Grounding 4,000 advisers with a RAG copilot under FCA scrutiny

The client is a large UK retail bank whose advice network of roughly 4,000 advisers sits at the sharp end of FCA conduct rules. Every recommendation an adviser gives is a regulated act, so the accuracy and provenance of the answer behind it are not service niceties but supervisory obligations.

Product, policy, and pricing guidance had accreted over years into hundreds of documents that changed constantly. The bank competed on advice quality, yet the knowledge its advisers depended on lived in a form no one could search quickly or evidence cleanly.

Advisers answered client questions from memory and a sprawl of scattered PDFs, so the same question could draw different answers depending on who fielded it and how recently they had read the latest bulletin. That inconsistency was slow for the client and carried real conduct risk for the bank.

When second line or the FCA asked how a particular answer had been reached, there was no trail to point to. The reasoning lived in an adviser's head, which is precisely what a regulated advice business cannot rely on.

Off-the-shelf assistants were ruled out because compliance would not accept a tool that could improvise beyond approved guidance. Anything deployed had to be provably grounded in the bank's own documented policy and unable to invent an answer it could not source.

We built a retrieval-augmented copilot over six years of policy and product documentation, constraining it to ground every answer in source and cite the exact document behind each line. When it could not find support, it said so rather than guessing, the behaviour compliance cared about most.

Guardrails, a human-in-the-loop review step, and full logging were designed in from the first sprint, not bolted on later, so the tool could clear second-line scrutiny on its own evidence. The adviser stays accountable for the advice; the copilot does the reading and points to the source.

We tuned retrieval against real adviser queries and rolled it out in cohorts, monitoring answer quality throughout, so the tool earned trust on genuine work before it reached all 4,000 advisers.