What is a good AI resolution rate in 2026

Resolution rate is the number every AI support vendor now leads with. Fin, Sierra, Lorikeet, and most of the field put a big percentage on the page and let you assume higher is better. It usually is not that simple. The way that number is counted matters far more than the number itself, and a high rate from a bot that guesses is worse for your business than a lower rate from a bot that is honest about what it cannot do.

What resolution rate actually means

In plain terms, resolution rate is the percentage of customer conversations your AI handles from start to finish with no human agent involved. If the AI touched 1,000 chats last month and 700 of them ended with the customer helped and no agent stepping in, that is a 70 percent resolution rate.

The idea is sound. It is a proxy for how much work the AI takes off your team and how much you can grow support volume without growing headcount. The trouble starts with the word resolved. Every vendor defines it a little differently, and the loosest definitions produce the biggest, most marketable numbers.

How vendors inflate the number

Be fair to the leaders here. Fin, Sierra, and Lorikeet have built genuinely capable models, and a well tuned bot on any of them can carry real load. The problem is not the tech, it is the counting. Here are the common ways a resolution rate gets padded:

• Counting deflections as resolutions. If a customer opens a chat, reads an article the bot suggested, and never replies, some tools mark that as resolved. Maybe they got their answer. Maybe they gave up and went to a competitor. Silence is not success.
• Counting unresolved but closed tickets. A conversation that times out, or that the customer abandons in frustration, gets auto closed and quietly folded into the resolved bucket.
• Counting confident wrong answers. The bot gives a clean, assured answer, the customer accepts it, and the conversation ends. It only surfaces as a problem later when the order is wrong or the refund never comes. On paper it was a resolution.
• Excluding the hard cases. Some reports only count conversations the AI was allowed to attempt, so anything routed straight to a human never drags the average down.

None of this is necessarily dishonest by the vendor. It is just that the metric rewards volume closed, not customers helped, and those two things come apart fast.

Why a high rate can be worse than a lower one

Here is the uncomfortable part. A bot that guesses will always score a higher resolution rate than a bot that admits uncertainty, because guessing closes the conversation and admitting uncertainty opens a handoff. If you optimize purely for the number, you are training your AI to bluff.

Picture two bots on the same 1,000 chats:

1. Bot A answers everything, never says it is unsure, and posts an 85 percent resolution rate. Twelve percent of those answers are quietly wrong. Those customers come back angry, or worse, they do not come back at all.
2. Bot B answers what it knows, hands the rest to a human with full context, and posts a 68 percent resolution rate. Almost none of its answers are wrong, and the handoffs feel seamless.

Bot A looks better in the sales deck. Bot B is the one that keeps your customers and protects your brand. The higher number bought you nothing except deferred damage. This is why we wrote the AI that refuses to lie in the first place.

What a genuinely good rate looks like

A trustworthy resolution rate is one where the definition is strict and you can verify it. Look for these properties before you believe any number:

• Resolved means confirmed helped, not merely closed or gone quiet. The best signal is the customer saying thanks or completing the action, not the absence of a reply.
• Every conversation is in the denominator, including the ones the AI hands off. No cherry picking the easy set.
• The handoff rate is treated as healthy, not shameful. A clean pass to a human with full context is a good outcome, not a failure to hide.
• Accuracy is tracked alongside it. Resolution rate without a wrong answer rate next to it tells you nothing.

Measured honestly, a resolution rate in the 60s or 70s for a broad support queue is genuinely strong. A claimed 90 plus usually means either a very narrow use case or a very generous definition. Ask which one it is.

How ZenTalk counts a resolution

ZenTalk answers from your own knowledge, not the open internet, and it is built to only claim an action it truly performed. If it can quote your real price and take payment in the chat, it does. If it cannot help, it says so and hands to a human with the full thread, rather than inventing an answer to close the ticket.

That means our resolutions are real ones. The bot would rather send a clean handoff than fake a win, so the number you see reflects customers who actually got what they came for. It costs us a few points on the headline metric and earns back trust that a bluffing bot quietly burns. You can see how the knowledge grounding works on our AI knowledge page, and if you are still shopping the field, read our honest comparison of the 2026 tools.