AI Ticket Deflection Explained - Rates, Claims and What They Count (2026)

The keyword adjacent to this one, how do AI agents help in customer support, carries a cost per click of $65.21. Somebody is paying sixty-five dollars for a single visitor asking a general question about support automation.

That tells you more about this category than any vendor page will. The margins are large enough that a $65 click makes sense, which means the claims on the pages you land on are worth reading carefully.

Five vendors, five numbers, no shared definition

Here is what the major platforms currently claim about how much volume they take off your queue.

That is a threefold spread for what sounds like the same capability. The explanation is not that MavenAGI is three times better than Intercom. It is that none of these numbers measure the same event, and only Freshworks bothers to qualify what it is counting.

Three mechanisms, only one of which is unambiguously good

Ticket volume goes down for three quite different reasons, and vendors report all three as the same win.

• The AI resolved it. Customer had a problem, got a correct answer, went away satisfied. This is the one you want.
• The customer self-served. They found the answer without opening a ticket at all. Also good, and frequently uncounted.
• The customer gave up. They tried, got nowhere, and did not escalate. This reduces ticket volume identically and is a failure.

Deflection metrics cannot distinguish the first from the third. Neither can most dashboards. A support queue that halved because customers stopped bothering looks exactly like a support queue that halved because the AI worked.

Where the volume actually goes

In practice, the reductions come from a narrow band of ticket types, and knowing which ones lets you predict your own result better than any vendor benchmark.

• Order status. The highest-volume ticket type in retail and almost fully automatable, provided the agent can query the order system rather than only the help centre.
• Password and access issues. High volume, low complexity, but the exceptions are exactly what reaches support, so expect less than the headline.
• Policy questions. Returns windows, shipping times, what is covered. Genuinely solved by documentation retrieval.
• Onboarding questions. Predictable, repetitive, well documented, and expensive to leave unanswered.

If your queue is mostly these, a high deflection figure is plausible. If your queue is mostly technical debugging or account-specific disputes, treat any headline number as marketing regardless of who published it.

The cost inversion nobody mentions

Reducing ticket volume is supposed to save money. On per-resolution pricing, every ticket the AI removes from your human queue is a ticket you are billed for.

At $0.99 per resolution, cutting 10,000 tickets from your queue costs $9,900. The saving is real only if that exceeds what those tickets cost you in human time, and at high volume, frequently it does not. This is the single most important piece of arithmetic in the category and it appears on no vendor page.

What to ask

• What event triggers your headline number? Answered, deflected, or confirmed resolved?
• What is the figure for resolution confirmed by the customer rather than inferred from silence?
• How is an abandoned conversation counted — success, failure, or excluded?
• What does this look like on a ticket mix like mine, not on your best customer's?

Every percentage on this page is a vendor claim, quoted as such. We have not run our own ticket set through any of these platforms, so we cannot tell you which figures survive contact with real traffic. That testing is scheduled.

The numbers, and what they are measured on

Published deflection figures cluster far above what first-year deployments achieve, and the gap is definitional rather than dishonest.

That bottom row is the one to plan against. Gartner's own survey puts full self-service resolution at 14% across all issues and 36% on the ones customers call very simple, with the top failure cause being content people could not find — so the teams currently deploying are underperforming a business case built on the top rows.

Volume does not vanish, it moves

Three destinations, and only one is a genuine reduction.

• Genuinely resolved. The customer got what they needed and did not come back. A real reduction.
• Displaced. Deflected from chat, arrives by email on Thursday. Your chat metrics improve and your total does not.
• Abandoned. The customer gave up. Counted as deflection, and it is the one that shows up later as churn rather than as a ticket.

The only way to tell them apart is to track total contacts across every channel, not per-channel deflection. Teams measuring one channel routinely report success while their aggregate volume is flat.

What actually reduces volume

Ranked by effect, and the software is third.

• Fixing the product problem generating the tickets. If 12% of contacts trace to one confusing checkout step, changing the step removes them permanently.
• Rewriting the twenty most-viewed help articles. Gartner surveyed 5,728 customers in December 2023 and found the most common reason self-service fails is that people cannot find content relevant to their issue, in 43% of cases.
• Deploying an agent against the documented remainder.
• Buying a more expensive agent.

Most teams start at four and work upward. The order matters because the first two are cheap, permanent, and make the third work better.

Where the documentation ends and the software begins

A distinction worth making because vendors have no reason to make it. Published benchmarks put deflection from a well-maintained knowledge base alone at 30-50%. Layering AI on top takes it to 40-70%.

Read those two ranges together and the AI is contributing perhaps ten to twenty points on top of documentation that was already working. That is real and it is not what a 70% headline implies you are buying.

It also explains why deployments on stale documentation disappoint. There is no baseline for the software to add to, and the vendor's benchmark was measured on companies that had done the unglamorous half first.

The two-tier pattern, and the number it actually produces

The configuration most teams end up with: the agent answers when its confidence exceeds a high threshold, and routes everything else to a human without attempting it.

Published figures put that at 40-55% overall deflection. Lower than the marketing, higher than a first-year deployment, and defensible in a review — which is usually the number you want to have promised.

The threshold is the lever, and it moves two things at once. Raise it and quality goes up, deflection goes down, and under per-resolution billing your invoice goes down with it. Lower it and all three reverse. Nobody tells you the confidence threshold is a pricing control, but it is.

The bit worth arguing with

AI agents genuinely reduce ticket volume, mostly on order status, access issues and policy questions. The published figures between 30% and 93% are not comparable because they measure different events, and the highest ones are usually counting the loosest definition. Ask what triggers the number, ask for confirmed resolution, and check what removing those tickets costs you before treating the reduction as a saving.

WRITTEN BY AR · UPDATED 2026-07-28

I read the fine print. Vendor pricing pages, billing definitions, terms, funding filings and acquisition notices — then I do the arithmetic nobody publishes: what a platform actually costs at your volume, what its headline metric is really counting, and who owns it now. I do not run benchmarks, and no page here pretends otherwise.

Editorial policy