A Reality Check on AI Deflection Rates

Vendor demos imply that most of your queue is automatable. Reported figures in 2026 put the median deflection considerably lower, somewhere around a fifth of ticket volume. That gap is worth understanding before you build a business case.

Why the gap exists

Selection. Vendors quote their best deployments. Those are real, and they are not the median.

Definition. Deflection counts avoided contacts, including customers who gave up. The honest number, genuine resolutions where the customer got an answer and did not come back, is lower than the deflection figure for the same deployment.

Content. The deployments that hit high numbers usually had good documentation before they started. The technology did not create that, and a vendor cannot sell it to you.

Topic mix. This is the big one. A business whose queue is dominated by order status and password resets has a completely different ceiling from one handling bespoke technical faults. Neither is doing anything wrong.

Your ceiling is your topic mix

The number that matters is not an industry median. It is what share of your own queue is documented and repetitive.

Measure it like this. Take a thousand real tickets, group them by topic, and mark each topic one of three ways:

• Documented. A current, clear article answers it today.
• Documentable. It could be documented but is not yet.
• Judgement. It needs a person: complaints, exceptions, anything involving a decision about money or a relationship.

Sum the volume in each column. The documented column is your realistic ceiling right now. The documentable column is what content work would add, and it is usually larger than people expect. The judgement column is permanent.

That exercise takes an afternoon and it is worth more than any benchmark, because it is about your business rather than an average of businesses unlike yours.

What the three columns usually look like

Patterns we see described repeatedly, which are not a substitute for measuring your own:

Businesses with high documented-plus-documentable shares tend to be e-commerce, subscription media, and self-serve software, where order status, access, billing policy and how-to questions dominate.

Businesses with large judgement columns tend to be B2B services, regulated industries, and complex technical products, where most contacts involve a decision or an investigation rather than a lookup.

If you are in the second group, automation will help less and you should know that before a sales process rather than after one.

The number to use in the business case

Not deflection. Use cost per resolution, with reopen rate and satisfaction as guardrails.

Deflection can be improved by hiding the human option, which raises the metric and damages the business. Cost per resolution paired with reopen rate cannot be improved that way, which is precisely why it is the number worth committing to in front of a finance team.

Build the case on the documented column of your own analysis, priced at the vendor’s published rate, compared against your current fully loaded cost for that same volume. That number is defensible. A vendor’s median is not.

Setting expectations internally

The most common way these projects fail politically is an overstated first projection.

Project from your documented column, not your documented-plus-documentable column, because the second requires content work that may not happen. Deliver against that, then expand as the content improves. A programme that beats a conservative projection gets more budget. One that misses an ambitious projection gets cancelled while working.

We wrote up the queue analysis method, what AI actually resolves, and why deflection is the wrong headline metric.