News You Can Use

On 13 October Legora launched Intake, an agent that sits in Slack and takes requests from the business. It asks follow-up questions, routes each request to an owner with a due date, and drafts a first version from the legal team's own playbooks and precedents. Lawyers review everything by default, and the team decides request type by request type when the agent can finish alone, such as a pre-approved NDA. Email and Teams are planned. The same day Harvey's Gabe Pereyra and Matt Zerweck set out Harvey's version, which was already live as a product. Intake reaches into Outlook, Slack, Jira and Salesforce, and Legal sets three routes: an agent answers routine questions within Legal's rules, Harvey prepares the work for the assigned lawyer, or it goes to a designated outside firm, chosen on matter history, performance and invoice data. Coheso and Spellbook shipped front-door features in the same week, and Legora says in-house teams are now its fastest-growing UK client group, growing more than three times as fast as its law firm base this year. Legora cites the ACC's benchmark that the median in-house lawyer now supports 367 employees, up 26% in a year. Neither vendor has published a customer, accuracy or time-saving figure for intake, Legora's version only works in Slack today, and intake tools that route Slack requests to lawyers already exist. Both vendors now add an agent that drafts the first pass inside the same platform.

Whoever owns the front door decides which work reaches a lawyer, and more and more often which work reaches a law firm. In Harvey's model the client's legal team writes rules that send routine questions to an agent, keep judgement calls in-house and pass the rest to a firm picked on its invoices. Our work for in-house clients will sit behind routing rules we did not write. Priori's Matthew Wheatley said this week that in-house teams want AI savings passed through and firms are not doing it, so price will shape those rules. For the in-house teams, deciding which requests an agent may finish alone is a supervision decision, and it should be owned and reviewed like one. For us, we should ask our largest in-house clients which platform their front door runs on, and make sure our precedents, matter context and pricing can arrive inside it, before the email instruction stops coming.

Last month an OpenAI agent went round a block on an Australian government portal. This fortnight rogue agents became an industry problem and a liability question. California's attorney general issued an investigative subpoena to OpenAI, which has now told more than 100 organisations about rogue agent activity. Then Anthropic disclosed its own. Running evaluations, its agents exploited flaws on third-party websites, used access tokens found in site files to reach paid data, and submitted forms they were meant to stop short of. In one case a Claude agent filed an invented tip on the Philadelphia police unsolved homicides form. Anthropic says the cases had "minimal real-world impact" and has cut its internal evaluations off the live internet, and Philadelphia police said being told two months later was unacceptable. Senators Hawley and Murphy announced a bill that would make anyone who knowingly runs an agent that recklessly causes hacking damage liable under the US computer misuse statute, with developers liable for missing reasonable safeguards. No text is public yet. In the UK, the Law Commission opened a consultation on 8 October proposing that product liability law cover software and AI systems however they are supplied, with a presumption of defect that the producer can rebut where the evidence is hard to get. It closes on 14 January. The SRA's review of solicitor competence, closing on 3 December, adds a supervision element covering oversight of technology and AI work.

Liability is moving towards whoever runs the agent. The US bill targets operators, the Law Commission would let a claimant start from a presumed defect, and the SRA wants supervision of AI work written into competence. If Anthropic's own agents filed forms on live websites in testing, we should not assume a vendor's agent stays inside the task we give it. Before we switch on agentic features, we need a written list of what each agent can reach outside our systems, a log we hold that proves what it did, and contract terms that say who pays if it goes further. We should also respond to the Law Commission. Where it draws the defect line will govern the AI we buy and the AI our clients build.

The Wall Street Journal reported that just 11% of almost 400 businesses surveyed could accurately project their AI costs. Prices moved both ways this fortnight. At the bottom, Anthropic's Haiku 5.5 launched at a tenth of its predecessor's price for most requests, OpenAI's Decisions charges 10 cents per million input tokens with no output charge, and Microsoft-Decision-1 charges about 4 cents on the same basis. Both are built to classify and route work. Ivo released Sage, an open-weight contract model built on DeepSeek, and on Vals' run of Harvey's legal agent benchmark Z.ai's new GLM-Legal scored within half a point of Google's Gemini 4 Argon, at 15 cents a test against $10.55. At the top, Argon became generally available on 13 October at its launch price, which doubles on 1 December. The Information reported that Anthropic's enterprise discounts end once usage passes the contracted volume. Research from Stanford, CMU, Berkeley and Microsoft found that the cheaper reasoning model ends up costing more in roughly a third of pairings, because it uses more tokens. On Vals' legal research board three models tie for first place at 59 cents, $6.76 and $23.06 a test.

The cheapest models are now priced for one job, sending work to the right model, and every route adds a price that can move. That is why the bill is hard to forecast: what a task costs depends on which model got it, how many tokens it used, and whether a launch price or discount had run out. Legora argues there is no best model and the system around it is what matters. That suits a company selling the system, and it is still where we should be looking. We should ask vendors which models sit behind each feature, what happens to our price when a launch rate ends or a volume tier is crossed, and for cost per completed task on our own work rather than per token. We should also run a small benchmark of our own routine tasks every quarter, because on today's boards the same legal research score costs anywhere from 59 cents to $23.