Deflection rate counts the conversations that ended without a human. Resolution rate counts the conversations where the customer's problem was actually solved. They sound like the same thing. They are not, and the gap between them is where AI support quietly goes wrong. This is how to read deflection rate vs resolution rate together, so one number cannot hide behind the other.
We have argued before that deflection rate is the wrong metric to lead with. This piece is the companion. Not "stop measuring deflection" but "never measure it alone."
What is the difference between deflection rate and resolution rate?
Both metrics start from the same pool of AI conversations. They count different things about them.
Deflection rate is the share of conversations that ended without a human agent getting involved. The AI replied, the customer did not reach a person, the conversation closed. That is a deflection. Notice what the definition does not mention: whether the customer's problem was solved. Deflection is about the absence of a human, nothing more.
Resolution rate is the share of conversations where the customer's actual problem was fixed. The refund was issued. The address was changed. The order status was confirmed against live data and the customer left with the answer. Resolution is about the presence of an outcome.
A conversation can be deflected and unresolved at the same time. The customer asked a question, got a vague answer, shrugged, and closed the chat. No human was involved, so it counts as deflected. Nothing was solved, so it does not count as resolved. That conversation is the whole problem in miniature.
Why deflection rate alone misleads
Deflection rate has one quality that makes it dangerous as a headline number. It is easy to move in the wrong direction.
You can raise deflection rate without helping a single extra customer. Make the AI answer more questions it should escalate. Hide the request-a-human option one click deeper. Count a conversation as deflected if the customer does not write back within 48 hours, which also counts everyone who gave up. Every one of those moves makes the deflection chart go up. None of them solves anything.
Resolution rate resists all of that. You cannot fake a refund into a customer's account. You cannot fake an order that actually shipped. To move resolution rate you have to do the real work, which is why it is the harder, more honest number, and why far fewer dashboards lead with it.
Here is the trap in one line. A support tool can show you a 90% deflection rate and a 45% resolution rate at the same time. The first number says automation is working. The second says less than half your customers got helped. Same conversations, two stories.
Deflection rate vs resolution rate: how to read them together
The numbers are useful as a pair. Apart, each one lies a little.
| | Deflection rate | Resolution rate | |---|---|---| | Counts | Conversations with no human | Conversations with a fixed problem | | Easy to inflate? | Yes, in several ways | No, the outcome is real or it is not | | What it tells you | How much volume the AI absorbed | How much volume the AI actually closed | | Fails by | Hiding give-ups as successes | Being harder to measure | | Read alone | Misleading | Incomplete but honest |
The signal is the gap between them. If deflection is 80% and resolution is 75%, the AI is doing close to real work, and the 5-point gap is conversations that ended without a human but without a clear fix. Worth sampling, not worth panicking over. If deflection is 80% and resolution is 40%, you do not have an automation win. You have a backlog of unhappy customers who have not contacted you yet. They will.
Track the gap over time. If deflection climbs while resolution stays flat, someone has tuned the AI to answer more and help the same. That is the exact failure the dashboard is designed not to show you.
How to measure resolution rate honestly
Resolution rate is only useful if you measure it strictly. A loose definition turns it back into deflection with a nicer name.
Three rules keep it honest.
Tie it to an outcome, not a reply. A resolved conversation has a verifiable result. An action was taken in a connected system, or the customer confirmed they got what they needed. "The AI sent a plausible answer" is not resolution.
Check for the customer coming back. A conversation that looks resolved but generates a re-contact about the same issue within 48 hours was not resolved. It was postponed. Fold re-contact into the definition or resolution rate drifts upward on its own.
Split it by topic. A blended resolution rate hides which work the AI is good at. Order status, store policy, and account questions usually resolve cleanly. Complex returns and exceptions usually do not. Reading resolution by topic, the way our glossary entry on deflection rate describes, tells you what to automate further and what to route straight to a person.
Resolution rate also depends on whether the AI can act, not just talk. An AI agent that can read live data and complete a task in a connected system will resolve things a text-only bot can only describe. Grounding the answers in your real systems, which our grounding glossary entry covers, is what turns a confident reply into an actual resolution.
What good looks like
There is no universal target, because resolution rate depends on what mix of questions you point the AI at. A team that sends only order-status questions to the AI should expect high resolution. A team that sends everything, including complex disputes, will see lower resolution, and that can still be correct.
So judge the pair, not the absolute number. Healthy looks like this: deflection and resolution move together, the gap between them is small and stable, and resolution holds up when you slice it by topic. CSAT for AI-resolved conversations sits close to CSAT for human-resolved ones, which our CSAT glossary entry explains how to compare fairly. Unhealthy looks like a wide, growing gap and a resolution rate that collapses the moment you break it out by topic.
The ambition is real. Gartner predicts that by 2029 agentic AI will autonomously resolve 80% of common customer service issues without human intervention. Note the verb. Gartner said resolve, not deflect. The forecast everyone quotes is a resolution forecast, which is the right one to be held to.
How Keloa approaches this
We report both numbers, and we show the gap between them on purpose. Keloa's AI agents resolve by acting. They read live order and account data and complete the task, so a resolution is an event that happened, not a sentence that sounded right. When the AI cannot resolve something, it hands off rather than padding the deflection number with a guess.
We would rather show a lower deflection rate next to an honest resolution rate than a high deflection rate that falls apart under one question. Our pricing is billed per reply, not per claimed resolution, so we have no reason to inflate either number. The metric we want you watching is the gap, because the gap is the truth.
Frequently asked questions
What is the difference between deflection rate and resolution rate? Deflection rate counts conversations that ended without a human. Resolution rate counts conversations where the customer's problem was actually solved. A conversation can be deflected without being resolved, which is why the two numbers can tell opposite stories.
Is a high deflection rate good? Only if resolution rate is high alongside it. A high deflection rate on its own can mean the AI is answering well, or it can mean customers are giving up and not coming back. The deflection number cannot tell you which, so it is unsafe to read alone.
Why is resolution rate harder to measure? Because it requires a verifiable outcome. You have to confirm an action was taken or the customer got what they needed, and check they did not re-contact about the same issue. Deflection only requires noticing that no human joined, which is much easier to compute.
What is a good resolution rate? There is no single target. It depends on which questions you route to the AI. A narrow set of factual questions resolves high, a broad set including complex disputes resolves lower, and both can be correct. Judge resolution next to deflection and by topic, not as one headline figure.
Should I stop tracking deflection rate? No. Track it, but never alone. Deflection rate next to resolution rate is genuinely useful: the gap between them shows how much "deflected" volume left without a fix. Deflection rate by itself is the number that lets a problem hide.