Anthropic's release of 20+ legal integrations, adopted by household names like Freshfields and Quinn Emanuel, signals that AI-assisted legal work is no longer optional or experimental. For UK mid-market firms—law practices, insurers, fintechs, accountancies—this is a moment of genuine pressure. You can see the productivity gains. You can see competitors moving. But Fortune's headline contains the actual warning: hallucinations are still happening in filed court documents. The SRA Code of Conduct for solicitors (outcomes 1.1 and 1.2) demands that you maintain competence and act with integrity. Deploying a large language model into live M&A due diligence or employment law work without understanding its failure modes does neither.

This story sits within a pattern we've watched accelerate since 2024. Generalist LLMs (Claude, GPT-4, Grok) are fast and cheap. They're also prone to confident errors when asked to perform specialist tasks at scale—extracting clause dates, identifying conflicts of interest, spotting missing schedules. Harvey and Legora built specialist legal models; Luminance focused on document risk flagging; Microsoft Copilot pushed integration over specialisation. What's changed now is volume: Anthropic isn't offering a single tool, it's offering 12 role-specific plugins that suggest the problem is solved. It isn't. The pattern is that firms with strong governance adopt AI within bounded use cases with human review. Firms without it deploy broadly and discover problems when clients or regulators ask questions.

Here's Trovix's direct position. Claude and its competitors are useful, but they are not fit for unsupervised live matter deployment in regulated sectors. The FCA Consumer Duty (PS22/9) and equivalent frameworks across SRA, FRC and PRA all demand that firms understand the tools they use and can evidence proper control. A plugin that drafts an employment handbook is only safe if you've defined: what it is allowed to do, what human review it requires, what error thresholds trigger escalation, how you log and version its outputs, and how you reverse a decision if the AI was wrong. Most firms deploying Claude right now have not done this. Trovix Sift exists specifically because document AI needs to be transparent about confidence, traceable to source, and governed by humans—not because we think AI shouldn't be in legal workflows. We think it should. We think it should be controlled.

If you're a mid-market law firm, insurer or accounting practice, do not wait for Anthropic to issue a special 'UK-compliant' version. Act now. Identify 1–2 specific, bounded use cases where AI would materially improve speed (routine document review, contract data extraction, intake triage). Define exactly what success means and what errors cost you. Implement with human review as standard. Use tools that show working and leave an audit trail. Train your fee-earners to recognise what AI handles well and where it fails. Big Law's move to production deployment should prompt you to move faster—not recklessly, but faster than firms still debating whether to start. Your regulators will care less about whether you used Claude than about whether you knew what you were doing when you did.

Source: Fortune