TLDR Tech
AI Agents Aren't Failing, Your Infrastructure Is
The term 'agentic AI' is doing a lot of heavy lifting right now, and most of it is dishonest. What enterprises are actually running are glorified chatbots with a few API calls bolted on. They call it an agent. It isn't. And the gap between that fiction and genuine orchestrated agency is where most digital transformation budgets are quietly disappearing.
For UK consumer finance specifically, this matters more than in most sectors. We operate inside a regulatory perimeter that demands explainability, audit trails, and documented decision logic. An AI agent that books meetings or drafts emails in a low-stakes SaaS environment is one thing. An agent that touches affordability assessments, customer communications, or credit decisions is operating in a completely different risk environment.
The infrastructure gap is the real story here. Three things consistently break agentic deployments before they leave the pilot stage:
- Nonhuman identity governance. When an agent acts, who is it? Most organisations have no answer.
- Logging that satisfies a regulatory audit, not just an internal retrospective.
- Data foundations clean enough that the agent isn't confidently hallucinating against stale or inconsistent records.
The FCA's existing guidance on model risk, combined with the Consumer Duty obligation to demonstrate good outcomes, creates a specific problem for any lender or broker pushing agents into customer-facing workflows. You cannot point at a black-box agent and say 'it acted in the customer's best interest.' You need to show the reasoning chain, the data it used, and why that output was appropriate for that customer's circumstances.
Orchestration frameworks like LangGraph or Microsoft's AutoGen are maturing quickly. The tooling is genuinely getting better. But tooling doesn't fix a data warehouse that hasn't been properly governed, or an identity and access management setup that was designed for humans.
The question worth sitting with is whether your organisation is calling something an agent because it genuinely operates with autonomy and judgement, or because the word sounds more impressive in a board update than 'workflow automation.'
- agentic
- AI agents
- AI