Explicit feedback is heavily biased
The people who rate an interaction are disproportionately the delighted and the furious. A satisfaction score built on that is a measure of intensity, not of typical experience, and it moves for reasons unrelated to quality.
1. Did they come back about the same thing?
A repeat contact within a few days is the clearest available signal that the first attempt failed, and it requires nobody to answer a survey. This is resolution confirmed downstream rather than assumed.
The strongest satisfaction signal you have is whether they had to ask twice.
2. Did they ask for a human?
Explicit escalation requests are a direct measure of the moment the system stopped being acceptable. Rising requests mean scope is wrong, and by the time someone types it the escalation should already have fired.
3. Did they abandon mid-conversation?
Silent drop-off is the invisible failure — no complaint, no rating, just a customer who gave up. It is the same signal as form abandonment and deserves the same attention.
Ask, but ask specifically
If you survey, ask whether the problem was solved rather than how satisfied someone is. Solved is a fact the customer knows; satisfaction is a mood that varies with everything else in their day.
Compare against your human baseline
Measure the same three signals for conversations your team handles. Teams frequently discover the AI is close to their own numbers, or worse in one specific category worth pulling back — neither of which is visible without the comparison.