Ask customer service leaders how their new artificial intelligence is performing and you will usually hear one number: deflection rate. It is the figure vendors put on the first slide, the one dashboards track in bold, and the one leadership repeats in the quarterly review. It is also the wrong number around which to build a strategy.
Deflection rate measures how many contacts were kept away from human agents. It says nothing about whether the customer's problem was actually solved. A ticket that an AI deflected but did not resolve does not disappear. It comes back as a repeat contact, an escalation, a chargeback, or a one-star review. The cost was not eliminated. It was moved somewhere your dashboard was not looking.
For anyone making CRM and contact center decisions, that distinction is the difference between an AI program that pays for itself and one that quietly generates work while reporting success.
Deflection won by being easy. It is simple to count: A contact either reached an agent or it did not. It is available on day one, needs no follow-up window, and always moves in the flattering direction when you add automation. It also happens to be the metric that makes the software look most valuable, so there is little incentive on the vendor side to complicate the story.
Resolution is harder. To know whether a problem was truly solved, you have to look past the single interaction: Did the customer come back? Did he escalate? Was she satisfied? That takes time and attribution, so it tends to get deprioritized in favor of the number that is ready now. Easy and immediate beats accurate and slow, until the repeat contacts start piling up.
The resolution rate is not one number, it is a small set of signals that together tell you whether the automation is doing the job or just hiding it.
First-contact resolution by channel shows whether the AI closed the issue on the first try and where it does that well or badly. Repeat-contact rate within 72 hours is the honesty check: If a resolved ticket generates a second contact three days later, it was not resolved. The escalation rate on AI-handled tickets tells you how often the automation reached its limits, which is fine in itself as long as you are counting it. And customer satisfaction on AI-resolved interactions, measured against agent-resolved ones, tells you whether customers experienced a real fix or just a faster brush-off.
None of these are exotic. Most contact centers already collect the raw data. The shift is deciding to look at them together and to treat a resolved-and-stayed-resolved ticket, not a deflected one, as the win.
If resolution is the goal, it has to show up in how you evaluate and buy, not just how you report after the fact.
Start with the vendor conversation. Ask how the product measures resolution, not just deflection or containment. Ask what happens to a ticket the AI cannot resolve and how that handoff reaches an agent with full context. Ask to see repeat-contact and satisfaction data from existing deployments, not only deflection figures. A vendor confident in outcomes will have answers. One that steers you back to deflection every time is telling you for what it optimizes.<
Then protect your baseline. Before anything goes live, capture your current first-contact resolution, repeat-contact rate, and satisfaction by channel. Without that starting point you will never be able to separate real improvement from a deflection number that simply learned to route people away. The baseline is cheap to capture on day one and impossible to recover later.
Finally, align the incentives in the contract. If a vendor's commercial model rewards volume deflected rather than issues resolved, the product will be tuned to deflect, and you will feel it in your repeat-contact rate. Where you can, tie success criteria and renewals to resolution and satisfaction metrics on which you both agree up front. Incentives shape behavior, in software as much as in people.
None of this means deflection is useless. It is a fine operational signal, a way to see how much load the automation is carrying. The mistake is letting a load metric stand in for an outcome metric and then making budget and vendor decisions as if the two are the same.
The teams getting real value from AI in customer service changed the question: not how many contacts did we keep away from agents but how many customers left with their problems actually solved, and did they stay solved. That number is harder to produce and less flattering in a quarterly review slide. It is also the only one that reflects what your customers experienced and the only one on which it is worth building a strategy.
Count resolutions. The deflection number will still be there when you need it, but it should never again be the headline.
Ralf Klein is founder of Triad, an artificial intelligence automation agency based in the Netherlands that builds operational AI agents for support and operations teams. He writes about measuring AI by outcomes rather than activity.