How GPT Business is tested.

Every number on this page comes from the GPT Business test records, with the date it was run and how it was measured. Nothing here is a projection.

79,633automated tests in our November 14, 2026 full run, 99.7% passing
4,404independent AI-judge ratings of simulated customer conversations
97.4%of those ratings were 9 or 10 out of 10 (simulated text rehearsals, not live calls)
29 of 29end-to-end stages passed in each of 7 recorded runs on our own test copy, with simulated phone, payment and email providers

Results by category

CategoryResultWhat Is MeasuredBasis
Escalation and safety96 of 96Hands the customer to you the moment it should: emergencies, complaints, anything outside its rules.Judged ratings of 9 or 10, simulated calls
First response402 of 402Opens the conversation the right way: who it is, which business, how it can help.Judged ratings of 9 or 10, simulated calls
Privacy and consent99.2%Respects opt-outs and keeps private details private. A STOP removes consent in the launch tests.134 of 135 ratings of 9 or 10
Empathy99.2%Reads the customer's mood and answers like a person who cares, not a script.131 of 132 ratings of 9 or 10
No invented facts98.7%Quotes your prices, hours and services as you set them, and asks when it doesn't know.397 of 402 ratings of 9 or 10
Clarity98.5%Short, plain answers a customer understands the first time.Judged ratings of 9 or 10, simulated calls
Honest about actions98.3%Only says booked, sent or paid when it actually happened.Judged ratings of 9 or 10, simulated calls
Listening98.0%Picks up what the customer actually asked, including details they mention once.Judged ratings of 9 or 10, simulated calls
Recovery96.4%When something fails, it says so, fixes what it can and keeps the customer moving.Judged ratings of 9 or 10, simulated calls
Handling objections96.1%Answers price and timing worries with your real options, without pressure.Judged ratings of 9 or 10, simulated calls
Getting it done93.4%Moves the conversation toward what the customer came for: a booking, a quote, an answer.Judged ratings of 9 or 10, simulated calls
No double actions0 doubles25 injected faults (timeouts, retries, crashes) caused 0 duplicate texts, 0 double bookings and 0 double charges.Fault-injection run, 12 scenarios

The whole loop, end to end

The end-to-end test plays a full customer journey through the real platform: a call, text or chat comes in, the AI qualifies it, books it, reminds the customer, sends a quote that gets signed and paid, chases what's overdue, reports to the owner and hands off to a human. It also checks that a STOP blocks all contact and that one business can never see another's data.

All 29 stages passed in each of 7 recorded runs between October 24 and November 13, 2026: 203 stage checks, 0 failed. These runs use a test copy of the platform with simulated phone, payment and email providers.

Launch scenarios

11 of 11 golden launch scenarios pass, 47 of 47 checks, re-run on November 14, 2026.

  • STOP removes consent
  • A price that isn't in your price book is refused
  • Prompt injection in a customer's message is blocked
  • An angry customer is handed to a person
  • …and 7 more

Judged conversations

4,404 ratings of simulated customer calls by independent AI judges, including a 96-scenario run across 16 industries; 4,291 (97.4%) scored 9 or 10 out of 10, with 0 critical failures in both full runs.

The production agent is judged separately, in blinded rounds against the scenarios customers actually bring: 44 conversations, 3 judges, 0 critical failures (no wrong price or hours, no invented price, no ignored danger, no contact after an opt-out).

How much gets tested

Testing isn't a phase here; it's most of the work.

  • 79,188 automated tests in the CRM and 445 on this website in the latest full run (November 14, 2026)
  • 7,143 browser tests defined across desktop, Android and iPhone
  • 3,023 labelled safety messages (attacks, look-alikes and clean texts) plus 44 red-team cases
  • 12,510 of 21,958 code changes (56.9%) added or updated tests
  • 6,033 test files added since September 6, 2026
  • Every change to the website passes its full test suite before it can be published

What these numbers are, and aren't

The numbers are here to be trusted, not to impress. The conversation scores are simulated calls rated by AI judges, and the end-to-end runs use simulated phone, payment and email providers. Results from real customers will be published here once there are some, and this page is updated with every major test run.

Conversation scores come from simulated customer calls rated by independent AI judges. End-to-end runs use our test copy of the platform with simulated phone, payment and email providers. Last run November 14, 2026.