As a diff the caller can hear: the old value, the new value, and which one is now on file, 'I had L-E-E, you said L-E-I-G-H, I now have LEIGH.' Sierra's τ-Entity benchmark, as Soham Ray reports it on October 22, 2026, found that of the wrong captures a voice agent goes back to verify, only 24% to 37% end up corrected; the rest repeat the mistake or replace the wrong value with a different wrong one. A read-back that speaks only the new value sounds the same either way. The spoken diff is the design handbook's rule that a product never silently replaces a wrong value, applied to audio, and it is the only form in which a swapped error is audible before the agent submits.
As a diff the caller can hear: the old value, the new value, and which one is on file.
The benchmark numbers, the pilot's opening metric, and the copy rule are in Sierra's Voice Agents Fix a Third of the Details They Catch. Read Back the Diff.. The rule the script comes from is in the design handbook chapter Recovery and Trust Repair, and the argument for scoring per definition before pricing is in One Support Agent, Three Resolution Rates.
Last reviewed 2026-10-29 · 1 min read