Track six customer support metrics to judge AI support: cost per ticket, first contact resolution, resolution rate (not just deflection), repeat-contact rate, hand-off quality, and the share of eligible tickets resolved with no human touch. Together they show whether customers' problems were solved, not just whether a chat was closed.

Most vendor pitches for AI customer support lead with a big adoption number: some percentage of support orgs "using AI" in some form. That number is probably true and also not that useful. Using AI for drafting internal replies is a different thing than letting it resolve a billing dispute on its own, and lumping them together flatters everyone.

If you're evaluating AI customer support for your own team, here's a more honest set of numbers to anchor on, measured on your own tickets, because benchmarks vary widely by company size and vertical.

Why is deflection rate not the same as resolution?

Ticket volume has shifted channel over the last decade. Chat and in-app messaging have overtaken email as the primary support channel for a lot of B2B and consumer software, and much of that chat volume is repetitive: password resets, plan changes, "where is this setting," refund status. That's the segment AI should be eating first. It's also the segment vendors love to report deflection numbers against, because it's the easiest to look good on.

The problem is that deflection and resolution are not the same metric, and conflating them is how a lot of AI customer support tools end up looking better in a sales deck than in a support queue. We've written about why the deflection metric hides more than it shows, but the short version: a closed chat isn't a solved problem if the user just gave up or filed an email ticket five minutes later.

What customer support metrics should you track?

  • Cost-to-serve per ticket. Work out what a human-handled ticket costs your team today, by ticket type. That is the baseline any AI support investment needs to beat, not just match.
  • First-contact resolution (FCR). Compare the AI's first contact resolution with your human team's on the same ticket types. An AI that can't get close on its claimed scope is just adding a slower first step.
  • Deflection rate vs. resolution rate. Vendors report the first because it's bigger. Ask for the second, specifically whether deflected users reopened a ticket, emailed, or churned within a week.
  • Repeat-contact rate. If the same account files three tickets about the same stuck point in a month, that's not a support problem, it's a product problem the support tool should be surfacing, not just answering around.
  • Hand-off quality. When AI can't resolve something, does the human agent get full context, or do they start from zero? This single number predicts more about customer sentiment than resolution rate does.
  • Adoption depth, not adoption breadth. "We use an AI tool" means nothing. What share of eligible tickets actually get handled end-to-end, with no human touch, is the number that reflects real capability.

That hand-off point deserves its own paragraph because it's where most AI customer support tools quietly fail. A chatbot that can't solve a problem and dumps a transcript on a human agent isn't escalating, it's forwarding. Real hand-off means the agent gets the session, the account state, and what was already tried, which is the whole idea behind hand-off with full session context rather than a bare transcript.

How should these metrics change how you buy AI support?

The honest framing is that AI customer support tools are good at a specific, large, and growing slice of ticket volume: the repetitive, well-documented, low-ambiguity stuff. They're bad, right now, at anything requiring judgment, account-specific nuance, or an action the tool wasn't explicitly built to take. The gap between those two categories is also where most of the cost-to-serve actually lives, because the easy tickets were already cheap.

This is the gap we built Peeve around. Peeve is an AI customer support agent that resolves each request in one of four ways: it answers from your docs, shows the user where to click, takes an account action with the user's confirmation on Growth plans and above, or hands off to a human with the full session attached. It is part of a wider shift in AI customer support from answering questions to resolving them. See how Peeve works and compare plans.

If you're comparing vendors, the resolution and hand-off quality numbers above are a better diligence checklist than any adoption statistic a sales rep quotes you.

FAQ

What's a realistic resolution rate for AI customer support?

For well-scoped, repetitive ticket types, a capable AI customer support tool can resolve a meaningful majority without human involvement, but blended across all ticket types, resolution rates are usually much lower than vendor-reported deflection rates suggest. The gap between the two numbers is the most useful diligence question to ask any vendor.

Does AI customer support actually lower cost-to-serve?

It can, but only if it reduces human-handled ticket volume on the expensive, repetitive categories rather than just adding an extra step before a human still has to get involved. Compare cost-to-serve before and after on a per-ticket-type basis, not as a single blended average, since averages hide where the tool is and isn't working.

How is Peeve different from a typical AI support chatbot?

Peeve is an AI customer support agent, not just a chat window. Besides answering, it can show the user the exact element on screen, take a backend action with the user's confirmation on Growth plans and above, or hand off to a human with the full session attached. That action layer, described at /product/do/, is what turns deflection into resolution.