Prepared for Example Air, Digital Care · every number below opens to the conversations behind it
Illustrative sample · daily volumes and the weekly human checks are simulated to show the format; four days open real test conversations
conversations resolved correctly, all three checks passed
times a person agreed with the AI reviewer, checking blind
days graded, every conversation sealed the moment it happened
1
What we promised
One outcome, agreed with Example Air before the month began, broken into three yes/no questions a person could answer by reading the conversation.
🎯
The outcome
The customer’s request is resolved correctly, within airline policy
A conversation counts as resolved correctly only when the answer to all three is yes. One “no” and it counts as a miss.
2
How we checked
An AI reviewer grades every conversation every day. A person re-grades a sample every week without seeing the AI’s answer, so we know how far to trust the AI.
3
What happened
Click any bar or reason to see the conversations behind it.
Why conversations missed
4
Day by day
Each tile is one day’s sealed grading. The purple tile at the end of each week is that week’s blind check by a person. Click any tile.
resolved correctlymissedperson’s blind check◆ airline policy changed that dayREAL opens real test conversations
5
Can you trust the grading?
The AI reviewer is itself checked. Each week a person graded a random sample blind, then we compared.
graded by both, where the person and the AI agreed
Click to see every conversation the person checked.
What that means for the headline
50%75%100%
Days nobody re-checked stay graded by the AI only; the report says so rather than hiding it.
6
What ran
Everything that produced these numbers, pinned before the month began. Change any of it and this report no longer checks out.
🔒 Locked on Aug 31, 2026, before the first conversation of the month. The lock itself is a sealed record.
7
Check it yourself
You don’t have to take our word, or the agent vendor’s. Anyone can check this report with free, open-source tools and no account.
Pick any conversationOpen it from any number above and take its record id.
Confirm it hasn’t changedThe verifier recomputes the record’s fingerprint and checks it against the month’s published summary.
Confirm when it was sealedThe summary was timestamped by an independent witness on Sep 30; nothing can be added or edited after that without it showing.
Technical details
If a record doesn’t match, or isn’t in the summary, the report is wrong, and anyone can show it.