Anthropic announced ten AI agents this month designed to draft pitch decks, review financial statements and escalate compliance cases across banking, insurance and asset management. On the surface, this looks like the answer to the compliance bottleneck plaguing mid-market UK regulated firms. The reality is messier. These agents are pattern-matching tools trained on public data. They will handle routine document flagging and assembly work competently. But they cannot replace human judgment on whether a client transaction breaches FCA Consumer Duty PS22/9, whether a fee arrangement satisfies SRA Code transparency rules, or whether a file triggers PRA SS1/23 regulatory capital thresholds. Regulated firms that deploy agents as decision-makers rather than as research assistants will create liability, not reduce it.

This news sits within a larger pattern: the industry is moving fast from 'ChatGPT can do compliance' to 'agentic systems will do compliance unsupervised.' Harvey, Luminance and others have built document-first AI products that embed legal and regulatory knowledge. That approach has merit. But agents that operate autonomously across multiple tasks without explainability are a different animal. The EU AI Act, now shaping UK regulatory expectations, classifies high-risk AI as anything that materially affects legal or financial rights. Compliance escalation done by an opaque agent qualifies. So does financial statement review that feeds into credit decisions. The industry has not yet faced the ICO, FCA or SRA enforcement action that will follow the first major failure. We will.

Trovix's view is this: agents should augment structured workflows, not replace judgment gates. A mid-market law firm should use agentic systems to extract data from disclosure bundles, flag keywords, surface precedent and suggest next steps. But the escalation decision—what leaves the firm and what gets reviewed by a partner—must remain human-controlled and auditable. That is why governance matters more than capability. Trovix Audit exists to make that audit trail and governance logic transparent and defensible to regulators. Tools like Copilot give you agent capability without compliance scaffolding. Anthropic's agents give you speed without explanation. Neither solves the real problem: how to stay compliant while automating repetition.

What should a financial services practice do now? Do not wait for your vendor to build 'compliance agents' and assume they work. Instead, map where compliance judgment actually lives in your workflows. Where do you need speed? Document triage. Where do you need judgment? Every decision that moves money, takes on client risk or shapes advice. Build your agent strategy around that split. Use agentic AI to feed human decision-makers better information faster. Use Trovix Sift or similar tools to extract and structure what agents surface, so your partners can see what the system saw and why it matters. Test agents on historical cases where you know the right answer before they touch live work. Document your testing, your controls and your escalation rules. When the FCA asks—and they will—you will have evidence that you deployed AI in service of judgment, not in place of it.

Source: Bloomberg News