- In a June human validation of NatWest's payment agent, 83% of test conversations passed overall. Topic-change scenarios, where a customer switches subject mid-conversation, scored only 43%.
- Jonathan Osborne said the single headline score hid a 54-point spread. He said digressions are common in real conversations, so an average can hide where customers do worst.
A single pass rate can hide where a system fails. Jonathan Osborne, lead AI engineer at NatWest, said on the AI Engineer show that a payment agent's 83% pass rate in a human review concealed a 43% score on topic changes.
What was said
Osborne described a June human validation of the bank's payment agent. Three independent human scorers marked each of 111 conversations. They covered 37 payment journeys across four brands. The overall pass rate was 83%.
He then split the result by scenario. Happy-path scenarios scored 97%. Edge cases scored 76%. Topic changes scored 43%. His example: a customer asks whether they can overpay a mortgage without a charge, then asks to close the account instead. He said: "So that single 83% score was hiding that 54% spread."
Osborne said a traditional decision tree is "really, really bad" at such digressions. He also said an agent that can answer any question can answer it wrong. If the overpayment allowance it quoted did not fit that customer's deal, their real money would be at stake.
Why it matters
Our reading: the 83% pools three kinds of scenario with very different scores: 97%, 76% and 43%. A pooled pass rate also depends on how many conversations of each type are in the test. Many easy conversations lift the average. Few hard ones barely move it.
Osborne said digressions are common in real conversations. In this test, the customers who digress were served worst.
For anyone buying or approving an AI agent, one pass rate is not enough. Ask for results by scenario. Ask which conversations sit in the weakest slice.
The other side
The excerpts give no breakdown of the 111 conversations by type. So the 43% has no stated sample size, and we cannot tell how much each type moved the 83%.
This was a human validation of one payment agent, not a measure of live harm. The excerpts do not say why topic changes scored lowest. Osborne said the two biggest causes of failure were engineering issues, not the model. The review still caught them.
The scoring was also soft. Humans agreed 95% of the time on bias and fairness. On whether the bot understood what the user wanted, they agreed only 64% of the time. The excerpts do not say whether the topic-change score improved afterwards.
Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.
The conversation this talking point comes from
- AI Engineer: Building the AI Agent Is Easy. Proving It's Safe Isn't - Jonathan Osborne, NatWest (2026-11-04)
