What "AI resolved 70% of tickets" actually means
A field guide to reading an AI resolution-rate claim — the denominator tricks, the definition games, and the arithmetic vendors hope you skip.
"Our AI resolves 70% of tickets." You will hear a version of this in every vendor pitch, and it is almost always technically true and practically meaningless. The number is real. What it counts is the question — and the answer is usually engineered to flatter. Here is how to read one without getting fooled.
Trick one: the definition of "resolution"
Start with the word itself. "Resolution" sounds like the customer's problem was solved. Often it means something much weaker.
When "the customer stopped replying" can count as a resolution, the headline rate quietly includes everyone who got a useless answer and gave up, everyone who rage-quit to a competitor, and everyone who found the answer elsewhere out of frustration — all scored as wins. Before you believe any resolution rate, get the definition in writing, and check whether assumed resolutions (silence) are bundled in with confirmed ones (the customer said it helped).
Trick two: the denominator
The second question is resolved out of what? A resolution rate is a fraction, and the vendor controls the bottom of it. Let the bot decline the hard questions and route them straight to a human, and those never count against it. Now "90% of the tickets it chose to handle" becomes "90% resolution" in the deck.
Watch what happens to a single vendor's number as the denominator changes:
Same product, same year. The all-customer average is 76% — by the vendor's own generous definition. The marketed number is 85%, drawn from the ten industries where the product performs best. Neither figure is fake; they're just different denominators, and the bigger one is on the billboard. If your business isn't a top-ten-industry, high-volume, well-documented-knowledge-base account, neither is your rate.
There is only one honest way to read a resolution-rate claim: ask what counts as a resolution, and over what.
Trick three: a forecast dressed as a fact
The third move is subtler — quoting a prediction as if it were a measurement.
"AI will resolve 80% of issues by 2029" is an analyst's forecast about the future, not a result anyone has booked. It's a reasonable thing to plan around and a dishonest thing to place next to a measured number without a label. Sort every resolution figure you're shown into one of three buckets: measured (someone counted real outcomes), self-reported (the vendor counted, by its own definition), or forecast (nobody has counted anything yet). Most of what you'll be shown is the middle one, dressed as the first.
How to read the next one you're shown
Four questions, in order:
- What counts as a resolution? Confirmed by the customer, or assumed from silence?
- Resolved out of what? All contacts, or the subset the AI chose to attempt?
- Whose contacts? Your industry and complexity, or a cherry-picked best case?
- Measured, self-reported, or forecast? Label it before you compare it to anything.
Do this and the impressive number usually shrinks to an honest one — often still good, just not 70%. A vendor that answers all four cleanly has earned some trust. One that can't is showing you how the number was built. For the full pre-signature checklist, use the vendor scorecard; for why the "deflection" cousin of this number is just as slippery, read deflection is a vanity metric.