What Does a Business Outcome Really Cost with AI?

The CFO opens the monthly expense report. AI spending is rising, adoption looks healthy, and business units report hours saved. Then comes the question nobody has answered: what did we get for the money? Usage tells us how widely a tool has spread. It does not tell us what changed in the business, or what that change cost.

From a usage meter to a business outcome

In his October 2026 annual letter, Satya Nadella proposed weighing the cost of tokens against the outcomes they produce. That is the Cost-to-Outcome perspective. Tokens are units of information processed by a model; inference is the work of generating an answer; marginal cost is the incremental expense of another operation. These matter when buying technology. They are not the same as the cost of finishing a business process.

Licenses measure purchasing. Usage measures adoption. Hours saved may indicate productivity. None, on its own, proves financial value. A widely adopted AI system can still cost more than the benefits it delivers.

Define the outcome before calculating the cost

In customer service, an outcome might be a ticket resolved without reopening within seven days. In finance, an invoice correctly checked and recorded. In sales, a lead that becomes a genuine opportunity. A drafted response is a step; a customer problem resolved is an outcome. Start by measuring the existing process: cost, handling time, waiting time, errors and rework.

Cost per valid outcome is the full process cost divided by completed outcomes that meet an agreed quality standard.

Include models, licenses, integration, maintenance, training, human review, corrections and an allocated share of implementation. Count only successful completions in the denominator, while keeping failed attempts in the cost base. Otherwise, a cheap generated answer can hide an expensive unfinished job.

A thirty-cent answer is not a thirty-cent resolution

Consider an insurance service center handling 10,000 tickets a month. This is a fictional illustration, not company data or an industry benchmark. A vendor offers AI at 30 cents per ticket. The table compares the entire process under a consistent quality standard and assumes non-overlapping cost categories.

Approximately 11.8% lower cost per valid outcome; $13,000 less in monthly spending, or $156,000 annually, only if the assumed reduction in staffing cost is actually realized.

AI introduces $42,000 in new costs, while the table assumes a $55,000 reduction in staff spending. Without that reduction, there is no budget saving: total monthly cost rises to $162,000. Human review alone costs $18,000, six times the $3,000 model bill. The economics depend less on the price of the technology than on how the organization redesigns the work around it.

This is where three different kinds of value must be separated. Hard savings are expenses that actually disappear from the budget. Cost avoidance means not incurring a credible future expense, such as a planned hire that becomes unnecessary as demand grows. Capacity creation means releasing time for other work. All three can be valuable; only the first is an immediate budget reduction. The others require evidence, not optimistic assumptions.

In organizations where staffing is difficult to adjust quickly, the most credible benefit may be faster service or absorbing growth without adding headcount. Those are legitimate outcomes. They should not be reported as savings already visible in the P&L.

Now halve the model bill to $1,500. Total cost drops from $107,000 to $105,500, an additional improvement of only about 1.4%. A dramatic discount on the model barely changes process economics. Reducing rework or unnecessary review may matter more.

A cheaper outcome is not necessarily a better one

Cost per outcome is not the same as value per outcome. Two resolved tickets can look identical in a spreadsheet even if one prevented a valuable customer from leaving and the other answered a routine question. Paying more may be justified by higher retention, shorter waits or lower risk. Cost-to-Outcome is a useful economic measure, not a replacement for quality, revenue and customer value.

Errors differ in severity, too. Misclassifying a routine request may require a quick correction; making an incorrect insurance commitment may create legal or regulatory exposure. The cost of failure is not limited to the rework line in a table. Rare, high-impact errors and clear accountability belong in the decision.

The same principle applies to model selection. If a smaller model or conventional software rule meets the quality bar, paying for extra capability is wasteful. But a cheap model that needs repeated attempts and extensive review can cost more overall. For Hebrew-language workflows, test on your own representative data rather than relying solely on English benchmarks.

What should leadership decide?

Start with one process, jointly owned by a business leader and a finance partner. Agree on the outcome definition, measure the baseline and test the alternative under representative conditions, including exceptions and workload peaks. Set the date and conditions for scaling, improving or stopping before the pilot begins. A successful demonstration on curated data is not proof of viable economics at scale.

Before approving funding, I would ask three questions: What actually changed for the customer or the business? Which expense fell, which future expense was avoided, and which capacity was merely released? What are the quality and risk thresholds, and when will we stop?

The economic challenge of AI is not simply making work cheaper. It is distinguishing work done faster, work that costs less and work that creates more value. Those are three different outcomes. Leadership teams that confuse them risk making the wrong investment decisions.

Concept source: Microsoft 2026 Annual Report · Shareholder Letter. The management approach and example are an independent application; all figures are illustrative.