Method
How the check decides.
Fixed rules decide every result. An AI model only pulls out what the agent said, word for word; it never judges whether that was right.
The steps
Read, pull out, judge, sign.
Read
The call comes in from Vapi or Retell. Personal data is masked before any model sees it.
Pull out
An EU-hosted model lists the promises: the time, the price, the discount, “you’re booked”, “we’re fully booked”. Each keeps the exact words.
Judge
Fixed rules hold each promise against the client’s calendar, opening hours and price list. The same input always gives the same result.
Sign
The result, the rule that decided it and the evidence are signed together.
The results
What each result means.
Matched
The client’s records back what the agent said: the time is free or booked as promised, the price is on the list, the discount is within the limit.
Didn’t match
A fact in the client’s records contradicts what the agent said: a price that isn’t on the list, a closed day, a booking that never reached the calendar, a “fully booked” when there was room.
Couldn’t confirm from your systems
The agent said something was done, such as “you’re booked”, but nothing in the client’s systems shows it happened.
Couldn’t check
The records needed to decide aren’t there: the calendar isn’t connected or couldn’t be read, the term was never written down, or the words were too unclear to pin to one time or price. It is listed, never counted as matched.
The report today labels these Correct, Wrong, Not backed by the system and Could not check. They are being renamed to the words above.
The rules
15 named rules.
Bookings and times
- R1 “You’re booked” with no booking made.
- R2 A time that differs from the one booked.
- R3 A booking in the past.
- R4 A booking outside opening hours or on a closed day.
- R5 A weekday that doesn’t fit the date, such as “Monday the 14th” when the 14th is a Saturday.
Prices and discounts
- R6 A price with nothing on the price list behind it.
- R7 A price that differs from the price list.
- R8 A discount above what the business allows.
- R9 Policy promises (cancellation, deposits). Being reworked: until these terms can be typed in, treat a policy result as not checked.
Against the calendar
- R10 A booking that never reached the calendar.
- R11 A booking in the calendar at another time.
- R12 A double booking.
- R13 A cancellation the calendar never got.
- R14 A move where the old slot was kept.
Refusals
- R15 A “sorry, we’re fully booked” or “we’re closed” that the records contradict.
The calendar is read at audit time. A booking made or changed after that read can differ from what the report shows.
Time zones and daylight saving
One policy, everywhere.
Every time is read in the business’s own time zone; it is never guessed. A time that doesn’t exist because the clocks went forward is rejected. A time that happens twice because the clocks went back is read as the earlier one and marked ambiguous.
Limits
What we don’t know yet.
Not yet measured
- False-flag rate: not yet measured. No real call has been checked, so we have no measured accuracy and quote none.
- How many promises the model misses on real calls.
Not yet built
- Typed cancellation and deposit terms, so promises about them can be checked.
- Anchoring the signed record to an outside log. Production signing keys are pending.