What Gartner said
Three Gartner analysts, Emily Potosky, Patrick Quinlan and Brett Overbey, wrote in CX Dive on 5 November 2026 that the prices customer service teams pay for AI today are subsidised rates set to drive adoption, not the real cost of running the models. Their 2026 C-suite AI survey found that 42% of service leaders did not know what they were spending on AI, what they were getting back, or both. About a quarter reported a positive return, about a quarter a negative one, and 78% still planned to spend more in 2026. They point at GitHub Copilot, which moved from a flat per-user price to access plus consumption in June 2026 and raised some clients’ monthly bills by three to ten times, and at ChatGPT Enterprise, which moved to usage-based pricing in July.
The real problem is the unit, not the price
Is there a real need to worry? Yes. Vendors have been pricing below cost, and prices will rise. That part is true and worth planning for.
What has changed is smaller than the headline suggests. Usage pricing has always existed. Hosting, SMS and payment processing are all billed on use, and nobody calls them a crisis, because the unit is predictable and the cost per unit is small and falls over time. The trouble starts when the vendor picks a unit that grows with your success.
In customer service, that unit is the resolution. Intercom, Gorgias and Zendesk bill per resolution. The Gartner piece even recommends negotiating outcome-based pricing as a fix. I think that is the wrong direction, and I want to explain why with a real shop in mind.
Per resolution means the better it works, the more you pay
Take an online shop that gets a steady stream of “where’s my order” questions on WhatsApp. It switches on AI customer service. In month one, the AI resolves a modest share, and the bill is small. By month three, the shop has tuned its rules, connected its order data, and the AI resolves most of those questions without a person. The bill has gone up, because every one of those resolutions is charged.
Then the festive peak arrives. Enquiries double. The AI resolves them all, which is the whole point of having it. The bill doubles in the month when margins are already thin from discounts and shipping.
Nothing went wrong in that story. The automation did its job. The shop grew. And the pricing model punished both. My view is that a competent AI customer service setup can handle 80 to 90% of enquiries. If you pay per resolution, that percentage is the size of your exposure, and the vendor sets the price of your own success.
The same thing is happening one layer down. The subsidised token prices Gartner describes are usage pricing at the model level. Per-resolution billing passes it straight through to the operator, with a mark-up on the outcome.
Price it like infrastructure
When I set up AeroChat’s pricing, I chose flat monthly plans. Each plan includes a set number of AI responses, and a merchant who uses more tops up, the way a mobile data plan works. A shop owner can look at last month’s conversations and know what next month will cost. For one bakery merchant it handles 87 to 94% of conversations with no person, through festive peaks, and the bakery’s bill does not spike because the AI had a good month.
Gartner’s other recommendations tell service leaders to manage cost with prompt caching, token monitoring and routing simple questions to cheaper models. Read that list again from the point of view of a clinic manager or a three-person ecommerce team. None of it is their job. Those are the hard, messy parts that sit underneath, and they belong to the vendor. If your automation vendor asks you to monitor tokens, they have handed you their work along with the bill.
Judge it against people, not against last month’s invoice
The 42% who do not know their spend or their return are looking at the wrong comparison. A per-resolution bill that went up looks like a cost problem. It might be the best news the business had all quarter.
The comparison that matters is simple. Count the conversations that were resolved with no person this month. Ask what it would have cost to have people answer them, including the ones at 11pm and the ones during the peak. That figure, set against a flat monthly plan, is the return. If the plan costs less than the people it replaced and the customers got answered faster, the automation is paying for itself. If the vendor’s bill moves with every resolution, you cannot make that comparison honestly, because the cost side keeps moving.
What to do
- Ask the vendor what the billing unit is: resolution, conversation, ticket, seat or response. If the unit is a resolution, the vendor controls the price of your success.
- Ask for the bill in a quiet month and in your busiest month. Get both numbers in writing before you sign.
- Count the conversations resolved with no person each month, and work out what the same conversations would have cost in people’s time. That is the return. Compare it against a flat monthly figure.
- Ask what you are expected to manage yourself. If the answer includes tokens, models or caching, the vendor has moved their problem onto your desk.
- Treat the subsidy ending as a given. Pick a model whose price you can forecast from your own volume, and plan for growth, because growth is the point.
Source: CX Dive: The AI customer service boom is hiding a cost crisis