Jimini AI
Jimini AI is a legal AI assistant built for continental civil law, sold to French law firms, individual avocats and corporate legal departments. It covers three jobs — analysing documents, drafting, and legal research grounded in recognised sources — and adds agentic workflows that plan and run multi-step legal tasks end to end, with more than 200 built and validated alongside in-house lawyers. It works inside Word and Outlook through Microsoft add-ins included in the subscription, and connects to document stores such as Google Drive and SharePoint. The platform is hosted in France on Scaleway and S3NS infrastructure and all customer data is stored in France, while model inference runs inside the European Union across Mistral AI, OpenAI's Europe deployment and Anthropic Claude models reached through Google Cloud Vertex AI, each under a zero-retention agreement. The published terms state that uploaded documents and prompts are never used to train models, and a dedicated clause addresses the avocat's professional secrecy under article 66-5 of the law of 31 December 1971 and article 2.3 of the Règlement Intérieur National. The product is published by Odysai SAS, registered in Nanterre with offices in Paris, and carries partnerships with the Paris, Hauts-de-Seine and Nantes bars. An individual plan is available self-serve with a two-week free trial; organisation pricing runs through the sales team.
Capability grades
All 15 axes, graded from public sources on the date shown. Hover a grade to see what the letter means on that axis.
AI Centrality
How much of the product is actually AI. Whether the machine learning is the mechanism the buyer is paying for or a feature layered onto conventional software, and whether the vendor is specific about which is which.
The CGU defines the thing being sold as the "Système d'IA" and the Site as an interface for interacting with it, which is as literal as this axis gets. Analysis, drafting and research are all model outputs, and the agentic workflow layer plans and orchestrates multi-step tasks. The home page describes a multi-model architecture in which Jimini selects, orchestrates and continuously optimises third-party models per task and builds legal reasoning, source handling and workflows on top. Remove the models and there is no document system or workflow engine left underneath. Checked 4 September 2026.
Citation Accuracy and Hallucination Disclosure
Whether the vendor publishes measured accuracy on citations and assertions, grounds output to primary sources, and says plainly what its system does when it does not know. Legal has a documented public record of fabricated citations reaching filed briefs, so an untested claim of accuracy is not evidence.
Grounding is real and documented rather than asserted. The research product is described as returning answers anchored in recognised sources of law, Legal Data Hunter is named in the CGU subprocessor table as the legal research provider located in the European Union, and the privacy policy describes the retrieval mechanism directly: uploaded documents are indexed and split into paragraphs to support search and retrieval. What keeps this off the top band is the absence of a measured figure a reader can interrogate. The home page carries a precision claim in a hero counter, presented with no methodology, no test set and no date; a number in that form is a claim shape rather than a published accuracy figure, so it would not lift this axis even if it were read. No failure modes are named anywhere. Separately, and as a limit on this record rather than on the vendor, the counter's numerals render through an animated component that did not resolve on fetch. Checked home page, product pages, workflows page, CGU and privacy policy on 4 September 2026.
Autonomy and Oversight Model
What the system decides on its own, what a lawyer must approve, and whether the vendor documents where the review point sits. A tool that drafts under review and a tool that files without one are different products and different risks.
The CGU preamble is emphatic and sits in bold at the top of the agreement: outputs are not legal advice, are not final, must be validated by a legal professional such as an avocat, and the user assumes entire responsibility for their use. Against that, the product markets agentic systems that plan, orchestrate several steps and carry legal tasks through from end to end, with more than 200 workflows designed and validated with in-house lawyers. Those two statements are in tension and nothing reconciles them. What the tension reveals is the limb that is missing rather than a reason to downgrade for its own sake: no threshold is published at which an agent stops and hands back, no review surface is described inside the workflow, and nothing states what happens after an agent is wrong. The contractual commitment to human validation is real and the control structure around it is not published.
Operational and Outcome Evidence
Named, dated evidence that the product works in production at real firms or legal departments. Case studies with figures and identified customers count. Unattributed testimonials and launch announcements do not.
Named deployment evidence with substance, short of measurement. A case study on the vendor's own blog dated 26 June 2025 names Hoche Avocats, describes a three-month trial followed by adoption across corporate, tax, employment, technology and litigation departments, and sets out how the rollout was co-managed by partners and associates and how prompts were built from the firm's own agreement frameworks and M&A audit reports. Attributed quotes name Louis Oudot de Dainville, partner at Gide Loyrette Nouel, and Jacques Perotto, partner at Alerion. The Paris Bar partnership is dated May 2025 with its terms stated, three months' free access for firms of one to twenty lawyers, and Pierre Hoffman is quoted as Batonnier. A user count of more than 6,000 avocats and juristes is published. What is absent is a documented outcome: the Hoche study reports no measured result, and the hours-saved-per-lawyer figure on the home page sits in a hero counter with no method, no cohort and no date behind it, which is a claim rather than a measurement and would not carry this axis to the top band in any event. As a limit on this record rather than on the vendor, that counter's numerals render through an animated component that did not resolve on fetch on 4 September 2026.
Privilege and Confidentiality Posture
How client confidences are handled: attorney client privilege and work product treatment, segregation of one client matter from another, whether client data trains any model, and what the vendor commits to in writing rather than in marketing.
Every limb is met and read before signing. Training is prohibited in the CGU itself and the prohibition extends to suppliers. Segregation is documented at both levels the buyer segments need: data is logically partitioned by user and by organisation, personal and enterprise libraries are stored in isolation, and access runs through Auth0 by Okta on a least-privilege basis with access logged and audited. Professional secrecy is addressed in a dedicated clause rather than by implication, citing article 66-5 of the law of 31 December 1971 and article 2.3 of the Règlement Intérieur National, acknowledging the absolute and public-order character of the obligation, and stating that no data in submitted documents is accessible to third parties or used for any purpose other than providing the service. Retention and deletion are published as a per-category table with a certificate of destruction available on request. The position on model providers is the strongest part: each inference provider is named with its location and the specific basis of its zero-retention engagement. The middle band's two named gaps, silence on segregation between users and silence on what the model provider may retain, are both false of this record, which is what decides it.
UPL and Professional Responsibility Posture
Whether the vendor is clear that it supplies a tool rather than legal advice, who its audience is, and how it addresses unauthorized practice of law, competence and supervision duties, and jurisdiction limits. ABA Formal Opinion 512 is the reference point.
The CGU preamble states in bold that outputs are not legal advice and must be validated by a legal professional, that Odysai is not a law firm and provides no legal advice, and then names the governing statute: under the law of 31 December 1971 as amended, Odysai, its officers and its employees do not carry on legal consultation, the drafting of private deeds, or representation. Naming the French UPL statute directly is rare. The audience is limited expressly to users acting in a professional capacity, and consumer mediation under articles L.611-1 of the Code de la consommation is disapplied on that basis, so there is no consumer surface to disclose on. The jurisdiction is stated throughout as continental and French law. The vendor's own help centre article on déontologie engages the CNB's recommendations and the EU AI Act compliance timetable. Competence and supervision are carried by the validation requirement rather than by a separate competence statement, which is the one place a reader must infer.
AI Governance and Bias Disclosure
Published governance over model behaviour: who owns it inside the vendor, what is tested before release, and what is disclosed about disparate output across matter types, parties, or populations.
This is the weakest disclosure on an otherwise strong record. There is no AI-specific governance framework published: no ISO 42001, no responsible AI page, no named owner accountable for model governance inside the company, no pre-release testing regime described, and nothing at all published about uneven output across matter types or populations. The vendor's help centre article on déontologie engages the EU AI Act timetable and states that technical documentation, the DPA and the security policy are available on request, which places the governing artifacts behind a request rather than in public. ISO 27001:2022 is real and certified but is a security management standard and does not answer this axis. The DPO, Auriane Roussel, is named for data protection, which is a different accountability from AI governance. Checked home page, CGU, privacy policy, trust centre and help centre on 4 September 2026.
AI Safety and Data Stewardship
Retention, deletion, access control, and what happens to prompts and documents after they are processed. Whether the vendor states its subprocessors and its incident practice, or leaves the buyer to assume.
The full set is published and specific enough to hold the vendor to. Retention is a per-category table: document library and AI prompts and responses for the life of the account plus 30 days, account data one year after last connection, billing data ten years, Word and Outlook add-in document and email content not stored at all, connector data processed in real time with zero storage. Deletion is specific, with account deletion executed within the day, aggregated usage data removed within the month, and a certificate of destruction issued on request. Access control names Auth0 by Okta, least privilege, logged and audited access, AES-256 at rest and TLS 1.3 in transit, logical partitioning by user and organisation, and vulnerability testing every six months. Subprocessors are fully named in a table with role, location and zero-retention status, with 30 days' advance notice of any addition or replacement, a right to object, and a right to terminate without penalty if an objection is not resolved. Incident practice is stated on both sides: notification to the customer within 24 working hours of qualifying a breach and to the CNIL within 72 hours under article 33.
AI Liability and Recourse
What the vendor stands behind contractually when its output is wrong. Indemnities, caps, carve outs, insurance, and whether any of it is published or only reachable through a negotiated agreement.
This is the sharpest gap in the record and it sits against an otherwise exceptional set of published commitments. The CGU's limitation of liability clause is an exclusion and nothing more: the company is not liable for damage arising from unavailability caused by factors outside it, from use of outputs without prior validation by a qualified legal professional, from user breach, or from external attack. The intellectual property clause adds that the company cannot be held responsible for outputs or their use, expressly including error, approximation, or content a third party may claim. There is no indemnity running to the customer, no liability cap expressed as a figure or a multiple, no warranty of any kind on the service or its output, and no insurance position was located on the home page, CGU, privacy policy, legal notices or trust centre on 4 September 2026. What is published is a clear allocation that the user bears the loss, and a buyer can read it before signing, which is why this sits at C rather than at the floor. The exposure the product creates is erroneous legal output and the clause disclaims precisely that.
Practice Systems Integration Depth
How deeply the product reaches into the systems legal work already lives in: document management such as iManage and NetDocuments, Word and Outlook, contract lifecycle management, matter management, e-billing, and court filing systems.
Integrations are into the systems legal work actually lives in, and the depth is documented rather than listed. Jimini pour Word and Jimini pour Outlook each have their own product page and are included in the subscription, with functions named per surface: drafting, analysis, correction, comparison and translation in Word; drafting, thread summarisation and text improvement in Outlook. Connectors run to Google Drive and SharePoint. The privacy policy then documents each integration's data flow to a level most vendors never publish: for the add-ins, document and email content including attachments is processed and discarded rather than stored, with only the prompt and the generated response retained; for connectors, only files the user selects are accessed, processing is real time with no storage beyond immediate use, access is revoked immediately on disconnection or account deletion, and Google connector handling is stated to follow the Google API Services User Data Policy. What syncs, in which direction, and what the firm must enable are all answerable from published material. No legal-specific document management system such as iManage or NetDocuments is named.
Deployment Model and Data Residency
Where the software runs and where the data sits. Multi tenant cloud, single tenant, private deployment, on premises, and whether region of residence is a published option or an enterprise conversation.
Residency is stated flatly — all customer data is stored in France — and then evidenced rather than asserted. The platform at app.jimini.ai is hosted by Scaleway and S3NS, both named with Scaleway's registered address given and both stated to hold servers physically located in France, Paris region. Processing location is published as distinct from storage location and is named per inference provider: Mistral AI in France, OpenAI's Europe deployment in the European Union, and Anthropic Claude models through Google Cloud Vertex AI in Belgium at europe-west1. Tenancy is stated separately as logical partitioning by user and by organisation. The marketing site itself is documented as setting no audience-measurement cookie and storing nothing on the visitor's device. The band's limb on what changes between tiers has nothing to bite on, because Jimini publishes a single hosted deployment in a single country rather than a set of options, and the limbs that do apply are answered to the cloud region.
Security Certifications and Trust Center
Independent attestation a buyer can pull without a sales call: SOC 2, ISO 27001, penetration test summaries, a trust center with current reports and named scope rather than a badge image.
Better evidenced than most records in this band and short of the top on two limbs. A trust centre is live at trust.jimini.ai, publicly reachable with no sales call, no email capture and no NDA click-through, listing the entity as ODYSAI, showing ISO 27001:2022, enumerating 114 controls grouped under business operations, governance, asset management and risk management, and reporting a last refresh a day before the check. The DPA goes further than the portal by publishing the certificate number IC-IS-2506288 and its validity to 17 June 2028, which lets a buyer verify the certificate independently. Two limbs are missing: no certification body or auditor is named on any surface read, and no attestation report was located as obtainable — the portal lists a single document without identifying it in what rendered. This is the R16 band gap appearing from the opposite direction, a record with a public portal, a certificate number and an expiry date but no named auditor and no reachable report. Checked 4 September 2026.
Model Supply Chain Disclosure
Which models sit underneath, whose they are, where they run, and whether the vendor commits to telling customers when that changes. A legal buyer inherits every dependency it cannot see.
The most complete supply chain disclosure this pull has seen. The CGU publishes a subprocessor table naming every provider with access to customer data, its role, its location and the specific basis of its zero-retention engagement: Mistral AI SAS for inference in France on documented activation, OpenAI's Europe deployment in the EU under a signed contractual amendment, and Google Cloud Vertex AI carrying Anthropic Claude models in Belgium europe-west1 under written supplier approval, alongside Scaleway and S3NS for hosting, Gradium SAS for voice transcription, DeepL for translation, Legal Data Hunter for legal research and Cloudflare for transit only with no storage. The privacy policy repeats the inference provider list independently. Change notification is committed at 30 days with a right to object and to terminate without penalty. One limb is only partly satisfied and is named here rather than glossed: Anthropic Claude is identified as a model family, but the specific Mistral and OpenAI models are not named, and the home page describes continuous selection and orchestration of the best models per task, so which model handles a given request is not disclosed.
Commercial Transparency
Whether a buyer can learn what this costs without entering a sales process: published rates, the unit being charged, what sits behind an enterprise tier, and what implementation adds.
Two named plans are published with their feature splits: an individual plan carrying all tools, the Microsoft extensions, AI workflows and data confidentiality, reachable through a self-serve signup with a two-week free trial and no sales contact; and an organisation plan adding collaboration features, training and onboarding, a usage dashboard and centralised billing, routed to the sales team. The CGU establishes the charging terms directly — self-serve rates are those displayed on the site at subscription, expressed in euros excluding tax, billed monthly or annually at the subscriber's choice, with 30 days' notice of any price change and a right to cancel before it takes effect. The individual figure itself was not read: the pricing page renders its numerals through the same animated counter component that obscured the home page statistics, and a search of the indexed page did not recover them. The CGU confirms figures are displayed at the point of subscription, so this is a limit on the retrieval rather than a gap in the publication, and it is recorded as such. Real pricing published for part of the range with the enterprise tier withheld is the middle band exactly.
Firm and Practice Coverage
Who the product is actually built for. AmLaw, midlaw, small firm and solo, in house departments, government and courts, and which practice areas are supported rather than merely claimed.
Segment coverage is described with real substance. The pricing page addresses independent avocats and organisations, splitting the latter into law firms and corporate legal departments, and the bar partnerships size the segment precisely, with the Paris agreement covering firms of one to twenty lawyers. The customer set spans both sides, naming law firms such as Gide Loyrette Nouel, Capstan, Hoche Avocats, FTPA and Franklin alongside corporates including CMA CGM, Petit Forestier, Léon Grosse and In Extenso. Practice coverage is evidenced through the agentic workflow library, which is filterable by affaires, civil, immobilier, pénal and public. The jurisdictional boundary is stated plainly and repeatedly as continental and French law. What is left open is the far edge: no statement says which practice areas are not supported, government use is not addressed, and the five workflow filters are a library taxonomy rather than a declared scope of support.
Legal Signals
What each signal meansA signal records what public sources say on the date shown. It is not a grade and it is not a recommendation. Where a signal reads Not addressed, it means the index did not locate the material in public sources on that date, which is a statement about disclosure rather than about the product.
Client Data in Training
Can material a lawyer puts into this product be used to train a model?
The published terms prohibit training on customer content. Not a policy page, the agreement.
The data processing agreement inside Part III of the CGU states that data is in no case used to train artificial intelligence models, neither Odysai's own nor those of its providers. The privacy policy repeats the prohibition separately for library documents, for prompts and responses, and for files reached through the Google Drive and SharePoint connectors. Every named inference provider is bound by a zero data retention agreement covering both retention and training, with the nature of each engagement specified per provider in the subprocessor table.
Prompt and Output Retention
How long does the product keep what a lawyer typed, and can that be set to zero?
A specific retention period is published and the customer cannot change it.
The retention table in the CGU publishes a fixed window the customer does not set: prompts and AI responses, and the document library, are kept for the life of the account plus 30 days. Word and Outlook add-in content is not stored at all, only the request and the generated response, and connector data is processed in real time with zero storage. A user can delete an individual conversation at any time from their personal space and can request early deletion, but the retention period itself is published rather than configurable. The Zero Data Retention badge on the home page describes the commitment binding the model providers, which is that no customer data is kept or trained on after inference; it does not describe Odysai's own storage, which the table sets out separately.
Ethical Walls and Matter Segregation
Does retrieval respect the firm’s ethical walls, or can the model read across them?
The product maintains its own permission model, documented, requiring the firm to keep it aligned.
The CGU security clause records logical partitioning of data by user and by organisation, and the privacy policy adds that raw files are stored in isolation per user or per organisation with access controlled through Auth0 by Okta on a least-privilege basis and access traced and audited. The permission model is Odysai's own rather than one inheriting a document management system's access control at query time. Segregation is documented at user and organisation level; no material describes permissioning between matters within a shared enterprise library.
Third Party Request and Subpoena Notice
If someone subpoenas the vendor for a firm’s data, does the firm hear about it first?
Published terms or policy address disclosure to authorities or in response to legal process, and no commitment or reservation regarding customer notice is located anywhere. The vendor has told the customer that data can leave and has said nothing about whether the customer hears of it.
The privacy policy addresses disclosure under legal obligation three times, stating for library documents, for Word add-in data and for Outlook add-in data that nothing is shared with third parties except for specific technical hosting needs or in the event of a legal obligation. No commitment to notify the customer of such a request was located, and no discretion over notice is reserved either. Searched the CGU including its Part III data processing agreement, the privacy policy, the legal notices and the trust centre on 4 September 2026; the professional secrecy clause states that no data in submitted documents is accessible to third parties, which addresses access rather than compelled disclosure.
Primary Law Corpus Provenance
Where does the law in this product come from, and does the vendor have the right to use it?
Coverage is described by jurisdiction with no identification of the underlying corpus.
The research product is described as returning answers anchored in recognised sources of law for continental and French law, and Legal Data Hunter is named in the CGU subprocessor table as the legal research provider located in the European Union, which identifies a supplier behind the corpus. No primary law database, publisher or collection is named, no licence or public domain basis is stated for any source, and no update cadence is published. A vendor blog post states that the models are fine-tuned on French legal information, without identifying that material.
Good Law Verification
Does the product tell you when the authority it just cited has been overruled?
The product instructs the reader to check the citation without performing the check.
The CGU preamble requires that results be validated by a legal professional and states they must not be treated as final, and the product is marketed on verifiable sources with answers anchored in recognised sources of law. Nothing on any located surface describes a check for subsequent history, or for whether an authority remains in force. The product retrieves primary law, so the question applies, and the published answer places the check on the reader rather than performing it.
Refusal and Uncertainty Behaviour
What does the product do when the answer is not in the corpus?
No located public material addresses what the product does when it cannot ground an answer.
No located public material describes what the product does when it cannot ground an answer. The CGU states that the AI system is in a development phase and that outputs may be neither reliable nor final, and the product pages claim verifiable sources, but no abstention path, no no-answer behaviour and no confidence or grounding score exposed to the user is described anywhere. Searched the home page, the product pages, the workflows page, the CGU, the privacy policy, the trust centre and the help centre on 4 September 2026.
Fabricated Citation Record
Does a public court record exist involving output from this product?
No court order, opinion or disciplinary record naming this product has been located as of the date shown. This is a statement about the public record, not a finding about the product.
The AI Hallucination Cases database maintained by Damien Charlotin was searched on 4 September 2026 on both the product name Jimini AI and the corporate name Odysai. No court order, opinion or disciplinary record naming the product was located. This records the state of the public record on that date and is not a finding about the product.
Bar Guidance Alignment
Has the vendor engaged in public with the ethics opinions its buyers are bound by?
Public materials engage with at least one named ethics opinion.
The vendor's help centre publishes an article on déontologie stating alignment with the recommendations of the Conseil National des Barreaux and setting out the ethical principles it considers applicable to AI use by avocats, alongside the EU AI Act compliance timetable. The CGU and the privacy policy both engage the professional secrecy obligation by citation, naming article 66-5 of the law of 31 December 1971 and article 2.3 of the Règlement Intérieur National. Partnerships with the Paris, Hauts-de-Seine and Nantes bars are published with their own pages. The guidance engaged is that of the French profession only; no other jurisdiction's ethics guidance is addressed and no clause-by-clause mapping of product behaviour to named provisions is published.
Billing and Fee Posture
Does the vendor address what happens to the bill when the work takes an hour instead of six?
Public materials claim time savings without addressing billing or disclosure.
Public materials claim time savings without addressing billing. The home page publishes an hours-saved-per-week-per-lawyer counter, and the Paris Bar partnership material is framed on immediate time savings on time-consuming legal tasks. Nothing on any located surface addresses what happens to a client's bill when AI-assisted work compresses billable time, and no per-matter record of AI-assisted work is described in the product or the documentation.
Outside Counsel Guideline Readiness
Can a firm get this vendor through a client’s AI clause without a bespoke negotiation?
A subprocessor and model provider list plus client facing disclosure material is published or available without an agreement in place.
Each of the three limbs is evidenced separately. Subprocessor list: the CGU publishes a current table naming every provider with access to customer data with its role, location and zero-retention basis. Model provider statement: the same table names the three inference providers, Mistral AI SAS, OpenAI's Europe deployment and Google Cloud Vertex AI carrying Anthropic Claude models, and the privacy policy repeats that list independently, so which providers see client content is stated rather than inferred. Forwardable client-facing material: the data processing agreement is published as the final section of Part III of the Conditions generales d'utilisation, headed 'Donnees personnelles, Accord de traitement (DPA)', reachable at https://www.jimini.ai/conditions-generales-utilisation without executing anything or contacting sales, and it carries the Article 28 terms, the subprocessor tables, retention periods, breach notification and the professional secrecy clause. All three artifacts are published in French only, the vendor's own language.
Court Disclosure Support
If a judge’s standing order requires an AI disclosure, can the product produce one?
Some elements of the record are available, short of a document level export.
The product is marketed on answers anchored in recognised sources of law with verifiable sources, so retrieved sources are addressed. No export of a per document record is described, nothing states that the model used for a given request is recorded or disclosed to the customer, and no record of human verification is mentioned. The multi-model orchestration architecture means the model behind any particular output is not surfaced. No disclosure guidance or template for a court was located on any surface checked on 4 September 2026.