How to Measure the ROI of AI: Metrics and a Worksheet

Measure the current process before you start. Choose two or three metrics per use case, covering quality, speed or volume, and one that turns into money. Count every cost, including your own people's time, and compare the change with a group that didn't use AI. ROI is realized value minus total cost, divided by total cost. Count time saved as money only when it becomes something real: spending you stop, a hire you no longer need, or more output you sell.

• Baseline first. Without a "before" number measured the same way, no "after" number proves anything.

• Time saved is capacity, not cash. In a Danish study linking surveys of about 25,000 workers to employment records (NBER working paper, revised March 2026), 85% of chatbot users said they put the time they saved into other work tasks.

• Measure, don't ask. In METR's July 2025 trial, 16 experienced developers believed AI had made them 20% faster, while their measured task time was 19% longer. METR notes these early-2025 results are now dated, but the gap between feeling and measurement is the lesson.

• Company-level returns are still rare. In McKinsey's State of AI 2026 survey (August 2026), 37% of respondents attributed at least some EBIT impact to AI. High performers were twice as likely as others to have defined processes for measuring AI's impact.

• Measure success, not only money. Quality, adoption and risk tell you whether value is coming before finance can see it.

• The gains are spread thin: a few minutes per task across many people, which no budget line shows.

• Saved time gets absorbed. People do more of other work, which is real value but rarely shows up as lower cost.

• Self-reports are unreliable. People overestimate or underestimate how much faster they are.

• Other things change at the same time: new staff, seasonality, a process fix. Without a comparison group you can't tell what caused what.

• Costs hide in other budgets: training hours, IT time, reviews.

• Costs come before value. Early measurements catch the learning dip.

Read famous AI figures carefully

The best-known claim is that 95% of companies get zero return from generative AI. It comes from MIT NANDA's State of AI in Business 2025 (July 2025). What it measured is narrower: for embedded or task-specific enterprise AI tools, 60% of organizations evaluated one, 20% piloted one, and 5% reached production with what users or executives described as marked and sustained impact on productivity or profit. The data are interviews with 52 organizations, 153 survey responses and a review of over 300 public initiatives, with impact judged six months after the pilot. The authors call the figures "directionally accurate", and the same report found general-purpose chatbots in wide use. It is a useful warning, not a measure of your project.

Single-company figures need the same care. In February 2024 Klarna reported that its AI assistant handled two-thirds of customer service chats in its first month, the work of 700 full-time agents, with satisfaction on par with people. In May 2025 its CEO told Bloomberg the company would make sure customers can always reach a human, as Entrepreneur reported, and spoke of investing in the quality of human support. Early volume figures are not the whole story. Track quality and customer outcomes for longer than a month.

• Pick the unit of work: one email, one invoice, one report, one ticket.

• Measure for two to four weeks: volume, time per unit, quality (error rate, review score, rework) and cost per unit.

• Use system data where you can, such as ticket timestamps. Where you can't, use time sampling: ask ten people to log the time of 20 tasks each.

• Keep a comparison group: people or teams who don't use the tool during the pilot.

• Write the hypothesis: "With AI drafts, handling time per email falls from 6.0 to 4.5 minutes by week 8, with the quality score unchanged."

Use one metric from each column. The guardrail is a number that must not get worse.

For AI agents that act on their own, add the share of tasks finished without a hand-off, the cost per task, and the cost of wrong actions.

Use loaded cost for people's time: salary plus benefits and overhead, per hour.

Running costs matter more than they used to. In McKinsey's 2026 survey, about one in five respondents said AI operating costs, including token costs, were limiting their AI use. If your team builds agents, track cost per run.

• Count time saved as money only where it converts: overtime or contractor spend that stops, a planned hire you no longer need, more billable output, or a faster cycle with a price (such as getting paid sooner).

• Report the rest as capacity. "2,652 hours freed, of which about 2,000 converted to savings" is honest and still useful.

• Use the difference from the comparison group, not the before-and-after change alone.

• Value quality at the cost of fixing errors: rework time, credits issued, complaints handled.

• When in doubt, leave it out, and list it as a benefit you didn't put a number on.

Harbor Logistics is an example company. Its pilot of AI-drafted replies to shipment-status emails comes from how to choose AI use cases. The team of 30 handles about 1,800 such emails a week, or 93,600 a year.

• Measured result: in the pilot group, handling time fell from 6.0 to 4.2 minutes. In the comparison group it fell from 6.0 to 5.9. The AI's share is (6.0 − 4.2) − 0.1 = 1.7 minutes per email.

• Capacity: 93,600 × 1.7 ÷ 60 = 2,652 hours a year. At a loaded $40 an hour, that is $106,080 of capacity.

• Realized value: the team absorbs next year's volume growth without the extra hire it had budgeted ($72,000 loaded), and overtime falls by 200 hours ($8,000). Total: $80,000.

• Guardrails: the review score held at 4.2 out of 5, and the reopen rate fell from 8% to 7%.

Year-one costs:

ROI = ($80,000 − $36,520) ÷ $36,520 = 119%. Counting every freed hour as money would give 190%, which is the number to avoid reporting as savings. Payback: one-time costs of $17,400, divided by a net $5,073 a month ($80,000 ÷ 12 minus running costs of $1,593), is about 3.4 months. If the avoided hire would only have started mid-year, count the payback from then.

A prompt engineering initiative might be training staff to write better requests, building a shared prompt library, or engineers improving the instructions inside an AI product. Prompt engineering changes quality first, so measure it in layers:

Make each criterion specific and measurable. Anthropic's guide to success criteria gives the example of replacing "good performance" with a precise target on a defined test set. Applied here, "the drafts should be good" becomes "at least 90% of drafts pass the five-point checklist on our 50 test emails, editing takes under a minute on average, and no draft invents a delivery date". Prompt metrics are early signals. Prompt engineering ROI is the change in the business metric, priced with Steps 4 and 5, so it counts only when that metric moves.

Review the worksheet at week 8, again at six months, and at every renewal.

What is a good ROI for AI?

There is no reliable universal benchmark, and published averages often lack a method. Compare an AI project with the return your company expects from any investment of similar size and risk, agree that threshold before the pilot, and judge on realized value, not capacity.

How long does it take for AI to pay back?

A focused pilot with a clear unit of work can pay back within months if the value converts to real savings, as in the worked example. Company-wide impact takes longer: in McKinsey's 2026 survey, 37% of respondents attributed any EBIT impact to AI. Track payback per use case, not for "AI" as a whole.

What are the most important AI ROI metrics?

For each use case, pick one quality metric (such as error rate), one speed or volume metric (such as time per task), one money metric (such as cost per unit or spend avoided), and one guardrail that must not get worse (such as customer satisfaction). Compare all of them against a baseline and a comparison group.

How can organizations measure the success of prompt engineering initiatives?

Measure in layers: pass rate on a fixed test set of real cases, the share of AI outputs used without heavy editing, time and cost per task, adoption of shared prompts, and errors caught in review. Then connect those to the business metric of the use case, because prompt improvements count as ROI only when that metric moves.

Should we count time saved as money?

Only the part that turns into money: overtime or contractor spend that stops, a planned hire you don't need, or extra output you sell. Report the rest as capacity. In a large Danish study, 85% of chatbot users said they put saved time into other tasks, which is valuable but doesn't lower costs on its own.

How do we measure the ROI of AI agents?

Use the same method, and add the share of tasks the agent finishes without a hand-off, the cost per task including model usage, and the cost of wrong actions. Compare against the baseline cost per task of the current process. How to evaluate AI agents shows your technical team how to build the test set.

• Pick the use case to measure: how to choose AI use cases.

• Compare build and buy costs: build vs buy AI agents.

• Measure adoption alongside value: how to roll out AI to your team.

• Test prompt changes properly: prompt evaluation for beginners.

• Track model costs per run: AI agent cost tracking.

• Explain agents to your leadership team: AI agents explained for business leaders.