Outcome-based pricing for AI

I have been working through how organisations should build budgets and forecasts for AI use, where the familiar mix of seat licences and token estimates is tied directly to measuring ROI, benefits and value.

A cost-in approach works neatly on a spreadsheet: estimate users, apply an adoption curve, add token consumption and give finance a number. Yet that precision conceals the part that matters: nobody can reliably know what a particular prompt, agent run or task instruction will cost before committing to it when the work involves retries, tool calls and (potentially) several models.

Although the speed of new models and capabilities makes every forecast perishable, the deeper problem is incentive design. Under token billing, providers earn from input and output whether the task succeeds, fails or has to be attempted again. The customer carries the cost of inefficiency while the provider controls much of the system that creates it.

If, as I expect, the venture-funded subsidy phase is ending, that misalignment becomes more alarming. The need to monetise will create pressure to grow margins, so organisations that treat today's unit economics as a stable planning assumption are building budgets on a commercial settlement that will not hold.

There is already evidence of a different settlement. Zendesk charges only when AI resolves an interaction end to end, while Pegasystems charges per completed case and absorbs the underlying model cost. OpenAI is reportedly allowing selected large customers to pay only when its AI completes the job, although the arrangement is unannounced and the report, from The Information, has not been independently verified. Salesforce is also negotiating contracts tied to revenue growth or service-cost reduction.

This does not mean outcome pricing has already won. Gartner says only 19% of services buyers and 13% of seller-side service agreements use it, while HP does not expect outcome-linked options for most early adopters until mid-to-late 2027. Contract practice still lags because defining an outcome, attributing value and preventing gaming are governance problems rather than billing details.

However, the direction of risk is changing before the market has settled the mechanics. When vendors absorb the cost of failed attempts and get paid from successful work, they finally have a reason to choose efficient models, limit unnecessary tokens and improve the whole system rather than sell more computational activity.

Organisations should stop trying to perfect cost-in forecasts for AI. They should set the value of the outcome, agree how it will be measured and shared, and make AI partners earn more only when the organisation does. A win-win contract is not a softer alternative to cost control; it is the only cost control designed for a technology whose inputs cannot be forecast with confidence.