Decide the number before the tool, then track time returned, error and rework, and cycle time or revenue effect against a baseline. Leave out vanity metrics and model benchmarks.
Executives should measure AI ROI by deciding the number before choosing the tool, then tracking three measures against a baseline: time returned, error and rework rate, and cycle time or revenue effect. Usage counts, enthusiasm, and model benchmarks are context at best. The return is real only when a named workflow costs less, moves faster, or produces a better result than it did before, and the person accountable for that workflow agrees the change counts.
The question matters because access has outrun evidence. Stanford's 2026 AI Index reports organizational AI adoption at 88%. Adoption is not return. Knowing how many people can open an AI tool does not tell you which piece of work became cheaper, faster, or more reliable because of it. The dream outcome for an executive is simpler than a dashboard: being able to say, with confidence, something like "this workflow now takes half a day instead of two, with fewer corrections, and here is the before and after."
Start with one recurring workflow, one owner, and one number the business already cares about. Record that number before any AI is involved. Without a baseline, every later claim is an estimate, and estimates drift toward what people hope is true.
Take a hypothetical example. A finance lead spends two days each month writing variance commentary for the board pack. Before changing anything, write down the hours per cycle, how many corrections the CFO makes in review, and how many days the pack takes from close to distribution. Then agree on what improvement would count as success and what result would make you stop.
Only then choose the tool.
Measure hours per cycle before and after, and include the time a person spends checking and correcting the AI's output. Review time is part of the cost. A draft that arrives in minutes but takes an hour to fix has returned less than it appears.
Measure the work itself rather than asking people how much time they saved. A study published by the National Bureau of Economic Research followed 5,179 customer support agents and measured productivity as issues resolved per hour: access to an AI assistant raised it 14% on average, with a 34% improvement for novice workers and minimal impact on the most experienced. The gain was real and uneven, which is why one average across a team can mislead.
Self-reports can be wrong in the other direction too. In METR's 2025 study of 16 experienced software developers, participants expected AI to make them 24% faster and afterward believed it had made them 20% faster, while their measured completion times were 19% longer. METR now describes those results as out of date. In our judgment, the measurement lesson still holds: a feeling of speed is not a baseline.
Count the corrections, the errors caught in review, and the times work goes back for another pass. Faster output with more rework can produce no return at all, and in finance, legal, or customer-facing work, one uncaught error can outweigh months of saved hours. If the rework rate rises, treat it as a result, not noise.
Cycle time is the elapsed time from start to decision: days to close the books, hours to a client proposal, time from question to board-ready answer. It often matters more to an executive than hours saved, because it changes when a decision can be made.
When a workflow touches revenue, such as proposal turnaround or response time to inbound leads, you can track the revenue effect as well. Be careful with attribution. Revenue moves for many reasons, so compare against the same period before the change and name the other factors that shifted.
Some numbers feel like progress and say little about return:
Ninety days covers several cycles of a weekly or monthly workflow:
Our guide to what five hours of private AI coaching should produce applies the same standard to coaching itself: a reviewable result, honest constraints, and a clear next step. For executives who want a structured 90 days with our team, Aravise 90 is the 90-day private AI program for executives and owners.
AI can gather the numbers, compare periods, and draft the summary. It should not decide what counts as value, which trade-off between speed and quality is acceptable, or what the board is told about the return. NIST's AI Risk Management Framework treats measurement as continuing work: systems should be tested before deployment and regularly while in operation. A measure that nobody owns tends to disappear when the quarter gets busy.
Our team at Aravise AI helps an executive choose the workflow, set the baseline, and keep the number visible between sessions, so the measurement survives a full calendar. The effort asked of you stays small: the outcome you care about, access to the work, and your judgment on what the result is worth. You keep the decision, the sign-off, and every statement about return.
There is no universal number. A good return is one that clears the threshold you set before you started, after counting the tool, the review time, and any rework. Judge it against your own baseline, not against a vendor's average.
Time returned and rework usually show within a few cycles of a recurring workflow, so a monthly process needs at least two or three months. Revenue effects take longer and need more care to attribute. No responsible adviser can promise a date for a result that has not been measured yet.
Per workflow. A tool can return a great deal in one workflow and nothing in another, and per-tool totals hide both. Measure the workflow, then ask which tool, if any, made the difference.
Stop or adjust, and treat the result as useful. Check whether the cause is the workflow choice, the quality of the underlying information, or the review burden. A clear "no" after 90 days costs far less than an unmeasured "maybe" that runs for a year.
If you have a workflow in mind and are unsure what to measure, bring it to the free, private 15-minute AI Starting Point with our team. We will look at where you are with AI today and find your next step together.
Where this kind of work shows up across a company, scored by ease, impact, and risk.
Turn closed, reconciled actuals into first-draft variance commentary for the board pack, so your finance team edits a draft instead of a blank page.
See the use caseFinance & AccountingUse AI to match transactions, surface only the real exceptions and draft reconciliation write-ups each close, with a person reviewing every result.
See the use caseOperationsTurn meeting transcripts into clear action items and well-formed tickets in your tracker, with a person approving before anything is created.
See the use caseBrowse all AI use cases by department
A short video that shows the steps on screen.
Technology Leadership
AI can help a CIO prepare an evidence-anchored investment decision brief while business priority, risk acceptance, accountability, capital allocation, and approval remain human-controlled.
Read the answerExecutive AI Coaching
What makes a short paid coaching engagement worth the time: a reviewable result, honest constraints, current-tool fit, and a clear next step.
Read the answerStart with a free, private 15-minute conversation with our team. Bring a task or question; we’ll look at where you are with AI today and find your next step together.