News You Can Use

Benedict Evans refreshed his twice-a-year "AI Eats the World" deck and put legal second in enterprise AI spend (~$2.5bn on a16z's numbers, behind only coding) and dead last on Bain's pilot-to-production funnel. Two surveys this fortnight show why the gap is so wide. Ironclad's State of AI in Legal (800+ professionals) reports adoption leaping to 91.6% from 69%, 99% now trusting AI as their roles expand (95% say it has made them more valuable), and 96% saying their organisation now expects more of legal than it did two years ago - yet 87.9% say their workloads have gone up, not down. Litera's Spring sentiment report adds the client angle: 85% feel direct client pressure on AI strategy, yet 32% cannot confidently demonstrate AI value to their most important client. Underneath both sit Kili's benchmark data (single-run consistency of 60% falling to 25% over eight runs) and Mary O'Carroll's demand-side argument that most legal execution is becoming a commodity.

Spend and adoption are racing ahead of anything you could call production value, and a third of firms cannot prove the return to the client paying for it. The win is showing up as throughput and capacity, not as hours saved, so any firm scoring success on hours-saved will read AI as failing while the value leaks out as rising expectations. That is the verification-tax dynamic, and it is the unsolved pricing problem behind every client conversation: clients are now driving the spend (Litera's 85%), yet the firms taking their money largely cannot show what it bought. The work that survives is the judgement clients always felt like they were paying for; the execution layer is being repriced towards zero whether firms are ready or not. Jamie Tso's piece ranks three client models - hidden leverage, a quiet pricing margin, and shared AI infrastructure - and argues the last is the only one that genuinely serves the client: lawyers, clients and AI in real-time loops under human supervision, rather than what he states as "AI slop moving back and forth between two professional parties, with legal fees layered on top". For an innovation function the job has become the adoption, measurement and proof layer rather than another rollout. Clients are forcing the spend, and the firms that can demonstrate value (and reprice around judgement rather than the hour) are the ones who will keep the work.

Accountability was a key theme this fortnight. Legora published the first vendor response to US v. Heppner, arguing enterprise legal AI under attorney direction sits very differently to consumer AI on privilege. The High Court admonished Pinsent Masons in Cork v Smith, a junior used the firm's AI tool to draft letters that twice put a fabricated insolvency rule to the court as if it were statute. The firm has self-referred to the SRA. This is the first reported English case where a firm-procured, enterprise-grade tool produced the hallucination. Bloomberg Law reports an AI-fuelled surge in self-represented litigants, with the Charlotin database now past 1,433 cases. Icertis found 47% of corporate legal teams could not detect an unauthorised AI action until after the fact, and only 23% have a documented agentic-AI policy.

The verification duty does not transfer to the tool just because the tool is enterprise-grade - Cork v Smith is the clearest English authority yet that buying a serious platform changes nothing about the lawyer's obligation to check. Read alongside Legora's privilege post, the vendors are now competing on legal-rather-than-technical positioning. Systemically this is showing that AI lowers the floor on who can file and how, which is genuinely good for access to justice and simultaneously floods courts with volume and error, because the same mechanism produces both. Governance is lagging behind autonomy - half of in-house teams have no real-time oversight of what their agents are doing. This is still a people / training problem and not necessarily a technology problem, the Pinsent's example is a great example of the tech doing its job and the lawyer not following guidance or best practice. We cannot let people get complacent when using these tools and there must be an intentional effort made by Innovation, L&D, Knowledge and supervisors on ensuring solid verification workflows.

Kirkland & Ellis, the world's highest-grossing firm, committed $500m over three to four years to build its own firm-owned platform (more than $100m this year, 180 tech professionals, designed off 250 lawyers, not for sale to anyone). This is spread across three to four years and, on a reasonable reading, partly notional - 250 fee-earners' time over four years could be north of $500m of billable hours on its own, so the figure is as much a statement of intent and an internal repricing of lawyer time as it is fresh capex out the door. And Kirkland was not the only firm, days earlier Fried Frank rolled out FundAssist, a proprietary, internally-built platform that produces first drafts of long-form PE fund-formation documents from client-specific precedent libraries. Osborne Clarke spun out its regulatory-monitoring tool Justima as an independent company it part-owns. Around them the platforms raced up a layer: Harvey shipped Command Center (adoption analytics and peer benchmarking off 1,500+ deployments), a DeepJudge partnership for institutional-knowledge grounding, and Contract Intelligence for in-house teams; Legora pushed its agentic "aOS" and a Datasite diligence integration, and crossed $100m ARR. However, Anthropic's Claude Opus 4.8 scored 10.4% on Harvey's Legal Agent Benchmark at the all-pass standard, up from 4.7's 7.1% and the first model to break 10%, which also means the best legal-agent model on earth still fully completes only about one in ten multi-step legal workflows end to end (it took 91.1% on the easier BigLaw Bench).

The question has shifted from "which legal AI vendor" to "what do we build, buy, internalise, or spin out", and there are now four answers on the board: buy-and-pool (Harvey, turning its install base into a benchmarking moat), buy-and-deepen (Legora, co-development and lower price), build-and-wall-off (Kirkland and Fried Frank, proprietary and exclusive), and build-and-spin-out (Osborne Clarke, externalise and take equity). Many firms are not in a position to build, whether due to talent or spend capabilities. The value for firms is still their institutional knowledge - so no matter the strategy taken the key element to work on is codifying your knowledge and using it alongside your AI Strategy. Ideally firms will protect what is genuinely core and differentiating, and productise or buy what is commoditised / peripheral. Whichever path you pick, autonomy is still running well ahead of reliability, so the human judgment layer is not optional yet and training and hand holding with lawyers to ensure verification and sensible usage is still important.