AI Support Metrics: Why Your Deflection Rate Is the Number to Stop Trusting

Your AI support agent deflected 80% of tickets last quarter. The dashboard was a wall of green, someone dropped it into the board deck, and nobody in the room asked the one question that actually matters: did those thousands of people get what they came for, or did they just give up?

That gap, between a customer who was helped and a customer who went away, is the whole problem with how most teams measure AI support right now.

The number that flatters everyone

Deflection rate is the headline metric for almost every AI support tool on the market. It counts the share of conversations that never reached a human. You can see why it caught on. It is easy to calculate, it always trends in the direction you want, and it maps neatly onto the thing finance cares about, which is fewer tickets landing on a paid support team.

But look at what it actually measures... Most tools define a deflected ticket as any conversation that did not get escalated to a person. So the customer who got a perfect answer and the customer who rage-quit the chat window and emailed your CEO instead are both, as far as deflection rate is concerned, a win. One number, two completely different human experiences, and it cannot tell them apart.

Deflection rate is the most flattering number in support, and that is exactly why I do not trust it. It measures the thing you want to be true. You want fewer tickets, so you track the metric that goes up when tickets go down, and then you stop looking. The problem is that a customer can be deflected and furious at the same time. The bot solved nothing. It just wore them down until they closed the tab. On your dashboard that reads identically to a brilliant answer. If a metric cannot tell the difference between helping someone and losing them, it should not be the number you put on the wall. Ian Naylor, Founder of SaaSToolkit.ai

Deflection is not resolution

The distinction that gets blurred is simple to say and easy to forget. Deflection means a human did not get involved - Resolution means the customer's problem actually got solved. They sound like the same thing, and they are not, and the distance between them is where your support reputation lives.

Picture the same thousand conversations measured two ways. A tool reporting deflection might tell you 80% never reached an agent. A tool reporting genuine resolution might tell you that only around half of those people actually walked away with a working answer. Same conversations, two very different stories, and only one of them is honest about whether your customers are okay.

Even the vendors selling this software are starting to admit the gap. By most industry estimates a mature AI agent resolves somewhere between 50% and 80% of routine, repetitive tickets end to end, and that is the easy, well-documented stuff. The messy, specific, angry tickets, the ones that actually damage a relationship when they go wrong, are a different matter entirely. Deflection rate flattens all of that into one cheerful percentage and hands it to you like good news.

Customers can tell the difference

People are not fooled by a bot that is good at dodging them. Gartner found that 64% of customers would prefer companies did not use AI for customer service at all, and a large share said they would consider switching to a competitor if they knew a business was leaning on AI in support. Sit with that for a second. The majority would rather you did not. Not because AI can never help, but because so much of it has been deployed to deflect rather than to resolve, and people feel that difference the moment it happens to them.

That is the real cost of optimising for deflection. You can hit your number and lose the customer in the same motion, and your metrics will congratulate you both times.

Deflection rate is a cost metric wearing a customer-experience badge," says Hannah Okonkwo, a customer experience consultant. "It was invented to show the finance team that automation was paying for itself, and for that one job it is perfectly fine. The mistake is letting it leak into how you judge whether support is any good. Those are different questions. Cheaper and better sit on different axes, and the moment a support team starts chasing deflection as a quality goal, the quality starts slipping and nobody notices, because the only number they are watching goes up either way.

What to measure instead

So what do you track? Not one number. A handful that, taken together, tell you whether your AI support is helping people or just hiding them from you.

• Resolution rate. The share of conversations where the issue was genuinely solved, confirmed by asking the customer where you can, not assumed from the absence of a follow-up. This is the number deflection rate has been pretending to be all along.

• Escalation quality, not just escalation rate. A handover to a human is not a failure. A clumsy, late, context-free handover is. When the bot gives up, does the agent inherit the full conversation, or does the customer start again from scratch?

• Repeat contact rate. If someone comes back a day later about the same thing, the first conversation resolved nothing, whatever the bot logged. High deflection sitting on top of high repeat contacts is a tool creating work and hiding the evidence.

• Customer effort. How hard did the person have to work to get helped? Effort is one of the best predictors of churn there is, better than satisfaction scores, and a bot built to deflect tends to push effort straight up.

• Post-resolution CSAT, split by path. Ask for a rating after AI-handled conversations specifically and compare it to the human-handled ones. If the two are close, good. If there is a gap, your deflection rate was busy hiding it.

When a team shows me a 90% deflection rate, my first question is always the same: deflected to where? Deflected to a solved problem, or deflected to a frustrated customer who gave up and told three colleagues your support is useless? The number cannot tell me, so I make them go and look. Read fifty deflected conversations end to end, and you will usually know within the first twenty whether that 90% is something to celebrate or something to panic about. Becky Halls, Strategist at SaaSToolkit.

Why the flattering version keeps winning

There is a practical reason deflection rate hangs around when half the people reporting it already know it is flawed. It is simply the number that is easy to get. Your support tool counts the conversations that did not escalate, and there it is, ready for the slide. Measuring real resolution is harder, because it means connecting what the customer did inside the conversation to what they did next, in the product, over the following days. Did they come back the next morning about the same thing, or did they finally finish what they were stuck on?

That is behaviour data, and it usually lives in a different system from your support logs. So the honest version of the metric is the one nobody has time to assemble, and the flattering version wins by default.

Closing that gap means putting your support conversations and your product behaviour in the same place, so a resolved ticket becomes something you can verify by what the customer actually did afterward, rather than something your bot marked as closed and moved on from. That is the difference between measuring whether a human was avoided and measuring whether a person was helped.

Deflection rate is not useless - As a cost measure it does its job. Just stop letting it answer a question it was never built to answer. Your customers already know the difference between being helped and being handled. Your AI support metrics should know it too!

FAQ

What is deflection rate in AI support? It is the share of support conversations that an AI agent or self-service tool handled without the customer reaching a human. Most tools count any conversation that did not get escalated as deflected. It is popular because it is easy to measure and it maps to cost savings, but it says nothing about whether the customer's problem was actually solved.

Is deflection rate the same as resolution rate? No, and treating them as the same is the core mistake. Deflection means a human was not involved. Resolution means the issue was genuinely fixed. A customer who gave up in frustration counts as deflected but not resolved, so you can have a high deflection rate sitting on top of a low resolution rate and never see the difference from the dashboard alone.

What AI support metrics should I track instead? Track a small set together: true resolution rate, escalation quality (whether handovers carry context), repeat contact rate, customer effort, and post-resolution CSAT split by whether AI or a human handled the conversation. No single number tells the whole story, but together they show whether your AI is helping people or just hiding them.

Does AI customer support actually hurt satisfaction? It can, when it is deployed to deflect rather than to resolve. Gartner found 64% of customers would prefer companies did not use AI for customer service at all. That does not mean AI support is doomed. It means a lot of it has been pointed at the wrong goal, and customers notice when a bot is built to avoid them rather than help them.

How do I know if my AI support is really resolving issues? Stop trusting the closed-conversation count and look at what happened next. Did the customer come back about the same thing? Did they complete the task they were stuck on? Did they rate the conversation well? The honest signal lives in behaviour after the chat ends, not in whether the bot marked the ticket as handled.

How does SaaSToolkit help with AI support metrics? SaaSToolkit captures product behaviour and support interactions from one JavaScript snippet, so a resolved ticket is one you can verify by what the customer did next, in the product, rather than one your bot simply closed. Support and behaviour sit in the same place, which is the only way the honest version of resolution rate ever actually gets measured.