Grounded knowledge copilot for a retail bank contact centre

The client is a UK retail bank running a large contact centre of roughly 2,600 agents handling everything from routine servicing to sensitive complaints. Every interaction is subject to conduct rules, so consistency and accuracy are not nice-to-haves but regulatory obligations.

High agent turnover and a sprawling, fragmented policy estate made it hard to keep answers consistent, and the cost of getting them wrong (in complaints, remediation, and conduct risk) was significant.

Agents hunted across a dozen disconnected policy systems mid-call, keeping customers on hold while they pieced together an answer. The experience was slow for the customer and stressful for the agent, and it produced inconsistent outcomes that carried real conduct risk.

New joiners took months to get fluent in where information lived and how to apply it, so the fragmented estate was also a training and retention problem. Every departure reset the clock on hard-won knowledge.

Compliance would not accept a bot that could improvise beyond approved guidance, which ruled out an off-the-shelf assistant. Anything the bank deployed had to be provably grounded in policy and unable to make things up.

We built a retrieval copilot over the approved policy corpus that refuses to answer when it cannot ground a response in source, rather than guessing. That refusal behaviour was the feature compliance cared about most: silence beats a confident wrong answer on a regulated call.

Every answer cites the policy reference it came from and hands the agent a ready-to-read response they can confirm before speaking, so the human stays accountable for what the customer hears. We tuned retrieval against real call transcripts and wired in a feedback loop, so gaps in the corpus surfaced and got fixed quickly.

Access rolled out team by team with handle-time and quality monitored throughout, so we could prove the tool improved outcomes before scaling it across all 2,600 agents.