The AI Legal Index Brief

The week in one line

Harvey named the model behind its product and published the retention and residency caveats with it, Wordsmith and Everlaw wired legal AI into the systems that hold the documents, and Avvoka started building templates from what a firm actually signed rather than what a partner remembered. The first issue of this log is mostly about where the work already lives.

This is the first issue of the Brief. It covers August 27 to September 5: 12 entries across nine vendors, ten Verified at source and two Partially Verified from trade press. Funding rounds, valuations, and awards are not logged, here or anywhere on this index.

One vendor published the uncomfortable half

Harvey added opt in access to Claude Fable 5.1, named the model and its version, and positioned it as strongest on complex document heavy analysis, with the largest gain over the previous version in litigation and dispute resolution.

It cited public leaderboard results rather than an internal benchmark, which is already unusual. Then it published the parts that do not help the sale.

The same post discloses that Fable models may involve data processing practices differing from existing customer agreements, that zero data retention is available to eligible customers only and under a time bound exemption, and that there is no regional processing for this model, so all data is handled in the United States.

For a firm with residency obligations that is not a footnote, it is the answer. A managing partner in London or Frankfurt now knows, before a procurement call, that this particular toggle sends work to a United States processor.

Harvey also rebuilt Playbook Review as a multi agent system, an architectural change with no user facing surface, and released Horizon Scanning in early access, which monitors regulatory and legislative change on chosen jurisdictions and topics and turns a surfaced change into a memo or policy update inside the platform.

Regulatory monitoring is one of the few legal workflows still routinely done by hand, usually by an associate reading newsletters on a Monday. Moving it into the tool that drafts the response is the part that matters, because the monitoring was never the expensive step.

Our read

Publishing an architecture change and a residency caveat in the same week is a posture, and it is the posture this index was built to make visible. Most vendors in legal AI will not tell a buyer which model is running, which means the buyer cannot reason about retention, residency or what happens when the model is swapped. A vendor that names it has not proved the product is better. It has made the product checkable, which is a different and more durable claim.

Buyer question

Ask which model answers each workflow, who provides it, in which region it runs, and what notice you get when it changes. Then ask for the retention position in writing, with the eligibility conditions and any expiry attached. A vendor that can only answer the first question has told you about a demo rather than about a service.

The AI moved to where the documents already are

Wordsmith released an iManage connector whose agentic workflows can find, open, draft, check and file documents back into the document management system without a person moving files between the two.

The file back is the load bearing verb. A legal AI product that cannot write into the document management system leaves the firm with a manual filing step and two versions of the truth, and naming the round trip is what separates a shipped connector from an announced relationship.

Everlaw launched a Model Context Protocol integration with Harvey, so evidence, project data and reporting held in Everlaw can be queried conversationally from inside Harvey rather than by moving between two systems.

That closes a specific and expensive gap between the layer a team drafts in and the platform that holds the underlying record. It also means two vendors have agreed that neither of them owns the whole workflow, which is a more useful admission than most partnership announcements contain.

Relativity shipped two connections of its own. RelativityOne now integrates with Microsoft Copilot over MCP for administrative reporting in natural language, and it can collect Claude Enterprise data directly from the source through Anthropic's Compliance API, mapping and normalising prompts, responses, chats and uploaded files into a reviewable form.

The second one is the entry to notice. Any organisation that has rolled out an enterprise assistant has created a new category of discoverable record, and until there is a defensible collection path, that record is either unreachable or reconstructed from exports by hand. This is the plumbing arriving slightly ahead of the first dispute that needs it, which is the right order and not the usual one.

Our read

Three of the four integration entries this week point the same way: the AI layer is conceding that the system of record stays where it is. That is good for buyers, because the alternative was a migration nobody wanted. It also means the integration is now the product boundary, and a connector that reads but cannot write is a demo rather than a workflow.

Buyer question

For any legal AI tool connected to a document management or eDiscovery system, ask whether it writes back, what metadata survives the round trip, and who the audit trail records as the author when an agent files a document. If the answer to the last question is the service account, the matter file no longer tells you who did the work.

The knowledge work turned into the product

Avvoka launched Curate, a standalone product that reads a firm's historical transaction documents, extracts the clause variations that performed best, and generates templates from them, feeding the output back into its automation and drafting workflows. Recorded Partially Verified from trade press.

Template maintenance is the unglamorous cost centre of knowledge management, and most template sets drift from what the firm actually negotiates. Building them from executed documents rather than from a partner's memory is the interesting claim, and it is testable against a firm's own back catalogue, which is the best kind of claim.

LegalOn published a set of workflow additions: Cross Document Review for analysing a whole contract family together, searchable negotiation history through its assistant, and an AI Revise upgrade that drafts and inserts language for missing provisions inside the Word add in.

Two of those solve the same problem from opposite ends, which is that negotiation context lives in people's memory and in old email threads. The smallest item is the one in house teams ask for constantly: a department level editor name on redlines in Word, so a document does not tell the counterparty which junior lawyer made the edit.

Wordsmith also shipped enterprise controls aimed at letting business units outside legal run legal workflows, adding the access and approval mechanisms that widen who can do the work without widening who can approve it.

That bottleneck is familiar to every in house team: routine requests queue behind the only people permitted to touch the tooling. Delegating it is an oversight problem before it is a capability problem, and this release is aimed at the oversight half, which is the harder half to sell and the one that decides whether the rollout survives.

Our read

The pattern under these three is that the asset being productised is the firm's own history: its executed documents, its negotiation record, its approval structure. That is the one thing a vendor cannot ship in the box, and a tool that reads it well becomes very hard to remove. Ask what happens to the extracted clause library and the negotiation history if the contract ends, because that is the switching cost being built here and nobody prices it at signature.

Market notes

CUBE added three agentic coworkers to RegPlatform: one scoring regulatory updates for relevance, one answering natural language questions against laws and rules, and one consolidating global enforcement actions. The agents run on Azure and are listed on the Microsoft marketplace, so the deployment path is documented rather than negotiated.

Relevance scoring is the load bearing feature. Regulatory monitoring platforms have never struggled to find updates, they have struggled to tell a compliance team which of several hundred this month actually apply, and a score that cannot be explained is one nobody can rely on the day they skip an update.

DataGrail added automatic high risk detection for data protection impact assessments, filling Records of Processing Activities forms by identifying AI use, biometric data and targeting of vulnerable individuals, alongside custom system naming in its live data map and a Request Manager Agent release.

Records of processing upkeep is the compliance task most teams are quietly behind on, because it is tedious rather than difficult. Automatic detection of the categories that trigger an assessment is worth more than the form filling, since missing a trigger is the failure that carries the penalty.

What the week says about the category

For an opening issue this is a remarkably consistent set. Nine vendors, and the through line is that legal AI has stopped trying to be the place where the work happens and started trying to be the layer that reaches into it.

The documents stay in the document management system. The evidence stays in the eDiscovery platform. The templates come from what the firm already signed. The regulatory feed lands in the tool that drafts the response. Even the model is somebody else's, which is exactly why naming it matters.

That leaves a clear question for buyers in this category, and it is not which product is cleverest. It is which vendor will tell you what sits underneath it, where your material goes when it is processed, and what you keep when you leave. This week one vendor answered the first of those in public. The index will keep count of who follows.

Index Answer

Which legal AI vendors tell you which model is running?

Very few, and the shortfall is not evenly spread. Of the 164 legal AI vendors the AI Legal Index has recorded on Model Supply Chain Disclosure, 5 answer the whole question: which models sit underneath the product, whose they are, where inference runs, and whether the vendor commits to saying so when any of that changes.

Another 53 name models or providers without completing the set. The remaining 106, 65 at C and 41 at D, either gesture at the underlying technology without identifying anything a buyer could check, or publish nothing at all about the supply chain a firm inherits when it signs.

A vendor that says it is powered by leading large language models has named nothing. A vendor that says which model, from which provider, running in which region, with notice when it changes, has told a firm what it is actually buying.

This week made the distinction concrete rather than theoretical. Harvey named Claude Fable 5.1, gave the version, made it opt in, cited public leaderboard results rather than an internal figure, and published the awkward parts in the same post: that zero data retention is available to eligible customers only, under a time bound exemption, and that there is no regional processing, so the data is handled in the United States.

None of that is favourable to the vendor and all of it is decision useful to a firm with residency obligations. That is the difference between a disclosure and a launch post, and it is why the axis grades what is published rather than what is said in a deal room.

The full grading, the axis definition and where each vendor sits are on the model disclosure page.