The AI Legal Index Brief

The week in one line

LegalOn and CaseBlink shipped agents that start work on their own when a matter changes or a deadline closes in, and Legora let in house teams name the requests, like a standard NDA, that its agent may finish without a lawyer. Microsoft announced Copilot for Legal, and court data, dockets and precedent moved in the same day. CaseMark put confidentiality on its price list, making its plan page the rare one that doubles as a privilege memo.

This issue covers October 10 to 19, with 67 changes across 67 vendors. Nobody shipped twice, which in this market counts as restraint. Eight of the changes are dated late September and early October and missed the last issue. Sixty were fully verified and seven partially, twenty were rated high impact, and eighteen landed on October 13 alone.

The agent stopped waiting to be asked

For most of this year a legal AI agent did nothing until somebody typed. This period several started on their own.

LegalOn opened a beta of Agent Builder in Japan. An agent runs a chain of tasks by itself when conditions the customer sets are met, and the first triggers are matter events: a new matter, a change of status, a new assignee. The early jobs are handling legal consultations and assigning matters, with automatic first pass contract review planned.

Admins decide what the agent may draw on, including whether it can search the web, and can block actions outright, such as emailing the person who made the request. The announcement comes from the Japan site and does not mention the United States.

In immigration practice, CaseBlink keyed the same idea to the calendar. A firm skill can now start on its own when an RFE or NOID response is coming due, a document is about to expire, or a priority date comes current on a chosen Visa Bulletin chart.

Each automation is written as a sentence. Before it is saved, a preview shows which cases it would pick up or skip today, and why. Upcoming runs can be skipped, each case header shows the automated runs waiting on the attorney, and every run records why it started and how it ended.

The preview is the clever part. A trigger that shows its work before it has done any is a trigger a supervising attorney can actually sign off on.

For employment law, SixFifty moved from answering questions to acting on them. Its agentic HR compliance specialist takes instructions in chat, drafts handbooks and policies and sends them for signature, distributes labor posters and new hire notices by location, assigns required harassment training, and files state government documents when a company hires somewhere new.

Filing with a state is a different kind of output from drafting a memo. SixFifty says each action follows strict workflows grounded in the employment law database its compliance team maintains, and buyers will still want to know which steps wait for HR or counsel before they run.

BRYTER’s BEAMON 2.0 added an agentic mode that plans a multistep task, pulls from several sources and works it through to the end, such as a long pleading drafted to a detailed specification. Eudia took the gentler route with Horizon Scanning, a monitor that runs daily, weekly or monthly against sources a team names and emails a report with every update dated, summarized and linked.

Our read

An agent that waits for a prompt borrows its judgment from the person typing. An agent that starts on a trigger needs that judgment written down in advance: which events count, which actions are allowed, which are blocked. The vendors doing this well publish the rule, the preview and the log. That is the difference between automation a firm can supervise and automation a firm finds out about.

Who decides when the lawyer is skipped

Starting on its own was half the story. The other half was a set of settings that decide when an agent may finish without a lawyer.

Legora launched Intake in Slack. An employee mentions Legora in a channel, the Agent asks for any missing details in the thread, and the request reaches the right legal team as a tracked task with an owner and a due date. The Agent then prepares the work from the team’s own Skills, playbooks and precedents.

A lawyer reviews every output by default. The exception is the request types a team approves in advance, such as a standard NDA, which the Agent can complete and return to the business on its own. It runs only in Slack for now, with Microsoft Teams, email and MCP planned.

The last issue covered Legora’s Skills, the files that teach its agent a team’s house rules. Ten days later those rules have a job: deciding which requests go back to the business without a lawyer reading them first.

Thirdfort brought the same choice to money laundering checks. Its AI Screening Agent reviews sanctions, PEP and adverse media hits in about 35 seconds each, comparing the client’s verified identity with the hit, applying the firm’s own screening policy and checking public sources such as Companies House.

Firms pick a mode. One sends every recommendation to a reviewer. The other lets the agent dismiss hits it is confident are not the client, with scheduled random sampling and an audit log of those reviews. Thirdfort reports early adopters agreed with its dismissals in 99.9% of nearly 2,000 live hits.

The money laundering reporting officer’s question used to be whether a hit was a match. Now it is which mode to run and how often to check the agent’s homework. The 99.9% comes from Thirdfort’s early adopters, and the sampling schedule is how a firm finds out whether it holds on its own client base.

Centari put the threshold on confidence. Agentic Review, in preview, learns from how a team approves extracted deal data and verifies high confidence answers on its own, so reviewers only see the ones that need a judgment call. Administrators control where it runs, and every automatic verification stays in the audit log.

Harvey went the other way. Legal Work Orchestration lays out a matter as phases, tasks, assignees and review gates inside a Space. Agents run defined tasks with their scope, inputs and reviewer set in advance, and a named lawyer approves each output before the next step starts.

Status can be updated by replying to an email, the one interface every partner has already mastered. Tasks and notifications are rolling out now, and reusable matter templates are in early access.

In mass torts, Pattern Data built the same kind of record into a new version now live in select litigations. Each case opens with every litigation criterion showing what the record establishes, what is missing and the step that closes the gap, graded critical, major or minor.

Reviewers accept, reject or adjust each suggested value with its reasoning and source document. The change history then shows whether a value came from Pattern or from the firm, and who changed it when.

Coheso applied the approval step to precedent itself. Memory reviews matters as they close, proposes positions worth repeating with the reasoning, conditions and exceptions behind them, and shows that precedent to the attorney on the next comparable request. Coheso states that nothing Memory proposes takes effect until an attorney approves it. Recorded Partially Verified.

Smaller versions of the same dial turned up in contract operations. SpotDraft’s new metadata roll up, which copies terms such as expiry date from an executed amendment to its master agreement, can run automatically or wait for a user to confirm. Lexzur now fills a new matter form from uploaded documents and highlights every field its assistant filled for review before the matter is saved.

Our read

Last issue put the lawyer at the start, approving the plan. This period the question moved to which work needs the lawyer at all, and the answers arrived as settings rather than slogans: a list of request types, a confidence level, a sampling schedule, a named reviewer at each gate. A setting can be audited and a slogan cannot. It also moves the hardest judgment in the product from the vendor’s roadmap to the customer’s admin panel, and someone at the firm now owns it.

Copilot opened a legal shelf

A month ago fourteen vendors on this index plugged into ChatGPT in two days. This time it was Microsoft’s turn.

Microsoft announced Copilot for Legal on October 15, with partner plugins beginning to roll out through its plugin registry. DeepJudge is one of three specialized legal plugins listed, bringing a firm’s prior work and precedent into Copilot, and it says existing permissions and ethical walls carry over. Recorded Partially Verified.

Trellis brought its state trial court data into Microsoft 365 Copilot as part of the same initiative, through the Microsoft Marketplace. Attorneys can find successful motions, research judges and opposing counsel, and draw on trial court filings while drafting in Word. It needs a free Trellis account and an eligible Copilot license.

Bloomberg Law announced an MCP connector the same day that brings its federal and state dockets and litigation intelligence into Copilot. Users can search dockets, pull filing details and monitor matters with alerts without opening Bloomberg Law. The announcement does not say which subscriptions include it.

The other assistants kept filling up too. Rocket Lawyer listed a connector in Claude’s directory that hands a legal question to one of its attorneys, books a 20 minute consultation and takes documents through a secure upload link, with the attorney rather than the model giving the answer.

IP Author arrived in ChatGPT with prior art search across 183 million patent publications and research papers, US freedom to operate checks and examiner statistics. Its own note carries the sentence every patent team should read twice. Content reached through the plugin falls under IP Author’s terms, but the conversation itself runs in ChatGPT, so firms should check those terms before discussing an unpublished invention there.

Legal AI also started plugging into legal AI. Midpage released a connector, listed in Harvey’s Connector Library, that grounds Harvey threads in its US case law, statutes and regulations for firms holding both subscriptions. Hebbia brought in every Octus product a firm subscribes to through the Octus MCP Connector. The supply chain behind a single answer now has more layers than most corporate structure charts.

Our read

The ChatGPT wave raised the question of whose permissions a plugin uses. Copilot sharpens it, because Copilot already lives inside the firm’s Microsoft tenant with administration and access rules of its own. A legal plugin there is joining a house that already has locks. The vendors that explain how their rules and Microsoft’s meet will have the shorter procurement conversation.

The connections started writing

Last issue’s connections were mostly read only, and the Brief called that the right default for a minute book. This period the doors opened the other way.

Conga CLM released a CLM AI Library that opens agreement lifecycle actions to Agentforce and Claude over MCP. From a prompt, users can create agreements, generate documents, import offline paper, send agreements for review or signature and activate them. Administrators install the package and choose which capabilities to switch on.

Malbek made its MCP Server generally available, so any assistant that speaks the protocol can discover, draft and negotiate contracts while Malbek’s lifecycle rules stay in force. Every action an agent takes through it is recorded on the contract’s timeline.

Avvoka opened early access to headless templating, where an assistant connected to Avvoka builds the template rather than only drafting from it: the questions, the conditions, the repeating sections and the calculated values. Claude and Copilot are named first, with Harvey and Legora to follow, though early access so far lists customers already on the Claude integration. Nothing reaches end users until the team reviews and publishes the draft.

Template building has long been the slow, expensive part of document automation, and the part knowledge teams quietly dread. Handing it to an assistant changes who gets a template at all.

August gave its assistant the Outlook calendar. It can create, move and delete events and accept or decline invitations, and every change waits behind a card showing the subject, time, location and attendees until someone approves it. Recorded Partially Verified.

Smith.ai’s receptionists now book the consultation onto a firm’s Lawmatics calendar during the call and log the lead with the call summary attached. Perspective AI let people answer its intake conversations inside Claude or ChatGPT, which changes where a prospective client’s answers pass through.

Our read

Reading is a privacy question. Writing is a malpractice question, and the vendors have noticed. Every write path this period shipped with a brake: capabilities an admin chooses, lifecycle rules that still apply, a card that waits for a click, a draft that waits to be published. The brakes are in sensible places. The question for a firm is who holds them, and whether that person knows.

Confidentiality got a price tag

CaseMark relaunched its Operator agent as Para and split confidentiality by plan. The $20 per seat Standard plan runs on a low cost model whose provider may train on what it is sent, and CaseMark tells firms to choose Pro for confidential client work.

The $100 per seat Pro plan offers a choice of models, frontier ones included, all under zero data retention, with a BAA on annual Pro. Enterprise adds single sign on and the option to run Para through the firm’s own model provider.

Most vendors leave the training question to a data processing addendum. CaseMark put it on the plan page with a recommendation attached, which is more candor than most procurement teams get without asking twice. The risk now sits with the Standard seat bought on a corporate card by someone who never read that page.

In Germany the same question turned into paperwork. AnwaltGPT now publishes its confidentiality agreement under § 43e BRAO as a download, so a firm fills it in, signs it and emails it back for countersignature instead of asking for it. German lawyers must bind any IT provider to secrecy in text form, and the agreement covers the operator, its staff and every named subprocessor.

HAQQ closed a quieter gap. Removing a member from a workspace now removes their access to its AI chats, which can hold privileged content, and each workspace shows its data region on its settings page. Steno opened a public beta of Litigation Genius, an AI workspace over whole case files with every answer cited to file and page, and states that its model providers work under zero data retention agreements.

Price moved in a second way. Briefpoint put Find Answers, which drafts interrogatory answers from case files, on credits, where one credit covers one document checked against one interrogatory. Each attorney seat adds 20,000 credits a month to a pool the firm shares, and use beyond that is billed at 1 cent a credit from October 20, with a usage report by user, case and document.

A flat seat price grew with the size of the team. A credit grows with the size of the record, which is a very different number on a large docket.

Our read

For two years the confidentiality answer was a sentence in a trust center. This period it became a choice at checkout, a form a German firm can download and sign, and a chat history that leaves with the lawyer who leaves the workspace. Each turns a policy into a setting someone has to pick. That is better for the buyers who read the plan page, and riskier for the ones who do not.

The software started admitting what it doesn’t know

The failure lawyers fear most is the confident wrong answer. Several releases this period traded a little confidence for a lot of candor.

Aline’s chat now answers questions across a whole contract repository, such as which governing law most NDAs use, with a counted answer linked to the contracts behind it. When some contracts have not been analyzed yet, it says so and fills the gap.

TrustArc’s AI Control Review now names each evidence item it read, labels evidence it could not read as unreadable rather than implying none was needed, and warns when a document was skipped for its size.

CounselPro’s Daystrom asks one clarifying question when a request is ambiguous, adds up table totals in code with a note of what went in, and names the account and months missing a statement. Those are the behaviors that matter when a figure ends up in a support affidavit.

Saying “I could not read this one” takes most first year associates a while to learn. It is a good sign that software is learning it too.

Agiloft put confidence icons beside the clause in its document viewer, so a reviewer sees where the model is unsure at the moment of reading. Recorded Partially Verified. Libra now builds Excel workbooks from a chat request and checks every formula before handing one over, while leaving errors in a user’s own file exactly where they were. Smokeball’s Archie highlights the placeholders in its drafts and asks for the missing facts.

In insurance defense, Wisedocs opened early access to Legal Intelligence, which reads deposition transcripts and demand letters against the medical record and flags contradictions, missing records and key admissions, each cited to both the testimony and the record. Phaselaw now surfaces badly made redactions already sitting in uploaded documents, so a reviewer decides on each one before export.

Research tools made the same move on sources. Lucio opened a free public beta of Lucio Research across US, UK and Indian case law, with an All sources view that labels every source it considered as highly relevant, relevant, flagged or discarded. Jus Mundi released Jus AI 3, which checks what it has found and runs follow up research on the gaps before it answers, and reports a faster first answer and more answers faithful to their sources on an internal set of 114 prompts.

Definely took the strictest line. AI Search in its Vault finds the clause a lawyer means even when the wording differs, and Definely states that every result is an existing firm document. Vault never generates answers or clause language.

Our read

A tool that shows what it discarded, what it could not read and where the record has gaps is handing the lawyer the material for the check rather than asking for trust. That is less impressive in a demo and much more useful on a Thursday night before a filing. The confident answer with no seams is starting to look like the riskier product.

Market notes

Some research products grew into whole workspaces. LexisNexis launched the Intelligize+ AI Disclosure Management Suite, which moves the product from SEC research into reviewing a company’s own 10-K and 10-Q filings against SEC rules, GAAP guidance and peer filings, with a Regulation S-K and S-X checklist that carries sign off. LexisNexis states the findings are preliminary and the reporting professional stays responsible for every call.

Lefebvre made GenIA-L the entry point to all of its services in Spain, folding in the LexON practice management software, Client Link and Centinela, so matters, calendar and billing sit beside research and drafting and the assistant decides which module answers each request. eBrevia launched Command, a contract management platform built around an AI repository, which takes it from analysis into obligations, renewals and requests.

Others widened their practice coverage. NetDocuments added six prebuilt apps to ndMAX Studio, taking the catalog to 45 and reaching into family law and estate planning, including a prenuptial agreement enforceability analyzer. Turbo Law extended its litigation workspace to all civil litigation and connected it to NetDocuments, iManage, Clio and the usual cloud drives.

DecoverAI moved from litigation review into transactional diligence, from $299 per seat per month with AI usage billed by the token, and Casepoint launched a public records product for state, local and education bodies.

The last issue’s Skills went international. Jimini launched Competences, reusable task methods its agent applies on its own, with more than 200 written by its lawyers for French law across about 50 specialties and data hosted in France.

Governance tools kept building the record of what agents did. Trustible’s Agent Scanner finds agents built on platforms such as Amazon Bedrock AgentCore, Microsoft Copilot Studio and Claude Enterprise, adds them to the AI inventory and flags any that changed since their last review. Ethyca now records why each access decision was allowed or denied, and OneTrust’s assessment APIs now show, answer by answer, whether a person reviewed the AI’s work.

Laurel published results from Berger Singerman, where attorneys recovered nearly 15 minutes of billable time a day and finished time entries about two days sooner. Fifteen minutes is two and a half tenths of an hour, a conversion every lawyer reading this made before the end of the sentence. Recorded Partially Verified.

In the plumbing, Relativity retired its legacy Transcripts application in favor of a new transcripts experience, Reveal began preserving Zoom content in place for legal holds, Litera added Okta sign in for new tenants, and Nexl completed a SOC 2 Type II audit. Cypris added Google’s Gemini 4 Argon to its model picker and named the provider, which still counts as news in this market.

What the ten days say about the category

Sixty seven changes, and the theme is the threshold. Agents started without a prompt. Assistants gained the right to write. Confidentiality became a plan choice, and the better tools started saying what they had not checked.

Each of those moves puts a dial in front of the buyer: when the agent starts, what it may finish alone, which model sees the file, how sure it has to be. A dial is only useful if the vendor says where it is set and what happens past the line.

That is the question this index has graded from the start, and a published answer is less common than this period’s launches suggest. The count is below.

Index Answer

Which legal AI vendors publish where the agent stops and the lawyer takes over?

Not many publish the whole answer. The AI Legal Index grades every vendor it covers on Autonomy and Oversight Model, which asks what a system does on its own, what a lawyer has to approve, and whether the vendor documents where that review point sits.

Of the 463 vendors graded, 55 publish the full structure: the modes, the thresholds, the review surfaces and the route a matter takes back to a lawyer’s judgment.

The large middle is 302 vendors that commit in writing to a supervising lawyer and show real review surfaces, but stop short of the full set. What is usually missing is the point at which the system stops on its own, or what happens after it gets something wrong.

A further 96 assert human oversight without describing a mechanism, so human in the loop appears as a phrase rather than a control. The remaining 10 publish no oversight structure at all for a system that drafts, advises or acts on client work.

This period’s launches show what that missing point looks like once a vendor writes it down. Legora lists the request types its agent may finish alone, Centari sets a confidence level above which answers verify themselves, and Thirdfort samples the hits its agent dismissed. Each of those is a threshold, and publishing it tells a firm where the agent stops. These records describe each product as last researched, so a launch from this period may sit ahead of its record.

A vendor that says a lawyer stays in the loop has described a hope. A vendor that says where the loop closes, and what happens once the agent is past it, has described a control.

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.