AI Token Costs: How CPOs and CTOs Stop Budget Overruns

Over Budget, Ask for More, Repeat.

The token pack ran out before the project went live.

It was a platform implementation, and the pack had been sold by the vendor and the implementation partner together. Our Finance Director did what a Finance Director should do. They asked the sales representative to explain what a token actually is, how we would use them, when we should buy more, and how to calculate it.

I remember it vividly, because the answers were far thinner than the question.

In my experience, most of the people who sell these packages do not fully understand them either. To our surprise, the vendor could not estimate our usage. What they could do was suggest we keep buying more.

That is when I stopped seeing AI token costs as a technology topic. It is a procurement problem, and it is also a governance problem.

It Is Not Just Our Story

After ProcureCon Europe in Cologne and DPW in Amsterdam, this was the subject I heard least about. Lots of use cases, plenty of roadmaps. Almost nothing on what happens to the AI budget once usage-based pricing meets a real programme.

Over coffee, Jaime H Paiva, Senior Procurement Leader gave me a rule I have not forgotten: budget three times for tokens and ancillary costs on every AI project.

It is a practitioner's rule of thumb, not a benchmark. But the data points the same way.

"72% had a surprise AI bill in the past year."

-Harness, State of AI in FinOps 2026

"46.9% of enterprises are over their AI budget. 5.6% are under."

-Futurum, September 2026

Futurum then asked what the over-budget organisations do. 47.6% ask for more budget. So the cycle in my headline is not a joke. It is the single most common response.

One caution. These surveys talk to IT and engineering leaders, not CPOs. Which tells you something by itself. The people who sign the contract are often not the people who see the bill.

Why Usage-Based Pricing Breaks Traditional AI Budgeting?

Seats were easy. You counted heads. Tokens and credits are a different animal, and the unit itself can move.

GitHub moved Copilot to token-based AI credits on 1 June 2026. The seat prices on the business and enterprise plans did not change. What you are billed on did.

Gartner flagged the risk back in 2024: credit and token pricing models are not standardised, they rely on unclear multipliers, and they lead to overspend.

Add the ownership gap. In the Harness study, 52% said nobody clearly owns AI cost, and only 45% of engineers know what the AI features they build cost to run.

Procurement signs, IT builds, Finance pays. And no one answers for the total.

Five Steps for AI Cost Governance Before You Sign

1. Name one owner before you sign.

   One person, by name, answers for the total AI spend. Not a committee. The CPO owns the commercial terms. The CTO owns usage tracking.

2. Ask the vendor for the maths in writing.

   What consumes a token. A worked estimate for your use case, not their demo. How the unit is defined, and what happens if they redefine it.

3. Add it in the contract.

   This is where AI contract negotiation earns its keep: shorter terms, consumption caps, limits on changing credits or multipliers mid-term. In an NPI poll of 89 IT sourcing leaders, 56% named unpredictable usage and scaling costs as their top AI spend challenge, and these were the protections they described.

4. Budget for the overrun now.

   Add a buffer to the project and set an AI token budget for each business unit. If three times feels too high, start from your pilot data and add a cushion you can defend.

5. Measure cost per result, not usage.

   Usage tells you what was consumed. Cost per result tells you whether it was worth buying. Put that number where the decisions are made.

Who Should Own What?

I would split it like this.

The CPO takes the contract and the vendor's usage model.

The CTO takes usage tracking and cost visibility.

The internal token budgets belong to both, with Finance watching. That is my recommendation, not a finding from a study. I would rather you challenge it before the contract than after the invoice.

Questions CPOs Ask About AI Token Costs

What are AI token costs?

The usage-based charges for the text a model reads and writes, billed in units called tokens or credits. The cost depends on how much your teams use the tool, not on how many licences you hold.

Why do AI budgets overrun?

Because usage is hard to estimate before launch, the billing unit can change under a stable price, and often no one owns the total.

What should a CPO ask an AI vendor first?

What consumes a token, a worked usage estimate for your own use case, and what happens when the unit definition changes.

The Question to Take Back

Who owns your AI bill today: procurement, IT, or nobody?

If the answer is nobody, you are already in the cycle. Get the usage maths in writing, name one owner, and plan the buffer for 2027 before the next request goes to Finance.

If your AI pilots are not showing savings either, I wrote about where procurement's AI savings are hiding. If you want a second pair of eyes on your 2027 AI budget and vendor terms, take a look at the AI readiness assessment or book a 30-minute discovery call with ProcureSynth.

Sources and References

Harness, State of AI in FinOps 2026 (700 engineering leaders and practitioners, May–June 2026). Vendor-sponsored and self-reported.

Futurum, press release, 10 September 2026 (n=1,636 IT decision-makers).

NPI, peer poll of 89 IT sourcing leaders, 7 May 2026.

GitHub blog, Copilot usage-based billing, 27 April 2026.

Gartner, abstract on credit and token pricing models, 6 June 2024.

Google Gemini for image creations

Foundation Assessment

Are you Ready to have This Conversation?

ProcureSynth assesses your foundation based on predefined metrics and delivers a concrete plan for resolution

Email: connect@procuresynth.com

© 2026 ProcureSynth. All Rights Reserved. ProcureSynth® is a registered trademark of Hemangi Tawade AI Transformation Services.