Six months ago, asking how much AI a company used could feel like a reasonable first question. The answer was usually a usage chart, a departmental budget, or a provider invoice. Those are useful inputs. They are not a value argument.

Heading into 2027, the harder question is what changed because AI performed a piece of work, and whether the company can defend the cost of that work. A leader who still treats activity as evidence has mistaken the meter for the result.

Usage is not the unit of value

Tokens, requests, seats, and enabled features describe system activity. They may help explain a bill, but none is necessarily the unit a business sells, budgets, or reports. If an assistant handles part of a support interaction, the useful unit may be a resolved case. If a model contributes to a product feature, it may be a customer, transaction, or completed workflow.

The unit has to match the work. Otherwise, a lower cost per token can coexist with a higher cost per customer, and a rising usage graph can be presented as progress even when no outcome improved.

The test: Can the person accountable for the result name the work unit, show the AI-delivered portion, reconcile its cost to the invoice, and explain what decision the result changed?

Six months is enough time to change the question

In April, it was easy to mistake visibility for control. A team could see which model was called and still fail to distinguish experiments from production. It could know an application was enabled without knowing whether its AI feature had become part of a customer promise. It could calculate a cost and still not reconcile that figure to what the provider charged.

The useful change is not a more confident forecast. It is a narrower claim with a traceable basis. Which workload produced this number? Which costs are included? What was excluded? Does the total tie to the invoice? What outcome is being compared with it, and who owns that comparison?

If those questions cannot be answered, do not polish the dashboard. Record the gap, assign an owner, and narrow the claim until the evidence supports it. Precision about uncertainty is more useful than an unsupported return estimate.

Keep two exposures separate

AI used inside the company is an operating expense and governance question. AI embedded in a product sold to customers is a cost-of-goods-sold and margin question. They may draw on the same providers, but they do not have the same owner, decision, or financial treatment.

Combining them into one company-wide figure hides both. Internal experimentation should not be judged as if every test were a production promise. Product workloads should not disappear inside a broad engineering budget when they determine the cost to serve a customer.

For each pool, define the work unit, attach the relevant provider and infrastructure costs, and state the boundary between research and production. A boundary that cannot be explained is not yet a measurement standard.

What a defensible 2027 review contains

Start with a small number of workloads that matter to a real decision. Name their owners. Reconcile observed usage with invoices and explain the difference between allocated and directly metered cost. Select an outcome measure the business already understands. Then show how cost and outcome move together over time, including the limits of the comparison.

This does not prove that AI caused every improvement. It gives leaders a disciplined basis for deciding what to expand, redesign, price differently, or stop. A credible review says where attribution is strong, where it is directional, and what evidence would improve it.

Ask the same question of every attractive claim: what would make this conclusion false? If no evidence could change the answer, it is not measurement. It is a slogan with a spreadsheet behind it.

The challenge for the next planning cycle

Take the AI slide that would have satisfied you six months ago. Remove the usage total. What remains that supports a funding, pricing, governance, or staffing decision? If the answer is nothing, the old slide measured motion. It did not measure value.

That is fixable, but it starts with an honest boundary and a real unit of work. The companies best prepared for 2027 will not be the ones with the most elaborate AI scorecard. They will be the ones that can show what they counted, why it matters, and what they changed when the evidence disagreed with the plan.