How we measure the 95.4%: one figure about our decisions, stated in full
2026-09-24
On 3,248 production cases handled by one build over three days in September, the system's recommendations and case notes were consistent with our own written collections policy and grounded in the invoice data in 95.4% of cases, with a 95% confidence interval of 94.6% to 96.1%, about three fifths of that traffic coming from demonstration and test organisations.
It measures adherence to a written policy and consistency with the invoice data. Every decision is scored — no sampling and no human labelling — and for any case you name we can show the rule it was tested against.
- Nothing we publish claims you will be paid faster until customer results show it
- Want to check it on your own ledger? Mark a week of recommendations agree or disagree — your disagreements become your settings
- Coverage: 90% of the collections work that isn't reserved for a person, in assist mode (356 of 396 steps)
The full method is in our performance paper on the Safe AI & Security page.