D

Doctrine

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading; this axis does not turn on it. The models are the engine of the search and drafting layer and sit on a product that predates them. Underneath is a corpus the vendor describes as 80 million interconnected French legal documents, searchable by keyword, court, company, advocate, date and decision number, which was the product for years before generative features arrived. On top of it sit capabilities that are model work: a legal chatbot answering in natural language from the corpus, a documents chatbot answering only from files the user supplies, a company chatbot over corporate and litigation records, decision and instrument summarisation, extraction, comparison, conformity checking and translation, and the two Flow products, Flow Litigate assembling classification, chronology, statement of facts and bordereau for a contentious file and Flow Counsel analysing contracts for risk. Strip the models out and a large searchable legal database remains, which is what most of this lane's incumbents sold before 2023. Vendor blog, home, features and pricing surfaces read via the search index 7 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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading; this axis does not turn on it. Grounding is real, documented and unusually specific about its boundaries, short of any published measurement. The vendor's own blog states that the Chatbot Juridique answers exclusively on the basis of the legal corpus while the Chatbot Documents relies only on the documents supplied, and that both always cite their sources. It sets out the reasoning as well as the claim, arguing that a professional legal tool must let the user verify the origin of every assertion rather than merely answer, and that the way to get the best from these systems is to confine them to precise legal corpora they can then cite so verification is easy. That is a described retrieval method plus linked primary sources, which is what this band asks for. Two further things are recorded because they are rare. The vendor publishes an analysis by Damien Charlotin, the researcher behind the AI hallucination database against which every record in this index is checked, of the first French decision involving AI hallucinations, on its own marketing blog. And it publishes an argument that general-purpose AI is unsuitable for legal work because a fabricated authority or a stale rule has concrete consequences. What is absent is measurement: no accuracy figure, no test set, no evaluation and no error rate anywhere located. Vendor blog posts read via the search index 7 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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading. Autonomy is claimed across substantial work and the only oversight offered is source citation. What the product does unattended is described plainly: Flow Litigate takes a contentious file and produces classification, a chronology, a statement of facts and a bordereau automatically; Flow Counsel analyses contracts, identifies risks and returns what the vendor calls legal recommendations in a few clicks; the company chatbot condenses each corporate instrument into a summary described as clear and exhaustive. The oversight side rests entirely on verifiability of sources: both chatbots cite what they relied on, and the vendor's stated position is that the user must be able to check the origin of every assertion. That is a real review surface and it is credited. What is not published anywhere readable is a control structure: no statement of what a person is expected to check before a generated chronology or bordereau is used, no threshold at which the system defers, and nothing on what happens when a summary or a risk recommendation is wrong. Vendor blog, home and features surfaces read via the search index 7 September 2026.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading; this axis does not turn on it. Real deployment evidence with substance, short of figures for what changed. The scale is first party and recent: the RELX completion release of 3 September 2026 states that Doctrine is used every day by 27,000 legal professionals across France, Italy, Germany, Spain and Luxembourg, and describes its customers as law firms of all sizes, large multinational corporations and public sector organisations including French government ministries, local authorities and universities. That is a named sector spread rather than a logo wall. An institutional partnership with the Paris Bar gives members preferential subscription terms, which is a third party with its own reputation choosing to stand behind the product. A named advocate in company law is quoted on the record, with a linked interview, saying he uses the platform systematically in both advisory and contentious work to get a precise and exhaustive picture of a company's position. What is missing is the quantified half: no figure for time saved, matters handled or outcomes changed, and no method behind any claim. RELX release, Paris Bar partnership page and vendor features surface read 7 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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable under R85, and this axis turns on it. Confidentiality is asserted in general terms in the material that can be read, and the terms that would substantiate it could not be reached. What is readable is a marketing statement rather than a commitment: the vendor describes the platform as reconciling public legal data with the customer's own documents in a secure workspace. That matters because the product does ingest client material, through the documents chatbot which answers only from files the user supplies and through connectors into a firm's document management system. Nothing readable states who at the vendor may access those documents, whether they are segregated between customers or matters, how long they persist, or what any model provider may retain. The conditions générales at /cgu and the personal data guide both return only a meta description through the search index and the site disallows automated access, so this grade records what could be established rather than a finding that the vendor is silent: the surfaces exist and were not readable. On a reading of those pages this row is rebuilt and redated. Vendor home and pricing surfaces read via the search index 7 September 2026.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable under R85, and the disclaimer that would ordinarily sit in it could not be read. What is readable describes the product in advice terms while nothing addresses where its output stops and a lawyer's judgement begins. Flow Counsel is described as analysing contracts, identifying risks and returning legal recommendations; the legal chatbot answers substantive questions of French law; and the marketing addresses advocates and in-house counsel directly. Against that, the vendor's published position on verification is real and is credited here rather than on the citation row: it argues that a professional tool must let the user check the origin of every assertion, which places responsibility with the reader. What is absent is any statement that an output is not legal advice, that the subscribing lawyer remains responsible for the determination, or what the tool does not cover. The audience is unambiguous and professional, which is recorded rather than credited. On a reading of the conditions générales this row is rebuilt and redated. Vendor blog, home and features surfaces read via the search index 7 September 2026.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading. Responsible-AI positions are published without a mechanism a buyer could audit. The vendor writes seriously and publicly about the failure modes of legal AI: that hallucinations have concrete consequences for a strategy or a file, that general-purpose models are unsuitable for professional legal work, that reliability depends on the quality of the sources a system draws on, and that context management is the key to getting useful output. It publishes an outside researcher's analysis of the first French hallucination decision on its own blog. All of that is a stated position on how legal AI should behave. None of it is governance of this vendor's own models: nothing names who is accountable for the chatbots' behaviour, describes what is evaluated before a change ships, publishes any result, or discloses whether output quality is uneven across areas of law, jurisdictions or the five languages the platform serves. No responsible-AI policy, model card or bias statement was located on any readable surface. Vendor blog posts read via the search index 7 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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable under R85, and this axis turns on it. A general assertion of security is readable and none of the substance behind it is. The vendor describes the platform as a secure workspace in which public legal data is reconciled with the customer's own documents, and that is the whole of what could be reached. None of the five things this axis grades is establishable: no retention period for documents uploaded to the documents chatbot or drawn through the document management connectors, no deletion commitment, no access control statement, no subprocessor list, and no incident notification practice. The conditions générales and the personal data guide at /centre-donnees-personnelles/guide/donnees-personnelles both exist and return only a meta description through the search index, and the site disallows automated access, so this records a limit on the reading rather than a finding that the vendor publishes nothing. That distinction matters more here than on most axes, because a French platform serving government ministries under the GDPR is very unlikely to be silent on these points in documents that could not be opened. On a reading of those pages this row is rebuilt and redated. Vendor home surface read via the search index 7 September 2026.

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.

Not yet assessed.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading; this axis does not turn on it. Real integrations into the systems the work already lives in, named and briefly described, short of documentation. Two are published. The platform connects to a firm's own databases and document management systems, which the vendor frames as activating the knowledge already stored in the GED and limiting interruptions to the work, so a user's own precedents and files are searchable alongside the public corpus. And Doctrine runs inside Microsoft Word, described as enriching documents without leaving the working environment, which is where drafting actually happens for the buyer. Those are connections into the right two places for a legal research and drafting product rather than a list of unrelated logos. What is not published on any readable surface is depth: no named document management systems, no statement of direction or sync behaviour, no configuration requirements, and no implementer documentation. The integrations page itself could not be reached. Vendor pricing and home surfaces read via the search index 7 September 2026.

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.

Not yet assessed.

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.

Not yet assessed.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading. The vendor describes its architecture and identifies nothing underneath it. What is published is the retrieval design rather than the models: the legal chatbot is confined to the legal corpus, the documents chatbot to the user's own files, both cite sources, and the vendor explains the general principle of supplying a model with specific relevant material to suppress hallucination. Third-party profiles describe a proprietary retrieval-augmented architecture. No model, model family, provider, hosting arrangement or inference location is named on any readable surface, and no commitment is given to notify customers when any of that changes. No subprocessor list was located, and the personal data guide that might carry one could not be read. This is graded as it stands rather than as a finding of silence, since the surfaces most likely to carry a supply chain disclosure are the ones the site refuses to serve. Vendor blog and product surfaces read via the search index 7 September 2026.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading. The shape of the offer is published and no figure could be established. The pricing page presents three buyer profiles rather than a single enterprise funnel, being the individual advocate, the law firm and the in-house legal department, and it answers the question that matters most about a research subscription: each subscription gives unlimited access to all Doctrine Pro features across every area of law, without segmentation, so a buyer is not paying per practice area or per module. A free trial is offered. The Paris Bar partnership publishes a real commercial term rather than a rate, giving members several free months and a discount over a twelve-month commitment, with a second licence free in the first year for two-advocate firms. No price, band or unit of charge could be established from any readable first-party surface; third-party sources state that Doctrine quotes on request, which is not first party and is not written. No VendorPricing row is written, since a row belongs to vendors graded A or B on this axis. Pricing page and Paris Bar partnership page read via the search index 7 September 2026.

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.

Agreement shape: the customer agreement exists and was unreadable this session, a researcher limit rebuttable on a later reading; this axis does not turn on it. Coverage is described with substance across buyers, jurisdictions and subject matter, and the boundaries are left open. Buyer segments are published as the structure of the offer itself, with distinct propositions for the individual advocate, the law firm and the in-house legal department, and the RELX release adds public sector bodies including French government ministries, local authorities and universities. Jurisdictional reach is stated and repeated: France, Italy, Germany, Spain and Luxembourg, with the vendor framing it as a global view with local impact. Subject-matter coverage is stated in the strongest available form, that a subscription reaches every area of French law without segmentation, with a tax encyclopaedia called out separately. The corpus is quantified at 80 million interconnected legal documents. What is not stated is any limit: no area of law, jurisdiction or matter type is named as unsupported, no depth comparison is offered between the five countries, and nothing describes what the platform does not cover. Pricing, home and features surfaces and the RELX release read 7 September 2026.