AI agent · Law firms
It takes new-client requests, identifies the area of law and the urgency, checks for conflicts of interest before booking, then prepares the file for the lawyer. It never gives legal advice.
Area of law, urgency, documents requested: the agent does the first sort and passes on a clear summary. The first meeting starts on substance.
Client names and the lawyer's calendar are covered by secrecy. The agent only reads free or busy slots and never reveals a conflict to the person asking.
The agent identifies the area of law, the urgency and the documents needed, and never gives legal advice. The lawyer receives a file ready to work on.
The agent offers the right lawyer's slots and sends the confirmation with the list of documents to bring. The conflict-of-interest check runs before the booking, and a conflict is never revealed to the person asking.
The agent reads incoming documents, renames them and files them in the right folder, so the team stops losing time on filing.
Clients ask where their case stands. The agent answers from the file and passes any question of substance to the lawyer.
Gmail or Outlook
Reads the request sent to the firm's address and pulls out the essentials.
Website form
Asks the intake questions, and never advises.
Cal.com
Offers the right lawyer's slots and sends the confirmation.
Secib, Kleos, Diapaz
Prepares the matter-opening sheet for the team to approve.
The firm's calendar
Reads free or busy slots only, never the content.
Document base
Files incoming documents in the right matter.
What goes through the agent, where it is hosted and who can access it are decided before deployment, following the guidance of the CNB and the Paris bar.
AI in law firms and professional secrecyThe agent takes the request, routes and books. Legal advice, the fee agreement and any negotiation stay with the lawyer.
AI tools for law firms, comparedSince 2 August 2026, Article 50 of the AI Act requires telling people they are talking to an AI from the start of the conversation.
The AI agent transparency obligationThe full guide: AI and legal document analysis
Legal Guide
What an AI agent does in a French law firm, from a client's first request to reading their contracts, and the RIN, AI Act and GDPR rules that apply at each step.
The working rule
The agent checks for conflicts before booking and reads free slots only. The lawyer opens the matter.
A law firm's AI agent takes new-client requests, checks for conflicts of interest, books the first appointment and files the documents clients send. The lawyer keeps the advice, the fee agreement and the decision to take the matter.
Section 01
For a new client, the agent receives the request, identifies the area of law and the urgency, then checks for conflicts of interest before booking into the right lawyer's calendar. It sends the confirmation and reminders. It gives no legal advice and negotiates no fees. A conflict hit, an emergency or a question on the merits goes to a person.
Three articles of the Règlement intérieur national (RIN, the national conduct rules of the French bar), in the consolidated version published by the CNB (Conseil national des barreaux) on 7 July 2026, set this path. Article 2.2 puts "the names of clients and the lawyer's diary" under professional secrecy. Article 4.1 bars taking on a new client if a former client's confidences are at risk. Article 19.2 requires a lawyer solicited online to check the person's identity in order to protect secrecy, avoid conflicts of interest and apply anti-money-laundering rules.
The request arrives by e-mail at the firm's general address, through the website form, or as a transcribed voicemail. The agent introduces itself as an AI in its first reply.
Area of law from the firm's routing table, urgency signals (police custody, hearing, deadline mentioned), individual or business. No analysis of the merits.
The requester's name and the other parties' names are compared with current and former clients and recorded opposing parties. Any resemblance stops the path.
Possible conflict, emergency, area the firm does not handle, question on fees or on the merits: the assistant or a lawyer takes over, with the summary.
Slot taken in the calendar of the lawyer in charge of that area, under the firm's rules: duration, slots kept for new clients, in person or by video.
E-mail naming the lawyer, with the address or video link, the documents to bring, the price of the first consultation and, if the firm wants it, a payment link.
Reminder the day before with a link to reschedule or cancel. A reschedule goes back through the calendar and the same rules; a cancellation frees the slot and notifies the assistant.
Structured summary sent to the assistant, or a draft matter created in the practice-management software when the vendor opens an API for it.
The lawyer confirms there is no conflict, meets the client, proposes the fee agreement and decides whether to take the matter.
The conflict check comes before booking because a first meeting is already an exchange of confidences. If the opposing party is a client of the firm, the lawyer must not hear their opponent's side, and the firm may have to step aside. The agent checks the names before the requester sets out the facts. The names compared, the software queried, the reply on a match and the log are covered in section 02.
The confirmation goes out after booking because it carries two pieces of information the RIN requires early: the name of the lawyer acting (article 19.2) and how fees are determined, due "from the moment the lawyer is engaged" (article 11.1).
The RIN puts the lawyer's diary under secrecy: the booking agent reads availability and nothing else. Google Calendar exposes freeBusy.query, which returns only busy periods, with a dedicated scope (calendar.freebusy). Microsoft Graph offers getSchedule, which accepts an application permission (Calendars.ReadBasic) suited to an agent running with no signed-in user.
The sample getSchedule response in the Microsoft documentation includes the subject and location of appointments. Check what comes back under your sharing rights, and mark client appointments as private if the agent should only see free or busy.
Cal.com offers a booking API (POST /v2/bookings) that accepts custom fields: area of law, other parties, urgency.
Since 2 August 2026, Article 50 of the AI Act requires people to be told they are dealing with an AI. According to the European Commission FAQ updated on 24 July 2026, the information must be given clearly at the start of the first interaction. The agent writes it in its first reply.
The same reply asks the person not to describe the facts. The CNIL (the French data protection authority) recommends that warning in its advice on chatbots (page of 19 February 2021): where sensitive data may be entered without the service having planned for it, users should be asked not to share it, and a purge should be in place.
For a first appointment, the agent needs the name, contact details, the other parties' names, the area of law, the urgency and the slot. The client tells the facts to the lawyer.
The general rules on AI and secrecy, with the CNB and Paris bar texts, are in our guide to AI in French law firms and professional secrecy.
Section 02
Before offering a slot, the agent compares the requester's name, the opposing party's name and the names of any other parties with the firm's list of clients, former clients and opposing parties. On the slightest resemblance, it books nothing, says nothing about any other client and passes the request to a person the firm has named.
The agent asks for three things before any slot: the requester's name, the opposing party's name and the names of anyone else involved (employer, business partner, spouse, other party to the contract). For a company, it takes the registered name and, if the requester knows it, the SIREN number (the French company registration number). It does not ask for the facts.
It compares those names with current clients, former clients and opposing parties already on record. Article 4.1 of the RIN bars a lawyer from taking a new client's matter "if the secrecy of information given by a former client is at risk of being breached" (our translation, as for every RIN quote in this section): closed matters are part of the comparison.
The comparison covers the whole firm. The same article applies these rules to the practice structure "as a whole and to all its members", and to associate lawyers' personal files. If those personal files are kept outside the firm's database, they have to be added, or the check has a blind spot.
Names are normalized before the comparison: accents, capital letters, company forms (SAS, SARL, the French equivalents of a corporation and a limited company), first name and surname order. The agent looks for a resemblance and does not require an exact match.
On namesakes, the rule is asymmetric: the agent flags every resemblance ("SAS Batiplus" and "Batiplus Construction"), a person clears the false positives. A false positive costs the assistant one e-mail, a missed conflict can cost the matter.
The check queries the firm's client and matter database. A firm that records opposing parties in a dedicated field can query them. A firm that writes them in free-text notes has no usable base, and that has to be said before the project starts.
On Secib, the connector published in Microsoft's documentation has no search-by-name action. Its "Retrieving view's data" action returns the content of a view identified by its name. The firm creates a view listing the parties to each matter, the agent reads it and runs the comparison on its side. That is our reading of the documentation, to confirm with the vendor.
This setup sends the firm's entire client list through the agent. Article 2.3 of the RIN requires the lawyer to ensure secrecy is respected "by any person who works with them in their professional activity", and makes the lawyer answerable for breaches. We count the agent's provider among those people: the list stays in hosting the firm controls, under a contract that says so.
On Kleos, the API add-on sold by Wolters Kluwer is the only documented route. Ask in writing whether it allows a search by name across clients and opposing parties before relying on it.
Diapaz describes, in its KYC module, recording the role of each client and contact in each matter, and a view of potential conflicts "at a glance". The site mentions no API: the agent collects the names, a person runs the check in Diapaz, and the agent offers a slot only after that person clears it. The pattern is the same with Jarvis Legal.
Without an API, a regular export of the client and party list to a table the agent reads is enough. A matter opened since the last export will not be in it: the lawyer runs the check again when opening the matter, the last step of the diagram in section 01.
The agent stops. It offers no slot, books nothing and sends no payment link. The request goes through the firm's internal messaging to the person the firm has named for conflicts, with a stand-in for when they are away.
That person receives the names entered, the database entry that resembles them, the reference of the matter concerned and the source channel. The requester is never copied on that message.
To the requester, the agent replies that someone at the firm will pick up the request and get back to them. It uses the same sentence for an emergency, a question on fees or an area the firm does not handle. A reply used only for conflicts would give the match away through its wording alone.
What happens next is the lawyer's call: clear a false positive, obtain the parties' agreement where article 4.1 allows it, or decline. Any refusal and its wording are theirs. When in doubt about what to do, the lawyer consults their bâtonnier (the elected head of the local bar).
It does not say that a name was found, that the opposing party is or was a client of the firm, which lawyer handles them, or why the request is going to a person. It mentions no other matter, even to reassure. Replying "we already act for that company" would disclose a client's name.
It can tell every requester, before any result, that the firm checks for conflicts of interest before every first meeting. Because every requester hears it, that sentence reveals nothing about anyone.
The rule rests on the RIN. Its article 2.2 puts "the names of clients and the lawyer's diary" under professional secrecy. Article 66-5 of the Law of 31 December 1971, the statutory basis of secrecy, covers consultations, correspondence, meeting notes and the documents in the file, and does not mention client names: the RIN adds them.
Neither text deals with a refusal on conflict grounds. Announcing a conflict to someone who has just named their opponent still tells them that the opponent is, or was, a client of the firm. We draw the rule from that: nothing in the reply may allow that inference. This is our reading of the texts, and the firm's own rules and the bâtonnier have the final say.
Each check leaves one line in a log: date and time, channel, names as received and normalized names, source queried and the date it was last updated, result.
For a possible match, the line adds the reference of the resembling entry, the person notified and the time of the notice. The decision taken afterwards is entered by that person; the agent writes none.
The log holds neither the facts, which the agent never asked for, nor a copy of the client list. It stays in the firm's environment, open only to the people who handle conflicts, and the firm sets how long it is kept under the GDPR.
The RIN spells out a requirement for "adequate procedures to identify and manage conflicts of interest" only for members of a multidisciplinary network (article 16.4). For other firms, this log is a precaution we recommend: it shows which check preceded each meeting.
This section describes how the system works and is not ethics advice. The general framework for secrecy and AI tools, with the CNB guide and the Paris bar charter, is in our guide to professional secrecy and AI in French law firms.
Section 03
As of 3 November 2026, none of the four practice-management products examined publishes openly accessible API documentation that lets a third-party agent create a matter. Secib exposes a read-only Power Automate connector, Kleos sells an API add-on, and the Diapaz and Jarvis Legal sites mention none. The agent works through the calendar and the mailbox; a person opens the matter.
| Tool | What the vendor documents publicly | What an agent can do with it |
|---|---|---|
| Secib neo and Secib air (Septeo) | Power Platform connector "SECIB (Preview)": trigger when a case file is created, read the list of firms and the data of a view, 100 calls per minute per connection, access on request to support. SECIB+: no public API catalog, projects case by case. | Read a view, be notified when a matter is created. Create a contact or a matter: no documented action, a project to negotiate with the vendor. |
| Kleos (Wolters Kluwer) | Brochure: "Kleos API" to interact with your systems, calendar, task and e-mail sync with Microsoft Office, EU cloud certified ISO 27001. | The most explicit API offer of the four. Scope, documentation and price to obtain in writing. |
| Diapaz (Xelya) | View of conflicts of interest in the KYC module, e-mail and calendar on Exchange Online and Outlook. No API mentioned on the site. | Book in Outlook. Conflict check and matter creation in the interface, by a person. |
| Jarvis Legal (LexisNexis) | Native Microsoft 365 integration (Word, Outlook, Excel) and an Outlook add-in. No API mentioned on its product page. | Same pattern: Outlook as the entry point, matter opened by hand from the summary. |
| Outlook (Microsoft Graph) | getSchedule (application permission Calendars.ReadBasic), findMeetingTimes (delegated only), event creation. | Read availability and create the appointment. |
| Google Calendar | freeBusy.query, calendar.freebusy scope. | Read busy periods without reading appointment content. |
| Cal.com | API v2 POST /v2/bookings with custom fields, Stripe payment, e-mail and text reminders. | Full booking, reminders and payment, connected to Outlook or Google. |
The "SECIB (Preview)" connector, published by SECIB in Microsoft's documentation, offers a "When a case file is created" trigger for SECIB neo, SECIB air and Gestisoft, and two actions: retrieve the accessible firms and retrieve the data of a view. It acts on behalf of the whole firm, since version 1.0 does not support per-user authentication, and access is obtained by writing to SECIB support. It contains no action that creates a contact or a matter.
Septeo presents SECIB+ as the route for connecting third-party tools and states that no API catalog is currently public: projects are considered case by case. For a firm on Secib, creating matters automatically is a project to negotiate with the vendor.
The connector's read action could feed the conflict check, from a Secib view listing the parties to each matter. That is our reading of the connector documentation, to confirm with the vendor before any development.
The Wolters Kluwer brochure places "Kleos API" in the software's ecosystem, next to Microsoft Office sync. The vendor's API page blocks automated reading; according to the extract indexed by search engines, the API module is an add-on activated through the sales team, and creating matters from the firm's website is among the listed uses. Ask for the documentation and the price in writing before building on it.
Diapaz, published by Xelya, describes a visual conflict check in its KYC module and e-mail on Exchange Online and Outlook: the tool is designed for a person at the screen. Jarvis Legal, distributed by LexisNexis, highlights Microsoft 365 integration and an Outlook add-in. Neither site mentions an API. With these products, the agent books in Outlook and the assistant opens the matter from the summary.
Without an API, the agent sends a summary in a fixed format: identity, contact details, other parties, area of law, urgency flagged, slot booked, conflict-check result and source channel. The assistant copies or imports it into the software without re-reading the exchange.
Section 04
In a law firm, automated legal document analysis means an agent that reads the documents and deeds the firm receives, picks out what matters and files them in the right matter. It starts with two simple tasks: sorting the documents clients send and answering questions on how a case is progressing. The lawyer checks what the agent picks out.
After the first meeting, the client sends their documents by e-mail or through the website form. The agent reads each one, recognizes its type (employment contract, payslip, dismissal letter, judgment), renames it to the firm's convention and files it in the matter.
It then compares what it received with the list of documents requested in the confirmation e-mail, and reminds the client about the missing ones.
The documents go where the firm already keeps its matters: a document management system such as iManage or NetDocuments, a SharePoint site, or the practice-management software when the vendor opens an API for it (section 03).
A client who asks where their case stands gets the last step recorded in the matter: submissions filed, hearing date. Any question on the merits goes to the lawyer, with the client's message.
The agent first checks that it is talking to the client on the matter, through the e-mail address or phone number on file. Confirming to a third party that a matter exists would tell them the person is a client of the firm, and article 2.2 of the RIN puts client names under secrecy.
Section 05
On a contract, a lease or a ruling, the agent picks out clauses, dates, amounts and discrepancies, and every point links to its page in the document. The lawyer reads that summary, checks the points that commit the firm and keeps both the advice and the negotiation. The table lists the most common extractions.
| Document | What the agent flags |
|---|---|
| Commercial contract (services, distribution, framework agreement) | Unusual or missing clauses |
| Non-disclosure agreement | Duration, scope, penalties, restrictive clauses |
| Commercial or residential lease | Termination, indexation, charges |
| Corporate documents (articles, AGM minutes, shareholder agreement) | The same grid applied to every document, in due diligence |
| Court ruling | Parties, reasoning, operative part |
| Terms and conditions of sale | Compliance points, abusive clauses |
| Two versions of the same contract | Every change, flagged |
The agent reduces errors from fatigue, skim reading of a large volume, confirmation bias, and mistakes when copying dates, amounts and references by hand.
It adds others: an invented clause (hallucination), a wrong extraction from a poor scan. A lawyer's validation, described in section 06, is there to catch them.
Lumyniq does not publish a percentage of time saved: it depends on the documents, their scan quality and the level of review the firm keeps. Measure on your own files the average time per document type before and after, errors caught downstream, cases handled per period and the total cost of the tool, licenses and training.
Disregard any ROI multiple announced by a vendor that has not seen your files.
Section 06
Every point the agent flags links to its page in the document. A lawyer approves the ones that commit the firm.
A reliable analysis is prepared upstream. Documents are scanned cleanly, grouped by type and indexed. The agent then processes them in four layers, from raw text to summary. A lawyer finally approves the points that commit the firm, and every extraction keeps its source: document, page, paragraph.
A poor scan triggers cascading errors: OCR must be precise before any analysis. Convert documents to a structured format (PDF with OCR, Word, plain text) to process them in batches.
Then group them by type (commercial contracts, leases, NDAs): each family gets its own instructions and review grid. Index the date, parties, jurisdiction and sector so you can filter later.
From scan to summary
Layer 1
Raw extraction
Quality OCR and structuring of text and visual elements (tables, signatures, stamps).
Layer 2
Semantic understanding
Language models (Claude, GPT, Mistral) parse meaning: parties, obligations, dates, amounts, non-compete clauses.
Layer 3
Critical analysis
Comparison against benchmarks: abusive clauses, imbalances, risks, negotiation points.
Layer 4
Synthesis and reporting
Executive summary, due-diligence brief, compliance report, comparison table.
A lawyer always reviews points with high financial stakes or legal risk. At regular intervals, a sample of documents the agent has already processed is re-analyzed by hand.
Lawyers' corrections are fed back into the agent's instructions. Every extraction stays tied to its source (document, page, paragraph) for fast checking.
Section 07
Two families of tools read legal documents. SaaS products (Kira Systems, Luminance, LawGeex, ContractPodAi) deploy fast, on subscription. A custom agent, which is what Lumyniq builds, follows the firm's instructions and connects to its tools. Volume, document type, integration and how sensitive the matters are decide between them.
| Criterion | Turnkey SaaS | Custom agent |
|---|---|---|
| Deployment | Fast, limited initial training | Longer implementation, with a competent partner |
| Cost | Predictable subscription | Higher upfront investment |
| Language and formats | Often designed for English and US formats | Instructions tailored to French law and documents |
| Integration | API bridges | Native: iManage, NetDocuments, SharePoint |
| Customization and data | Limited customization, vendor dependency | Full control over data |
| Documents and volume | Standard documents | High, steady volume |
| Investment horizon | Short | Long |
For defense and strategic litigation, prefer in-house hosting, whichever family you choose.
The tool-by-tool comparison (research, drafting, contract review) is in our guide to the best AI tools for legal.
Section 08
A law firm's documents mix personal data and information covered by professional secrecy. Four sets of rules apply in France: the GDPR, the French Data Protection Act that complements it, the EU AI Act, which classifies some legal AI systems as high-risk, and the RIN, whose article 2 protects professional secrecy including towards technical tools.
From document to access log
Original document
PDF, Word or scan, entry on the user side
TLS 1.3 in transit
Network channel encryption
AES-256 at rest
File and database storage encryption
EU hosting
OVH, Scaleway, AWS Europe, Azure Europe
DPO and logging
Signed DPA, traced access, periodic audits
Host in the European Union (OVH, Scaleway, AWS Europe, Azure Europe), with no transfer outside the EU. Encrypt traffic with TLS 1.3 and storage with AES-256.
Use the enterprise versions of the models (Claude Enterprise, Azure OpenAI, Mistral Enterprise), the only ones that disable training on your data.
Have the Data Processing Agreement (DPA) set out purposes, retention periods and sub-processors. Log every access and every extraction for audits.
Anonymize documents when you can: it greatly reduces the GDPR risk. Forbid staff from entering confidential data into unsecured AI tools.
The processing terms of OpenAI, Anthropic and Mistral are compared in our guide to GDPR and AI agents.
Section 09
Start with new-client intake. The agent handles little data there (name, contact details, other parties, area of law, urgency, slot) and the conflict check protects the firm from the first message. First make sure your opposing parties are recorded in a dedicated field of your database, or the check has nothing to query.
Then add document sorting and case updates, which start from the matter the lawyer has opened. For contract review, test the agent on a sample of your own documents and measure time and errors before and after.
To choose the technical building blocks (automation platform, model, compliant hosting), our guides and comparisons lay out the trade-offs, with sources.
Section 10
Related guides
CNB 2026 guide, Paris bar charter and an architecture that protects secrecy.
Who is the processor, who trains on your data, where it is hosted, and when a DPIA becomes mandatory.
Research, drafting and contract review: what each tool does.
Obligations, deadlines, and what the 2027 deferral changes.
What must be disclosed when an agent speaks to a client.
Links verified at publication. Regulatory texts change — always defer to the official source.
Free audit
45 minutes with Zakaria El Asri: leave with an action plan on the AI automations most useful to your firm.