AI spend is starting to show up in boardrooms in a familiar way.
The budget was one number. Actual spend is another, often a much bigger one.
The obvious question follows:
Are we spending too much on AI?
It is a reasonable question. It is also incomplete.
If AI can automate work, increase output and improve decisions, boards should ask another question at the same time:
Are we spending enough?
An organisation can waste enormous amounts of money on AI. It can also destroy value by controlling expenditure so tightly that it underinvests in valuable uses.
A few months ago I wrote about the token bill coming due: organisations encouraging AI adoption without a good way to connect spend to value. The next question is harder. How do we know whether the bill is justified?
Cost is not the same as value
Imagine two organisations.
The first spends €10 million on AI and generates €2 million of measurable value. The second spends €100 million and generates €500 million.
Which has the AI cost problem?
Absolute expenditure would point us in the wrong direction. Yet governance discussions often begin and end with spend: how fast is it growing, how far is it above budget, and which team is responsible?
Those are useful accounting questions. They do not tell us whether the organisation is allocating resources well.
Every model call has a cost. Longer context, more reasoning, tool calls and retries have costs. But those costs may replace or augment human time, reduce errors and delays, or enable work that was previously too expensive to attempt.
The question is therefore not only how much AI are we consuming?
It is what are we getting for the intelligence we buy?
Three questions, not one
AI cost governance should separate three questions that are too often treated as one.
First, are we buying AI efficiently? Are expensive models handling tasks that smaller ones could perform just as well? Are we sending unnecessary context, repeating calls or building agents that loop? Could caching or better architecture lower the cost without reducing quality?
Second, are we using AI for things worth doing? An efficient model call is still wasteful if the underlying task produces little value. Making a low-value process cheaper does not necessarily make it worth running.
Third, are we using enough AI where the economics are compelling?
Suppose a workflow costs €5 to run and replaces an hour of manual work costing €100. Reducing the AI cost to €4 is useful. Preventing the workflow from running because its budget is exhausted is not.
Cost discipline is not cost minimisation.
Cheaper models do not settle the question
The market is already changing the price side of the equation. In September 2026, OpenAI introduced GPT-6 Sol and Luna at lower API prices than their GPT-5.6 counterparts. Anthropic introduced Claude Opus 5.5 with a lower price than Opus 5. xAI launched Grok 4.7 at $2 per million input tokens, though that price itself was not a cut from Grok 4.6.
Competition changes both the price per token and the number of tokens needed to complete a task. Providers claim gains in efficiency as well as lower prices. Neither guarantees a smaller total bill: lower costs make more use cases viable, and agents may perform more steps on each task.
It also means model policies need revisiting. A common response to a growing API bill is to tell teams to avoid the most capable model: use Opus, not Fable. That may be sensible for routine work. But if Fable can complete a harder investigation that Opus cannot, or materially reduces review and rework, choosing Opus by default could cost more in the end or leave valuable work undone.
I would test both on representative tasks, including the difficult cases that motivated AI adoption. Compare successful outcomes, total inference cost, human effort and the value of work that becomes possible. As prices and capabilities change, run that comparison again. A policy that was economical last quarter may now be holding back the business.
As intelligence gets cheaper, allocating it well matters more.
Tokens are not the KPI
Published API rates look simple: a price for input tokens and a price for output tokens. The economics of an application are less obvious.
A person asks a question or runs an agent. In the background, the system may retrieve documents, assemble context, call tools, invoke several models, retry failed steps and loop before it produces a result. The user sees one task; the infrastructure sees many interactions.
In managed AI products, the application or provider may determine some of those choices. The supplier can therefore influence how much of the metered resource is consumed. That does not imply wrongdoing, but a good price per million tokens gives little assurance if nobody understands why so many tokens were used.
Making token reduction the sole target creates its own problem. An employee who spends twenty minutes saving €0.10 on a model call has improved the token metric and worsened the economics.
The control objective should be to use the least costly combination of people and AI that reliably produces the required outcome at the required quality. A more expensive model may be cheaper overall if it needs fewer retries or less human review.
Build AI unit economics
Budgets still matter. Organisations need to forecast spend, approve it and explain variances. They also need to connect consumption to outcomes.
For each material use case, management should understand measures such as:
- Cost per successful customer interaction, investigation or audit
- Cycle time and human review time per outcome
- Error, rework and retry rates
- Additional revenue or capacity enabled
- Total cost of the workflow, including people and infrastructure
These measures will be imperfect. Productivity gains may be absorbed into better work rather than visible headcount savings. New capabilities may have no previous process for comparison. But imperfect unit economics are more informative than an invoice on its own.
The unit of analysis also matters. Comparing model prices per call misses the cost of failures, extra orchestration and downstream checking. A model that is twice as expensive per call could still deliver a successful outcome for less.
What should Internal Audit ask?
An audit that reports a significant budget overrun may be factually correct and still tell the board very little.
A useful audit would trace the increase to use cases and outcomes. Can costs be attributed to applications, teams and workflows? Can management explain a material jump in consumption? Are model and architecture choices tested against quality and total workflow cost? Do agents create unnecessary loops? Who decides when additional quality justifies additional inference spend?
The most difficult question may be: where should the organisation deliberately spend more?
Internal Audit does not need to decide the AI budget. It can assess whether decision makers have reliable information and a credible process for allocating it.
Cloud FinOps taught organisations to connect on-demand computing costs to ownership and value. AI adds a comparison with human work. For each task, the right answer may be a person, a smaller model, a frontier model, an agent, or some combination. The appropriate level of human review changes that calculation again.
This makes AI cost governance a question of operating model and resource allocation, alongside procurement and technical efficiency.
Ask both questions
Boards are right to question rapidly growing AI expenditure. Cost attribution, transparency and budget accountability matter.
But why are we spending so much? should lead to two more questions: what are we receiving in return, and where would spending more or less create greater value?
The organisations that get this right may not have the smallest AI bills. They will be able to explain what those bills produce, identify waste, and invest further where the return warrants it.
How is your organisation measuring the value side of AI consumption? If you are building AI cost controls or assurance around model spend, I’d like to hear how you are approaching it.