And why a falling unit price didn't stop it from happening.
The unit price of an AI token keeps falling. The invoice keeps rising.
By 2026, over 75%1 of IT spending is expected to be OPEX-driven. Most CIOs, CFOs and IT controllers still read this shift as a pricing problem: negotiate harder, get a better rate, move on. What actually broke is the fact that consumption became elastic and decentralised.
This diagnostic addresses 3 questions:
Why does the bill keep rising as the price falls? A falling unit price is mistaken for a shrinking bill, when the mismatch actually sits in the volume.
What happens if nobody catches it? Costs compound into a recurring pattern, and each recurrence gets more expensive than the last.
Where do you start? The first concrete moves toward steering consumption instead of just approving it.
Two prompts last week, a hundred this week
Nothing was approved in between. The model changed, the habit formed, and nobody was watching either.
This breaks a logic that held for a decade: negotiate the unit price down, lock in a volume commitment, defend the variance at quarter-end. That logic worked when consumption had a ceiling: a fixed footprint, a training requirement, a predictable trigger for change. AI consumption has none of those three things, which is exactly why it keeps outrunning every forecast built for the CAPEX era.
Here's the test I'd apply: if forecasting next month's AI spend still means guessing, the bill hasn't shocked the organisation yet. It will.
Here's why:
- No natural ceiling. Unlike a server with a fixed footprint, AI consumption expands with habit and use-case breadth, with no upper bound visible to Finance.
- A barrier lower than any tool before it. No training is required to open a chat window, and the habit is reinforced by using the same tools privately outside the office.
- Invisible triggers. A provider rolls out a new model overnight, and consumption shifts without a single policy changing. There's no event for anyone to notice.
Harder to explain every time
No single function has the visibility or the leverage to fix this alone.
Procurement gets pulled in, but large AI vendors set their own terms, with no key account manager to negotiate with. Accounting can't reallocate the cost either, since providers hand over an invoice and volume data with no department or employee breakdown. Attribution stays retrospective instead of live, so the pattern keeps recurring undetected.
Each recurrence costs the organisation more than money.
Left unaddressed, the pattern escalates. Teams overcorrect in both directions: some get too restrictive, and people avoid tools that would help them, while others stay too loose, and the habit compounds.
That habit turns into duplicated, untraceable work: a developer reprompts rather than reusing a colleague's output ("in blue this time, not green"), and as agents start talking to each other, no one can trace consumption back to a person at all. What started as a budget question becomes a reputational one: without shared visibility, explaining the recurrence upward gets harder each time.
Where steering starts
Usually, it takes a bill to get the rules written.
A German bank wrote its FinOps governance the same week it started its cloud rollout, and reached its first invoice already able to act on it. FinOps principles for cloud consumption, refined years ago, apply directly to AI. The starting signal always looks the same: nobody can forecast next month's spend with any confidence, no one can say which team drove last month's number, and no one has been formally asked to own that link.
Once that pattern shows up, the problem has stopped being a pricing conversation, and the first moves from there are concrete and immediate:
Build a rolling forecast, not an annual one. Updated monthly, driven by actual consumption patterns rather than last year's baseline, targeting under 8%2 variance on run costs.
Measure cost per outcome, not cost per token. That's how a transportation company finally justified its rising AI spend, and knew where to scale it.
Bring IT, Finance and Procurement into one room before setting any rule. A joint workshop across departments is where shared ownership starts.
Replace individual approvals with guardrails. Per-team usage thresholds and automated alerts catch a spike before it becomes an invoice, instead of asking a manager to bless each request one at a time.
The number worth watching
Every quarter, the unit price looks better on paper and the total on the invoice looks worse.
The price was never the hard part. The volume was. So how much of it can you actually change, and how much is already written into your vendor's contract?
Sources
- Gartner, cited in Sopra Steria Next, CIO Compass: Entering the Post-CAPEX Age, CIO Compass Series.
- Sopra Steria Next recommended target, CIO Compass: Entering the Post-CAPEX Age, CIO Compass Series.
