The AI meter: support pricing is quietly switching sides

There is a line on this year's support-tooling invoice that did not exist on last year's. It is not seats. It is not storage. It is a meter — resolutions, conversations, actions, credits — and it ticks every time the software does something an agent used to do.

For two decades support software had one price shape: the seat. You counted heads, multiplied by a list price, argued about the discount, and knew your bill for the year. That shape is dissolving in real time, and the story of how is worth reading closely, because the vendors themselves are still figuring out what they are selling.

The $2 lesson

Salesforce launched Agentforce in October 2024 with the cleanest outcome price imaginable.

Elegant on a whiteboard; unworkable in a forecast. What counts as a "conversation" when a single customer query triggers eight backend processes? Buyers could not model their exposure, and a price you cannot model is a price you do not sign. Seven months later Salesforce introduced Flex Credits — actions instead of conversations, sold in packs.

And by late 2025 the wheel turned fully: a per-user licence reappeared at $125-plus per month, because enterprises wanted what they have always wanted — a number they can budget and a contract shape they understand. Three pricing models now run side by side, and the approach is working commercially.

The lesson is not that outcome pricing failed. It is that the market is mid-negotiation about what an outcome is — and every contract you sign right now is a position in that negotiation.

The add-on that dissolved into a meter

Zendesk ran the same journey from the other direction. For years its AI story was a classic seat add-on: a flat fee per agent per month on top of your plan. In May 2026 the add-on was eliminated and the capabilities folded into every Suite and Support plan — which sounds like a price cut until you read the meter. Each plan includes only a small allotment of automated resolutions, on the order of five to fifteen per agent per month depending on tier. Past that, every automated resolution is billed on top of your seats.

A companion change in January 2026 made the meter uncapped, auto-charging overages above committed volume. Sit with the incentive for a second: the better your AI performs, the larger your invoice grows. Deflection — the thing you spent years trying to maximise — now has a unit price.

The pioneer and the holdout

Intercom saw this shape first and priced Fin per resolution from day one.

That definition does load-bearing work: a customer who goes silent counts as resolved. Whether you consider that a resolution or an abandonment is exactly the kind of question the order form answers and the marketing page does not.

Meanwhile Freshworks has kept the seat shape for its AI copilot.

The two shapes encode two different bets. Per-seat AI is priced as a person-amplifier: you buy it per human, and the vendor wins when your team keeps its size. Per-outcome AI is priced as a person-replacement: you buy it per ticket, and the vendor wins when the bot handles more of your volume. Read a vendor's price sheet and you can read their thesis about your headcount.

What this does to your budget

Seats scale with your org chart, which you control. Meters scale with contact volume, which you mostly don't. A product incident used to cost you overtime; on a metered contract it costs overtime plus a software bill that spikes with the queue. Forecasting a tooling budget now requires a deflection forecast — you need to predict how many contacts you'll have and what share the machine will claim, just to model an invoice.

And there is a quieter problem: the vendor now gets paid every time the bot declares victory. When the entity that defines "resolved" is the entity that bills per resolution, the definition deserves an audit, not a nod. Does silence count? Do reopens within seven days claw back the charge? Does a customer who gives up and emails again register as one resolution and one new contact?

Negotiate the definitions, not the rate

The per-unit rate is the least important line in the contract. The definitions are the contract.

• Pin the billable event in the order form. What exactly triggers a charge, in writing, with the edge cases — silence, reopens, multi-contact issues — resolved in your favour or at least resolved at all.
• Cap the meter. Committed volume with a hard ceiling and a pre-agreed overage rate beats an uncapped meter at any discount.
• Demand the audit trail. Line-item billable events you can sample against real tickets, monthly, without a support case to get them.
• Model three volumes. Normal month, seasonal peak, incident spike. If the incident-spike number frightens you, that is the number to negotiate.
• Re-run the math at renewal. These models are a year old and moving; the shape you signed is not the shape you'll be offered next.

None of this is a reason to avoid outcome pricing — aligning cost with resolved demand is, in principle, saner than paying for chairs. But principle is doing a lot of work in that sentence. The sticker was never the price, as we argued in the true cost of support tooling, and your stack is a strategy document whether you meant it or not. The meter is just the newest paragraph. Read it before you sign it.