Guide

What Is a Good AI Deflection Rate?

The number your vendor quotes and the number that actually saves you headcount are rarely the same, and the gap is where support budgets quietly bleed.

The short answer

A good AI deflection rate is whatever your own re-contact-adjusted number is, measured the same way every month, trending up. That is not a dodge. It is the only benchmark that survives contact with reality.

The reported figure you will see quoted by vendors and dashboards usually lands somewhere between 40 and 70 percent. Treat that as a volume metric, not a value metric. It counts conversations the bot handled, which is not the same as conversations it resolved. Once you strip out the tickets where the customer came back within a few days, true deflection commonly sits 15 to 25 points lower. A headline 60 percent can be a real 38. These figures are illustrative, but the size of that gap is the whole story, and it is the part nobody puts on a slide.

Reported deflection is a volume metric in disguise

Here is what most tools actually count when they say deflection. A customer opens a chat, the AI responds, and the session ends without a human ever touching it. That counts as a deflection. The customer abandoning the chat in frustration counts. The customer getting a confident, wrong answer and giving up counts. The customer who got helped and the customer who rage-quit are scored identically, because the metric only knows that no agent was assigned.

That is why the number is so easy to inflate and so easy to love. It goes up when you route more traffic to the bot, regardless of whether the bot is good. It is a measure of containment, which is an operational fact about where the ticket went, and it gets sold as a measure of resolution, which is a claim about whether the customer's problem is gone. Those are different things, and conflating them is the single most common mistake I see support leaders make when they evaluate an AI agent.

If a vendor quotes you a deflection rate without telling you how they handle re-contacts, abandons, and escalations, you do not have a benchmark. You have a marketing number.

True deflection: closed and no re-contact

The honest definition is narrow on purpose. A conversation is truly deflected when the AI closed it and the same customer did not come back about the same issue inside a defined window. Pick a window that fits your product and hold it constant. For most B2B SaaS, three to seven days catches the bulk of the boomerangs.

The re-contact filter is the part that hurts, and it is the part that matters. When a customer gets a non-answer from the bot and reopens the next morning, that ticket got counted as deflected on day one and is now also a fresh human ticket on day two. You paid for the same problem twice and your dashboard gave you credit for solving it once. Multiply that across a quarter and the reported savings and the actual savings diverge fast.

You also want to watch the silent failures, the customers who never re-contact because they churned, downgraded, or just decided support was not worth the effort. Those never show up as a reopened ticket, so true deflection alone will flatter you there too. Pair the number with a periodic CSAT or thumbs-rating on AI-handled conversations so a falling satisfaction score can flag deflection that is technically real but practically a loss.

What the benchmark ranges actually mean

When you see 40 to 70 percent reported, read it as a band that depends almost entirely on ticket mix, not on how good anyone's AI is. A product with a heavy tail of password resets, plan questions, and how-do-I-find-this requests will post high reported deflection because those questions are genuinely easy. A product whose inbound is mostly account-specific debugging, billing disputes, and integration failures will post lower numbers no matter what tool sits on top, because those tickets need data, judgment, or a human with permissions.

So a peer's 65 percent and your 45 percent may both be excellent or both be mediocre. You cannot tell without the ticket mix, the re-contact rate, and the definition behind each figure. Comparing your headline deflection to someone else's is comparing two numbers that were calculated differently on two different populations. It tells you nothing actionable.

The useful comparison is internal. Take your re-contact-adjusted rate this month against the same measure last month, segmented by ticket type. That tells you whether the AI is getting better at the work you actually have, which is the only question with a budget attached.

How to measure your real number

Five steps, in order. First, define a deflection as AI-closed with no human touch, and log it. Second, define the re-contact window and subtract every conversation where the same customer reopened the same issue inside it. That difference is your true deflection rate. Third, segment by ticket type, because a blended average hides the categories where the AI is quietly failing and the ones where it is carrying you.

Fourth, layer a quality signal on top, a CSAT or thumbs rating on AI-resolved conversations, so you catch the confident-but-wrong answers that re-contact filtering misses. Fifth, track all of it as a trend line, not a snapshot. A true deflection rate of 35 percent climbing three points a quarter is a far healthier asset than a reported 60 percent that has been flat since launch and is propped up by abandons.

Run it this way and the vendor headline stops mattering. You will have your own number, defined honestly, and you will know whether it is going the right direction. That is the benchmark.

The takeaway

Stop asking what a good deflection rate is in the abstract. The reported rate is a containment metric that rewards routing volume, not solving problems. The number that maps to real headcount and real customer outcomes is true deflection: closed and no re-contact, segmented, quality-checked, and tracked over time. Expect it to come in well below the figure on the dashboard, and judge it against your own past, not a vendor's case study.

If you want to see the gap between your reported and true deflection without rebuilding your analytics stack, that is exactly what Throughscan's AI deflection audit is built to surface. It is a useful next step once you have decided the headline number is not good enough to plan a budget on.