Anthropic's release of 10 purpose-built AI agents for banking, insurance and asset management has spooked the incumbents — FactSet and Morningstar shares fell hard because institutions recognize the threat. On the surface, the agents promise exactly what mid-market regulated firms want: automated pitch drafting, financial statement review and compliance escalation. The problem is not whether these agents work; it is whether they work *safely* within an FCA regime that has made compliance accountability personal (Consumer Duty PS22/9, COBS 2) and an SRA that treats delegation as a governance failure, not a convenience. A compliance review that gets automated must still be *reviewable*, and reviewability requires humans who can explain what they checked and why they signed off.

This announcement is the inevitable outcome of the past two years of agentic AI hype. Since late 2024, vendors including Harvey, Legora and now Anthropic have promised that AI agents can handle autonomous workflows with minimal human oversight. The appeal is obvious: if an agent can draft a pitch deck or flag a regulatory breach, why pay a junior associate or compliance officer? But the pattern is also clear. Agents work brilliantly in sandboxed demos on clean data. In production, across real messy files, legacy formats and contradictory source documents, they fail quietly in ways that regulators and clients spot months later. The market is now waking up to the fact that 'autonomous' is a marketing word, not a compliance posture.

Trovix's view is unfashionable but necessary: agents are tools for *speed*, not for *safety*. An agent can process a financial statement 100 times faster than a human, but it cannot assume liability for the interpretation. For UK regulated firms, the question is not whether to use agents but how to build the governance around them so that automation increases throughput without erasing auditability. This is where Anthropic's announcement leaves a dangerous gap. Their agents are built to work at scale, not to integrate into the review chains that SRA and FCA require. Compare this to Trovix Audit, which treats agent output as a *candidate decision* that must pass through a human checkpoint and leave an audit trail. The agent drafts; the lawyer signs. The agent flags; the compliance officer decides. That is not innovation theatre — it is regulation-proof design.

If you are a mid-market law firm, insurer or financial services practice, the right move is not to buy Anthropic's agents and hope for the best. Instead, audit your current workflows to identify where human review is a genuine compliance requirement (it is almost everywhere in regulated work) and where it is just habit. Then look for platforms that automate the *candidate work* — the drafting, the extraction, the flagging — but leave the *decision* and the *sign-off* in human hands. Trovix Reach and Trovix Aria are built on this principle: agents should make fee-earners smarter, not replace their judgement. Build that discipline now, before the regulator builds it for you.

Source: Bloomberg News