The AI Legal Index Brief

The week in one line

Legal AI spent the run up to Halloween hunting ghosts. Legora, Filevine and Lulius shipped checks that catch a phantom citation before a judge does, Harvey made the ethical walls a firm keeps in iManage govern its own agents, and DiliTrust wrote down that hosting data in your country and processing it there are two different sentences. Twenty three connectors arrived, and most of them led with whose permissions they use.

This issue covers October 19 to 30, with 77 changes across 77 vendors. Once again nobody shipped twice, which is either discipline or a very well run industry calendar. Fifty eight were verified at the source and nineteen partially, fourteen were rated high impact, and twenty nine of the 77 landed on October 21 and 22. Two entries dated October 19 arrived after the last issue closed. It is Halloween, and the period obliged: the biggest theme was ghosts.

Hunting the phantom citation

The case that does not exist is legal AI’s ghost story, and unlike most ghost stories it comes with docket numbers. This period thirteen vendors shipped something to catch one before a judge does, and the check moved from a warning at the end to a mark on the sentence.

Legora took Ontological Legal Research live in the United States, the approach it described in September. It pairs primary law Legora collects itself, including slip opinions from court websites within 24 to 48 hours, with a prebuilt map that links defined terms to the cases that interpret them, overruled cases to the decisions that overruled them, and rules to their exceptions.

Legora says the structure makes its Agent 40% faster on an internal set of thousands of research questions, and that its AI native citator comes next. The figure measures speed, on Legora’s own questions, so accuracy is the result to ask for. Research that knows which case overruled which is a direct answer to AI citing law that is no longer law, the failure a hallucination check cannot see.

Filevine changed what a warning does. Case names in LOIS drafts that fail its anti hallucination check now appear as clickable highlights instead of a static notice. Clicking one opens a case law lookup listing the full citations of opinions that might match, so the lawyer confirms the reference without leaving the draft. It is live on all US production environments.

A flag that opens the candidates gets checked. A flag that says “verify your citations” hands the lawyer their own job back, usually at eleven at night.

In Germany, Lulius added a one click citation check to its editor. It matches every statute section, court file number and ECLI in a draft or uploaded Word file against its corpus of federal law and federal court decisions since 2010, and marks each one found, not found, or outside the data set, such as state law, EU law or lower courts. Running it uses no query credit.

That third mark is the clever one. A citation the tool could not check is a different thing from a citation it could not find, and a lawyer who knows which is which knows where to spend the next ten minutes.

Litmas AI made its Agentic Associate publicly available for the first time. It chains an evidence mapper, jurisdiction specific research with two layers of citation checks against US court opinions, and drafting into one workflow that produces summary judgment, motion to dismiss and discovery drafts with pincites to the record. Each project starts with a session plan the attorney must review before anything is drafted.

Clearbrief’s Word analyzer now recognizes the public domain citations used by Illinois, Ohio, Pennsylvania and Louisiana courts and formats them correctly in the Table of Authorities, links every title of the California Rules of Court, and finds the Lexis source for more Westlaw citation variants. It also reworked its quote and data mismatch checks so fewer flags fire, each with a clearer reason. Fewer, better explained flags are the ones that get acted on.

The same move reached review and investigations. Relativity’s new agent for Assist attaches citations to individual sentences, with working links and highlighting in the document, and shows its plan, searches and retrieved documents live while the answer is generated. Nuix launched the Nuix Neo AI Agent, which answers by calling more than 20 Neo tools, links each answer to its source documents, records every step it takes and lets the user export the whole trail, which is the shape a regulator or a court will eventually want it in.

Patents got their own version. XLSCOUT’s Deep Research 2.0 adds an evidence review agent that checks, claim element by claim element, how well the cited product evidence supports each claim chart scoring 0.70 or above, then rates each report High, Moderate or Low confidence. Reports below the line are marked not reviewed rather than quietly passed. It now runs on up to 20 patents at once.

Contract and privacy tools attached the source to the cell. Pactly rows can hold a schedule or side agreement beside the main contract, extraction reads them together, each cell records which document its answer came from, and adding or removing a supporting document marks the row’s answers outdated. DataGrail’s Vera now links each assessment suggestion to the documents it came from, with page numbers. Recorded Partially Verified. Chamelio takes a user from researching a requirement to a saved company policy, with the sources cited and a published list of the regulators, courts and official journals it draws on, by jurisdiction. Luminance’s External Sources lets Lumi search curated legal research sources, the web and LexisNexis Protégé from inside the contract workflow, with citations attached.

Perspective AI went furthest. Anything in quotation marks in an analysis session is now exactly what a participant said, paraphrases appear without quotation marks and still link to the moment in the conversation, and it removed its Generate Infographic feature because it could show findings participants never said. Removing a feature on purpose is the rarest release note in software, and this one went for putting words in someone’s mouth.

Our read

Two years ago the check was a disclaimer. Last year it was a warning. This period it became a highlight that opens the candidates, a mark that says outside my data set, a cell that names its document and a trail that exports. Each of those hands the lawyer the material for the check instead of asking for trust. What none of them hands over is a rate. Legora’s 40% is speed, Ruli’s benchmark below is a score, and nobody this period published how often the check finds something. That number is the count at the end.

Every door in the house is open

Trick or treating in legal AI now runs in both directions. The assistants knock on the vendors’ doors, the vendors knock on the assistants’, and twenty three of this period’s 77 changes were a door opening. What matters is what each one lets through.

Harvey made its Connector Library generally available on October 21, and three vendors on this index walked in the same day. iManage’s connector, built on the iManage MCP Server, lets joint customers search and open iManage documents, matters and work product from inside Harvey, with existing iManage permissions limiting what each user, and Harvey, can see.

Mary Technology brought in its record of a litigation matter, the documents it has mapped plus the team’s findings and review history, so litigators can test witness accounts against documents and weigh evidence for and against individual allegations from inside Harvey. Recorded Partially Verified. PointOne brought time and billing data into Harvey’s agents, so lawyers review time entries in conversation, supervising partners see pre bills and matters nearing budget, and firms can compare AI assisted work with the rest of the matter.

That last clause is the record this Brief counted in September and found almost nobody keeping: one that joins what the AI did to what it cost. Harvey has become the place other legal tools go to be used, a sentence its competitors will read twice. Eight days later Harvey closed the loop on walls, which is in the next section.

The MCP servers kept coming, and they kept the decisions at home. PERSUIT’s MCP brings its Persi agent into Claude, ChatGPT Business and Enterprise and Microsoft 365 Copilot, where users can get price benchmarks for a matter, draft a request from their own templates, analyze firm proposals, compare firms and panels and check invoices against agreed terms. Each user signs in with their own account, and requests are still reviewed, published and awarded inside PERSUIT.

Precisely released an MCP server in beta that lets Claude, ChatGPT and other assistants search projects, look up templates, draft contracts from a template questionnaire and generate documents. The assistant acts with the connecting user’s role permissions, an admin must allow MCP integrations first, and the audit log marks which entries an assistant created. Luminary opened its trust and estate client data to ChatGPT, Claude and Copilot, with prepared skills that answer who serves as trustee, project estate tax for a future year of death and build estate flow diagrams from its own entity values and tax logic. Recorded Partially Verified.

PatSnap put its Life Sciences and Patent MCP servers inside EMET, BenchSci’s research environment for preclinical drug discovery, so scientists can map the patent landscape around a target while they are still choosing one. IP teams at drug developers should expect earlier and more informal freedom to operate questions as a result.

Claude collected two more. Aptus.AI published a connector that passes legislation and case law into Claude with short verbatim quotes, the in force status and a link to the full text, and retrieves judgments by ECLI or citation, including from the Court of Justice of the EU, read only through the user’s own account. Recorded Partially Verified. Patlytics’ existing connector now feeds the ready to use artifacts Anthropic added to Claude for Legal on October 28, including a patent claim comparison built on verbatim claim text rather than a model’s paraphrase.

Avvoka extended the headless idea from the last issue to volume. Headless Massdraft lets lawyers run its bulk document generation and signing workflow from Harvey, Legora, Claude or Copilot. The assistant turns the unstructured deal data a client sends into the spreadsheet Massdraft expects and builds the template from the pro forma, while Massdraft still does the assembly and tracks the signatures.

Legal tools also opened to each other. Everlaw users can now search their Everlaw projects from inside Legora through a live, read only connection that turns plain language into Everlaw searches, including Bates lookups, but cannot change, export or delete a document. An Everlaw admin switches it on, and it works in the US, UK, EU, Canada and Australia environments. Recorded Partially Verified. Libra became an add in inside Wolters Kluwer’s Benelux research databases, InView Legal, InView Tax and Legal Intelligence, answering beside the source a lawyer is reading. CoCounsel Legal’s classic version gained an Otto Schmidt connector that pulls more than 120 modules of German commentary into its research, with the rebuilt agentic version due to get it in early 2027.

Lexis+ AI made the biggest single source move. Lexis+ with Protégé can now pull Lex Machina litigation analytics, Intelligize disclosure data, CourtLink dockets and Law360 reporting into one prompt, alongside LexisNexis content and firm documents from connected document systems. Users pick the sources or let the orchestration layer choose, it is rolling out to US customers this month, and access follows what the firm already licenses.

The document systems held their line as the record. DeepIP and Solve Intelligence both shipped NetDocuments integrations where attorneys pick matter files in, each on their own NetDocuments account, and save finished work back into the matter as a new version or, at Solve, a new file. Hebbia’s Datasite connector brings data room documents into Hebbia on the user’s own Datasite credentials, syncing only what each person may see, and turns a full diligence question list into a memo with every point cited to the source page. Ordalie now opens a file from a connected source with that connection’s permissions and never makes a copy.

And the smaller doors. Wordsmith brought playbook review into Google Docs, with clauses outside agreed positions, proposed fallbacks and amendments as tracked changes. Irys One Intake now watches Slack channels, read only, posting nothing and pulling no history from before setup. Billables AI and VXT both started landing time entries in Filevine without a manual step.

Our read

Three issues ago the question was whose permissions a plugin uses. This period almost every launch answered it in the first paragraph: the user’s own account, the source system’s permissions, read only unless someone chooses otherwise. The newer question is where the work finishes. PERSUIT still awards in PERSUIT, Avvoka still assembles in Massdraft, Everlaw stays the record, NetDocuments keeps the version history. The assistant is becoming the front door, and the vendors are taking care to keep the furniture in the house.

Who else is in the house

The last issue watched confidentiality turn into a price plan. This period it turned into controls a firm can point to.

Harvey now enforces ethical walls and information barrier policies synced from iManage Security Policy Manager, generally available after an early access period. With enforcement on, the policies govern the Threads, Vaults, Review Tables and Spaces tied to a client matter, iManage stays the system of record, and audit logs record wall syncs, access attempts and enforcement changes.

Walls were the most common reason firms kept AI away from conflicted matters. The wall a firm already maintains now governs the agent too, and the risk team gets a log showing it held, which is what they will need when the conflict question comes up a year later.

DiliTrust published documentation that separates where customer data is stored from where AI requests are processed, and names four processing options: frontier models run by Google Cloud in Europe or the US by default, open weight models DiliTrust operates in France, a specialist EU provider such as Mistral AI, and a customer’s own model through an OpenAI compatible API. It commits to zero retention and no training for the first two, and states plainly that customers hosted in Canada, the UAE, Morocco and Saudi Arabia have AI requests processed outside their country. Recorded Partially Verified.

Hosting your data in your country and processing it there turn out to be two different sentences. DiliTrust wrote both down, which changes the answer a legal department gives to its own data transfer question.

AnwaltGPT added a workspace governance layer. Administrators choose which chat models every member may use, and the choice covers helper calls and fallback models too. They can make pseudonymization in chat mandatory, and Deep Research through a separate provider is blocked when it would break the model approval. A firm now decides centrally which models see client matter text instead of trusting each lawyer to pick correctly.

Gaius-Lex made encryption of files on My Drive mandatory for every cloud account, encrypting existing files in the background at next login. Once a user saves eight backup codes, Gaius deletes its own server copy of the file key, so a copy of the database plus the server keys is no longer enough to read the files, and shared folder keys moved to a hybrid post quantum scheme. Recorded Partially Verified. A vendor that deletes its own copy of the key has written a frightening story for its support desk and a comforting one for its customers, and firms should decide now where lawyers keep those codes.

Opus 2 stated in a hearing AI webinar that client documents, transcripts and anything AI Assist produces are not used to train or improve any model, that AI Assist can be switched off for a whole workspace on request, and named the models behind it: Amazon Nova and Anthropic’s Claude Haiku and Sonnet, chosen per task. Recorded Partially Verified. Legito integrated Claude into its data extraction and Kedy assistant with a setting to choose which Claude model runs, which is still a question most document automation vendors leave unanswered.

Access controls got more granular. Ivo admins can restrict which documents individual users see inside a repository room, off by default, and a new document audit ledger records every upload, deletion and room change. Athennian admins can build dedicated entity pages for teams such as Tax or Finance with role based access down to the data point, on an add on rather than the base plans. Lawcus’ Matter Access Center lists everyone connected to a matter with their access and invitation status, and lets a firm assign a task to an outside contact without opening the file.

The record of who did what got names. Every Checkbox matter now has a History tab where each change names who or what made it: a person, a workflow, the Checkbox AI agent or an integration. Ethyca’s Astralis 3.0.6 writes a parseable audit line for security relevant actions, never logging identity values or request contents, and ships switched off. TrustArc added an admin setting that turns off AI completion for assessment respondents without a TrustArc account, so an outside vendor answering a privacy questionnaire does not get a model answering for it. Relyance AI can now trace user input all the way to the LLM vendor it is sent to, a path that data maps built before AI agents cannot show.

Our read

Every confidentiality release this period is a sentence a firm can put in its own documentation instead of pointing at a vendor’s trust page: the wall is the one we keep in iManage, the models are the ones we approved, the key is one the vendor no longer holds, the ledger says which agent touched which document. That is progress. Two of the best controls, Ivo’s visibility rule and Ethyca’s logging, ship switched off, and a control that ships off is a control somebody has to go and find.

The agent dressed as the law firm

It is Halloween, and the agent spent the two weeks in costume. In several launches it answered the phone, worked the inbox or faced the client under the firm’s own name.

August launched August Products, which lets a firm package a legal service as a fixed fee product that clients buy and use through a site in the firm’s own branding. Clients sign in, upload documents and receive finished work produced by the firm’s agents, built from its precedents and playbooks or adapted from August’s library. Lawyers review output at checkpoints the firm sets, and their corrections can be saved as skills that later runs reuse. Recorded Partially Verified.

That moves AI from a tool inside the firm to a service clients use directly, so the checkpoints are the firm’s main control over what goes out under its name. Where a lawyer must sign off before a product goes live is a decision the firm makes once, and it had better make it before the first client logs in.

LegalZoom opened its AI Assistant to all business owners for free, including those who formed their business elsewhere. It answers status and compliance questions from the customer’s own entity record, flags filing deadlines, links the filing product needed and, when judgment is required, routes the customer to a licensed attorney or specialist. The handoff is where a consumer legal tool stays clear of unauthorized practice, and LegalZoom built it into the flow rather than the footer.

At the front desk, Smith.ai’s AI Receptionist now recognizes ten caller types instead of four, adding other law offices, government agencies, case related third parties, opposing parties, media and incarcerated callers, each with its own routing to the AI or a live agent, required intake fields and call outcome. Every new type is on by default. Recorded Partially Verified. Opposing counsel reaching the front desk and being greeted like a prospective client is the haunted house every firm has walked through once.

Litify launched Call Intelligence inside its case management platform. Intake staff get live caller identification, transcription by speaker and case type playbooks, questionnaires prefilled with suggested answers linked to the transcript and the audio, and leads scored against firm defined criteria with cited reasons. It connects to more than 20 telephony providers. Lead Docket shipped Lead Viability, which evaluates each incoming lead against positive and negative indicators for 24 case types. A tool that decides which prospective clients get a call back first deserves a look at what it scores low, because a missed viable case is lost revenue and a person who never heard from the firm.

In house, SpotDraft released an auto triage agent that runs the moment an intake arrives from email, Slack or the web form, setting assignee, priority and due date and creating subtasks, following plain language instructions the admins wrote. Each decision is logged with its reason, and handled intakes carry a badge. Tomorro gave its contract review agent Oro an email address. Forward a Word or PDF contract and Oro replies with the review, or a redlined Word file for a deep review. It is in beta, and legal should decide which contract types may be reviewed this way, since the reply goes straight back to whoever sent it.

Hona began rolling out Global Lia, where personal injury firm staff ask firm wide questions and get reports on message volume, response rates, cases in progress and open escalations. Users grant Lia permission to act as them for 90 days, and it sees only what their own account can. Recorded Partially Verified. Casefleet’s new Case Stages Board shows each matter as a card in a litigation stage lane, and Casey, its assistant, can move a case forward or back or close it on request.

Our read

When the agent faces the client or the caller, the review point is the firm’s reputation. The launches that got this right carry a reason with every decision: SpotDraft logs why it routed, Litify links each answer to the transcript, LegalZoom routes to a lawyer when judgment is needed, August lets the firm set the checkpoints. The costume is the firm’s. The log is what proves who was inside it.

The house rules, now with an approval step

The October 10 issue watched a firm’s house rules become files its agent reads. This period the files got owners.

Newcode launched Configurable Skills, which turn a firm’s processes, playbooks and risk positions into reusable agent capabilities written in plain language, built from instructions, approved knowledge, templates and examples of work to avoid. The agent can turn a finished conversation into a skill, and administrators control who can create, edit, approve, publish and share them, down to the team or practice group.

The approval step decides whose judgment the agent repeats. Publishing a skill is publishing a precedent, and treating it that way keeps one group’s shortcut from becoming the whole firm’s habit.

Simpliciter added personal skills. A user asks the assistant to save a reusable method for recurring work, such as comparing contracts or preparing client updates, and can customize an existing skill without changing it for anyone else. Personal skills sit outside central review, so firm standards still need a shared home. Vesence shipped Library, one place for a user’s skills, integrations, agents and apps, including any MCP server, where each item stays private or is shared across the organization for colleagues to add in one click. Allowing any MCP server is the setting IT will want to govern.

And August’s corrections saved as skills, above, close the loop. The lawyer’s red pen on this run becomes the agent’s rule on the next one.

Our read

Skills started as a way to make the agent sound like the firm. They are turning into the firm’s precedent system, with the questions precedent always raised: who approved it, who may use it, and what happens when it is wrong. Vendors that ship the approval step with the feature have noticed. Personal skills on their own hand every lawyer a private playbook, which is how firms worked before AI, with better tooling.

Market notes

Patent work had a busy two weeks beyond the connectors. IPRally released Drafting, which plans and writes a patent application from invention materials, seed documents and known prior art. A prior art check analyzes the ten most relevant documents feature by feature and suggests amendments, four quality reviews cover clarity, claim support, sufficiency of disclosure and reference numbers in EPO and USPTO variants, and every AI change appears as a colored diff the user accepts or rejects. A search vendor moving into drafting means the model that finds the prior art now shapes the claims. Patent Bots introduced Sprocket, a drafting assistant in its Word add in that works step by step rather than producing a whole application at once, carries the attorney’s corrections forward and inserts claim amendments with strikethrough and underline already applied. MarqVision rebuilt Seller Intelligence around the operator rather than the listing, clustering storefronts that share a business registration or VAT number, ranking each infringing seller on five stated factors and proposing a next step.

Practice coverage widened. At ClioCon, Clio announced a drafting experience in Clio Work that combines generative drafting, automation from firm approved templates and a library of more than 40,000 fillable court forms covering all 50 states, federal courts and USCIS, with live redlining, a full document editor and a Clio for Word add in, available now to all US Clio Work customers according to LawSites. Recorded Partially Verified. Small firms that buy court forms from one vendor and drafting AI from another can now do both inside the practice system, and coverage of a firm’s own courts is the part to test.

MinuteBox extended Automated Compliance to corporations and LLCs in every US state, DC and the inhabited territories, and for the first time to trusts, nonprofits, partnerships, professional corporations and cooperatives, with deadlines calculated from each entity’s formation date, fiscal year end and jurisdiction and existing entities rechecked automatically. Lexroom launched Lex Desk, a practice management module that brings case files, the client register and Italian civil electronic filing together around the AI, with calendar, deadlines and fee quotes due in November and billing after that. Sapient-IA, Lefebvre Giuffrè’s Italian assistant, added case law, though the page does not say which courts or years. Recorded Partially Verified.

In review and contracts, Litera’s Kira added three generative smart field types, including an Extraction type that returns verbatim excerpts, with a preview before a field is applied to a worksheet, and a Lito skill that drafts a first red flags report from a grid export, the most time consuming deliverable in a deal review. Oneflow launched AI Insights Premium, where customers define their own clause rules, fields and thresholds and flag what is missing or off standard in team views. Recorded Partially Verified. Parambil now reads DICOM medical images, CT, MRI, radiograph and ultrasound, in a built in viewer, and its Ask AI compares them with the written radiology report, which can surface what the report left out and still needs a qualified expert before it shapes a demand. Wisedocs made API v1 generally available, with an OAuth scope per operation and server side encryption required on uploads. Lawmatics brought its Merlin assistant to the mobile app, for reading and reporting only.

Vulcan Technologies described Justinian 3.0.0 as an AI software engineer for government. Every chat gets a workspace with a full development environment where the agent writes code, runs tests and pushes changes to a connected repository for the team to review, plus browser use the user can watch and take over and full computer use inside FIPS encrypted sandboxes managed by tenant administrators. Recorded Partially Verified. A legal AI vendor shipping a software engineer is the plot twist nobody put on a roadmap slide, and the review before a merge is the control to examine.

Evidence arrived with numbers attached, most of them the vendor’s own. Ruli published results from the 2026 Legal Benchmarks evaluation, a lawyer designed test on real in house contract and extraction tasks: an overall 72.0 against a legal AI average of 63.7, a full task pass rate of 31.1% against 24.7%, and 41.7% of extraction tasks fully correct against 31.3%. Recorded Partially Verified. A shared benchmark is more useful than a case study because buyers can compare it across vendors, and an industry average that passes roughly one task in four makes the review step look less like a formality and more like the product.

BigHand reported that Penningtons Manches Cooper sees 12% better recovery on matters run through Matter Pricing, on the firm’s own reporting, with the Ayora AI pricing capability described as planned rather than shipped. CUBE reported that CMC Markets cut time on regulatory monitoring by 50% under a license covering 49 issuing bodies. BoostDraft reported that Kawakin Holdings’ legal group cut its final contract format check from up to 30 minutes to about 5 to 10, and a comparison table from more than an hour to about 5 minutes. All three are recorded Partially Verified, and all three are the customer’s own figures.

What the two weeks say about the category

Seventy seven changes, and the theme is the check. The citation check moved onto the sentence and into the editor. The permission check moved into the first paragraph of every connector launch. The wall check moved from the firm’s document system into its agents, and the model check moved into an admin’s approval list.

The things that go bump in legal AI are a case that does not exist, a data path nobody drew and a caller the front desk misread. This period the vendors went looking for all three with the lights on, and most of them published where they looked.

What almost none of them published is how often the check finds something. That is the number a lawyer would most like to have, and the count below shows how few vendors give it.

Index Answer

Which legal AI vendors publish how often their citations are wrong?

Almost none, which is the quiet part of this period’s ghost hunt. The AI Legal Index grades every vendor it covers on Citation Accuracy and Hallucination Disclosure, which asks whether the vendor publishes measured accuracy on citations and assertions, grounds its output to primary sources, and says plainly what the system does when it does not know.

Of the 466 vendors graded, 16 publish measured accuracy or hallucination evidence an outsider can check. That is the A, and it is the only grade that answers the question in the heading.

Another 245 document grounding to primary sources and a cite check without publishing a number. That is where most of this period’s launches would land: a highlight that opens the candidate cases, a mark that says a citation sits outside the data set, a cell that names its document. Those are real checks. None of them comes with a rate.

A further 182 claim accuracy without documentation, and the remaining 23 publish nothing on the question at all.

On the related signal, 82 of the 466 document what the product does when it cannot ground an answer, which is the design decision every fabricated citation passed through on its way into a brief.

Legora’s 40% is a speed figure on its own test set, Ruli’s 72.0 is a score on a shared benchmark, and this period produced no new accuracy rate on citations from anyone. These records describe each product as last researched, so a launch from this period may sit ahead of its record.

A vendor that shows the lawyer where to check has done half the job. A vendor that says how often the check finds something has done the other half, and that half is still rare.

Every vendor’s grade on this axis, with the good law and refusal signals beside it, is on the accuracy disclosure page, and the grading method is on the methodology page.