Measuring AI ROI comes down to six moves: set a baseline before the pilot starts, measure each kind of payoff separately, subtract the work AI adds, check that quality held up, review a short scorecard every month, and decide on a date whether to widen, change or stop.
Skip the first or third step and you can end up unable to say whether AI is working, even when everyone feels busier and faster. The hard part is that saved time is the benefit people notice, and saved time isn’t money until something changes because of it.
This is for owners, CEOs and the finance and operations leaders who are paying for AI tools and have to decide whether to expand them, change how they’re used or stop paying.
Why AI ROI is so hard to pin down
Three findings show why.
People feel more productive than the numbers show. McKinsey’s 2026 survey found that about 80% of respondents said AI had improved their individual productivity, while only about 37% of organizations reported any positive contribution from AI to EBIT, their operating earnings.
Feeling faster and being faster aren’t the same. In a 2025 randomized study, METR found that experienced software developers using early-2025 AI tools took 19% longer to finish real tasks, while believing afterward that AI had made them about 20% faster. When METR tried to repeat the study in late 2025, it ran into a new problem: many developers no longer wanted to work without AI, and some chose not to submit tasks they didn’t want to do without it. METR called the new data an unreliable signal. Once a team won’t work without the tool, there’s no clean “before” left to compare against.
Time saved doesn’t automatically become money. A large 2025 study of Danish workers in AI-exposed jobs by economists Anders Humlum and Emilie Vestergaard, published by the National Bureau of Economic Research, found that people reported saving time, and employers created new tasks like overseeing AI output, but there was no measurable effect on earnings or hours worked two years after ChatGPT launched.
AI can still pay off. The payoff just has to be measured deliberately, because it won’t announce itself.
Step 1: Write down how the job works today
If you haven’t chosen which job to try AI on yet, how to implement AI in your business walks through picking one.
Before a pilot starts, write down how the job works today, described by the person who does it best:
- Time per job, timed on several real instances.
- Review time, including whoever checks the work.
- Error and rework rate, how often the work comes back.
- Cycle time, how long from start to finished, in calendar days.
- The quality standard, what good looks like, in writing.
Do this before anyone starts using the tool on that job. Afterward, people’s memory of how long things used to take is unreliable, and as METR found, a team that has adopted the tool may not want to go back long enough to measure.
Step 2: Measure the payoffs separately
They get blurred together, and they shouldn’t be.
Capacity back. Hours freed up. This is a common place for AI value to show up first, and it turns into money only through what the hours go to: a hire you didn’t make, a contractor you stopped paying, more time actually spent selling or on billable work. Record what the freed hours went to, as well as how many there were.
More output. More proposals sent, more customers served, more billable work delivered, without adding the same amount of labor. Count the extra output first, then connect it to revenue or gross profit where you honestly can.
Cash saved. A cost that actually leaves the books, like a subscription you cancel or outside spend you no longer need. This one is easy to verify.
Better work. Higher quality, fewer errors, better decisions. Measure it with a sampled review by whoever owns the standard: pull a handful of finished pieces each month and judge them against the written standard from step 1.
Faster delivery. Work that finishes sooner on the calendar. When my agency started building website wireframes with AI in a working session, it didn’t save hours on paper, but calendar time compressed from one to two weeks of scheduling gaps down to a focused session and a few days of refinement.
The biggest leak is between the first two. BCG’s 2026 survey of workers found that 66% got limited or no guidance on what to do with the time AI saved them. If nobody decides what the time is for, it disappears into busywork, and the capacity never turns into anything you can find on a financial statement.
Step 3: Count what AI adds
AI creates work as well as saving it. Count it as line items:
- Setup. Building, configuring and connecting the tool.
- Review. Time spent checking AI output. Google’s DevOps research team calls this the “verification tax.”
- Corrections. Fixing what the AI got wrong.
- Training. Time people spend learning to use it well.
- Upkeep. Maintaining the instructions, examples and connections as things change.
- Licenses. The subscriptions themselves.
Expect some of the costs to come first. Google’s 2026 DORA guidance on AI-assisted software development describes a J-curve: productivity can dip while teams learn the tools and workflows adjust, before the gains show up. A cost line that only counts licenses undercounts. In BCG’s 2024 research, most of the challenge in implementing AI came from people and process, not technology.
Step 4: Time it instead of asking about it
The easiest way to fool yourself is to survey people. In a nationally representative U.S. survey published by the St. Louis Fed in 2025, people who used generative AI reported saving about 5.4% of their work hours. That’s what people estimated, which makes it a starting signal and nothing more.
Time the job instead, the same way you did for the baseline, on real work, including review and corrections. If the measured number and the reported number disagree, believe the measured one.
Step 5: Check quality every month
If the finished work is faster and worse, it doesn’t count.
A 2023 study of AI in customer support by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, covering more than 5,000 agents, found that an AI assistant raised the number of issues resolved per hour by 14% on average and 34% for newer, less-skilled agents, with little effect on the most experienced. Two lessons for measurement: the researchers tracked quality signals as well as speed, and an average can hide big differences between people. Measure by role as well as in total.
The opposite case is just as instructive. Klarna said in 2024 that its AI assistant was doing the work of 700 customer service agents. By 2025 its chief executive said the chatbots were cheaper than human staff but produced “lower quality” work, and the company began hiring people again.
Two quality warning signs to watch for inside your own company:
- Approval turning into a rubber stamp. As volume rises, the person approving AI output may stop really checking.
- Quiet rewrites. If people are taking AI drafts and rewriting them from scratch, the time savings on paper aren’t real.
The ROI math, once the measuring is done
Once you’ve measured, the formula itself is ordinary:
- AI ROI = (financial benefit you can honestly attribute to AI, minus total AI cost) ÷ total AI cost × 100
- Payback period = total AI investment ÷ monthly net financial benefit
Total cost means everything from step 3, with people’s time valued at what it costs you. Financial benefit means cash saved plus the gross profit on any extra output. Keep capacity, quality and cycle time on the scorecard even when you can’t responsibly turn them into dollars. Leaving them out of the formula is more honest than forcing them in.
A worked example
Here’s an illustrative example with made-up numbers, showing the common mistake and the honest version.
A salesperson writes ten proposals a month. The baseline shows each takes about six hours, or 60 hours a month. With an AI tool, drafting drops to about three and a half hours, plus an hour of review, so four and a half hours per proposal, or 45 hours a month.
The calculator version: 15 hours saved, times a loaded cost of $100 an hour, equals $1,500 a month in savings.
The honest version: The tool frees about 15 hours a month. Subtract about two hours of upkeep to keep its instructions and examples current, and you have roughly 13 hours of net capacity. Separately, the company pays $100 a month for the license. Now ask what happened to the 13 hours. If the salesperson used them to send two more proposals a month, and your close rate turns some of those into deals, the gross profit on those deals is the financial benefit, and the license plus the upkeep time, valued at what it costs you, is the cost side of the formula. If the hours went to email, the savings are on paper only. And if a sampled review shows the proposals got worse, the number doesn’t count at all.
The difference between those two versions is the whole discipline.
A simple monthly scorecard
You don’t need software for this. One page, reviewed monthly, with a handful of lines:
- Hours freed on each job in the pilot, measured against the baseline.
- What the hours went to, in plain words.
- Cash saved, if any.
- Quality, from the sampled review.
- Cycle time against the baseline.
- Costs added: licenses, review time, upkeep.
- Usage, as context only. People logging in isn’t value.
Give each line an owner and a written definition, so the numbers mean the same thing every month.
When to stop
Set the stop rule before the pilot starts: a review date after the learning period, and the threshold that would justify widening it. Then stop or change course if:
- review and corrections eat most of the time saved
- quality drops and doesn’t recover
- nobody uses it once the novelty wears off
- the freed time isn’t going anywhere you can name
A stopped pilot still tells you which job AI isn’t ready for, and that’s worth knowing before a wider rollout.
Common questions about AI ROI
Is there any ROI on AI?
Yes, but less than the headlines suggest and more slowly. In IBM’s 2025 survey of chief executives, only about a quarter of AI initiatives had delivered the expected return. In MIT Sloan Management Review’s 2026 interviews with more than 30 senior leaders, companies without an explicit way of measuring AI’s return rarely saw credible, lasting gains.
How do you calculate AI ROI?
Measure the job before the pilot, then compare. ROI is the financial benefit you can honestly attribute to AI, minus its total cost, divided by that cost. Count cash saved and the profit on extra output as benefit; count setup, review, corrections, training, upkeep and licenses as cost. Don’t multiply hours saved by a wage and call it savings.
How long does it take for AI to pay off?
You may see whether a narrowly defined pilot is improving the job within weeks, as long as there’s enough real work to compare. Financial payback usually takes much longer. In Deloitte’s 2025 survey of senior executives in Europe and the Middle East, most reported satisfactory returns on a typical AI use case in two to four years, and only 6% saw payback within a year.
What is the 10/20/70 rule for AI?
It comes from BCG’s 2024 research: roughly 10% of the challenge in getting value from AI is the algorithms, 20% is the technology and 70% is people and process. The practical point is that most of the cost and effort is in changing how people work.
Do most companies measure AI ROI?
Most large ones say they do. In a 2025 Wharton survey of senior leaders at large U.S. companies, 72% said they formally measure the return on generative AI. Measuring it well is another matter.
Where to go from here
Pick one repeatable job. Time ten real examples done the current way, then the next ten with AI, including review and corrections. Compare time, quality, cycle time and cost, and write down what happened to any time you freed. Do that before the tool goes any wider, then keep a one-page scorecard every month until you can say clearly whether it’s paying off.
If you’d like help choosing the right first jobs, measuring them properly and deciding where AI earns a wider rollout, that’s the work I do as an AI implementation consultant. You can start a conversation on Work With Me.
