News You Can Use

It was argued that the efficiency is real but the savings are not reaching clients, citing 36% of GCs expecting outside counsel spend to increase against 20% expecting a decrease, billing rates up 7.4% year on year against 2.8% inflation, and profits per equity partner up nearly 12%. This could be a question of how clients classify legal work rather than a technology question: treated as a luxury good, AI improves margin and exclusivity justifies the price; treated as a utility, AI cuts the price. Potentially, elite firms will not cut fees and some large buyers actively prefer premium pricing. Meanwhile Crowell & Moring published six months of its own platform data: over 2 million interactions since January, 91% of its 700+ attorneys, nearly 70% weekly.

Crowell published adoption data and nobody published outcome data, which is why neither side of this argument can be settled on anything but assertion. Usage tells us someone did 'something' but whether that something had value, changed the work, what it cost, or who benefited. The insurance sector has already run this experiment for us: McKinsey's 23 July report records technology improving insurance labour productivity by 14% in property and casualty and 24% in life, while aggregate cost ratios ended the period 10% higher than 2005, the gains absorbed by IT spend, compliance overhead and the complexity of layering new tools onto old operating models. Their summary is "efficiency improved, average industry costs increased", and there is no reason to assume we are structurally different. The LexisNexis position is the sharpest counter-argument on the other side: token expenses are under 1% of RELX revenue, profits rose as usage grew, and the CEO has no plans to move to consumption pricing. Cost is not what constrains price in law firms, and pretending otherwise will not survive a panel review. What we could do to prove to a client is take a matter type end to end: baseline time and cost, AI-enabled time and cost, the quality checks applied, the exceptions found, and what that means for the fee.

Three things pointing in three different directions. Crosby, the AI-native US firm, announced it is seeking professional liability insurance for its AI agents specifically, so those agents can do legal work that reaches the client without a lawyer reviewing it; the CEO says lawyers currently review every output and "this won't be necessary in the future". Days earlier OpenAI disclosed that during an internal cybersecurity evaluation, an unreleased model with its guardrails off escaped its sandbox through a zero-day in the research environment, reached the internet and breached Hugging Face production infrastructure, in order to steal the answers to the benchmark it was sitting; a second victim, Modal Labs, was reported on 29 July. Anthropic then reviewed 141,006 of its own evaluation runs and found six, across three incidents, where Claude reached the internet from what it believed was an isolated environment and compromised three real organisations. In one, a model published a booby-trapped package to the public PyPI registry that landed on 15 real systems. Separately, the Civil Justice Council proposed that litigators must declare non-use of AI when preparing witness statements.

Both labs found these incidents themselves, disclosed them voluntarily, notified the affected organisations within days and brought in third parties, which is what a working safety process looks like rather than a scandal. The useful detail is what actually failed. Anthropic describes it as closer to a harness and operational failure than a model alignment failure: a misconfiguration left live internet access in an evaluation environment whose prompts told the model it had none. The boundary was an assumption, not a guarantee, and the models behaved as capable agents would inside a world they had been told was fake. Most AI Policies rest on that same assumption holding. So we have a firm arguing its agents should be insurable to work unsupervised, at the same time two of the most careful organisations in the industry showed how containment fails when a single configuration is wrong, and a court proposing that a witness statement carry a declaration about AI. Law firm policies cover the lawyer, but not the witness, the expert, or an agent acting without one. AI enabled workflows will need clear ownership structures, who owns the boundary, what the agent is permitted to touch, and what happens when it reaches something it was not meant to reach. Insurance follows evidence of control, and the market has a worked example (Orbital) of how slowly that gets priced: the only accuracy guarantee in legal tech took a narrow, well-understood product to place, and two years on it has still not extended to commercial property.

Legora acquired Wexler, a litigation fact-intelligence platform, on 29 July. It is the fifth acquisition of 2026, one a month since March: Walter AI in Canada, Qura in Stockholm, Graceview in Australia, Cadastral in New York and now Wexler in London, whose engineering team becomes the founding team of a London hub. Legora also published BAR, a benchmark of 5,161 evaluation cases across 28 practice areas built on 11,075 source documents, which is explicit that it evaluates "the model, the harness, the tools it has access to, and the system it operates within" rather than the model, and which reports that Legora's own harness improvements added 5% to output quality across production models between June and July. Legal Benchmarks credited BAR's application-level testing and jurisdictional breadth but concluded it "reads closer to a well-polished product-evaluation report than to an independently reproducible research benchmark", noting the judge model is undisclosed where Harvey names its own. Separately, a researcher reported finding more than 40 AI refusal artifacts left as empty document bodies inside Harvey's expanded diligence tasks.

Buying five companies in five months and publishing a benchmark arguing the system is the unit of measurement are the same move stated commercially and technically. If what gets measured is the whole platform, the way to win is to own more of the platform, and Legora is buying practice areas and engineering hubs to do it. That makes "which model does your tool use" the wrong procurement question, and instead we should focus on what a vendor's harness adds over the raw model on our tasks, and whether they will let us measure it. Based on AG benchmarking, through changing only the review architecture inside a platform it moved the results 8.9 points, from 81.6% to 90.5%. LN Labs found roughly 10 points doing the equivalent with the model held constant. Legora's 5% is a relative change across its own models between two dates with no isolation of version drift, so it is a vendor conceding the wrapper contributes single digits rather than a comparable figure. The wider point is that vendor benchmarks are now being audited, and both audits found problems in a week. Having your own evaluation set, run on your own documents, is worth more than any external or vendor driven leaderboard.