The week in one line
Advocacy launched an agent that proposes litigation work and charges only for the parts a lawyer accepts, which is roughly how clients have always wished associates were billed. August let a client’s software agents put legal questions to their law firm’s agents, and Relativity took away the option to switch AI off in breach response. The billable hour has had quieter weeks.
This issue covers September 19 to 26, with 58 changes across 44 vendors. Forty six were verified directly at the source and twelve through other reporting. Thirteen were rated high impact.
The agent proposes, the lawyer accepts
Advocacy launched Promptless, and the name is the pitch. It proposes litigation tasks on its own when a document production arrives, when opposing counsel sends a letter, or when a deadline gets close.
It classifies incoming documents and flags contradictions with the case theory. It runs research, drafts outlines, and calculates deadlines from docket entries while showing the governing rule and the arithmetic.
Every proposed action stays unconfirmed until an attorney accepts it. Teams decide, matter by matter and client by client, which work the AI may do, where approval is required and which sections stay lawyer only.
Then the pricing. A firm opens a matter for a nominal fee and pays only for the AI contributions it accepts into that matter. Rejected drafts are free, which may be the first time that sentence has been true anywhere in a law firm.
Legartis added agentic workflows that carry a contract from intake through negotiation, approvals and signature. Legal teams set the playbooks and decision boundaries, and anything beyond a boundary goes to the responsible person with proposed solutions.
Casepoint released its first two agents for document review, one scoring relevance with its reasoning and one coding documents against matter issues. Both require the team to validate performance on a sample before the agent runs across the full set, and the issue coding agent can report recall and precision per issue.
That validation step is the detail that will matter in court. A review agent that has to prove itself on a sample first is producing the evidence a party will one day be asked for.
The same instinct reached Word and Outlook. DocJuris launched a Word add in whose proposed edits are written only on click, as native tracked changes. StrongSuit’s new Outlook add in drafts replies from the matter’s documents and states that it never sends an email or changes its recipients. Recorded Partially Verified.
Filevine brought its assistant into Outlook too, with an Insert button that opens the draft for editing before anything goes out. And Conga CLM added supplier sanctions screening that never blocks a supplier on its own. An authorized user blocks with a recorded reason, and only then do the linked contracts pause. Recorded Partially Verified.
Our readThe accept button is becoming the unit of legal AI. It is where responsibility passes to a lawyer, where the audit trail starts, and at Advocacy it is now where the invoice starts too. Vendors that design around that click will find their risk committee meetings much shorter. Vendors that treat it as friction will keep having those meetings.
The client’s agent called the firm’s agent
August launched August for Agents in beta, and it describes a new kind of legal service. A client company’s software agents send a bounded legal question to a service that the client’s law firm builds and supervises on August. Recorded Partially Verified.
The firm’s agent checks the relevant agreement and playbook, then returns a sourced answer with an execution record. When judgment is needed, it routes the matter to a lawyer. Tools, escalation rules and retention are set per client, per firm and per service.
The billable hour has survived a great deal. It has not yet had to survive a client that is itself a piece of software.
August also became an MCP client the same week, connecting to other vendors’ hosted servers with each user’s own credentials, stored encrypted and not shared with the firm. It is switched on by request rather than self serve.
The connections kept running both ways. Irys made a two way Clio connector generally available. Documents come in from the Clio matter, and finished drafts, notes, tasks and time entries go back to it. Harvey extended its agentic search to SharePoint, OneDrive and Google Drive, so it can find a user’s documents without the user choosing them first. Recorded Partially Verified.
BigID launched AgentIQ, which lets privacy and security teams run data work from Claude, Copilot, GPT, Gemini or their own agents. The agents inherit the requesting user’s permissions, every action is attributable to a person, and revoking access or deleting data can be set to require approval. DataGrail extended IP allowlisting to agents and outside MCP clients, off by default until an admin switches it on.
Our readLast issue, the systems that hold legal work plugged into one assistant. This week the traffic started running between agents, including agents that belong to different organizations. That changes what a firm can sell. A playbook an agent can query at any hour is a product, and a firm that builds one has turned part of its practice into a service. The questions that follow are old ones in new clothes: who the client is, who is supervising, and what the record shows when the answer was wrong.
The walls got taller
The last issue asked whether a firm’s ethical walls survive once client material leaves the system built to hold them. Several vendors answered in product this week.
NetDocuments now lets a user share an AI Assistant thread with colleagues, who can read it and carry on the conversation. The documents travel with the thread, access controls stay in force, and the sharer is warned when a recipient cannot see a document in it.
That warning is the wall working at exactly the point where it is most likely to fail.
Smokeball reworked matter permissions in a beta release, so a matter can be Open or Restricted. A new option keeps staff who join the firm later out of a restricted matter by default. Recorded Partially Verified.
GC AI limited sharing of vault chats to full members of that vault, with no public links, and gave admins upload metrics that reveal no document or vault names. Athennian let admins restrict individual structure chart views to specific teams.
Noxtua published the data path behind each of its model options. The standard option runs open weight models on Deutsche Telekom and IONOS infrastructure, with no US provider in the chain. The new frontier option reaches a closed model through Deutsche Telekom into an EU data center, which Noxtua says reduces, rather than removes, exposure to the US CLOUD Act. Recorded Partially Verified.
A vendor saying out loud that one of its own options only reduces a legal exposure is rare. It is also exactly the sentence a firm with European clients needs to read.
Our readWalls used to be something a firm bought once, from one vendor, for its document system. They are now a property every legal AI product has to carry, down to the shared chat thread and the chart view. The vendors worth watching describe the failure case, the recipient who cannot see a document or the colleague who joins next year, rather than only the happy path.
The model became a menu item
Merlin Search Technologies published the most useful evaluation in the log this week. It had seven models from OpenAI, Anthropic and Google write cited reports from a public opioid litigation document set. It then checked 777 citations and 454 cited sentences one by one, with graders from other providers auditing the grader.
It found no fabricated facts. Six of the seven models were statistically tied on accuracy, and the most expensive cost more than nine times the cheapest without ranking higher. That is the most useful sentence a legal procurement team will read this month.
Other vendors turned the model into a setting. Noxtua shipped a model picker in chat, Word and Outlook that lets users switch categories mid conversation. Casefleet made Claude Opus 5.5 its premium model and says it costs 40 percent less than Opus 5. Everchron moved its summaries to GPT-5.4, and Perspective AI made GPT-6 Sol its default.
Perspective also fixed something quieter and more important. A question a participant never answered now stays empty, instead of being filled in as a blank, a zero or a No. In legal intake, the difference between no answer and No is the difference between a follow up call and a declined case.
Research kept getting deeper and more local. Harvey says its agent can now search US case law by publication status, court and citing references, and follow citations between opinions before it answers, with treatment warnings described as coming later. Recorded Partially Verified.
Vincent AI added Canadian case law and legislation, more than 567,000 decisions. Lex Machina added analytics for more than 576,000 Miami-Dade civil lawsuits, and Jus Mundi added 35 arbitration publications to its library. GC AI now opens the cited passage when a user clicks a citation, and separates case law, legislation and web sources in its citation panel.
Our readWhen six of seven models tie and the price differs ninefold, the model stops being what a legal AI vendor sells. What is left to compete on is the evaluation, the data path, the corpus and the citation a lawyer can click. Merlin and Noxtua are two answers to the same shift. One published how its models performed, and the other published where each one sends the data.
Breach response lost its off switch
Relativity removed the option to turn off generative AI mode in Data Breach Response. Every ingestion now runs through the AI workflow that finds personal information and links it to individuals.
Four days later Epiq acquired Canopy, maker of a data breach response platform with an agentic first level review. Epiq says Canopy will be folded into Epiq AI and a new offering for cyber matters.
eDiscovery AI made its CaseBot assistant available by default on every completed job, with no activation step. Recorded Partially Verified. And Reveal brought Reveal Hold, Onna and Logikcull together as one corporate eDiscovery suite, and added Slack and Zoom collection to Reveal Hold.
Removing an off switch is the most confident thing a software vendor can do. It is also a message to every customer who was still using it.
Our readBreach response is where legal AI has gone furthest toward being the default rather than an option, because the job is volume against a deadline and the regulator counts in days. The next question for this corner of the market is the one a notification letter will eventually raise: how the AI decided which person a record belonged to, and what the reviewer checked before the letter went out.
Market notes
Credo AI added the AIUC-1 standard for AI agents to its governance platform, mapping each requirement to the agents it applies to, and was careful to say it does not issue the certification itself. Relyance AI tightened its Strict InHost mode so sampled record values stay out of scan results. TrustArc let the clock on a data subject request start from either receipt or verification.
In practice management, Lawmatics began selling its Merlin Copilot as Lite and Pro seats at $29 and $49 a month, and added routines that run a request on a timer. CosmoLex let Canadian firms set province and tax rules per matter. Mitratech’s Collaborati now links a resubmitted invoice to the one it replaces.
Case Status connected to Confido so settlements can reach clients digitally rather than by paper check. Recorded Partially Verified. Vesence gave every task its own desktop, so an agent can keep working on one redline while the lawyer moves to another. It is multitasking with a colleague who never needs a coffee.
Law Insider launched Risk Review 2.0, which flags nonstandard terms against its own contract database and offers a redline for each. Recorded Partially Verified. Alt Legal released an Outlook add in that files emails straight into an IP docket, and Rocket Lawyer added video to its attorney consultations.
What the week says about the category
Fifty eight changes, and the theme is the handoff. The agent proposes, the lawyer accepts, and the record starts at that click. It showed up in litigation, contracts, review, sanctions screening and email, from vendors that sell to very different buyers.
Around the handoff, the rest of the market moved. Agents started talking to agents across organizations. Walls started following the text into shared threads. Models became a menu, and breach response lost its off switch.
The billable hour has always been a record of what a lawyer did. This week several vendors started building a record of what the AI did and what the lawyer kept. The two records will have to meet on an invoice eventually, and the count below shows how few vendors are ready for that meeting.
Which legal AI tools help a firm bill for work the AI did?
Hardly any, and the gap is one of the widest this index measures. The AI Legal Index records how 330 vendors handle the bill when AI does the work. 244 of them sit inside a fee relationship between a lawyer and a client, where an hour saved changes the invoice.
Of those 244, 200 publish claims about hours saved and say nothing about what the saving does to the bill. Another 29 say nothing about billing at all, and 10 reach the client as a pass through cost, which is a disclosure question of its own.
Only 5 produce a per matter record of the work the AI did, and 1 of those also publishes guidance on how to bill for it and disclose it. The rest of the index sits outside the question. 73 work before an engagement exists or for teams that bill no client, and 13 exist to review legal bills rather than generate them.
The imbalance matters because ethics guidance is plain that a lawyer bills for time actually spent. A tool that turns six hours into one has changed the invoice, not just the workflow. A firm that cannot show what the tool did has nothing to put in front of the client who asks.
This week’s log showed what the missing record can look like. Advocacy keeps an audit trail of human and AI contributions for each matter, and prices on the contributions a lawyer accepts. Whatever a firm makes of the pricing, the record underneath it is the artifact a fee conversation needs.
200 vendors will tell a firm how much time it is about to save. Very few will help it explain the saving to the person paying the bill.
Every vendor’s billing record, alongside its court disclosure posture, is on the billing posture 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.