Clio reports that firms using its AI bill more hours, not fewer, while New York's courts say AI use needs no disclosure but does need verification, Anthropic shipped ready-made legal workflows, and Google says its new model makes fewer unsupported statements.
At its ClioCon conference in Boston, Clio published its 11th Legal Trends Report and announced it is buying Page, a 12-person regulatory-monitoring startup, for an undisclosed price. Clio says firms using its Clio Work and Clio Manage products billed 6.5% more hours and earned 10% more revenue per lawyer than comparable firms without its AI. It also expects billable time to reach 42% of the working day in 2026, up from 39%.
Read as a buyer, the pattern is the reverse of the one a legal department wants. The firms in Clio's comparison used AI and billed more, so the gain landed on the firm's side of the invoice. The report does not say whether any of it was passed on in price, or which tasks the extra hours went to. The Page deal points the same way: regulatory monitoring is sold as a product for lawyers, with a lawyer reading the output.
This is the problem statement showing up in a vendor's data. Inexpensive work is done by expensive resources, and when AI makes that work faster, the hourly model can turn the saving into more hours. A department only keeps the saving if the routine work stops going to the hourly desk at all. Outsourcing it to supervised agents, with a named lawyer reviewing, is how the saving stays on your side.
New York's court system published a guide on October 27 from its Advisory Committee on Artificial Intelligence and the Courts. Lawyers do not have to tell judges they used AI, but they stay fully accountable for every factual and legal statement in a filing. The guide urges checking citations against authoritative research tools, reading the authorities themselves and documenting the checks.
The guide is explicit about where the risk now sits. Heidi Brown of New York Law School, quoted in the report, objects that without disclosure nobody downstream knows to look harder, and warns about "misgroundings": real cases that misdescribe their holdings and are harder to catch than invented ones. The principle is the one England's Law Society set out last week: the tool does not carry the responsibility.
If outside counsel is not required to tell a court it used AI, will it tell you? Ask your firms whether they use AI on your matters, what they verify, and whether the verification record is something you can see.
Anthropic added a gallery of "Artifacts" to Claude for Legal on October 28, which Artificial Lawyer describes as essentially pre-made workflows that users can edit. The first set includes playbook deviation review and tabular review, and Goodwin, PwC, Foley & Lardner and Harvey have each published one. Anthropic says more than 60 legal connectors have been added since Claude for Legal launched in May.
Artificial Lawyer's view is that the launch is probably not that large, since a skilled legal engineer could already build comparable workflows, and that the aim is to keep legal users on Anthropic's models. Sharing is also more constrained than it sounds: one guide to the launch notes that connected data uses each viewer's own access, and that data stored in a shared artifact can be visible to everyone who opens it.
A playbook deviation board is a good example of the pattern: the agent flags where a contract departs from your playbook, then a lawyer works through the flags. That is useful, and it also keeps the NDA on an expensive person's desk. Tools are becoming a commodity. What a legal team needs is the routing: standard agreements handled end to end by supervised agents, with only genuine exceptions escalated to a lawyer.
Google announced Gemini 4 Carbon on October 28, two weeks after Gemini 4 Argon, and calls it its strongest model for coding, agent tasks and long-horizon work. Google says internal testing showed about half as many unsupported claims as Argon. Access starts with cybersecurity defenders, with developers, businesses and consumers to follow.
These are vendor-reported figures against Google's own earlier model, not an independent benchmark, and the model is not yet generally available. Even taken at face value, fewer unsupported statements means fewer errors to catch, not none. It does not tell a legal team who checks the output, or what happens when the agent acts on something it should have escalated.
As models get cheaper and more reliable, the model stops being the scarce part. What remains is your templates, your escalation rules and a named reviewer. Which of those do you have written down today?
A vendor's data shows AI adding billable hours, a court system keeps accountability with the lawyer, and two technology companies shipped workflows and models. The capability is arriving. The routing of routine work away from expensive hours is not.
Four stories, one gap: AI is getting better at the work, yet the routine part still reaches expensive people because nothing routes it elsewhere. Inexpensive work is done by expensive resources until someone owns that routing.
Outsource legal work to supervised agents. Flank's agents know your templates, terms and escalation rules, take the NDAs, standard redlines and procurement agreements, and hand finished work to a named lawyer for review before anything leaves.