Anthropic's new legal plugins are technically impressive — 12 role-specific integrations, grounding that pulls only from verified sources like Westlaw, and live use across Freshfields, Quinn Emanuel and Holland & Knight. This matters to UK regulated firms because it signals something powerful: the legal sector has stopped debating whether to adopt generative AI and started racing to embed it. But the Fortune headline buried the real issue: hallucinations are still showing up in legal filings. Grounding reduces the risk. It does not eliminate it. And that distinction is exactly where most UK law firms, insurers and accountancy practices are getting this wrong. The SRA Code expects regulated practitioners to act with integrity and in the public interest. Pushing AI through live matters because your US competitors are doing it is not integrity — it is liability transfer.

This story is part of a pattern we have watched for 18 months. Harvey, Legora and Luminance all claimed their models would solve hallucination through better training, better architecture or better data. None of them did. Instead, the industry has moved the goalpost: first it was 'we will eliminate hallucinations', then 'we will reduce them', now it is 'we will mitigate them with grounding'. That is honest progress, not victory. Meanwhile, the real winners are the firms that have stopped talking about the model and started talking about the process: governance, audit trails, human sign-off workflows, output verification, and clear accountability for decisions. The EU AI Act is coming. The FCA Consumer Duty (PS22/9) already applies. The ICO's guidance on generative AI makes clear that 'technical measures alone are insufficient'. Big Law is betting that if Westlaw data goes in, Westlaw data comes out. UK mid-market firms cannot afford that bet.

Trovix's view is clear: grounding is table stakes, not strategy. It matters. It is not enough. A law firm using Claude through verified legal sources still needs to answer three questions that Anthropic's plugins do not address: (1) Who owns the decision if the AI output is wrong? (2) How do clients know whether their document was reviewed by a human or a model? (3) What happens when a client sues because the AI missed something a junior associate would have caught? These are not technical questions. They are governance questions. And they are why Trovix Audit exists — to map exactly where AI is used in your workflows, who is accountable at each step, and how you prove compliance to the SRA, FCA or your insurance panel. Grounding tools like Claude's reduce hallucination risk. An AI governance dashboard tells you whether your deployment is actually compliant. The two are not the same.

If you run a mid-market law firm, insurer, financial services business or accountancy practice, the message from Big Law should make you move faster, not copy them faster. Yes, adopt Claude or another advanced model. Yes, use grounding where it is available. But before you push it live on client work, map your AI workflows, assign accountability, document your sign-off process, and audit your outputs. Your bigger competitors have in-house AI legal teams and client relationships strong enough to survive a mistake. You do not. The regulatory risk from cutting corners on AI governance is now higher than the business risk of moving slightly slower than Freshfields. Use the tools Anthropic is building. But own the governance yourself.

Source: Fortune