When Freshfields, Quinn Emanuel and Holland & Knight announced they are using Claude on active client matters, they did not say how they are catching the hallucinations that Anthropic itself has documented. They did not publish their governance frameworks. They did not explain what happens when an AI confidently cites a non-existent case in a brief. Anthropic's release of 20 new legal plugins — from M&A due diligence to employment handbook drafting — is technically impressive. It is also a live demonstration of the gulf between product capability and institutional readiness. For UK regulated firms, this matters acutely. The SRA Code demands that solicitors act with integrity and competence. The FCA Consumer Duty (PS22/9) extends similar requirements to financial services. Neither regulator has yet issued explicit guidance on AI use in live client work, but both will interpret existing duties to include responsibility for AI-generated outputs. Big Law's sprint towards deployment is not a permission slip for mid-market firms to follow.
This story is part of a larger pattern: the industry is fracturing into two camps. One camp — represented by Anthropic, Harvey, and now Freshfields — is betting that raw capability, scale and speed trump everything else. They are rolling out AI on live work and learning governance in real time. The other camp, including tools like Luminance and Legora, has built compliance and verification into the workflow from the start, accepting slower deployment in exchange for audit trails and explainability. Neither approach is risk-free. But the first approach is now running ahead of the second with no clear safety net. Meanwhile, traditional vendors like Microsoft Copilot are still positioning AI as an assistant, not an agent, keeping human lawyers in the loop. That design choice looks less conservative and more wise as hallucinations compound.
Trovix's view is direct: deploying generative AI on live client matters without robust verification, audit and rollback mechanisms is not innovation — it is regulatory exposure. Big Law can absorb reputational damage and settle claims. Mid-market firms cannot. If Freshfields makes a mistake with Claude, they defend it as a training exercise in a cutting-edge firm. If a mid-market practice makes the same mistake, it becomes evidence of negligence in a professional conduct hearing. The difference is not the tool. It is governance. That is why Trovix Audit exists: not to slow down AI adoption, but to make it defensible. Audit trails, output verification, model provenance, and version control are not optional extras — they are the conditions under which AI can legally touch client work. Anthropic's plugins are genuinely useful. But usefulness without governance is negligence waiting to happen.
What should a mid-market law firm, accountancy practice, insurance broker or financial services firm do right now? Do not follow Big Law into Claude on live matters yet. Instead, start with a clear governance framework: define which tasks can go to AI first, then build the verification layer around those tasks, then — and only then — run live work through it. Use Trovix Sift or similar to extract and verify key data before an AI agent touches it. Use Trovix Reach for client-facing AI so you control the boundary between firm knowledge and generative output. And use Trovix Audit to document everything: what the AI was asked, what it produced, who verified it, and what was changed before it reached a client. The SRA, FCA and ICO will not care what Anthropic says works. They will care what you can prove you checked. Big Law's speed is not your model. Big Law's liability insurance is not your budget.
Source: Fortune