← All issues

The AI Legal Index Brief

November 1 to 14, 2026 · Published November 15, 2026

The week in one line

Legal AI spent two weeks asking for write access, and mostly got it. Conga CLM’s agent can now terminate a contract, once you confirm you meant it. Relativity lets Copilot set up matters and decide who gets into a workspace, and Agiloft shipped an agent that edits records in bulk. Inhouse put attorney time behind what its AI drafts, and iManage published a bias test its own feature had to pass before launch.

This issue covers November 1 to 14, with 76 changes across 76 vendors. Fifty six were verified at the source and twenty partially, and eighteen were rated high impact. Thirteen landed on November 5 alone, a Thursday that a lot of release calendars seem to share.

The agents got the edit button

The October 19 issue watched the first connectors start writing. This period writing was the main event, and the actions got heavier.

Conga CLM went furthest. Its MCP server gained write tools and its AiMe assistant became a contracting agent. A user can now create, amend, renew and terminate contracts, generate documents and send for signature from a chat in AiMe or Claude.

Anything that creates, changes, deletes, terminates, cancels or notifies needs the user to confirm the contract, the action and its consequences first. Recorded Partially Verified.

Terminating a contract from a chat window is a sentence that would have ended a meeting two years ago. The confirmation step is the control to test, along with who in the company holds a role that allows it.

Relativity’s MCP server now takes write operations from Microsoft Copilot Chat. A user can create clients, matters and workspaces, run usage reports, and add or remove groups and users from a workspace. It is listed for Relativity aiR commercial environments only.

Adding a user to a workspace decides who sees the review data. That is an access decision that happens to be typed into Copilot, and admins should pick who gets to type it.

Agiloft added three agents to its Ask AI layer. One edits up to three fields across many contract records at once, one routes plain language requests to the right intake form, and one prepares Adobe Acrobat Sign workflows. It also opened an MCP server in beta, with Claude as the first client. Recorded Partially Verified.

A bulk edit is the fastest way to fix a thousand records and also the fastest way to break them. It wants a named owner before it is switched on.

LawVu customers can now connect Harvey and Legora to its MCP Server, alongside Claude, Copilot and ChatGPT. The server exposes more than 50 actions for reading and updating matters, contracts, spend and knowledge, and each user’s LawVu permissions carry over.

Wordsmith launched a Connector Library that reaches iManage Work, SharePoint, Box, Ironclad, SAP Ariba, DocuSign, ServiceNow, Salesforce and Workday. Depending on the connector, its agents save redlines back as new versions, reply on ServiceNow tickets, start Ironclad workflows and update SharePoint contract registers.

The smaller launches showed where the brakes go. Lawmatics’ Merlin Copilot can now create tasks and log calls and emails to a matter. It shows the planned action and waits for approval before saving, and it says so when no notes match a question instead of guessing. Vulcan Technologies’ Justinian can search a mailbox and draft replies, and any send needs the user to press Approve. Both are recorded Partially Verified.

Verbit’s new Claude connector lets attorneys question a transcript while the deposition is still going, and anything it schedules goes through a confirmation form. Avvoka’s ChatGPT plugin can be set to read only or read and write, and signed or locked documents cannot be changed either way.

Our read

The earlier question was whose permissions a connector uses. The new one is what it can change, and who clicks yes. The careful launches put a confirmation in front of the things that cannot be taken back: a termination, a sent email, a booked court reporter. The actions to watch are the ones that still look like admin work, such as a bulk edit or a new user in a workspace. Nobody ever had to explain to a client that the assistant read a contract too carefully.

Agents that start without being asked

The same October issue covered the first agents that begin work on a trigger. This period there were more of them, and they came with job titles.

Claren launched Envoys, always on agents for in house legal teams. They pick up requests arriving in Slack, Teams, email or the Claren app and work through them unprompted. Each has a job, such as contract review, intake, research, privacy or diligence.

Envoys work from the team’s playbook and past decisions, and escalate high risk items to the right person. Recorded Partially Verified, from trade press.

Leah launched Harness Builder. A business user describes a goal in plain language, and Leah’s Maestro orchestrator designs and builds the agents, tools, rules and governance to run it. Its legal operations example reviews outside counsel invoices against billing guidelines, watches budgets and routes exceptions for approval.

Genie AI launched Legal Specialist Agents. A lead agent plans each task and picks specialists for drafting, negotiation, risk analysis, research and proofreading, along with the quality gates each step must pass. A second agent reviews the work, and exceptions go to a named person.

A planner, specialists, a reviewer and a named human for the hard calls is roughly how a firm supervises junior lawyers. The agents have arrived at the org chart on their own.

Gatekeeper released an Agent Library where admins assign ready made agents to a workflow phase. Each one acts on every card entering that phase, and its findings quote the evidence and point to the page or clause.

Everlaw released Workflows, a drag and drop canvas for multistep document review, with gates that hold documents until a task is done or approved. The same release logs what connected AI tools did in a project.

DISCO made Cecilia Advanced Research, its multistep research mode, the default for new questions. Answers get longer and take longer by default, and reviewers should hear that before they notice it. MarqVision’s new operation agent for brand protection shows which brand data and past cases it relied on, and records a reviewer’s corrections so later reviews use them.

Our read

When agents start on their own, the lawyer’s job moves from doing the work to reviewing what gets escalated. That makes the escalation rule the most important setting in the product. An agent that only brings you the hard ones is a good colleague, as long as you both agree on what counts as hard.

Vendors put something behind the output

Inhouse added a guarantee to its terms for paid plans. If a contract, filing, formation document or demand letter its AI generated on or after November 1 has a covered defect, its affiliated law firm fixes it at no charge.

A citation to a law or case that does not exist, or that is described inaccurately, is named as a covered defect. The fix is capped at three hours of attorney time per membership year. Recorded Partially Verified.

Three hours is a backstop, not insurance. It is still rare for a vendor to put a lawyer’s time behind what its AI writes, and rarer to name the fabricated citation in the terms.

iManage published a test most buyers have not thought to ask for. It had 76 synthetic user profiles, varying name, gender and race signals and type of organization, ask a new Ask iManage feature identical questions about the same agreements. On the hardest questions, some profiles got less complete answers.

iManage limited what the user profile is used for, reran the test and found no meaningful gaps. It says the unfixed version never shipped, and that the test now blocks any release that fails it.

Answers that change with who is asking are a bias risk almost nobody in legal AI tests for. Finding the problem, fixing it and publishing all of it is the part worth copying.

Four vendors published numbers. Midpage ran Claude Opus 5.5 on 100 legal research questions, with and without its tools. A judge model preferred the Midpage answer 89 times, and the agent using Midpage had actually read 82% of the cases it relied on, against 33% with web search alone. Recorded Partially Verified.

The second figure is the useful one. A case the agent never opened is where a bad citation starts.

August published results from the seven benchmark suite that gates every model change in its product. One test asks questions whose answers are missing from the documents. GPT-6.1 Terra answered none of them, and Kimi K3 answered 15 to 21%. Recorded Partially Verified.

How often a model answers a question it cannot answer is a number few vendors print, and it is the one behind most fabricated citations.

StrongSuit’s update drafts motions, oppositions and memos from the full matter record, and its new Draft Studio has a different frontier model audit the draft for authorities, citations and record support. Its own benchmark scored its drafts at 96 out of 100, against 78 or lower for the frontier models it tested.

CUBE reported that its regulatory agent found three to four times more essential regulation than general AI models, on an answer key CUBE wrote and graded itself. The CUBE entry is recorded Partially Verified.

Three of the four are vendors marking their own work, or asking a model to mark it. That is worth knowing, and a published method still beats no method.

The product checks kept arriving too. Prudencia.ai’s Nova now checks cited statutory articles against their current text, and stops reusing a case reference once it has been corrected. LegalOn released Flags, which reads a whole contract for deal specific risks a playbook would miss, and tells users to check any law or case it cites. Both are recorded Partially Verified.

Smokeball’s Archie now tells the user when it cannot apply a requested format, and no longer drops or renumbers clauses it was not asked to touch. It also added GPT-5.6 Terra, Claude Opus 5.5 and Claude Sonnet 5.5 to its model list.

Our read

The last issue ended on the number nobody would publish, which was how often the check finds something. This period a few vendors got closer. Midpage counted how many cited cases its agent actually read, August counted how often a model bluffs, and Inhouse wrote down what it will do when the output is wrong. None of that is an accuracy rate yet. It is a lot closer to one than a disclaimer.

Where the data stays

HAQQ rebuilt its platform so that each region runs as a separate estate, with its own database, document storage, search index and logs. A central identity service holds no client documents, matters or chats. HAQQ commits that a workspace’s records stay in their home region and that no product action moves them. Recorded Partially Verified.

The logs are the detail to notice. Logs often carry client text, and they are where residency promises tend to leak.

PatSnap moved the OpenAI, Claude and Gemini deployments behind its EU environments into EU regions. That removes a transfer question for invention disclosures, which are about as sensitive as documents get.

Legistify published an AI usage framework approved by its CISO. It bars AI from making a final legal decision or sending an external message without human sign off, bars client data from training, and names the approved models and where they run: Claude models on AWS Bedrock in Mumbai. Recorded Partially Verified.

The last issue noted that storing data in a country and processing it there are two different sentences. HAQQ and PatSnap spent this period turning them into one.

Government buyers had a good two weeks. Thomson Reuters announced FedRAMP certification covering Westlaw, Practical Law and CLEAR. Ironclad’s Word add in now runs in FedRAMP environments on Microsoft GCC High. Aderant completed a SOC 2 Type 2 examination across five products, though the BillBlast review period covers only three months.

Patlytics shipped a NetDocuments integration that stores no credentials and uses no service account, so it can only reach what each attorney can already open. IPRally’s MCP server now gives assistants read access to a user’s saved work, and says search text and seed patents sent through it are not stored.

Ethyca hardened single sign on so an account can no longer be taken over through a matching email address. Ivo now sets edit access playbook by playbook.

Noxtua added a data export for chats and projects, for organizations on encrypted cloud storage only. Theta Lake now routes AI incidents detected by CrowdStrike to legal and compliance reviewers, with an audit trail. Both are recorded Partially Verified.

The connectors kept coming

The assistants collected more legal tools. Filevine shipped a LOIS Connector that brings a firm’s case files into Claude, with answers that cite their sources. Lexbe released Connect for Claude, which searches and drafts against eDiscovery cases without an export and shows the search terms it used.

Paxton AI’s research now works inside Claude, ChatGPT, Outlook and Word. It is read only inside the assistants, with every authority linked to its source and checked against Paxton’s citator. Recorded Partially Verified.

Litera put Compare and Firm AI Search inside Microsoft Copilot. The redline comes from Litera’s own comparison engine, not from a model’s guess at what changed, and the experience search respects ethical walls.

Legal tools also kept connecting to each other. Saga linked to Emma Legal, so diligence findings arrive in the draft with their evidence attached. DeepIP added iManage, with work saved back to the right matter. The DeepIP entry is recorded Partially Verified.

Jimini AI opened more than ten connectors, including Secib and iManage, and its agent now reads mail, calendars and time entries. Legora made its Google Workspace integration generally available, with admins choosing which of Drive, Gmail and Calendar to allow.

GC AI now syncs completed Docusign envelopes, read only, so signed agreements become searchable without a migration. Irys linked its Word add in to the matter file, and a new instruction no longer accepts pending changes without asking. CARET Legal added Outlook to its automated time capture, with staff reviewing entries before anything is billed.

Market notes

Contract review had a busy two weeks. Workday CLM added AI Playbook Review to its Word connector, and kept it separate from a second feature that suggests edits from general legal knowledge. A reviewer can tell company policy from a suggestion, which is a distinction worth keeping.

netLex launched redlining in its Review Agent, with more than 40 ready made playbooks for Brazilian legal teams. Recorded Partially Verified. Aline’s AI Chat now reads the open comment threads in a contract and makes the requested edits as tracked changes. Tomorro lets admins set up a new contract type by chatting with its assistant, which first checks that the type does not already exist.

Eudia launched Tabular Analysis, which pulls chosen terms from tens of thousands of files into a table and links each cell to its place in the source. Hebbia launched knowledge graphs that record which amendment changed which term. Pactly’s assistant can answer questions about a finished extraction grid and propose new columns without overwriting checked answers.

Contractbook now groups a contract with its amendments and renewals as one family. Anyone who has quoted the original agreement three amendments later will see the point.

In litigation, Doctrine added full document drafting. Flow Litigate drafts French pleadings from the case exhibits and the opposing filing in about 3 minutes, answering each argument raised. Neos made Demands Plus generally available. It builds medical chronologies and demand letters from records already in the case file, and an attorney approves every demand.

Briefpoint gave each attorney their own Discovery Playbook without changing the firm’s shared one. Phaselaw now transcribes audio and applies a redaction to the recording and the transcript together. Case Status added Call Intelligence, with a summary, a transcript and a sentiment flag on every recorded call.

Coverage widened. Harvey added a Peru knowledge source from Tirant lo Blanch, its fourth Latin American country since late October. Orbital signed a content license with the American Land Title Association, which says it is the first time it has licensed its content to an AI platform.

PatentSight+ added about 1.8 million Indian patent documents, without claims or descriptions. XLSCOUT can now limit infringement searches to one country and read product pages in the local language. Simpliciter added Italian and EU transport sources, and Aranzadi Allegra arrived inside a legal service for Spanish town halls.

On the consumer side, LegalZoom released a redesigned business account and a free mobile app with its AI Assistant built in, routing owners to an attorney when a question needs one. Jusbrasil rewrote the terms for Resolve, its assistant for people without a lawyer. Resolve can now decline work reserved to lawyers, refer a case to public defender offices, and point users to emergency channels when it detects risk.

For privacy and governance teams, OneTrust added a framework for Colorado’s automated decision rules and one for Canada’s OSFI E-23 guideline. Relyance AI now shows only the issues its AI review is highly confident about by default, so the low confidence filter deserves a regular look.

DataGrail’s Vera can now fill data map fields from attached contracts, with nothing changing until a user accepts. JUPUS can learn a lawyer’s writing style from their sent emails, which is a brave thing to hand over. Both are recorded Partially Verified.

Philips SpeechLive launched dictation that types at the cursor, with recognition tuned for citations and case names. Brightflag added a page for comparing outside counsel side by side, and Athennian added custom ownership views that leave the legal record untouched.

What the two weeks say about the category

Seventy six changes, and the theme is consequences. An agent that can terminate a contract, add a user to a workspace or edit a thousand records has moved past the stage where the worst outcome was a bad summary.

The vendors mostly responded the right way. They put a confirmation in front of the actions that cannot be undone, an escalation rule in front of the agents that start alone, and in a few cases a published test behind the claims.

What is still thin is the answer to a plain question. When the agent gets it wrong, who pays? Inhouse answered with three hours of a lawyer’s time. The count below shows what everyone else has put in writing.

Index Answer

Who pays when legal AI gets it wrong?

Usually the customer, and most vendors do not spell that out in public. The AI Legal Index grades every vendor it covers on AI Liability and Recourse, which asks what a vendor stands behind when its output is wrong, and whether a buyer can read that before signing.

Of the 534 vendors graded, 39 publish the specifics: what is indemnified, where liability is capped, what sits outside the cap, and any warranty or insurance a customer can actually call on.

Another 151 publish a real position that stops short of that, often an indemnity with no scope or cap attached. A further 219 rely on a standard limitation clause, and the remaining 125 publish nothing on who bears the loss.

Even the full versions tend to stop at the same place. The usual contract promises that the software will work as documented, and says the output comes as is.

That is why the Inhouse guarantee stood out this period. It names a fabricated citation as a covered defect and puts attorney time behind fixing one. These records describe each vendor as last researched, so a change from this period may sit ahead of its record.

A liability cap tells a firm how much it could recover. A guarantee on the output tells it how much the vendor trusts its own product. Both are worth reading before the demo.

Each vendor’s grade on this axis is on its profile in the vendor directory, and the grading method behind it is on the methodology page.

The AI Legal Index Brief is published by AI Legal Index, an independent reference for evaluating the AI software used in legal work. No vendor pays for inclusion, placement, or rating. Grades and signals for every vendor named here are on the vendor directory, and the grading method is on the methodology page.

Contact

Correct a record, or ask how something was graded

Every grade and every signal on this index is drawn from public sources and dated. If a record is wrong, out of date, or missing an artifact the index did not locate, send the source and it will be reviewed and the record redated. Vendors are welcome to submit documentation. Nothing on this index is for sale, including a listing, a placement, or a grade.

AI Legal Index

The AI Legal Index is an independent index that tracks changes to AI vendors in legal. It holds 535 vendors across 9 categories, each graded on the same 15 capability axes and recorded against 12 legal signals, from public sources. No vendor pays for inclusion, placement, or rating.

Index Status
Last index update
November 19, 2026
The AI Legal Index is an editorial reference. It is not a regulatory body, not a law firm, and nothing published here is legal advice or a recommendation to retain or avoid a vendor. Records are verified against published sources, bar guidance and public court records. Where a record reads not addressed, the material was not located in public sources on the date shown. See the Methodology page for evaluation standards and limitations.
© 2026 AI Legal Index
3801 N Capital of Texas Hwy, Ste E240 · Austin, TX 78746