Deflection rate counts how many tickets your AI closed without passing them to a human. It does not measure whether the customer's problem was solved. Those are two different things, and the gap between them is where this number misleads you.
Here is the gap in practice. On Tuesday, two customers ask the same question in chat. The AI answers both. The first customer is happy and closes the chat. The second customer is still stuck, so an hour later they send an email. The next morning the dashboard shows two closed chats, both marked as handled by AI. The deflection rate goes up by two. One of those customers is on the way out, and the dashboard cannot tell you which one.
Deflection rate is the number most support teams report first when they talk about AI. It is also the easiest number to read the wrong way. This article explains what it really measures, why a deflected customer often comes back through a different channel, and what to count instead.
What the deflection number really counts
The definition is narrow. Deflection rate counts interactions the AI closed on its own, with no human involved. That is all it counts. Whether the customer got what they needed is a separate question, and the number never asks it.
Michael Howlett, who writes the Customer Experience Decoded newsletter, calls this the AI metrics mirage: the contact centre numbers look great while customers are leaving. A bot that walks a customer in circles until they give up still counts as a deflection. So does a bot that genuinely solved the problem. The number treats the two the same.
You might expect a high satisfaction score to catch this. It usually does not. Howlett explains why. Customers tend to answer surveys after easy questions. The hard, upsetting cases either get passed to a human, where they count as human contacts, or the customer walks away without answering. The average looks calm because the worst cases are missing from it.
The deflected customer comes back through another door
Kevin Davis, who writes the KD Be Schemin newsletter, sums up the result in a title that says it all: your AI support agent closed the ticket, the customer left anyway. The question that matters for measurement is what the customer does next. Most of them do not disappear. They ask again somewhere else.
Call this cross-channel leakage. A customer is deflected in chat at 2pm, and the chat system records a clean deflection. At 3pm the same customer, still stuck, sends an email or picks up the phone. The email system records a brand-new contact. Two systems, two records, one problem that nobody solved.
This is normal behaviour. A person whose problem is still there will keep asking. The customer knows they were left stuck. The only part of the company that thinks the matter is closed is the dashboard.
Why your dashboard cannot see the leak
The reason is how the systems are built. Most contact centres measure each channel on its own. Chat reports its own numbers, phone reports its own, email reports its own. Each report is honest about its own channel. None of them follows a single customer across all three.
So when the customer comes back by email, the email team records it as fresh work. It can even make the email team look good: they solved an email on the first try. One channel's failure turns into another channel's success. Nobody is lying. The way the systems are wired does the lying for them.
This is also why adding a smarter bot to each channel does not fix it. A better chat bot closes more chats. It still loses the customer who was sent in circles and is now on the phone. The leak sits in the space between channels, and no single channel is in charge of that space.
What to measure instead
The fix is to follow the customer, not the channel. Count a deflection only when the AI closed a ticket and the same customer did not come back, in any channel, about the same problem, within a set window. Howlett suggests 24 to 72 hours. That is long enough to catch the answer that did not hold, and short enough to ignore unrelated new questions.
Measured this way, deflection rate becomes a count of problems actually solved. The number will drop. That drop is not bad news. It is the measurement finally telling the truth.
Mark Levy, who writes Decoding Customer Experience, asks a useful question: is your AI strategy improving the customer experience, or just making it cheaper to ask? Making it cheap to ask a question is not the same as giving an answer that holds. A deflection number that blurs the two will keep showing progress while the same problem keeps coming back through another door.
Where this leaves the deflection rate
The number still has a use. As a measure of workload, deflection rate tells you how much volume the AI absorbed, and that is worth knowing. The damage starts when a team reads it as a measure of how well the AI is doing.
A clear dashboard shows two numbers side by side: tickets the AI closed, and tickets the AI closed where the customer never came back. The first is the old number. The second tells you whether the customer is still a customer in three months. When the two numbers drift apart, that gap is the story worth your attention.
None of this is hidden behaviour. A customer whose problem is unsolved contacts you again. It looks invisible only because the second contact lands in a different channel from the first, and the deflection rate counted just the first.