Anthropic's announcement of 20+ integrations and a 90.9% BigLaw Bench score has triggered a gold rush. Global firms including Freshfields and Quinn Emanuel are running Claude on live client work. For UK regulated practices, this matters because it signals a dangerous normalisation of untested AI in billable work. The SRA Code of Conduct is explicit: you must act in the client's interests and not allow incompetence to harm them. A 90.9% accuracy rate sounds impressive until you realise that in a 50-page M&A due diligence report, that error rate means hallucinated findings embedded in client advice. Big Law can absorb reputational damage and litigation costs. Mid-market firms cannot.

This story is the third act of a predictable pattern. Harvey launched with strong benchmark scores and regulatory backing. Luminance marketed AI as a solved problem. Now Anthropic is doing the same—riding model improvements and integration breadth into market adoption before anyone has seriously tested what happens when these systems fail on live matters. The underlying problem is structural: vendors measure success against benchmarks, not against real-world deployment failure rates. A benchmark tests the model in isolation. A law firm tests it under pressure, with limited time for review, with partners billing by the hour, and with clients who will sue if advice is wrong. Nobody is publishing failure rates. Nobody is measuring the cost of hallucinations that made it past a rushed review.

Trovix's view is clear: integration breadth is not the same as integration safety. Connecting Claude to 12 role-specific plugins and Microsoft 365 does not solve the hallucination problem—it amplifies the surface area where hallucinations can hide. The difference between Anthropic's approach and responsible deployment is governance. A 90.9% model running unsupervised is not safe. A 90.9% model running with audit trails, output validation, human sign-off protocols, and real-time monitoring of error patterns is defensible under the SRA Code and FRC ISA UK. If you are using Trovix Audit, you can measure what Claude actually does in your workflows, flag systemic failure modes, and prove to the regulator that you did not just hope for the best. Firms using only the plugin integration cannot.

Here is what a mid-market UK regulated firm should do right now: Do not wait for Big Law to finish the experiment. If you are considering Claude or any large language model for live client work, do not deploy it without an audit framework. Establish a baseline—what error types matter most in your practice? Run the model on historical work, score it against your own standards, not against vendor benchmarks. Require human sign-off on any output that touches client advice, regulatory filing, or financial recommendation. Document the control environment so that when—not if—the regulator asks, you can explain why you chose this model and why you believe it meets your duty of care. If you lack the internal capability to audit AI output, you lack the capability to deploy it safely.

Source: Fortune