As AI token prices collapse, boards often confuse cheap raw intelligence with a cheaper business.
If any major enterprise input became 286 times cheaper in two years, every CEO, CFO, CIO or board member would expect the cost base to collapse. AI has delivered exactly that kind of price collapse.
Stanford's AI Index found that the price of accessing roughly GPT-3.5-level capability fell from about $20 to $0.07 per million tokens between late 2022 and late 2024. The trajectory remains downward.
Yet cheaper technology does not guarantee a cheaper enterprise. It invites more consumption, more use cases, more software and more operating complexity.
The input gets cheaper. The enterprise around it becomes more expensive.
Organisations do not buy the same amount of cheaper intelligence and bank the difference. They embed it in more processes, automate more decisions and create new dependencies across data, security, integration, assurance and support.
The boardroom mistake is to confuse a cheaper unit of intelligence with a cheaper enterprise.
The mistake begins with the unit
AI has made an old IT cost-management failure harder to hide.
Most AI conversations start with the price per token, model call or licence. Those measures are useful to engineers and procurement teams, but they are supplier units. They do not tell a CFO what it costs to complete a transaction, support an employee, run an application, serve a branch or keep a plant productive.
Technology leaders have relied on similar abstractions for years: IT spend as a percentage of revenue, spend per employee, tower budgets, peer benchmarks and broad run, protect, improve and innovate allocations.
These measures provide context. They may tell you whether spending appears normal. They do not explain why a service costs what it does, whether the difference is justified or where management should intervene.
Two enterprises can spend the same percentage of revenue on technology and have entirely different economics. One may be funding resilience, regulatory strength and differentiated digital capability. The other may be carrying duplicated applications, underused infrastructure and operational friction.
Benchmarks tell you whether you appear to be in bounds. Unit costs reveal what is happening underneath.
Peel back the assembled price
The discipline begins by naming the operating unit that matters.
That unit may be a supported employee, a resolved service ticket, a processed transaction, an application in production, a branch served, a plant-hour supported or a completed customer workflow. The unit changes with the service. The discipline does not.
Once the unit is visible, management can peel back the costs that create it. This mirrors the useful lesson from Elon Musk’s challenge to rocket economics, where he refused to accept the assembled price as the answer. He refused to accept the assembled price as the answer. He broke the rocket into its constituent materials and challenged the economic logic underneath.
IT needs the same engineering instinct.
What sits inside the cost of one resolved enquiry? Applications, infrastructure, licences, data, integration, support, security, people, rework and exception handling may all contribute. The cost of a production release carries a different mix: development, testing, assurance, tooling, release management, support and the cost of failure.
None of these relationships is visible in the aggregate bill. Naming the unit is what brings the constituent costs back into view.
AI makes the discipline more urgent
AI leaves traditional unit economics intact, but it changes the proportions and accelerates the volume.
Take a single resolved enquiry. An AI assistant answers it in seconds, and the manual handling cost all but disappears. But behind that one clean resolution now sits a retrieval pipeline feeding the model current data, logging to prove what was said, a monitoring layer watching for wrong answers, and a human reviewer for the cases the model gets wrong. The enquiry got faster and cheaper to answer. However, the service got more expensive to run.
This is the pattern beneath most AI-enabled work. It may reduce manual handling while increasing compute. It may shorten cycle time but require more verification. It may strip out process steps only to add integration, monitoring and assurance in their place. A pilot looks cheap because it proves that something can work. It does not price the full operational dependency that follows.
This is why the relevant economic question is broader than the price of the model. Did AI reduce the fully loaded cost of the operating unit? If not, did it create enough additional value to justify the increase?
The same distinction matters when management claims productivity gains. Faster work may create more output, better service, higher quality or additional capacity. Those outcomes can be valuable, but they are not automatically cash savings.
An hour saved becomes a saving only when expenditure is removed or a measurable future cost is avoided. Until then, it is capacity. The enterprise still needs to decide what that capacity will produce and whether the result matters.
Paying more is not the problem
Unit economics should not become a blunt cost-cutting exercise. A higher unit cost may be entirely justified because the enterprise is buying resilience, compliance, faster recovery, better customer experience, strategic differentiation or lower operational risk.
The objective is to understand why the enterprise is paying more, what additional value it expects in return and whether that value is being delivered.
A higher unit cost is not automatically a problem. An unexplained unit cost is.
Every material variance should therefore carry a label and an explanation. Strategic cost must demonstrate the value or differentiation it buys. Transitional or establishment cost must show a credible path to amortisation or decline. Waste and sprawl have no defensible economic story and should be removed.
AI provides a clean test. The first use case may reasonably carry the cost of shared identity, data access, evaluation, observability and governance. But the fifth use case should not cost as much to establish as the first.
If it does, the organisation is not scaling a shared capability. It is rebuilding the enterprise one pilot at a time.
Cost, value and governance belong together
The executive task is to connect cost to consequence: what the enterprise is paying for, what it gets in return, and whether that result is worth the cost, risk and complexity it must carry.
Once those answers are visible, governance stops being an approval ritual. It becomes the mechanism for deciding whether to accept the additional cost, invest further, redesign the service, consolidate it, decommission it or stop paying for it.
Cost without value is waste. Value without cost visibility is theatre. Governance forces the trade-off.
Economic ownership therefore means more than approving a business case. Someone must own the unit, understand its constituent costs, explain the value being created and have the authority to act when the economics no longer hold.
The boardroom test
Boards do not need hundreds of technology measures. They need the few unit costs that reveal the economic truth of the estate.
Which units matter most? What is driving each one? Where are we deliberately paying more for resilience, differentiation or lower risk? Which costs are transitional, and when should they decline? Who can restructure or stop a service when its unit economics no longer hold?
The boardroom mistake is to confuse a cheaper unit of intelligence with a cheaper enterprise.
The management failure is not knowing what any meaningful unit of the enterprise actually costs.
