OpenAI has reportedly started letting a handful of its largest customers pay only when its AI finishes the job. The Information broke the story on 30 August, and The Next Web picked it up the next day with the example of a customer support interaction handled end to end. OpenAI has not confirmed it, the terms and customers are unknown, and the arrangement covers select major accounts rather than the general price list. Keep the word "reportedly" attached to every part of that sentence.
Even so, it deserves your attention. The company that sells intelligence by the token (the purest input pricing in the industry) is experimenting with charging for results instead. Our standing argument at Project Flux is that professional services have billed for inputs, mostly hours, since the industrial era, and that AI breaks the link between effort and output. When that link goes, pricing migrates to outcomes, and firms still operating on the hourly model get squeezed. The AI vendors are now running that exact experiment on themselves, in public, with quarterly scorecards.
Three of those scorecards landed within a week of each other. They point in different directions, which is precisely what makes them useful.
What OpenAI is reportedly doing
Per The Information, and unverified by TNW or by us, select OpenAI enterprise accounts can now pay for completed tasks rather than tokens. The commercial logic is easy to see. TNW notes one developer ran up $1.3M in token charges in thirty days across a fleet of agents, with cost scaling increasing based on attempts rather than the results. A bill that only arrives when something worked is far easier for a finance director to sign off.
Customer support got there first because a resolution is one of the few AI outputs anyone can define. Intercom charges $0.99 per conversation its Fin agent resolves. Zendesk bills only for what it calls Verified Resolutions, confirmed by an LLM evaluation within 72 hours, at roughly $1.20 to $1.50 on committed volume. Notice what both vendors accepted along with the model: the failed attempts are now their cost, and a referee decides what counts.
Outcome-based pricing is becoming a market standard.
Kirkpatrick's firm found in May that 43% of buyers prefer consumption-based pricing and 27% prefer outcome-based pricing. Fewer than one in five still want to pay per user. His stronger finding, quoted by TNW, is that vendors offering only per-user pricing are being ruled out before evaluations even begin.
Snowflake had a very good week selling inputs
Here is the complication. On 2 September Snowflake – the archetypal consumption business, metered by the compute-second – reported product revenue of $1.49B, up 37% year on year, and raised full-year guidance to $6.07B. Its CoCo coding agent added more than 2,000 net new accounts in the quarter to pass 9,100, roughly 28% growth in three months. The shares closed 17.43% higher the next day.
"AI is compounding Snowflake's advantage," CEO Sridhar Ramaswamy told analysts, crediting AI products with about half of the growth acceleration. So metered inputs are hardly dead. A value-based model works when the meter reflects value: a query that runs is worth something. Counting attempts, as token billing does for agents, makes the cost harder to justify. For a consultancy, the key consideration is how closely the pricing model reflects the value delivered.
Salesforce keeps rewiring its own meter
Salesforce shows how messy the transition gets. Agentforce launched at $2 per conversation, billed whether or not anything was resolved, which customers found expensive and impossible to forecast.
Flex Credits followed, pricing individual actions at about 10 cents each. Now the company reports in "Agentic Work Units" – 3.2 billion delivered in its second quarter alone, up 97% on the prior quarter – while Agentforce annual recurring revenue passed $1.5B, up over 240% year on year after some product regrouping.
We just delivered one of our best quarters ever, outperforming across every key metric.
Three pricing constructs in under two years, from one of the most sophisticated software sellers alive. Nobody has this settled.
Questions a consultancy should be asking
If you run or price a QS, PM or advisory business, the vendors are stress-testing your future commercial model at scale. Watch them, and ask these questions of your own services now:
Which of our deliverables have an outcome a client and we would both recognise, the way a resolved support ticket is recognisable, and which are genuinely judgement calls?
Who verifies that the outcome happened – us, the client, or a third party playing the role Zendesk hands to an LLM evaluator within 72 hours?
Where does the risk of failure sit, and what would we charge to absorb it, given that per-outcome rates cluster around a dollar rather than a cent precisely because vendors price in their own failure rate?
Which services are better metered as consumption, where usage tracks value closely, rather than forced into an outcome frame that invites disputes?
What happens to our fee model when a client can see, from their software bills, exactly what a completed task costs?
Takeaway
One reported pilot at OpenAI does not end the billable hour. What ends it, eventually, is the pattern underneath: buyers rewarding sellers who take on delivery risk, verification systems making outcomes contractible, and even the input-pricing champions winning only where the meter tracks value. Snowflake's 37% quarter proves inputs can still sell brilliantly – when the input is a fair proxy for the result. Hours, increasingly, are not. The firms that thrive will have worked out their definable outcomes, their verification method and their risk price before a client asks. The vendors are publishing the playbook a quarter at a time. Read it.
Project Flux exists to help built environment professionals get ahead of exactly this kind of shift – the business models AI drags along behind the tools. If you want the commercial side of AI in construction covered every week, join thousands of readers at projectflux.ai.
Links and Stuff
All content reflects our personal views and is not intended as professional advice or to represent any organisation.
