Resources › AI accuracy
Our AI accuracy check, published in full
Test run: 16 September 2026 · Page last updated: 26 September 2026
A number is only useful if you can see what was measured, on how much data, how it was scored, and what it does not prove. This page shows all four.
Published readiness check
One customer. One assistant. 25 questions. 16 September 2026.
Answered from the knowledge base
20 of 20
Questions the assistant answered from the customer's own knowledge base instead of handing them to a person.
Escalation handled correctly
5 of 5
Too few cases to be a reliable rate
Test questions designed to require a human, where the assistant correctly handed over instead of answering.
Overall readiness score
100%
A combination of the two results above. It is a summary of them, not a separate measurement.
How we run this check
What is tested. The check runs against the customer’s real assistant, through the same live pipeline that answers website visitors: the same knowledge base, retrieval, confidence thresholds and escalation rules. Test conversations are flagged so they never reach the customer’s inbox.
The customer. One production customer in US company formation services, published with the customer’s written permission. The customer is not named.
Where the questions come from. No person writes the questions. An AI model reads the customer’s knowledge base and generates realistic customer questions, plus a smaller set of edge-case questions that should be handed to a human, such as billing disputes or requests for personal legal or tax advice.
How answers are scored. Scoring is automatic. The check records whether each question was answered from the knowledge base or handed to a human, and whether escalation-test questions were correctly handed over. No person reviewed or graded the individual answers.
How we got here. This was not the first check we ran on this assistant. The first automated check scored 86% overall, with 80% of questions answered from the knowledge base. We then added missing answers to the knowledge base and adjusted the escalation rules three times, and re-ran the check until it reached the result shown above. We think you should know that.
What these numbers do not tell you
A perfect score on a small sample is not proof of a perfect system. With 25 questions, a perfect result is consistent with a true underlying rate as low as about 89% at 95% confidence. The sample cannot tell an excellent assistant from a very good one.
The questions come from the knowledge base itself. Because the questions are generated by reading the customer’s content, they naturally lean toward things that content covers. Real visitors ask messier, broader and more unexpected questions.
“Answered” is not the same as “verified correct”. The check measures whether the assistant answered from the knowledge base or handed over. It does not have a person confirm that every answer was complete and correct.
The results reflect tuning. The knowledge base and escalation rules were improved after earlier checks on this same assistant, so this result shows what a tuned deployment can reach, not a first-day result.
Results depend almost entirely on the knowledge base. A thorough, current knowledge base produces good answers; a thin or outdated one does not. This result describes one customer’s content as much as it describes our software.
AI responses can still be wrong. AI output is generated by probabilistic systems and can be incorrect, incomplete or outdated even when the source content is right.
Escalation is not guaranteed in any single conversation. The assistant is designed to hand over when your content does not support an answer, but a confidence score is not a measure of correctness.
One customer is not every deployment. These results are not representative of GoBrise generally, of your industry, or of what your own assistant would score.
Our commitment
If a later check scores lower, we will not quietly switch back to an older, better run. When the published run changes, the test date on this page changes with it.
What you can ask us for
- The question set used in this check
- A description of how the readiness score is calculated
- A readiness check on your own content during evaluation
Write to sales@gobrise.com. If something is not documented yet, we will tell you that.
Run the same check on your own content
The readiness check is built into GoBrise. It runs automatically after you import your website or make larger knowledge base changes, and shows where your assistant answers confidently and where it should hand over.
It will show how your assistant performs on the questions it generates from your content. It will not reproduce the numbers on this page, and it does not predict live performance.
Related: Security & Trust · Terms of Service