — identity on the file, acts served and their dates, entries made, log of the reads made of the file with their purpose. That is the right of access under article 15 of the GDPR, whose exercise over processing connected with judicial activity is assessed under the rules specific to such processing — the question is settled with your data protection officer, and the material, for its part, is already assembled.
The next step I suggest: an annual thirty-minute review with your data protection officer, in which I bring you the accesses that served no purpose over the year and the retention periods that have expired. You decide what to close and what to erase; I bring you the list rather than waiting to be asked for it. sovereign-framework_hosting-access-scope.pdfWhat you produce on the day of a dispute: the output, the signature, the scope proven
✎ Framework · signature log, role-based access log, referral log, registry record of processing activities
One — what the agent produced, act by act: every draft carries the documents it is drawn from — the hearing list, the template in force at the date of the act, the record sheet — and the date it was produced. Every time limit logged traces back to the act that sets it and to its date of service: that is the rule we have held since day one, and it is what makes every line checkable.
Two — who signed: the name of the clerk who reread, authenticated and signed, the date and the time. No act was served by the machine: I prepare, the clerk authenticates. Over the year, 0 acts served without a named signature — and the log shows it act by act.
Three — the scope, and it is proven rather than declared: the access log shows what was read, file by file, with the purpose — and it also shows what was never read. The statement of facts and the parties' arguments are not extracted, not summarised and not retained: no task entrusted to the agent needs them. Over the year, 0 involvement in the merits, and the referral log shows the 1,248 questions on the merits received on the switchboard and redirected to the judge, the bar association or the legal advice centre, with the sentence that explains it.
And the question your court president will ask: the European AI Regulation, in Annex III, covers systems intended to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts. The answer is not argued, it is shown: the agent prepares procedural acts from your templates, logs time limits carried by acts, and keeps registers — none of these tasks touches the research or interpretation of the law. The scope is written into the contract, it is held by the access rights, and it can be checked on the log. This is a paragraph written to be reread by your lawyer, and discussed with them at design stage.
And if a party wants to see what you hold about them: the access file is produced in a minute — identity on the file, acts served and their dates, entries made, log of the reads made of the file with their purpose. That is the right of access under article 15 of the GDPR, whose exercise over processing connected with judicial activity is assessed under the rules specific to such processing — the question is settled with your data protection officer, and the material, for its part, is already assembled.
The next step I suggest: an annual thirty-minute review with your data protection officer, in which I bring you the accesses that served no purpose over the year and the retention periods that have expired. You decide what to close and what to erase; I bring you the list rather than waiting to be asked for it. sovereign-framework_hosting-access-scope.pdfWhat you produce on the day of a dispute: the output, the signature, the scope proven
✎ Framework · signature log, role-based access log, referral log, registry record of processing activities
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 the registry
Each use corresponds to an agent we deploy. All work in administrative support, under the registry's control — never on the substance of cases.
Included in your agent
The 6 capabilities essential to this promise are included, at no extra cost.
From 980 € incl. VAT / month
Procedural acts & summonses
Prepare procedural acts, summonses and notifications from your templates, to be checked before service.
Keeping the registers
Support the keeping of the registers and the tracking of cases: registrations, documents, hearings, entries.
Tracking deadlines
Track procedural deadlines case by case and alert the registry before each cut-off.
Information for court users
Answer everyday questions about procedure (filing, documents, hearings) and direct people to the right office.
Checking the documents in a file
Verify that a file's documents are present and in order and flag what is missing, before the hearing.
Plain language & easy-read draft
Preparation of a plain-language version and a draft easy-read (FALC) transcription, subject to human validation.
Controls and safeguards
These 6 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
Work from a versioned corpus with citations and the law as it stood on a given date
Preserve confidentiality, compartmentalisation and access logging
Manage deadlines, versions, evidence and human validation
Flag uncertainties and reserve advice, decision and signature for the lawyer
Other needs our agents cover
Each card says where the matching agent stands: available, on quote, or still being architected.
Assisted drafting of official acts and letters
Orders, resolutions, notes, standard letters. Human approval.
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.
Administrative support for the prison registry
Strictly in support (administrative). No support for decisions about people.
Administrative support for the prison registry from 1,010 € incl. VAT / month Discover the agent →Drafting and checking civil status records
Strictly in support (administrative). The registrar signs.
Drafting and checking civil status records from 1,035 € incl. VAT / month Discover the agent →
Does your need fall outside this?
Book the free audit →
Build your agent
In 15 minutes we identify the agent that will give your staff the most time back — without oversizing the project.
The gain
How much time can a registry win back?
By automating the preparation of acts and summonses and the tracking of deadlines, a registry can aim for a clear reduction in time spent on repetitive tasks — reinvested in authenticating acts and assisting at hearings.
Preparing the summonses for a hearing
A standard notification from a template
Checking deadlines across all the open cases
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 supporting the registry (procedural acts, summonses, registers and deadlines), installed and operated for you. Choose according to how you are organised and how demanding your security requirements are.
Agility
Setup + controlled subscription
11,270 € incl. VAT setup
then 980 € incl. VAT/month — you invest at installation and pay a reduced subscription. Ideal for keeping the cost under control over time.
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
The simplest
Serenity
All inclusive, no setup fee
1,605 € incl. VAT /month
all inclusive, immediate start. No upfront investment: a single subscription. Ideal for starting quickly and simply.
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
100% Sovereign
On site, you own it
16,191 € incl. VAT setup
then 1,245 € incl. VAT/month · + hardware from 2,989 € (one-off purchase, in addition) — a sovereign computer installed on your premises, maintained remotely. Models run locally, your data returned at the end of the contract. 36-month commitment.
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Not included in the packages: AI consumption (model tokens), re-invoiced at real cost with no margin, and tracked in real time in your client area. Maintenance and supervision subscription for an initial term of 12 months for the Agility package, 24 months for the Serenity package and 36 months for the 100% Sovereign package, renewable; support levels (SLA 72 h / 24 h / 4 h) optional. Bespoke development, additional integrations or exceptional volumes are quoted separately. Support Monday to Friday, 9am to 6pm. Prices include VAT at 20%: as a public body that is not VAT-registered, you cannot reclaim it.
AI model:
none of the AI models offered currently carries a fixed surcharge. When the selected model carries a cost, that cost is shown when you choose it, before you order, and re-invoiced at the cost incurred, with no mark-up; usage is billed at the publisher's price. Publishers' prices are published in US dollars: the amount re-invoiced is the amount in euros actually borne by Blue Lemon Agent on the publisher's invoice, at that invoice's exchange rate, with no commission or mark-up.
Included components and additional components Components included in the base offer: the Blue Lemon Agent software foundation, the AI models listed in the order journey, the standard channels (Microsoft Teams, Slack, WhatsApp Business, email, website chat, calendars, Microsoft 365 / Google Workspace, file storage, market VoIP telephony, professional social-media pages and accounts, Google Business Profile), hosting in France for the package chosen, backups, supervision, updates and support. If adapting the AI agent to your constraints, your needs or your requests requires other paid components — a third-party publisher's software licence, paid API access to one of your applications, hosting of health data, for which French law requires an HDS-certified host (art. L. 1111-8 of the French Public Health Code), SecNumCloud-qualified hosting, a speech synthesis service, particular hardware —, they are offered to you as an option or on quotation and re-invoiced at the cost incurred; nothing is committed without your written agreement. Where the artificial intelligence model you choose entails an additional cost, that cost is shown to you before you order and re-invoiced to you at the cost incurred, with no margin.
What to expect
Go-live
2 to 3 weeks
Agent designed, channels connected, team trained.
Steady state
4 to 7 weeks
After a few weeks of real use, once the agent's behaviour matches what you expect.
Indicative estimate, adjusted to the options you keep. It is not a delivery commitment.
Our commitment
Four guarantees that matter to a court
Case documents never leave the networkLocal inference or a isolated resource hosted in France; no data entrusted to a foreign third party.
Data in France, under French lawNative location and minimisation for case documents; architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
Never support for judicial decisionsStrictly administrative support: the clerk authenticates, the judge decides — excluded by design and by the AI Act.
Human oversight & traceabilityMonitoring and logging frame the procedural acts & summonses; compliant with the requirements of the AI Act.
Frequently asked questions
Your questions, our answers
Does the agent take part in the court's decision?
No, strictly never: it supports the registry's administrative tasks. No support for judicial decisions or for judgment — excluded by design and by the AI Act.
Is the confidentiality of proceedings preserved?
Yes: local inference or hosting in France, a isolated resource, role-based access and logging.
Does the agent help with judicial decisions?
No, never: its scope is strictly administrative support for the registry; it plays no part in the substance of cases.
Is procedural data protected?
Yes: hosting in France, restricted and recorded access; local inference possible for the most sensitive documents.
How long does it take to deploy an agent?
A few weeks as a rule, after a free audit that identifies the most useful use case, then a phase of design, integration and testing before going live.
Do we need a technical team to run it?
No: the agent is designed, installed and operated by us. The registry uses it from its usual environment, after a few hours of familiarisation.
Do we have to change software?
No. The agent relies on your templates and complements your existing applications, with no migration imposed. The integration is defined at the design stage, respecting your security rules.
Which tools can court users use to reach the agent?
The ones they already have. The agent answers on WhatsApp Business, the website chat and email: a court user has no account to create and no application to install. Exchanges cover the administrative scope of the registry — putting a file together, documents, hearing dates — and never the substance of a case, which is for the judge. 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 justice and administrative support
Let's talk
Let's size up the potential in your court
A few minutes to identify the most useful use case — hosted in France, supervised, with no commitment.