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 — Justice
AI agents for justice: time given back to prosecutors, judges and court clerks
Faced with an under-resourced justice system — 3.2 prosecutors per 100,000 inhabitants, against a European median of 11.2 (an indicative figure, to be confirmed) — summarising files, researching case law and preparing procedural documents take up a considerable amount of time. Your AI agent absorbs this supporting work. Hosted in France — on local inference or an isolated resource — with the confidentiality of proceedings preserved. The AI agent assists, the judge decides: no judicial decision is automated.
Hosted in France
Confidentiality of proceedings preserved
GDPR & AI Act: governed deployment
Reinforced human oversight
Updated on
Deployed in a few weeks
Justice assistant · hosted in France
On this 480-page investigation file, establish the chronology of the facts and list the documents cited, before I read it.
480 pages analysed. Chronology reconstructed of the facts, the interviews and the procedural steps, with a reference to the exhibit number of each document. 3 points flagged: one document cited but absent from the file, two interview dates inconsistent between two records.
A 2-page summary and a table of documents are ready — for you to read over.
⛓ Source · the investigation file, numbered exhibits
A 2-page summary and a table of documents are ready — for you to read over.
⛓ Source · the investigation file, numbered exhibits
Is there any recent case law on the classification adopted?
Search carried out across the applicable statutes and the case law of the last five years. Three relevant rulings identified, each cited with its reference and an extract from the decisive reasoning. Caveat: one decision has since been overturned — flagged.
I am preparing a sourced research note, for your judgement.
✎ Action · note ready to read over — the judge decides
I am preparing a sourced research note, for your judgement.
✎ Action · note ready to read over — the judge decides
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 judge decidesThe agent assists, it never judges
✦ In brief
Within a court, a Blue Lemon Agent agent assists prosecutors, judges and clerks with supporting tasks — summarising bulky files, researching statutes and case law, preparing procedural documents and templates, managing deadlines and receiving court users. It runs on local inference or is hosted in France: the data never leaves the European Union, the confidentiality of proceedings is preserved, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. The judicial decision remains entirely with the judge — the agent does not judge. Since the administration of justice is a high-risk use within the meaning of the AI Act, human oversight is reinforced. Live within a few weeks.
100%
hosted in France in the target architecture
0
transfer outside the EU in the target architecture
9
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 the courts — and why they hesitate
Delays are lengthening, files are growing, staffing stays constrained. AI can give judges and clerks back time on supporting work — but the data involved is among the most protected in law: the confidentiality of proceedings, personal data, the safety of individuals.
! The issue
The court is caught between a demand for justice that does not let up and human resources that are counted. Yet most consumer AI tools amount to entrusting case documents, criminal data and the confidentiality of proceedings to a third party, often hosted outside Europe and subject to the Cloud Act — which is quite simply out of the question for judicial data.
✓ Our answer
AI is only of interest to a court if it is sovereign and confidential by design. Local inference or an isolated resource hosted in France, reinforced human oversight, the decision reserved for the judge: time gained on supporting tasks is never paid for in lost confidentiality. The aim is not to replace the judge, but to give them back time to think — the AI agent assists, the public official decides.
The decisive point
Confidentiality of proceedings and judicial data: sovereignty & compliance
A court handles the most protected data in law. Here is how the architecture of our agents preserves it, file by file.
Local inference
The agent can run on a machine at the court: no case document leaves the network, nothing passes through a cloud.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law — the data: processing and access within the European Union targeted by the architecture.
Reduced extraterritorial exposure
Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity for case documents: our architecture depends on a subcontracting chain and remote access documented for the configuration chosen.
A resource isolated per court
No pooling of data: a strictly dedicated, partitioned environment, with access recorded.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
The administration of justice is a high-risk use: reinforced human oversight, transparency and traceability from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
For the most sensitive data, SecNumCloud and HDS options are available depending on your requirements. A single architecture is designed to answer both the GDPR and extraterritorial exposure. Designed for deployment in line with the GDPR and the AI Act, after the processing, roles and context-specific risks have been assessed.
Demonstration
See the agent at work
5 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
The public body in this demonstration
Fictional public bodyBréval-sur-Loing Court of First Instance — court of medium jurisdiction
- Sector
- Judicial court — civil division, criminal division, two investigating chambers, single reception desk for court users open 30 hours a week
- Headcount
- 44 staff — 13 sitting judges, 4 prosecuting magistrates, 18 court clerks and 9 administrative officers at the reception desk; the agent serves the 35 judges and clerks
- Public served
- 210,000 residents in the court's area — 6,400 new civil cases and 2,100 new criminal cases a year, 11,800 enquiries received at the reception desk
- Order of magnitude
- 47 investigation files under way, 480 pages on average for a large file, 38 hearings a month, 14 cases per hearing list, 1,240 procedural deadlines tracked
- Tools in place
- Judicial service applications, document management system, the court's reference library, hearing-room diaries — the agent plugs into them read-only, nothing is replaced and nothing is migrated
- Who decides
- The judge decides and gives reasons; the clerk authenticates the document; the head of registry sets how the service is organised; the presiding judge sets the hearing list
- Room for improvement
- Summarising a large file takes 5 hours of support work; 3 cited exhibits are missing per investigation file on average; 1,900 of the 11,800 reception-desk enquiries get no answer the same day; and 26 of the 1,240 procedural deadlines were checked less than 48 hours before they ran out
In Bréval-sur-Loing, demand for justice does not let up and staffing is tight: the support work — reading, matching, retrieving, formatting — eats into the time for analysis and hearings. The data involved is among the most protected in law: procedural exhibits, parties' identities, secrecy of the investigation. The agent runs on-premises on one of the court's machines, reads files without ever writing to them, and sends nothing out without validation by a judge or a clerk. The exchanges below cover one year, from the review of the investigation files to the review presented to the management committee.
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.
Justice assistant · hosted in France
I reread your 47 investigation files under way this weekend — 22,400 pages, 6,180 numbered exhibits — and I have drawn out a finding your files were carrying without anyone having the time to see it.
The gap measured, and it decides everything else: 141 exhibits are cited in a document in the file yet are absent from the file itself — 3 per file on average, and 11 files carry 5 or more. These are not lost exhibits: they are exhibits announced in a police record, a report or an interview, and never matched to their exhibit number — the exhibit number is the order in which a document is filed; it is how it is retrieved and cited.
What that costs today, across the three items you can check yourselves:
· Summarising a large file: 5 hours of support work — that is the unit of measurement for everything that follows, and I take it as 100 %.
· A case-law search on one point: 33 % of that scale, or 1 hour 40 minutes.
· Answering a court user who asks where their case stands: 10 %, or 30 minutes between the search in the application, the call back and the note to the file.
What I am proposing, and it is not a promise — it is already done: the 141 missing exhibits are listed file by file, with the document that cites them, the page and line where they are announced, and the body holding them according to the file itself. 117 of the 141 are held by an identified third party — investigating service, expert, bailiff, health establishment — and the 117 requests for production are written, on your templates, ready to go out with the registry's signature.
What that changes for the parties, and it is the measure that counts: an exhibit requested today comes back within eight days on average; an exhibit found missing at the hearing costs an adjournment. Last year, 9 hearings were adjourned in this court for that reason alone.
The next step, and it takes twenty minutes: the investigating chamber reviews the 11 files carrying 5 or more missing exhibits, and the requests go out the same day. I have sorted them by nearest hearing date — that is the order in which they win the most procedural days. cited-and-missing-exhibits_141-across-47-files.pdf141 exhibits, 117 production requests already written
⛓ Sourced · 47 investigation files, 6,180 numbered exhibits, documents and records citing the exhibits
The gap measured, and it decides everything else: 141 exhibits are cited in a document in the file yet are absent from the file itself — 3 per file on average, and 11 files carry 5 or more. These are not lost exhibits: they are exhibits announced in a police record, a report or an interview, and never matched to their exhibit number — the exhibit number is the order in which a document is filed; it is how it is retrieved and cited.
What that costs today, across the three items you can check yourselves:
· Summarising a large file: 5 hours of support work — that is the unit of measurement for everything that follows, and I take it as 100 %.
· A case-law search on one point: 33 % of that scale, or 1 hour 40 minutes.
· Answering a court user who asks where their case stands: 10 %, or 30 minutes between the search in the application, the call back and the note to the file.
What I am proposing, and it is not a promise — it is already done: the 141 missing exhibits are listed file by file, with the document that cites them, the page and line where they are announced, and the body holding them according to the file itself. 117 of the 141 are held by an identified third party — investigating service, expert, bailiff, health establishment — and the 117 requests for production are written, on your templates, ready to go out with the registry's signature.
What that changes for the parties, and it is the measure that counts: an exhibit requested today comes back within eight days on average; an exhibit found missing at the hearing costs an adjournment. Last year, 9 hearings were adjourned in this court for that reason alone.
The next step, and it takes twenty minutes: the investigating chamber reviews the 11 files carrying 5 or more missing exhibits, and the requests go out the same day. I have sorted them by nearest hearing date — that is the order in which they win the most procedural days. cited-and-missing-exhibits_141-across-47-files.pdf141 exhibits, 117 production requests already written
⛓ Sourced · 47 investigation files, 6,180 numbered exhibits, documents and records citing the exhibits
Take the 480-page file listed before the investigating chamber next month. I have to read it this weekend.
It is mapped: chronology rebuilt, table of numbered exhibits, and 3 points that call for your eye — you read to decide, no longer to retrieve.
What I am handing over, and where it comes from:
· The chronology of the facts, the interviews and the procedural steps, each line pointing to the exhibit number and page that establishes it. 314 dated events, and the 26 that rest on a single exhibit are flagged as such — that is exactly where the argument will be.
· The table of exhibits: 128 numbers, nature, date, author, and the document that first cites it.
· A two-page summary, in the form your chamber already uses.
The 3 points I flag without being asked:
· An exhibit cited as number 47 in a police record is absent from the file. The body holding it is named in the record itself, and the request for production is written.
· Two interview dates conflict between two records — the 14th and the 15th of the same month for the same person. I give you both pages side by side; one of the two carries a handwritten correction, and I tell you which.
· One exhibit is dated after the document that refers to it. I give it to you with its date and the date of the document, and nothing more: the significance of that gap is a matter of assessment, and the assessment is yours.
The time this moves: summarising a large file goes from 100 % to 12 % — from 5 hours to 36 minutes, and those 36 minutes are your reading, not the rebuilding. Across the 120 large files the court summarises each year, that is 528 hours going back to analysis and hearings.
The next step I propose: that every file over 200 pages arrives mapped as soon as it is registered, without anyone having to ask. Over the past year, 120 files were concerned — and the 3 points I have just shown you would have been flagged 31 days before the hearing on average, instead of the weekend before it. 480-page-file_chronology-and-exhibit-table.pdf314 dated events, 128 exhibit numbers, 3 points flagged
⛓ Sourced · 480-page investigation file, 128 numbered exhibits, interview records
What I am handing over, and where it comes from:
· The chronology of the facts, the interviews and the procedural steps, each line pointing to the exhibit number and page that establishes it. 314 dated events, and the 26 that rest on a single exhibit are flagged as such — that is exactly where the argument will be.
· The table of exhibits: 128 numbers, nature, date, author, and the document that first cites it.
· A two-page summary, in the form your chamber already uses.
The 3 points I flag without being asked:
· An exhibit cited as number 47 in a police record is absent from the file. The body holding it is named in the record itself, and the request for production is written.
· Two interview dates conflict between two records — the 14th and the 15th of the same month for the same person. I give you both pages side by side; one of the two carries a handwritten correction, and I tell you which.
· One exhibit is dated after the document that refers to it. I give it to you with its date and the date of the document, and nothing more: the significance of that gap is a matter of assessment, and the assessment is yours.
The time this moves: summarising a large file goes from 100 % to 12 % — from 5 hours to 36 minutes, and those 36 minutes are your reading, not the rebuilding. Across the 120 large files the court summarises each year, that is 528 hours going back to analysis and hearings.
The next step I propose: that every file over 200 pages arrives mapped as soon as it is registered, without anyone having to ask. Over the past year, 120 files were concerned — and the 3 points I have just shown you would have been flagged 31 days before the hearing on average, instead of the weekend before it. 480-page-file_chronology-and-exhibit-table.pdf314 dated events, 128 exhibit numbers, 3 points flagged
⛓ Sourced · 480-page investigation file, 128 numbered exhibits, interview records
All of this assumes you read procedural exhibits covered by the secrecy of the investigation. Where do they go?
They go nowhere. I run on-premises on one of the court's machines, and nothing I read leaves its network.
On-premises inference means the model computes on your machine: the text of a police record or an expert report crosses no external network to be processed. If the court would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to Bréval-sur-Loing — no pooling with another court, which is the condition for the continuity of your service.
What that changes, point by point:
· Procedural exhibits train no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no right to write to your applications — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, role-based access — rights follow the role: an investigation file can be opened only by the chamber seised of it and its registry. 9 roles for your 44 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· Logging: who asked for what, when, and what I produced. That is proof the court had on none of its tools.
What the law reserves for a person, and it is what gives your acts their value: the judicial decision belongs to the judge and the authentication of the document to the clerk. Everything leading up to it I have already done: the file arrives complete, the chronology rebuilt, the exhibits matched to their number and the sources cited — and the decision stays reasoned, traceable and open to challenge, which is exactly what a court user is entitled to expect.
The figure that sums all this up: 0 procedural exhibits left the court's network across the 22,400 pages reread, and processing in the EU targeted.
What I propose: that I keep up to date the file your data protection officer and your internal control authority will ask for — architecture, data processed, retention periods, who accesses what. It is asked for once a year and takes three days to rebuild; the first version is written and attached. technical-framework_where-procedural-exhibits-live.pdfOn-premises, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account rights, role-based access log
On-premises inference means the model computes on your machine: the text of a police record or an expert report crosses no external network to be processed. If the court would rather not host a machine, the other route is an isolated resource hosted in France, dedicated to Bréval-sur-Loing — no pooling with another court, which is the condition for the continuity of your service.
What that changes, point by point:
· Procedural exhibits train no model, neither ours nor a third party's.
· I work read-only, and the technical account I read through has no right to write to your applications — that is stronger than a promise, because it can be checked with one command.
· Encryption in transit and at rest, role-based access — rights follow the role: an investigation file can be opened only by the chamber seised of it and its registry. 9 roles for your 44 staff, and the log shows 0 out-of-role access since go-live.
· Hosting in France, under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
· Logging: who asked for what, when, and what I produced. That is proof the court had on none of its tools.
What the law reserves for a person, and it is what gives your acts their value: the judicial decision belongs to the judge and the authentication of the document to the clerk. Everything leading up to it I have already done: the file arrives complete, the chronology rebuilt, the exhibits matched to their number and the sources cited — and the decision stays reasoned, traceable and open to challenge, which is exactly what a court user is entitled to expect.
The figure that sums all this up: 0 procedural exhibits left the court's network across the 22,400 pages reread, and processing in the EU targeted.
What I propose: that I keep up to date the file your data protection officer and your internal control authority will ask for — architecture, data processed, retention periods, who accesses what. It is asked for once a year and takes three days to rebuild; the first version is written and attached. technical-framework_where-procedural-exhibits-live.pdfOn-premises, read-only, processing in the EU targeted
✎ Framework · deployment architecture, technical account rights, role-based access log
On the classification adopted in this file, is there any recent case law? I have no time to do the search myself.
The research note is written, sourced decision by decision, and I give you the state of the question rather than a list of references.
What the note contains:
· The provisions applicable to the classification, in the wording in force at the date of the facts — that is the first pitfall of a search done in haste, and it costs dearly: a provision amended since the facts does not apply to the file.
· The decisions retained over the last five years, each with its court, its date, its full reference and the extract of the reasoning that makes it relevant here — the extract, not my summary of the extract.
· The sorting into three families: decisions that support the classification, those that depart from it, and those whose scope was later discussed.
· What I flag at the top and not in a footnote: one of the decisions retained was followed by a reversal — a reversal is when a higher court abandons the solution it had been applying. I give you both decisions and their dates, in order, with the extract of the reasoning of each — you settle their scope in thirty minutes, sources open side by side, instead of half a day of searching.
The rule I hold without exception, and it protects you: a decision I have not read in the reference library, I do not cite. No approximate reference, no reconstructed number: every line of the note points to the document as published, and you open the source in one click from the note.
The time this moves: a case-law search on one point goes from 33 % to 10 % of the scale — from 1 hour 40 to 30 minutes, and those 30 minutes are your critical reading. Across the 380 searches the court carries out each year, that is 443 hours given back to the service.
The next step I propose: that the court's 9 most worked questions be monitored continuously, and that I warn you the day one of them moves. I identified them by rereading your last three years of research notes: those 9 questions account for 214 of the 380 annual searches. research-note_classification-and-state-of-the-question.pdfThree families of decisions, each extract of reasoning alongside
⛓ Sourced · the court's reference library, provisions in the wording in force, published decisions
What the note contains:
· The provisions applicable to the classification, in the wording in force at the date of the facts — that is the first pitfall of a search done in haste, and it costs dearly: a provision amended since the facts does not apply to the file.
· The decisions retained over the last five years, each with its court, its date, its full reference and the extract of the reasoning that makes it relevant here — the extract, not my summary of the extract.
· The sorting into three families: decisions that support the classification, those that depart from it, and those whose scope was later discussed.
· What I flag at the top and not in a footnote: one of the decisions retained was followed by a reversal — a reversal is when a higher court abandons the solution it had been applying. I give you both decisions and their dates, in order, with the extract of the reasoning of each — you settle their scope in thirty minutes, sources open side by side, instead of half a day of searching.
The rule I hold without exception, and it protects you: a decision I have not read in the reference library, I do not cite. No approximate reference, no reconstructed number: every line of the note points to the document as published, and you open the source in one click from the note.
The time this moves: a case-law search on one point goes from 33 % to 10 % of the scale — from 1 hour 40 to 30 minutes, and those 30 minutes are your critical reading. Across the 380 searches the court carries out each year, that is 443 hours given back to the service.
The next step I propose: that the court's 9 most worked questions be monitored continuously, and that I warn you the day one of them moves. I identified them by rereading your last three years of research notes: those 9 questions account for 214 of the 380 annual searches. research-note_classification-and-state-of-the-question.pdfThree families of decisions, each extract of reasoning alongside
⛓ Sourced · the court's reference library, provisions in the wording in force, published decisions
Those nine questions are indeed what occupies us. How do you monitor them without drowning us in alerts?
One note a month, six pages, and nothing in between — except what changes the answer to a question you have on your desk.
What the monitoring produced this month, and it is concrete:
· 4 new decisions on the 9 questions monitored. 3 confirm the state of the question — I cite each in one line. The fourth departs from it, and I give it a full page: the facts, the reasoning adopted, and the 3 files under way in the court where the same question arises. Those are the three you want to see, and they are named by their case number — the number under which a case is registered and tracked from beginning to end.
· 2 amendments to provisions now in force, with the date of entry into force and the pending cases whose facts predate it — 17 cases concerned, and for each the applicable wording is the earlier one.
· 1 of the 9 questions has produced nothing for four months. I propose taking it out of the monitoring and putting another in its place — I have three candidates, drawn from the count of your searches this quarter.
What I do on top, and that nobody has the time to do: I go back over the research notes already delivered in the light of what has just been published. This month, 2 notes delivered less than a year ago are affected by the decision that departs — they come back to you with the paragraph to reread highlighted, and the updated version already written. A research note that ages without saying so is the only flaw a well-made search can have, and it corrects itself once monitoring is in place.
The next step: the monthly note goes to the 13 sitting judges and the 4 prosecuting magistrates on the first Monday of the month, in one send. It is written, laid out and waiting in draft — the first can go out next Monday, and I suggest attaching the list of the 3 pending files concerned. monitoring_9-questions_note-of-the-month.pdf4 decisions, 2 amended provisions, 3 pending files concerned
⛓ Sourced · reference library, this month's publications, research notes delivered within a year, pending files
What the monitoring produced this month, and it is concrete:
· 4 new decisions on the 9 questions monitored. 3 confirm the state of the question — I cite each in one line. The fourth departs from it, and I give it a full page: the facts, the reasoning adopted, and the 3 files under way in the court where the same question arises. Those are the three you want to see, and they are named by their case number — the number under which a case is registered and tracked from beginning to end.
· 2 amendments to provisions now in force, with the date of entry into force and the pending cases whose facts predate it — 17 cases concerned, and for each the applicable wording is the earlier one.
· 1 of the 9 questions has produced nothing for four months. I propose taking it out of the monitoring and putting another in its place — I have three candidates, drawn from the count of your searches this quarter.
What I do on top, and that nobody has the time to do: I go back over the research notes already delivered in the light of what has just been published. This month, 2 notes delivered less than a year ago are affected by the decision that departs — they come back to you with the paragraph to reread highlighted, and the updated version already written. A research note that ages without saying so is the only flaw a well-made search can have, and it corrects itself once monitoring is in place.
The next step: the monthly note goes to the 13 sitting judges and the 4 prosecuting magistrates on the first Monday of the month, in one send. It is written, laid out and waiting in draft — the first can go out next Monday, and I suggest attaching the list of the 3 pending files concerned. monitoring_9-questions_note-of-the-month.pdf4 decisions, 2 amended provisions, 3 pending files concerned
⛓ Sourced · reference library, this month's publications, research notes delivered within a year, pending files
And when the question is new, with no settled case law? That is where I lose the most time.
That is where I serve you most, and I have prepared it: when the answer is not in the library, I give you a map of the ground rather than a silence.
What I hand over on a question with no settled answer:
· The applicable provisions, in plain words, with what they say and what they leave unsettled. On the question you raised last week, two provisions meet and neither settles the case expressly — that is exactly what makes it difficult, and it is written at the top of the note.
· Neighbouring decisions — 7 retained — with, for each, what brings it close to the case and what sets it apart, in two columns. It is the comparison that sheds light, not the number of references.
· Published academic positions, identified as such and never presented as law, with author and date.
· The 3 files in this court where a neighbouring question has already been settled, found in your own delivered decisions — and that is the benchmark that carries the most weight, because it is yours.
What that gives you, measured: of the 380 annual searches, 62 concern questions with no settled answer. They took half a day each; the map of the ground is ready in 30 minutes, and your half-day becomes an hour of thinking about the substance.
And the measure that counts for court users: of the 62 new questions this year, the court delivered 58 decisions within the time it had set itself, against 41 the year before. Those are not hours saved, those are cases judged on time.
The next step I propose: that the decisions delivered by the court itself become searchable, question by question. You have twelve years of reasoned decisions and nobody can retrieve them other than from memory — I have indexed one year to show you what it gives, and the 3 neighbouring files just mentioned come out of it in one sentence. question-with-no-settled-answer_map-of-the-ground.pdf7 neighbouring decisions in two columns, 3 internal precedents
⛓ Sourced · applicable provisions, 7 neighbouring decisions, published academic writing, decisions delivered by the court
What I hand over on a question with no settled answer:
· The applicable provisions, in plain words, with what they say and what they leave unsettled. On the question you raised last week, two provisions meet and neither settles the case expressly — that is exactly what makes it difficult, and it is written at the top of the note.
· Neighbouring decisions — 7 retained — with, for each, what brings it close to the case and what sets it apart, in two columns. It is the comparison that sheds light, not the number of references.
· Published academic positions, identified as such and never presented as law, with author and date.
· The 3 files in this court where a neighbouring question has already been settled, found in your own delivered decisions — and that is the benchmark that carries the most weight, because it is yours.
What that gives you, measured: of the 380 annual searches, 62 concern questions with no settled answer. They took half a day each; the map of the ground is ready in 30 minutes, and your half-day becomes an hour of thinking about the substance.
And the measure that counts for court users: of the 62 new questions this year, the court delivered 58 decisions within the time it had set itself, against 41 the year before. Those are not hours saved, those are cases judged on time.
The next step I propose: that the decisions delivered by the court itself become searchable, question by question. You have twelve years of reasoned decisions and nobody can retrieve them other than from memory — I have indexed one year to show you what it gives, and the 3 neighbouring files just mentioned come out of it in one sentence. question-with-no-settled-answer_map-of-the-ground.pdf7 neighbouring decisions in two columns, 3 internal precedents
⛓ Sourced · applicable provisions, 7 neighbouring decisions, published academic writing, decisions delivered by the court
I need the summonses for the parties for the hearing of 12 March. Fourteen cases on the list, and the template changes with the nature of the case.
The 31 draft summonses are ready, each on the template matching the nature of the case, and two addresses are flagged before a letter goes out for nothing.
The hearing list is the list of cases called at the same hearing; it governs the number of summonses and their order.
What each draft carries, and where it comes from:
· The identity and capacity of the parties, taken from the file, and the case number.
· The date, time and courtroom, taken from the list set by the presiding judge.
· The mandatory particulars of your template, verbatim — these are your templates, reviewed with the registry, and I write none of their lines on my own initiative.
· The interval required between the summons and the hearing, calculated case by case: all 31 comply, and 2 complied only by a single day — I write that to you before sending, not after.
The two points I flag without being asked: 2 addresses are incomplete in the file — a missing street number, an illegible address line on an exhibit. I do not complete them myself: in both cases I give you the exhibit in the file where the address appears in full, and the summons goes out right, first time. Last year, 74 summonses came back for want of an exact address, and each one cost an adjournment or a service of process.
The time this moves: preparing the summonses for a hearing took 2 hours; it takes 12 minutes of checking by the registry. Across 38 hearings a month, that is 63 registry hours a month going back to assisting at hearings and authenticating documents.
The next step I propose: that the drafts build themselves as soon as the list is set, the evening before, and wait for you at the registry the next morning. You approve the principle once; every hearing then arrives with its summonses ready, and the clerk only has to check and serve them — the clerk authenticates, and it is that act which gives the document its force. summonses_hearing-of-12-march_31-drafts.pdf14 cases, 31 parties, 2 addresses flagged
⛓ Sourced · hearing list of 12 March, registry templates, files of the 14 listed cases
The hearing list is the list of cases called at the same hearing; it governs the number of summonses and their order.
What each draft carries, and where it comes from:
· The identity and capacity of the parties, taken from the file, and the case number.
· The date, time and courtroom, taken from the list set by the presiding judge.
· The mandatory particulars of your template, verbatim — these are your templates, reviewed with the registry, and I write none of their lines on my own initiative.
· The interval required between the summons and the hearing, calculated case by case: all 31 comply, and 2 complied only by a single day — I write that to you before sending, not after.
The two points I flag without being asked: 2 addresses are incomplete in the file — a missing street number, an illegible address line on an exhibit. I do not complete them myself: in both cases I give you the exhibit in the file where the address appears in full, and the summons goes out right, first time. Last year, 74 summonses came back for want of an exact address, and each one cost an adjournment or a service of process.
The time this moves: preparing the summonses for a hearing took 2 hours; it takes 12 minutes of checking by the registry. Across 38 hearings a month, that is 63 registry hours a month going back to assisting at hearings and authenticating documents.
The next step I propose: that the drafts build themselves as soon as the list is set, the evening before, and wait for you at the registry the next morning. You approve the principle once; every hearing then arrives with its summonses ready, and the clerk only has to check and serve them — the clerk authenticates, and it is that act which gives the document its force. summonses_hearing-of-12-march_31-drafts.pdf14 cases, 31 parties, 2 addresses flagged
⛓ Sourced · hearing list of 12 March, registry templates, files of the 14 listed cases
The real issue is the deadlines. Which files have a deadline running out in the next fifteen days?
Six files, and for all six the step to be taken is already prepared — an alert list is worth nothing if all it does is alarm you.
The six, in order of expiry:
· 3 deadlines for filing submissions in pre-trial preparation — the stage during which the case is prepared, exhibits are exchanged and the file becomes ready for judgment. For all three, the reminder letter to counsel is written, with the exact date and the exhibit awaited.
· 2 appeal deadlines, with the date of service that starts them running and the exhibit in the file that establishes it — that is the exhibit you look for when the deadline is disputed six months later.
· 1 expert report awaited, outstanding for 94 days against a 90-day deadline: the reminder letter to the expert is ready, and I give you the file's two earlier reminders with their dates.
What I have done with the whole set, not just the six: the court's 1,240 procedural deadlines are checked every night, file by file, each line pointing to the case number and to the exhibit that starts the deadline running. Last year, 26 of those deadlines were checked less than 48 hours before they ran out; this quarter, none. Checking all pending files by hand took a clerk a full day; it takes 20 % of that — two hours of review.
And the figure that speaks to the parties rather than to the service: 9 hearings adjourned last year for a missing exhibit or step, 2 this quarter. An adjournment means three to six more months for court users who are already waiting — it is the only figure on this page they would see if they read it, and it is the one that has moved most.
The next step I propose: a statement to the registry every Monday morning — the fortnight's deadlines, the step prepared for each, and the files now waiting on a single exhibit. It is written and waiting; tell me what time you want it and it goes out on Monday. deadline-schedule_6-files-within-15-days.pdf1,240 deadlines checked nightly, the step ready for each
⛓ Sourced · case management application, 1,240 tracked deadlines, service exhibits, reminder log
The six, in order of expiry:
· 3 deadlines for filing submissions in pre-trial preparation — the stage during which the case is prepared, exhibits are exchanged and the file becomes ready for judgment. For all three, the reminder letter to counsel is written, with the exact date and the exhibit awaited.
· 2 appeal deadlines, with the date of service that starts them running and the exhibit in the file that establishes it — that is the exhibit you look for when the deadline is disputed six months later.
· 1 expert report awaited, outstanding for 94 days against a 90-day deadline: the reminder letter to the expert is ready, and I give you the file's two earlier reminders with their dates.
What I have done with the whole set, not just the six: the court's 1,240 procedural deadlines are checked every night, file by file, each line pointing to the case number and to the exhibit that starts the deadline running. Last year, 26 of those deadlines were checked less than 48 hours before they ran out; this quarter, none. Checking all pending files by hand took a clerk a full day; it takes 20 % of that — two hours of review.
And the figure that speaks to the parties rather than to the service: 9 hearings adjourned last year for a missing exhibit or step, 2 this quarter. An adjournment means three to six more months for court users who are already waiting — it is the only figure on this page they would see if they read it, and it is the one that has moved most.
The next step I propose: a statement to the registry every Monday morning — the fortnight's deadlines, the step prepared for each, and the files now waiting on a single exhibit. It is written and waiting; tell me what time you want it and it goes out on Monday. deadline-schedule_6-files-within-15-days.pdf1,240 deadlines checked nightly, the step ready for each
⛓ Sourced · case management application, 1,240 tracked deadlines, service exhibits, reminder log
After the hearing come the record of decisions and the minutes. That is what drags most, and it is what the parties are waiting for.
The record of decisions is available the same evening as the hearing, and the draft minutes the next morning — the registry corrects, it no longer drafts.
The record of decisions is the short list of what was pronounced, case by case; the minutes are the full document reporting how the hearing ran. The first matters to the service the same evening, the second stays in the file.
What I produce after a hearing of 14 cases:
· The record, in the order of the list: the case, its case number, the sense of what was pronounced, and the date of the next step — reserved judgment, adjournment, removal from the list. Across the month's 38 hearings, the « next step » line was the one missing once in five when the record was drafted eight days later.
· The draft minutes, in your registry's form, with the applications made and the decisions pronounced, each attached to the case concerned.
· The list of steps the hearing triggers — 17 for that hearing: services, summonses for the adjourned date, transmissions. All 17 are prepared the same evening, and the registry checks and authenticates them the next day.
One rule I hold, and it protects the court: what was pronounced is reported exactly, and what was exchanged at the hearing is reported without interpretation. I never summarise reasoning: I report what was said and I point to the decision once it is delivered — the reasoning belongs to the judge who signs it, and it is what makes the decision open to challenge, therefore defensible.
What that gives, measured over the year: 456 hearings, a record available the same evening instead of eight days later, and services on the parties going out 6 days earlier on average. Six days on an appeal deadline is not service comfort: it is time given back to the party to decide what to do.
The next step I propose: that the evening record go to the presiding judge and the hearing registry in one send, as soon as the hearing ends. It is written and waiting in draft; across last month's 38 hearings it would have won eight days for each of the 17 chains of steps a hearing triggers. hearing_record-of-decisions-and-minutes.pdf14 cases, 17 steps triggered, record the same evening
⛓ Sourced · hearing list and notes, registry minute templates, 456 hearings of the previous year
The record of decisions is the short list of what was pronounced, case by case; the minutes are the full document reporting how the hearing ran. The first matters to the service the same evening, the second stays in the file.
What I produce after a hearing of 14 cases:
· The record, in the order of the list: the case, its case number, the sense of what was pronounced, and the date of the next step — reserved judgment, adjournment, removal from the list. Across the month's 38 hearings, the « next step » line was the one missing once in five when the record was drafted eight days later.
· The draft minutes, in your registry's form, with the applications made and the decisions pronounced, each attached to the case concerned.
· The list of steps the hearing triggers — 17 for that hearing: services, summonses for the adjourned date, transmissions. All 17 are prepared the same evening, and the registry checks and authenticates them the next day.
One rule I hold, and it protects the court: what was pronounced is reported exactly, and what was exchanged at the hearing is reported without interpretation. I never summarise reasoning: I report what was said and I point to the decision once it is delivered — the reasoning belongs to the judge who signs it, and it is what makes the decision open to challenge, therefore defensible.
What that gives, measured over the year: 456 hearings, a record available the same evening instead of eight days later, and services on the parties going out 6 days earlier on average. Six days on an appeal deadline is not service comfort: it is time given back to the party to decide what to do.
The next step I propose: that the evening record go to the presiding judge and the hearing registry in one send, as soon as the hearing ends. It is written and waiting in draft; across last month's 38 hearings it would have won eight days for each of the 17 chains of steps a hearing triggers. hearing_record-of-decisions-and-minutes.pdf14 cases, 17 steps triggered, record the same evening
⛓ Sourced · hearing list and notes, registry minute templates, 456 hearings of the previous year
The reception desk is overwhelmed. People call three times to find out where their case stands, and our nine officers cannot hold it all.
11,800 enquiries last year, 1,900 with no answer the same day — 16 %. That is not a failure of welcome: it is a desk open 30 hours a week facing 210,000 residents.
What I have done with those enquiries, by rereading twelve months of the desk log:
· 7,100 of the 11,800 concern eleven subjects, always the same — where a case stands, hearing date, documents to provide, legal aid, joining as a civil party, consulting a file, a copy of a decision, criminal record, service addresses, appeal deadlines, contact details for counsel.
· 4,300 of those 7,100 ask only one thing: where does my case stand, and when am I summoned. The answer is in the application; it was availability that was missing.
What I propose, and you keep the key: I answer these enquiries at any hour, on the reception desk's phone line and on the site's form, and I say in my first sentence that I am a digital assistant of the court, not an officer. This is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be informed, and the caller can ask for an officer at any moment — I then take their details and leave a dated call-back for the service.
The rule that protects the court, and I hold it without exception: I give the state of progress and the timetable, never an assessment of the outcome. A question calling for legal advice goes back with the contact details of the bar, the legal advice point and the victim support service — the court user leaves with a complete route, not with an approximate answer given in the court's name.
The gain, measured: answering a court user who asks where their case stands goes from 10 % to 8 % of the scale — from 30 minutes to 24 minutes, and those 24 minutes are no longer officer minutes: they run on their own, Saturdays included. Across 4,300 enquiries a year, that is 430 hours given back to the reception desk, and enquiries with no answer the same day fall from 1,900 to 470 — those that remain are the ones asking for an officer, and they arrive with the subject already noted.
The next step I propose: that the reception desk review the eleven answers tomorrow morning, one by one — twenty minutes. As soon as they are approved, the phone answers that very night, and I hand over each morning the one-page list of what went out. reception-desk_11800-enquiries-11-subjects.pdf1,900 unanswered, 4,300 about where a case stands
⛓ Sourced · reception desk log over 12 months, site forms, case management application
What I have done with those enquiries, by rereading twelve months of the desk log:
· 7,100 of the 11,800 concern eleven subjects, always the same — where a case stands, hearing date, documents to provide, legal aid, joining as a civil party, consulting a file, a copy of a decision, criminal record, service addresses, appeal deadlines, contact details for counsel.
· 4,300 of those 7,100 ask only one thing: where does my case stand, and when am I summoned. The answer is in the application; it was availability that was missing.
What I propose, and you keep the key: I answer these enquiries at any hour, on the reception desk's phone line and on the site's form, and I say in my first sentence that I am a digital assistant of the court, not an officer. This is not an option you could switch off: the European regulation on artificial intelligence requires that anyone interacting with an AI system be informed, and the caller can ask for an officer at any moment — I then take their details and leave a dated call-back for the service.
The rule that protects the court, and I hold it without exception: I give the state of progress and the timetable, never an assessment of the outcome. A question calling for legal advice goes back with the contact details of the bar, the legal advice point and the victim support service — the court user leaves with a complete route, not with an approximate answer given in the court's name.
The gain, measured: answering a court user who asks where their case stands goes from 10 % to 8 % of the scale — from 30 minutes to 24 minutes, and those 24 minutes are no longer officer minutes: they run on their own, Saturdays included. Across 4,300 enquiries a year, that is 430 hours given back to the reception desk, and enquiries with no answer the same day fall from 1,900 to 470 — those that remain are the ones asking for an officer, and they arrive with the subject already noted.
The next step I propose: that the reception desk review the eleven answers tomorrow morning, one by one — twenty minutes. As soon as they are approved, the phone answers that very night, and I hand over each morning the one-page list of what went out. reception-desk_11800-enquiries-11-subjects.pdf1,900 unanswered, 4,300 about where a case stands
⛓ Sourced · reception desk log over 12 months, site forms, case management application
Answering directly commits the court. The registry cannot review every answer, and I will not sign a blank cheque either.
You do not have to choose between the two: I am proposing a written mandate, capped, dated and withdrawable in a word — and it is already drafted, attached.
What it earns, first, because that is what decides: 7,100 enquiries handled at any hour, the average time to answer a court user's e-mail cut from 11 days to 1, and an enquiry arriving on a Saturday no longer waits until Monday.
What the mandate says, and it fits in six lines:
· Exact scope: the eleven subjects, listed by name, and nothing else. Any enquiry outside that list reaches the service with a draft answer written, sourced and ready to go out from you.
· What stays outside the mandate, in every case: the content of a file, an assessment of the outcome of a case, and any information on an investigation file under way. There, I hand the registry the answer already drafted, exhibits alongside, and the registry sends it — over the trial quarter those answers went out within 3 hours instead of 11 days, and they went out from the court, which is what makes them binding.
· Every answer carries the mention of its source — the application, the date the information was last updated — and the mention that it was prepared by a digital assistant of the court.
· The registry receives each morning the one-page list of the answers sent the day before. An inaccurate answer is caught in an hour, not in three weeks.
· Duration: reviewed after three months, with the record of what it changed. Without an explicit decision at the review, the mandate stops — it is renewal that requires a signature, not termination.
· Withdrawal: one word from the head of registry, and direct answering stops within the minute. Answers go back to being drafts for approval; nothing else changes.
The decision belongs to the head of registry and the presiding judge — and it is taken on a text already written, in one signature. The mandate is drafted, and so is the information notice displayed at the reception desk, and the review is already in the diary for the 15th of the third month: you sign, and the service answers from the next morning. direct-answer-mandate_11-subjects-capped.pdf11 subjects, pending files excluded, review at 3 months
✎ Framework · drafted mandate, list of the 11 subjects, information notice for court users, enquiry log
What it earns, first, because that is what decides: 7,100 enquiries handled at any hour, the average time to answer a court user's e-mail cut from 11 days to 1, and an enquiry arriving on a Saturday no longer waits until Monday.
What the mandate says, and it fits in six lines:
· Exact scope: the eleven subjects, listed by name, and nothing else. Any enquiry outside that list reaches the service with a draft answer written, sourced and ready to go out from you.
· What stays outside the mandate, in every case: the content of a file, an assessment of the outcome of a case, and any information on an investigation file under way. There, I hand the registry the answer already drafted, exhibits alongside, and the registry sends it — over the trial quarter those answers went out within 3 hours instead of 11 days, and they went out from the court, which is what makes them binding.
· Every answer carries the mention of its source — the application, the date the information was last updated — and the mention that it was prepared by a digital assistant of the court.
· The registry receives each morning the one-page list of the answers sent the day before. An inaccurate answer is caught in an hour, not in three weeks.
· Duration: reviewed after three months, with the record of what it changed. Without an explicit decision at the review, the mandate stops — it is renewal that requires a signature, not termination.
· Withdrawal: one word from the head of registry, and direct answering stops within the minute. Answers go back to being drafts for approval; nothing else changes.
The decision belongs to the head of registry and the presiding judge — and it is taken on a text already written, in one signature. The mandate is drafted, and so is the information notice displayed at the reception desk, and the review is already in the diary for the 15th of the third month: you sign, and the service answers from the next morning. direct-answer-mandate_11-subjects-capped.pdf11 subjects, pending files excluded, review at 3 months
✎ Framework · drafted mandate, list of the 11 subjects, information notice for court users, enquiry log
There are also court users who do not understand our letters. Summonses, notifications: they call us back to ask what it all means.
They are the ones who decide whether justice is genuinely accessible, and it is the first place where I can serve without taking anything from anyone.
What is ready, and costs the registry not one hour:
· Every summons and every notification goes out with a plain-language companion sheet — the document itself is not touched by a comma, its particulars are what make it valid. The sheet says in ten lines what is expected of the person, what they must bring, where they must attend and what happens if they do not.
· These sheets exist in an easy-read version — short sentences, one message per sentence, everyday words rather than procedural ones.
· And nothing goes out without a sign-off: every sheet is reviewed and signed off by the registry director before the first mailing, with the original document attached alongside. A companion sheet that contradicted the document would be worse than no sheet at all — so a hearing date, an appeal deadline and a court address are copied verbatim, and only the sentence around them is rewritten. Of the 11 sheets, 3 were sent back for rewriting at that sign-off: they said “without delay” where the document said “within fifteen days”.
· The court's online procedure pages have been reworked against the public accessibility standard — the framework that applies to public-sector websites — on the points that are a matter of wording. The 11 points that belong to the code are written up as defect sheets ready to send to the supplier, each with the page, the criterion and the fix expected.
On online procedures, and this is where people give up: I guide step by step, screen by screen, with the documents to gather before starting — that is the commonest abandonment: the person starts, a supporting document is missing, they close the window and never come back. Of the 82 sessions this quarter, 71 went through to the end. On a court user's personal account, the fastest route is also the only one worth taking: I fill the screen in front of them from the documents they bring, I read out what is being asked and why, and they enter their own credentials and click submit — it is that click which makes the procedure theirs, and therefore valid.
What that gave, measured: call-backs to the reception desk asking « what does this letter mean » fell from 1,040 to 260 over the year, and 38 of the 82 sessions involved people who had never set foot in the courthouse.
The next step I propose: a one-hour help session for online procedures on Saturday mornings, run by me, with a reception desk officer reachable for anything outside the eleven subjects. Over a trial quarter, 29 people came. Say yes and the poster goes out with the next legal advice point newsletter. readable-notifications_easy-read-signed-off-accessibility.pdf1,040 call-backs down to 260, 11 defect sheets for the supplier
⛓ Sourced · sign-off record of the 11 companion sheets, reception desk call-back log, wording audit of the site
What is ready, and costs the registry not one hour:
· Every summons and every notification goes out with a plain-language companion sheet — the document itself is not touched by a comma, its particulars are what make it valid. The sheet says in ten lines what is expected of the person, what they must bring, where they must attend and what happens if they do not.
· These sheets exist in an easy-read version — short sentences, one message per sentence, everyday words rather than procedural ones.
· And nothing goes out without a sign-off: every sheet is reviewed and signed off by the registry director before the first mailing, with the original document attached alongside. A companion sheet that contradicted the document would be worse than no sheet at all — so a hearing date, an appeal deadline and a court address are copied verbatim, and only the sentence around them is rewritten. Of the 11 sheets, 3 were sent back for rewriting at that sign-off: they said “without delay” where the document said “within fifteen days”.
· The court's online procedure pages have been reworked against the public accessibility standard — the framework that applies to public-sector websites — on the points that are a matter of wording. The 11 points that belong to the code are written up as defect sheets ready to send to the supplier, each with the page, the criterion and the fix expected.
On online procedures, and this is where people give up: I guide step by step, screen by screen, with the documents to gather before starting — that is the commonest abandonment: the person starts, a supporting document is missing, they close the window and never come back. Of the 82 sessions this quarter, 71 went through to the end. On a court user's personal account, the fastest route is also the only one worth taking: I fill the screen in front of them from the documents they bring, I read out what is being asked and why, and they enter their own credentials and click submit — it is that click which makes the procedure theirs, and therefore valid.
What that gave, measured: call-backs to the reception desk asking « what does this letter mean » fell from 1,040 to 260 over the year, and 38 of the 82 sessions involved people who had never set foot in the courthouse.
The next step I propose: a one-hour help session for online procedures on Saturday mornings, run by me, with a reception desk officer reachable for anything outside the eleven subjects. Over a trial quarter, 29 people came. Say yes and the poster goes out with the next legal advice point newsletter. readable-notifications_easy-read-signed-off-accessibility.pdf1,040 call-backs down to 260, 11 defect sheets for the supplier
⛓ Sourced · sign-off record of the 11 companion sheets, reception desk call-back log, wording audit of the site
The management committee will ask what this has given us after a year. I need a figure I can defend.
1,401 hours given back to the service over the year — forty hours for each of the 35 judges and clerks — and the calculation is yours, it fits in three lines.
The calculation, item by item, so you can redo it on your own logs:
· Summarising a large file: 120 a year, 5 hours down to 36 minutes — 100 % → 12 % of the scale — that is 528 hours.
· Case-law search: 380 a year, 1 hour 40 down to 30 minutes — 33 % → 10 % — that is 443 hours.
· Answering a court user about where their case stands: 4,300 a year, 30 minutes down to 24 — 10 % → 8 % — that is 430 hours.
What those hours are, and it is what defends best before a committee: staff time given back to the service, at unchanged headcount — no post cut, no post created. It is not a saving on staffing and it must not be presented as one: it is time going back to analysis, hearings and reasoning, and it can be checked on your own logs.
What those hours became, according to your records:
· Hearings adjourned for a missing exhibit or step: 9 in the previous year, 2 this quarter.
· Procedural deadlines checked less than 48 hours before expiry: 26 → 0.
· Reception desk enquiries with no answer the same day: 1,900 → 470.
· Call-backs asking what a letter means: 1,040 → 260.
· Decisions delivered within the time the court had set itself, on new questions: 41 → 58 out of 62.
The figure that does not flatter me, published with the rest: of the 96 research notes of the first quarter, 11 had to be corrected on substance — 11.5 %. Better that it comes from me than from a party at a hearing. The cause, measured and not assumed: 8 of the 11 concerned decisions published after the last update of the court's reference library — I was faithfully citing a library that stopped at a date. What I did about it: the library is matched against the day's publications every night, and any note delivered within the year whose scope a new publication changes comes back to you, paragraph highlighted and updated version already written. The following quarter: 3 corrections out of 104 notes — 2.9 %, and not one due to the library.
What I propose for the committee: the calculation page is written and fits on one side — three lines of calculation, five delays, two framework measures. Send it with the convening notice: a figure read the day before is discussed better than a figure discovered in the room. year-review_1401-hours-given-back.pdfThree lines of calculation, 11.5 % → 2.9 % corrections
⛓ Sourced · production log, reception desk records, hearing register, research notes of both quarters
The calculation, item by item, so you can redo it on your own logs:
· Summarising a large file: 120 a year, 5 hours down to 36 minutes — 100 % → 12 % of the scale — that is 528 hours.
· Case-law search: 380 a year, 1 hour 40 down to 30 minutes — 33 % → 10 % — that is 443 hours.
· Answering a court user about where their case stands: 4,300 a year, 30 minutes down to 24 — 10 % → 8 % — that is 430 hours.
What those hours are, and it is what defends best before a committee: staff time given back to the service, at unchanged headcount — no post cut, no post created. It is not a saving on staffing and it must not be presented as one: it is time going back to analysis, hearings and reasoning, and it can be checked on your own logs.
What those hours became, according to your records:
· Hearings adjourned for a missing exhibit or step: 9 in the previous year, 2 this quarter.
· Procedural deadlines checked less than 48 hours before expiry: 26 → 0.
· Reception desk enquiries with no answer the same day: 1,900 → 470.
· Call-backs asking what a letter means: 1,040 → 260.
· Decisions delivered within the time the court had set itself, on new questions: 41 → 58 out of 62.
The figure that does not flatter me, published with the rest: of the 96 research notes of the first quarter, 11 had to be corrected on substance — 11.5 %. Better that it comes from me than from a party at a hearing. The cause, measured and not assumed: 8 of the 11 concerned decisions published after the last update of the court's reference library — I was faithfully citing a library that stopped at a date. What I did about it: the library is matched against the day's publications every night, and any note delivered within the year whose scope a new publication changes comes back to you, paragraph highlighted and updated version already written. The following quarter: 3 corrections out of 104 notes — 2.9 %, and not one due to the library.
What I propose for the committee: the calculation page is written and fits on one side — three lines of calculation, five delays, two framework measures. Send it with the convening notice: a figure read the day before is discussed better than a figure discovered in the room. year-review_1401-hours-given-back.pdfThree lines of calculation, 11.5 % → 2.9 % corrections
⛓ Sourced · production log, reception desk records, hearing register, research notes of both quarters
A committee member will ask whether you can assess the risk a person presents, or tell us which cases will go badly. What do I answer?
That I will not do it, that it is the one thing I refuse you, and that what he is really after, I already give him — with figures.
What is excluded, and it is not a setting of caution: assessing a person, giving them a score, predicting their behaviour or the outcome of their case. European law treats the administration of justice as a high-risk use and reserves the decision to a person; and scoring a court user would be precisely what makes a decision indefensible before the very person it concerns. I produce no profile, no score, no prediction — about anyone, on any request.
And here is what the committee gets instead, which answers its real question — why some cases stall:
· Deadlines met and missed, case by case, without a person's name on them. 26 tight deadlines last year, 0 this quarter.
· Files waiting on a single exhibit to be ready for judgment: 34 today, and for all 34 the request for production is written. They are what makes the months of waiting, and they can be counted.
· Grounds for adjournment, counted over twelve months: 61 adjournments, of which 9 for a missing exhibit, 14 for a returned summons and 38 for other grounds. The first two can be fixed, and they have been.
· Workload by division and by type of case, never by judge or by clerk — and I tell you why, mechanically: the day a named indicator is tracked, it becomes a target, and it is the way files are handled that bends to please it. The measuring instrument would destroy what it claims to measure.
What that gives the committee: three measured levers instead of an intuition — the 34 files waiting on an exhibit, the 23 avoidable adjournments, the 62 new questions that take the most time. On the first two, the court has already won 7 adjournments and 26 tight deadlines in one quarter.
The next step I propose: that these three levers become the committee's quarterly tracking page, in place of the activity return prepared by hand today. It is written, it fits on one side, and the first version is attached. scope-of-use_what-is-excluded-and-what-is-given-back.pdf3 measured levers, 0 scoring of people
✎ Framework · scope of use, adjournment register over 12 months, state of files awaiting an exhibit
What is excluded, and it is not a setting of caution: assessing a person, giving them a score, predicting their behaviour or the outcome of their case. European law treats the administration of justice as a high-risk use and reserves the decision to a person; and scoring a court user would be precisely what makes a decision indefensible before the very person it concerns. I produce no profile, no score, no prediction — about anyone, on any request.
And here is what the committee gets instead, which answers its real question — why some cases stall:
· Deadlines met and missed, case by case, without a person's name on them. 26 tight deadlines last year, 0 this quarter.
· Files waiting on a single exhibit to be ready for judgment: 34 today, and for all 34 the request for production is written. They are what makes the months of waiting, and they can be counted.
· Grounds for adjournment, counted over twelve months: 61 adjournments, of which 9 for a missing exhibit, 14 for a returned summons and 38 for other grounds. The first two can be fixed, and they have been.
· Workload by division and by type of case, never by judge or by clerk — and I tell you why, mechanically: the day a named indicator is tracked, it becomes a target, and it is the way files are handled that bends to please it. The measuring instrument would destroy what it claims to measure.
What that gives the committee: three measured levers instead of an intuition — the 34 files waiting on an exhibit, the 23 avoidable adjournments, the 62 new questions that take the most time. On the first two, the court has already won 7 adjournments and 26 tight deadlines in one quarter.
The next step I propose: that these three levers become the committee's quarterly tracking page, in place of the activity return prepared by hand today. It is written, it fits on one side, and the first version is attached. scope-of-use_what-is-excluded-and-what-is-given-back.pdf3 measured levers, 0 scoring of people
✎ Framework · scope of use, adjournment register over 12 months, state of files awaiting an exhibit
And if the court wants to stop in a year's time, how stuck are we?
The court is not stuck anywhere, and it can be checked before committing, not afterwards.
What there is to dismantle the day you stop:
· The index. It is deleted, and it holds no procedural exhibit — only what is needed to find them where they are. Your files have not moved by a single byte: same applications, same exhibit numbers, same rights.
· The log of enquiries and productions. It is handed to you in an open format, or destroyed — the court chooses, and the question is settled at go-live, not on the way out.
· What was built from your own material: the eleven standard answers and their plain-language and easy-read versions, the companion sheets for summonses and notifications, the 11 defect sheets for the site, the index of decisions delivered by the court and the hearing record templates. They are yours: they are made of your acts, they stay in your files, readable without us. That is the only asset this go-live will have created, and it would not be honest for it to stay with us.
What does not exist, and should be checked with everyone: no migration on the way in, therefore none on the way out. Your case management applications are not replaced, no format belongs to us, and no judge or clerk has changed the way they work other than reviewing instead of rebuilding.
On public spending, since the committee will ask: the subscription is annual and carries no automatic renewal clause — it is renewal that requires a decision, not termination — and the amount compares with what it frees up: 1,401 hours, without a post created or cut.
What I propose so this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that the court works exactly as before, we switch back on. The protocol is written, it fits on one page, and the date that costs you least is the first Friday in August — 3 hearings that day against 9 on average. The committee will know what the promise is worth before committing to a second year. technical-framework_where-procedural-exhibits-live.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · index architecture, export formats for the log and the productions, dry-run exit protocol
What there is to dismantle the day you stop:
· The index. It is deleted, and it holds no procedural exhibit — only what is needed to find them where they are. Your files have not moved by a single byte: same applications, same exhibit numbers, same rights.
· The log of enquiries and productions. It is handed to you in an open format, or destroyed — the court chooses, and the question is settled at go-live, not on the way out.
· What was built from your own material: the eleven standard answers and their plain-language and easy-read versions, the companion sheets for summonses and notifications, the 11 defect sheets for the site, the index of decisions delivered by the court and the hearing record templates. They are yours: they are made of your acts, they stay in your files, readable without us. That is the only asset this go-live will have created, and it would not be honest for it to stay with us.
What does not exist, and should be checked with everyone: no migration on the way in, therefore none on the way out. Your case management applications are not replaced, no format belongs to us, and no judge or clerk has changed the way they work other than reviewing instead of rebuilding.
On public spending, since the committee will ask: the subscription is annual and carries no automatic renewal clause — it is renewal that requires a decision, not termination — and the amount compares with what it frees up: 1,401 hours, without a post created or cut.
What I propose so this does not stay a sentence: a dry-run exit at the end of the first quarter, half a day: we switch off, we check that the court works exactly as before, we switch back on. The protocol is written, it fits on one page, and the date that costs you least is the first Friday in August — 3 hearings that day against 9 on average. The committee will know what the promise is worth before committing to a second year. technical-framework_where-procedural-exhibits-live.pdfReversibility: 0 migration in, 0 migration out
✎ Framework · index architecture, export formats for the log and the productions, 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 a court
Each use corresponds to an agent we deploy. All work in support, under the oversight of the judge or the clerk.
Included in your agent
The 7 capabilities essential to this promise are included, at no extra cost.
Summarising files
Reading bulky files, chronology of the facts and a table of the documents cited — for the judge to approve.
Preparing documents & templates
Drafting first versions of procedural documents, summonses and templates from your models, to be approved.
Managing deadlines & hearings
Tracking procedural deadlines and scheduling summonses and appointments at the registry.
Receiving court users
First-line information: the state of a case, the documents to provide, the steps to take — sourced answers.
Notes & minutes
Records of decisions and minutes of hearings or meetings, structured and faithful to what was said.
Plain language & easy-read draft
Preparation of a plain-language version and of a draft easy-read (FALC) transcription, submitted for human validation. Translation and voice output are not part of this offer’s core: they are handled by the Accessibility and inclusion agent.
Coordination of the court's sub-agents
Summaries, research, monitoring, documents, deadlines and reception are shared out between dedicated sub-agents, then brought together into a single file. The judge and the registrar keep the arbitration and the decision.
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 arbitration of conflicts and irreversible decisions
Observability of costs, timescales, quality, failures and safe stop
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
What the agent must be connected to
These 2 connections are required for the agent to work. They concern your information system and are scoped during the audit.
Register of authorised agents and interface contracts
Connect an authorised legal database: corpus and licences supplied by the client, access rights, update frequency, date of the law consulted, traceable citation of sources and responsibility for validation. Third-party subscription borne by the client, excluded from the price until it is known.
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.
Legal document agent (litigation)
RAG, deadlines.
Legal document agent (litigation) from 750 € incl. VAT / month Discover the agent →Support for the court registry and procedural acts
Strictly in support (administrative). No support for judicial decisions.
Support for the court registry and procedural acts from 980 € incl. VAT / month Discover the agent →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 →Document summaries
A three-hundred-page report, a study, a case-law file: reading it in full is not always possible in the time available.
Summarising bulky documents from 565 € incl. VAT / month Discover the agent →Legal AI agent
Searching legal databases, regulatory monitoring, preparing acts and pleadings: a sovereign AI agent, hosted in France, with sources cited.
Sourced business document search from 905 € incl. VAT / month Discover the agent →AI information & monitoring agent
Watching texts, case law and publications, sourced monitoring notes: a sovereign AI agent, hosted in France, with no data leaks.
Automated regulatory and legal watch from 825 € 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 court recover?
By automating the summarising of files and documentary research, a court can aim for an appreciable reduction in time spent on supporting tasks — reinvested in analysis, the hearing and the decision, which remain the judge's office.
Summarising a bulky file
Case-law research on a point
A reply to a court user on the state of a case
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
A business agent for justice
A supporting agent (summarising, research, preparing documents), installed and operated for your court. Rates suited to the public sector — for as long as it takes for the gains to settle in.
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 court
Case documents never leave the courtLocal inference or an isolated resource hosted in France; the confidentiality of proceedings is preserved.
Data in France, under French lawAs regards case documents: hosting in France under French law, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
The judge keeps the decisionThe agent produces summaries, research and templates that can be checked; no judicial decision is automated.
Reinforced human oversight & traceabilityA high-risk use within the meaning of the AI Act: monitoring, logging and transparency from end to end.
Frequently asked questions
Your questions, our answers
Does the agent hand down court decisions?
No, never. It assists (summarising, research, formatting); the direction of a case, the submissions and the decision remain with the judge. No judicial decision is automated. The AI agent assists, the public official decides.
Is this compatible with the AI Act?
The administration of justice is a high-risk use within the meaning of the AI Act: reinforced human oversight, traceability and transparency. The agent stays strictly in support, without ever standing in for the judge.
Is the confidentiality of proceedings preserved?
Yes: sovereign hosting in France, an isolated resource, access recorded; the data does not leave the European Union and stays covered by an architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity.
How can I be sure a case-law search is reliable?
Every answer is sourced: the agent cites the statutes and rulings it draws on, with their reference and an extract from the reasoning, and flags any caveats (a reversal, an isolated decision). The judge keeps the judgement and checks the sources.
Do we have to change our case-management software?
No. The agent connects to your existing tools (case-management applications, document management, documentary databases) and complements them, without forcing a migration. We adapt the integration to your environment.
How long does it take to deploy an agent?
A few weeks as a rule, after a free audit that identifies the most valuable use, then a phase of design, integration and testing before going live and training the public officials.
Related roles
Other professions in law and monitoring
Let's talk
Let's size up the potential in your court
15 minutes to identify the most useful use case — hosted in France, supervised, with no commitment.