By Michael Wei, founder, Aissist.io
Every AI customer service vendor now leads with a percentage. Sixty per cent automated. Eighty per cent deflected. Half your tickets, gone. Business leaders are being asked to sign contracts on the strength of these numbers, and almost none of them are measuring the same thing.
That is not a rounding problem. It is the difference between a programme that pays for itself and one that quietly moves cost from the support queue to the escalation queue.
Deflection is not resolution
The most common number quoted is deflection: the share of customers who opened a chat window and did not go on to create a ticket. It is easy to measure, it flatters everyone, and it counts the customer who gave up.
A resolution is something stricter. The definition worth holding vendors to is this: the issue the customer brought was finished end to end by the AI, with no human agent completing it, and no follow-up contact on the same issue. That last clause is where most published numbers quietly fail. A bot that answers a refund question, and is then contacted again two days later about the same refund, has deflected one conversation and resolved nothing.
Apply that stricter definition and the numbers come down, but they start to become comparable. In our own 2026 benchmark, published with the methodology attached, the median for first-tier automation sits at roughly 41 per cent, with a top quartile near 59 per cent. Less exciting than most vendor homepages. Also the figure a CFO can plan against.
The range matters more than the average
The more useful finding is how wide the spread is between industries, and how little of it is explained by the quality of the AI.
Ecommerce and retail run at 70 to 84 per cent. Consumer fintech sits at 60 to 75 per cent. SaaS lands at 50 to 70. Travel and hospitality, 45 to 70. Telecom, utilities, healthcare and insurance cluster at 40 to 60 per cent.
The same underlying models are doing the work in all of these. What differs is the shape of the request. “Where is my order” is a lookup against a system of record with one correct answer. “Why was this claim denied” is a judgement that depends on policy language, regulatory constraint and often on information the customer has not supplied yet. No amount of model capability collapses the second question into the first.
So a vendor’s reference case from another industry tells you very little. A 78 per cent figure earned in retail is not a forecast for an insurance book. Ask for the number in your category, on your ticket mix.
Watch the satisfaction gap, not just the rate
Resolution rate on its own can be gamed by letting the AI hold on to conversations it should hand over. The check against that is customer satisfaction, measured separately for AI-handled and human-handled contacts within the same team.
Across industries, AI-handled interactions typically score five to ten points below human-handled ones in the same organisation. A gap in that band is normal. A gap much wider signals automation reaching past its competence, and a resolution number bought with goodwill.
The pairing is what makes either number meaningful. Resolution rate tells you how much work moved; the satisfaction gap tells you what it cost in the relationship.
The economics are thinner than the headline suggests
Unit prices look remarkably cheap: about five to fifteen cents per interaction, and fifty cents to a dollar fifty per resolution on self-serve tiers.
All-in cost is a different figure. Once integration, knowledge preparation and the engineering time to keep it correct are amortised, a realistic planning number is closer to five dollars per AI resolution. That is still a large saving in most categories — but the business case rests entirely on the resolution number being real. At 40 per cent genuine resolution the maths works. At a 70 per cent deflection figure that is really 35 per cent resolution with a re-contact tail, it does not.
Four questions worth asking
Before signing anything, ask a vendor for these, in writing:
- How do you define a resolution, and does it exclude re-contacts on the same issue within seven days?
- What is the resolution rate in my industry, on a ticket mix like mine?
- What is the CSAT gap between AI-handled and human-handled contacts in your reference accounts?
- What is the all-in cost per resolution in year one, including integration, not the unit price?
A vendor who answers all four plainly is worth talking to, whatever the number turns out to be. One who cannot is selling a percentage, not an outcome.
The transformation is real. It is just smaller, slower and more industry-specific than the marketing suggests — and the companies getting the most from it are the ones that insisted on measuring it honestly before they scaled it.
