What AI Support Deflection Actually Costs, and Saves

For two years, "AI is transforming customer support" was the kind of sentence you could put on any slide and nobody would check. It meant everything, which is to say it meant nothing.

That era is quietly over. The numbers have arrived, they're specific, and they're a lot more interesting than the hype ever was. So let's skip the adjectives and look at the receipts.

A 24x price tag on the same piece of work

Start with the one statistic everything else hangs from. An AI-handled support ticket costs between $0.50 and $1.05 (Gartner, 2025). A human-handled one costs $8 to $12 (Forrester, 2025). Same ticket, same customer, same problem solved. Up to 24 times the price depending on who picks it up.

Gaps like that don't sit politely in a spreadsheet. When one method of doing identical work costs an order of magnitude less, the market notices fast, and it did. In the field, deployments shave support costs by about 30% on average (IBM, 2025), while the sharpest operators, the top quartile, cut 53% (McKinsey, 2025). Gartner tallies the global prize at $80 billion in savings by 2027.

Hold onto that distance between "average" and "top quartile." It's the whole plot.

Deflection, in plain terms

The savings come from a single mechanism with an ugly name: deflection. It's just the share of customer contacts that get fully resolved without a human ever touching them. A password reset answered by a bot at 2 a.m. is deflected. A furious enterprise escalation is not, and shouldn't be.

The starting line is lower than people assume. In the technology industry, average deflection sits around 23% (Pylon). Teams that deploy AI well climb to 40 to 60%. The best in the business deflect up to 85% of routine questions.

What separates them isn't a smarter model. Everyone has access to roughly the same models. The difference is the knowledge base underneath, the boring library of answers the AI actually pulls from. Which is why these systems compound in a genuinely satisfying way. Early on, an AI might resolve 42 to 50% of what crosses its desk. Feed it real tickets for a few months and that creeps toward 70%. The thing gets better at the job by doing the job, like an employee who actually remembers what happened last week.

And customers, it turns out, are not heartbroken about this. 86% say they expect self-service, and 69% would rather solve a problem themselves than talk to a living soul. Done right, deflection isn't a cost cut people grudgingly tolerate. It's frequently the experience they'd have chosen.

The headline everyone quotes, read carefully

Here's the line that shows up on every conference slide: by 2029, Gartner expects agentic AI to autonomously resolve 80% of common service issues with no human in the loop, trimming operational costs 30%. Nearer term, they figure 80% of support organizations will be using generative AI in some form by the end of 2025.

The load-bearing word in that forecast is "common." Read past it and the strategy isn't "fire the support team." It's "let AI own the high-volume, well-understood middle of the bell curve, and route everything that needs a human's judgment to a human." Autonomous coverage plus exception handling. The interesting work moves up, not out.

The plot twist in the data

Now the number nobody puts on the celebratory slide: 61% of AI customer-service projects miss their year-one targets (McKinsey).

Sit with that against everything above. The technology plainly works. And yet most deployments fall short of what they set out to do. That's not a contradiction, it's the actual lesson, and it's hiding in that gap between the 30% average and the 53% top quartile.

The teams that win do two deeply unglamorous things. They measure a real baseline before they switch anything on, so they know what "better" even means. And they scope the AI to the issues it can actually resolve, instead of pointing it at everything and act surprised when it fumbles the hard cases.

So the next time you read that some company halved its support costs with AI, resist the urge to ask which tool they bought. Ask what they measured, and which slice of tickets they aimed it at. That's where the results actually live, and it's the least sexy and most important answer in the whole field.