News You Can Use

The UK Jurisdiction Taskforce published its Legal Statement on Liability for AI Harms on 7 July. Its headline conclusion is reassuring: existing English private law is largely capable of allocating liability without a bespoke AI-liability regime. They also stated that inappropriate use of AI, choosing an unsuitable system, inadequate due diligence, or failing to validate hallucinated output may all amount to negligence, and, in the right case, a professional could fall below the standard of care by failing to use AI where a reasonably competent peer would have used it. You do not escape responsibility because a chatbot produced the statement. This is persuasive analysis rather than legislation, but previous UKJT statements have been influential with the English courts. Two days later the CPS admitted that two non-existent authorities had found their way into grounds of opposition and were carried into submissions before the High Court. The CPS said generative AI may have been the immediate source, but the operative cause was a human failure to verify; a review of 78 cases found no similar issue. The judge accepted that AI may be useful or even necessary, while stressing the need for proper oversight.

The useful move here is away from the vague instruction to keep a "human in the loop" and towards a defensible professional method. We need to be able to show why a given system was suitable for the task, what sources were checked, who owned the verification, and how an error would have been escalated. A nominal approval clicked at the end of a workflow is not that, and the UKJT has now given a court the language to say so. LLMs are still not perfect, an example from the Legal Benchmarking leaderboard showed that across 34 drafting and 30 extraction tasks, the leading model reaches only 67.6% drafting reliability, and the model with the highest usefulness score (GPT-5.5) manages just 41.2% reliability, illustrating the gap between polished and dependable output. Law Insider asked 534 AI-adopting transactional lawyers what they most want added to their tools, verification and sources came top at 20% of responses, ahead of everything else. We need to look to replace generic human-oversight language in our own guidance with a task-level verification record: approved system, permitted use, named reviewer, sources checked, exceptions found, and the escalation route. More questions will begin appearing focusing on how you prove that something has been verified, especially from clients with audit powers.

Norm Ai raised $120 million at a $1.2 billion valuation for a model that pairs software with human-supervised legal and compliance delivery through affiliated Norm Law. The company says its clients represent more than $30 trillion in assets under management, and it is explicitly selling outcomes rather than seats. That supply-side move matches what buyers are saying. Ford GC Steven Croley says the company's internal AI adoption is now outpacing what it sees from external counsel, spans the full range of practice, and will affect panel relationships: Ford intends to deepen relationships with firms that demonstrate productivity gains and reduce its reliance on laggards. The unresolved question is pricing, and Croley is blunt about it: clients are still seeing annual rate increases without a persuasive account of how firms' AI investment produces any rate relief. Axiom's survey of 528 in-house legal leaders points the same way. Among AI users, all plan budget growth, yet 83% lack established ROI measures and only 7% are optimising and measuring AI across the department. Asked who they prefer for AI-enabled legal work, respondents chose ALSPs over law firms by 52% to 24%.

The competitive unit is becoming the verified outcome, not access to a model. Clients are changing their provider mix before many traditional firms have changed their operating or pricing model, and the field they are choosing from now includes software vendors, law firms, ALSPs and supervised service businesses all competing across the same workflow. We cannot answer that by citing licence numbers or a handful of isolated pilots. It needs evidence that a defined matter type is genuinely faster, cheaper or better, plus a pricing mechanism that shares some of the gain with the client rather than banking all of it. The comforting half of the story is that the buyers have the same discipline problem we do: growing budgets with no task-level ROI leaves them unable to tell useful experimentation from expensive activity, which is exactly why the ALSP preference is a lead we can still close rather than a settled verdict. Law firms will need to first understand the impact of AI on matters and then transparently discuss this with clients - covering things like baseline time and cost, AI-enabled time and cost, the quality checks applied, the exceptions, and the pricing consequence for the client. One honest review of one matter type is worth more to a panel review than any amount of "we use AI extensively". New entrants such as Norm can burn through venture money to capture a market for only so long until they have to up their prices as well, law firms need to keep up or risk having sections of their work go to these aggressive new business models.

Harvey co-founder Gabe Pereyra said the platform's token consumption has increased 14-fold in six months, excluding embeddings. The underlying scale is roughly one trillion tokens in January and a June run-rate of about 13 trillion, and he gave indicative workload costs to go with it: a single assistant query can cost around $20, and reviewing 100,000 contracts can cost around $20,000. These are company examples rather than audited unit economics, and the $20,000 figure is a run across 100,000 contracts, not one contract review, but they make the production-cost question concrete for the first time. OpenAI's general release of GPT-5.6 makes the tiering just as concrete. Luna is priced at $1/$6, Terra at $2.50/$15 and Sol at $5/$30 per million input/output tokens. OpenAI reports that Sol improves quality on difficult professional work, and that programmatic tool calling can cut prompt-token use, supported by customer evaluations from Clio and Legora. Those are vendor and customer-reported evaluations, but on the DD benchmark GPT-5.6 Sol performed well, coming second (although under GPT-5.4 - and at a higher cost).

The wrong response to any of this is a flat token ceiling. A cheap task can waste real money at scale, and an expensive task can still be excellent value, so a blanket cap punishes the wrong thing. The measure that actually matters is cost per verified outcome, with review, rework and failed runs counted in, not the raw token number on the invoice. Benedict Evans supplies the wider market caution worth carrying: supply, demand, marginal cost and buyer ROI are all still unstable, and the current dynamics point towards foundation models becoming commodity infrastructure with more of the value captured by the products and services built on top of them. That strengthens the case we have made before for model portability and routing, and against letting our long-term operating model depend on today's token prices or any one provider's frontier tier. Up until now, businesses encouraged staff to offload the work they disliked, but often failed to redesign the jobs or decide what the recovered capacity was actually for. So token governance is really a workflow and workforce question. We need bounded experimentation, task-level budgets, and routing rules that send routine work to cheaper models while escalating the difficult or high-risk work, and we need to decide what people will do with the time this gives back. A useful action would be to measure one workflow end to end, route it by difficulty and risk, and record its token cost alongside the human review and rework it generated, rather than reporting tokens as a standalone infrastructure line. First, think about what we can charge for this work or what the value is - then compare numbers. Firms with solid data architecture and captured information on what work costs to deliver today and how that is recovered and charged for will be at an advantage when thinking about pricing the value of output against token costs.