The AI Legal Index Brief

The week in one line

Legora shipped Skills, plain text files that teach its agent how a team likes things done, which means a senior associate’s red pen can now be saved as a file. Harvey and LexisNexis built a motion to dismiss workflow where the lawyer approves the plan before any research runs, and every case in the memo has been through Shepard’s. Filevine added a Deal breaker flag, for the clauses where a Medium risk rating is itself the risk.

This issue covers two weeks, September 27 to October 9, with 25 changes across 20 vendors. Fourteen were verified directly at the source and eleven through other reporting. Nine were rated high impact, and thirteen of the 25 landed between October 5 and 8, as if someone had circled the week on a shared calendar.

The lawyer approves the plan first

Harvey released a Motion to Dismiss workflow built with LexisNexis, and the interesting part happens before any drafting. A planning agent reads the uploaded complaint and proposes up to eight grounds for dismissal, each rated by strength.

The lawyer approves or edits that plan. Only then do parallel agents research each ground in LexisNexis by jurisdiction, producing a memo of three to five cases validated through Shepard’s.

The draft motion arrives with its tables of contents and authorities, and it follows the forum, whether that is Rule 12(b)(6) in federal court, a California demurrer or a New York CPLR 3211 motion. That removes one of the oldest reasons to still be at the office at nine.

Access requires Ask LexisNexis or Lexis+ with Protégé alongside Harvey, and a summary judgment workflow is next.

The order of operations is the point. The lawyer decides which arguments go in before the agents spend any time on them, and the authority underneath is checked by a citator rather than remembered by a model.

Opus 2 applied the same shape to depositions. An agent reads a transcript against the issues in the case and flags the passages that matter most, and a lawyer accepts, rejects or adjusts each one. Approved designations land in the matter record linked to their issue, time stamped and attributed to the lawyer who approved them. Recorded Partially Verified.

That attribution will earn its keep the first time a designation is contested, because who made the call is the first thing opposing counsel asks.

Filevine closed a smaller gap that every litigator has hit. When its LOIS assistant cites a spreadsheet inside Word, a built in viewer now opens the file and highlights the rows the answer relied on. Damages figures and billing data can finally be checked the way a quoted paragraph can.

Our read

Last issue the lawyer accepted the agent’s work at the end. This time the lawyer approves the plan at the start, and the check on the authority happens before the draft exists. That is a better place for the human to stand. A rejected ground costs nothing, while a rejected brief costs a weekend.

The house rules became files

Legora released Skills to all customers. A Skill is a Markdown instruction file, which can bundle templates and examples, that teaches the Legora Agent a team’s standards, processes and preferences.

The Agent applies a relevant Skill on its own, or a user calls one with a slash. More than 250 ship in a library on day one. Admins can restrict sharing, a shared Skill has to be installed before it runs, and matter permissions still apply.

That install step is quietly important. It stops one enthusiastic lawyer’s preferences from becoming the whole firm’s house style overnight, which is a conversation every knowledge management team would rather have before it happens.

Ivo came at the same problem from the playbook side. Its Playbook Builder now keeps the positions in an uploaded playbook unchanged rather than rewriting them, and its Assistant gained a thinking effort setting with four levels, from Fast to Max. Ivo also launched Collaborate for managing contracts before signature, which moves it from review into negotiation. Recorded Partially Verified.

A playbook that comes back reworded is no longer the firm’s playbook. Keeping the positions intact sounds like the least a tool could do, and it is also the first thing a reviewer has to trust before anything else works.

Filevine added a Deal breaker flag to LOIS for Word. A team can mark any playbook check as a term where any change should stop the deal, and the flag sits beside the High, Medium and Low ratings LOIS assigns rather than replacing them.

It exists for the clauses where a Medium rating is itself the risk. Nobody gets to wave a change through on those because the AI sounded relaxed about it.

Ironclad turned its Assistant into Ironclad Agent on October 8, a single entry point that routes requests to agents for intake, review, archive, renewals and analytics. They draw on a Contract Knowledge Graph of clauses, obligations and negotiated decisions, so answers reflect what was actually agreed rather than only what the template says. A Playbook Agent is in early access.

Our read

For two years legal AI vendors sold how much their models knew about law. This period several of them shipped ways for a firm to tell the model what it knows about its own practice: its positions, its red lines, its house style. That knowledge has always lived in partners’ heads and in documents nobody updates. Turned into files a tool reads, it becomes easier to share and easier to audit. It also becomes the hardest thing to leave behind when a firm changes vendors.

Everyone opened a door, and most were read only

The connections kept arriving, and this time the fine print was more careful than the announcements were loud.

Athennian launched Athennian MCP, a read only connection that gives ChatGPT, Claude, Gemini and Copilot live entity data while applying each user’s existing roles and permissions. Admins choose which entities and documents the AI can reach, and can exclude individual documents. It shipped alongside Athennian Intelligence, which answers plain language questions over corporate records with links to the source.

Read only is the right default for a minute book. A corporate secretary wants the assistant to know who the directors are, not to appoint new ones.

Hebbia launched Headless Hebbia, an API and MCP server that Hebbia builds into a firm’s own applications with each team. Claude and ChatGPT can call it and get answers with links back to the evidence. Days later it added a Microsoft Word plugin, version history for decks and documents, and memory for its Max assistant that admins switch on for the whole organization and users can review and delete.

TrustArc took MCP in both directions. Its Arc Intelligence can now be used inside Claude, Microsoft Copilot and ChatGPT, and it can draw context from a customer’s Salesforce, Atlassian and other systems, with every connection approved before access is granted.

Legora connected to iManage Cloud. Users link an iManage matter to a Legora project, and the Agent can search it by client, matter, author, date or the firm’s own custom fields. Edits made in Word show as tracked changes and go back to iManage only when the user chooses. Each user connects their own iManage account, so the walls set in the document system keep holding.

GC AI added a native Salesforce connector, so in house teams can question deal and contract records alongside their own documents. Libra let a conversation started in its web app carry on in Word or Outlook, with an icon showing where it began. Recorded Partially Verified.

One detail from Ironclad belongs here too. Switching off Ironclad Agent in company settings also switches off AI Search and MCP access. A single switch that closes every door is the kind of control an incident response plan likes to find already installed.

Our read

Two issues ago the story was fourteen plugins arriving in two days. The connections are still coming, but the language around them has changed. Read only, approved per connection, the user’s own permissions, nothing saved without a click. Those phrases used to surface in a security review after the sale. Now they are in the launch post, which means vendors have worked out what the first question in the room is going to be.

Governance moved from the policy to the runtime

OneTrust introduced CORIE, a shared intelligence layer that maps AI systems, models, data, vendors and identities to their governance requirements, applies policies and prior decisions, and keeps the audit trail in what it calls an Evidence Ledger. The same release names an AI Control Plane, a runtime layer that allows an agent’s action, blocks it, or routes it to a person for approval.

The control plane is the part to watch. Governance has mostly lived in documents that describe what agents should do. A layer that can stop an agent mid action is governance that actually runs. Some of the release, including an MCP gateway and posture management for GDPR and the EU AI Act, is in private preview.

TrustArc’s Ask Arc can now carry out multistep privacy tasks from a single prompt, through commands a person approves, rather than only giving guidance. TrustArc also states that customer data is never used to train external models, which is the sentence procurement looks for before pointing anything at regulated data.

Nuix released a governed framework that connects a customer’s chosen model to investigation data, with audit telemetry, policy guardrails and token metering on every call. Its BYO AI workflows run across every item in a dataset rather than a sample and return cited reports, which answers the defensibility worry that sampling always raises. Recorded Partially Verified.

Noxtua added auto delete for chats and projects after a chosen period of inactivity, from 8 hours up to 180 days, and any activity resets the clock. It covers only items created after the setting is switched on, so a firm that wants it should switch it on before it needs it. Recorded Partially Verified.

The eight hour option suits sensitive one off work, the kind of matter nobody wants a tool to remember by tomorrow morning.

Our read

Every governance release in this period made the same move. It stopped describing what should happen and started deciding what does, with a person approving at the point that matters and a record kept of why. For legal and privacy teams that is a change in kind. A policy can be skipped on a busy afternoon. A control plane cannot, which is exactly why the people who write the policies will want a say in how it is set.

Market notes

PERSUIT connected with Wordsmith so in house teams can price the work they keep. When Wordsmith resolves a routine matter internally, PERSUIT estimates what outside counsel would have charged, from comparable proposals for the same matter type and jurisdiction.

The last issue counted how few tools help a firm bill for work the AI did. This one runs the other way, putting a dollar figure on work that never reached a firm at all. In house teams have always known roughly what they saved by not calling outside counsel. Now they have a number, which may make the next budget meeting shorter or considerably more interesting.

LawVu launched Lens, which analyzes whole contract portfolios at once and answers plain language questions with citations to the source contracts, such as which agreements need consent on a sale. Recorded Partially Verified. SmartAdvocate brought the AI features it announced in August into a named release, including demand letter and medical chronology drafting and an intake voice agent. Recorded Partially Verified.

Filevine’s LOIS Console can now assemble demand and exhibit packets, apply Bates numbers and edit PDF text from chat. Somewhere a paralegal’s least favorite afternoon just got shorter. PostSig expanded its post signature platform for finance and procurement teams, tying what a contract says to what is actually billed, with an MCP server exposing more than 30 tools and data hosted in the US or EU region the customer picks. Recorded Partially Verified.

Transcend solved a problem that has a sense of humor of its own. Corporate email security scanners open every link they receive, which meant a scanner could verify or cancel a data subject request on someone’s behalf. Verification links now need an explicit button click, a rare case of software being protected from software that is too diligent. Recorded Partially Verified. Sandstone made its DocuSign notifications name signers, recipients and signing order, and fixed several sync issues that could stall an intake run. Recorded Partially Verified.

What the two weeks say about the category

Twenty five changes, and the theme is instruction. Firms told the agent which arguments to pursue before it researched them. They wrote their positions, red lines and house style into files the agent reads. They told outside assistants exactly which doors were open, and governance tools started deciding in real time which agent actions need a person.

The most interesting work this period did not come from vendors claiming their models know the most. It came from vendors making it easy for a firm to say what it already knows, and moving the check on the work earlier: a plan approved before research, a spreadsheet row highlighted beside the answer, a case run through a citator before the memo is written.

That last check is the oldest one in legal research and still among the least common in legal AI. The count below shows how far the field has to go.

Index Answer

Which legal AI research tools tell you when a case is no longer good law?

Fewer than a buyer would hope. The AI Legal Index records, for every vendor it grades, whether the product checks the authority it cites for subsequent history. 83 of those vendors list legal research as a primary or secondary category.

Only 3 of them surface treatment from a named commercial citator next to the case. Another 15 compute subsequent history themselves and describe the method, which is a real check a firm can examine, even if it is not the one most litigators were trained on.

A further 19 tell the lawyer to verify the citation without performing the check. The remaining 46 say nothing about subsequent history at all. Some of those are docket, analytics or practice tools for which research is a side line, but many answer legal questions in prose.

So 18 of 83 research tools do the check, and 65 either hand it back to the lawyer or leave it unmentioned.

It matters because a citation can be real, correctly quoted, on point and overruled. That failure looks nothing like a hallucination, and it lands the same way in front of a judge.

This issue’s lead showed the check built into the workflow rather than left to the reader. Harvey’s motion to dismiss runs every case in its research memo through Shepard’s, for firms that also hold the Lexis subscription it requires. These records describe each product as last researched, so a workflow shipped this week may sit ahead of its record.

A tool that tells a lawyer to check is describing the lawyer’s job. A tool that checks is doing part of it.

Every vendor’s record on this question, alongside its published accuracy evidence, is on the accuracy disclosure page, and the signal itself is defined on the signals page.