OpenAI's launch of ChatGPT for Financial Services—targeting research, modeling and pitchbook automation—signals that generalist AI products are now moving directly into the core work of junior investment bankers and financial analysts. This matters urgently for UK regulated firms because it forces a hard question: if your competitive edge depends on junior staff doing repetitive analytical work, you have no competitive edge at all when a $20-per-month API makes that work commodity. Goldman Sachs partners have already flagged the cognitive atrophy risk—the danger that automating foundational work prevents junior talent from building the judgment and reasoning they need later. That's not a tech problem. That's a workforce problem. And the FCA Consumer Duty (PS22/9) already requires regulated firms to think about the impact of their operational choices on customer outcomes. If your junior analyst team loses the cognitive foundation to spot model errors or validate assumptions, that's not just an internal training issue—it's a conduct risk.

This story is part of a larger pattern: generalist LLM vendors are moving from task automation (email drafts, summarization) into judgment automation (analysis, recommendations, decisions). Harvey focused on legal document review and due diligence with a specialized approach. Luminance built on document intelligence with human-in-the-loop reasoning. But ChatGPT for Financial Services is arriving as a plug-and-play tool with no built-in guardrails, no audit trail for regulatory review, and no way to prove to the FCA that a junior analyst didn't just accept a model recommendation without checking it. The EU AI Act is watching this space closely, and the UK—post-Brexit and desperate to position itself as the global AI hub—is moving toward its own AI regulation with liability frameworks that will make firms answerable for what their AI systems do. The Goldman warning isn't a bug in OpenAI's product. It's a feature that regulators will eventually price into firms' compliance costs.

Trovix's view: automation without governance is liability without benefit. We've seen firms across insurance, legal and financial services deploy AI assistants, document extraction tools, and research platforms without first building a compliance framework to understand what the AI is doing, who is relying on it, and where the firm is exposed if the AI fails. That's the opposite of what the PRA SS1/23 guidance requires—it explicitly demands that firms assess AI risks, document their controls, and maintain human accountability. Generalist products like ChatGPT make sense for low-stakes work: drafting routine emails, brainstorming deal names, formatting documents. They are not suitable for judgment calls—model validation, assumption-checking, recommendation approval—without additional layers. Trovix Audit exists because we found that most regulated firms know they need AI governance but have no way to implement it without either hiring expensive compliance staff or blocking AI adoption entirely. You need to know: which AI tools are in use, what data flows through them, who is accountable for outputs, and how you'll prove to regulators that you're not just offloading risk to an API.

If you are a mid-market financial services, insurance or legal firm, you have 6 months to make a decision: either you build a simple AI governance framework now (user registry, output approval processes, audit logging, bias and hallucination testing) or you wait until the first junior analyst's missed error hits a client outcome, gets escalated to compliance, and forces you to audit your entire AI stack under regulatory pressure. The firms moving fastest are not those buying the most AI tools. They are those building controls first. Start by mapping which AI tools your team is already using informally—ChatGPT, Copilot, Claude, others—and who is responsible if the outputs are wrong. Then layer in basic approval workflows for any output that touches client advice, pricing, or modeling. This is not about saying no to AI. It is about saying yes to AI in a way that survives an FCA visit.

Source: CNBC