Seats, Consumption, or Outcomes: How to Negotiate AI Pricing You Can Forecast

Per-seat pricing had one great property: you could forecast it. You knew your headcount, you knew your rate, you knew your bill. That property is disappearing, and most revenue leaders are signing the replacement without modeling what it does to their budget in month nine.

Three models are now common in GTM AI, and each one shifts risk in a different direction. Knowing which risk you just accepted is the whole negotiation.

Per seat: predictable, and increasingly dishonest

Seat pricing prices access, not value. It worked when software was a place people went. It works poorly for agents, because the whole promise is that work happens without a person sitting in a seat.

Where it survives, watch for the quiet version of the same problem: seat minimums that assume every rep uses the tool, combined with usage caps that make heavy users expensive anyway. You get the rigidity of seats and the variability of consumption.

Negotiate for: true-down rights at renewal, not just true-up. Most seat contracts are one-directional by design.

Consumption: honest, and unforecastable by default

Consumption pricing charges for what the system does: messages, enrichments, agent runs, tokens, tasks. It is philosophically fair and operationally hazardous, because the thing generating cost is now a machine you have instructed to be productive.

The failure mode is well documented in adjacent categories. Usage grows fastest exactly when the tool is working, so success produces a budget overrun, and the person who championed the tool is now explaining an invoice instead of a result.

Under consumption pricing, the tool working well and the tool costing too much are the same event.

Negotiate for: a hard spend cap with a soft alert threshold, rollover of unused capacity, a documented definition of a billable unit that cannot be redefined mid-term, and a rate card that improves at volume tiers rather than staying flat.

That third term matters more than it sounds. If the vendor can change what counts as an agent run through a product update, your unit price is theoretical.

Outcome-based: attractive, and hard to make real

Pay per qualified meeting, per resolved ticket, per sourced opportunity. It sounds like perfect alignment and it is the model most likely to end in a dispute, for one reason: the definition of the outcome is the entire contract, and it is usually written by the vendor.

If the vendor's system decides what counts as qualified, you have outsourced your own quality bar to the party being paid by it. Meetings will be qualified. Many will not be worth having.

Negotiate for: a definition of the outcome that references your systems and your people. A qualified meeting is one your AE marked as qualified in your CRM within five business days, with a documented dispute path and a cap on disputes. Anything else recreates the problem outcome pricing was supposed to solve.

The four terms that matter more than the rate

1. Definitional control. Who defines the billable unit, and can it change without your written consent? If it can, the price is not the price.
2. Downside symmetry. You have upside exposure if usage grows. What is the vendor's exposure if quality falls? A quality obligation with a credit remedy is the answer.
3. Forecastability. Ask the vendor to model your bill at 2x and 5x current usage, in writing, before signing. Their willingness to do so is itself information.
4. Exit economics. What does it cost to leave mid-term or at renewal, including data export, retraining, and any minimum commitment left on the table?

How to bring this to finance

Finance does not object to variable cost. Finance objects to unforecastable cost. Bring three scenarios: expected, high adoption, and runaway. Attach the cap that makes runaway impossible. Attach the alert that gives you thirty days of warning before the cap.

Then own the operating discipline: someone reviews usage monthly against the model, and there is a written trigger for renegotiation if actuals diverge by more than a set percentage. That single practice separates teams who scale AI spend deliberately from teams who discover their AI spend during an audit.

The through line

Every pricing model is a statement about who carries the risk of the technology not working as promised. Seats put it on you. Consumption puts it on you. Outcomes theoretically put it on the vendor, but only if you own the definition.

Read the pricing model as a risk transfer, not a rate. Then negotiate the transfer, not the discount.