How to Measure AI ROI: A Five-Step Method for Smaller Businesses

To measure AI ROI, pick one workflow, measure what it costs and produces from your own records before and after the AI goes in, compare it against the same work done without AI over the same weeks, and count only the time savings that turn into money you can point to. Then subtract the full cost and read the result at fixed checkpoints. The formula is simple, (measured value − total cost) ÷ total cost, and almost all of the difficulty sits in the "measured value" half, which is where most businesses' numbers go wrong.

The numbers people usually reach for, like hours saved per week or a vendor dashboard's estimate, tend to overstate the return, because they count time that gets eaten by checking the AI's work or that never turns into anything the business can bank. That gap helps explain how widespread AI use can sit beside small reported gains. In a survey of nearly 9,000 executives across seven countries in summer 2026, firms reported that AI had raised their annual labor productivity growth by just 0.2% over the previous three years (Yotzov et al., "Firm Data on AI", NBER Working Paper 34836, 2026). Our guide to why AI projects fail lists "nobody defined the value" among the main reasons projects end.

Key Takeaways

• AI ROI is measured value minus total cost, divided by total cost. Measure value for one workflow at a time, from operating records such as timestamps, invoices, and payroll, never from how fast the tool feels.
• Count net time, not gross time. In Workday's survey of 3,200 employees at large organizations, nearly 40% of AI time savings were lost to rework, and only 14% of employees consistently got clear, positive net outcomes from AI.
• Compare against work done without AI over the same weeks. A before-and-after reading picks up seasons, price changes, and staffing changes that have nothing to do with the AI.
• Freed time is only a return when it turns into money. Danish research linking AI adoption to payroll records found "precise null effects on earnings and recorded hours," with employers absorbing AI through task reorganization instead.
• Read the numbers at 30, 90, and 180 days, and expect the early reads to look worse than the later ones, since the setup work comes first and the payoff follows.
• Executives in a 2026 survey of nearly 9,000 firms reported that AI had added just 0.2% a year to labor productivity growth over the previous three years, a small number set against widespread use and real spending.

Table of Contents

• What is AI ROI, and why is it hard to measure?
• Why most businesses can't tell whether AI paid off
• Step 1: Pick one workflow and measure it from records
• Step 2: Count net time, not gross time
• Step 3: Compare against work done without AI
• Step 4: Count freed time only where it turns into money
• Step 5: Read the numbers at fixed checkpoints
• An 85-person door service company measures its first workflow
• Common mistakes when measuring AI ROI
• Where Avolis fits
• Frequently Asked Questions
• Continue Learning

What Is AI ROI, and Why Is It Hard to Measure?

AI ROI (return on investment) is the value an AI workflow produces, minus everything it costs, expressed as a share of that cost. If a workflow returns $60,000 a year in measured value and costs $30,000 to build and run, the ROI is 100%. The cost side is easier to find on paper, since it shows up on invoices and timesheets, although your own staff's time is the line most often left out, and our guide to choosing an AI consulting service for business ROI covers what belongs in it and how payback periods get quoted.

The value side is hard for three reasons. Time saved is usually reported gross, before anyone subtracts the minutes spent checking and fixing what the AI produced. Changes in the business get credited to the AI when a busy season, a price increase, or a new hire would have moved the same number anyway. And hours saved get counted as dollars even when the freed time went into longer lunches, slower afternoons, or other work nobody valued. A method that handles those three problems gives you a number you can defend to a partner, a lender, or yourself.

Why Most Businesses Can't Tell Whether AI Paid Off

Most businesses can't tell whether AI paid off because they measure the activity instead of the result. They know how many people use the tool and roughly how much faster it feels, and neither of those is a return. The broad research shows the same pattern at every scale: AI use is widespread, and the measured gains are small so far.

In the summer 2026 survey of nearly 9,000 CFOs, CEOs, and senior executives in the United States, United Kingdom, Germany, Australia, Canada, Mexico, and South Africa, 78% of firms used AI, and U.S. firms expected to spend about $2,000 per employee on it in 2026. Yet the executives reported that AI had raised annual labor productivity growth by 0.2% over the previous three years, and had essentially no net effect on employment. They expected a larger boost of 0.6% a year over the next three years (Yotzov et al., 2026). A business spending real money on AI and seeing a fraction of a percent in its numbers isn't necessarily failing, but it can't tell which of its projects are working without measuring them one at a time.

Self-reports don't settle the question either. In a randomized trial run by the research nonprofit METR, experienced software developers took longer to finish tasks with AI tools while believing the tools had sped them up, a finding our page on why AI projects are failing now covers in detail. When people tell you how much time AI saves them, treat the answer as a signal about adoption and measure the return from records.

Step 1: Pick One Workflow and Measure It From Records

Measure AI ROI one workflow at a time, using numbers your systems already record, because a company-wide "AI ROI" mixes a few working projects with several that aren't and hides both. A workflow here means a repeated piece of work with a clear start and finish, such as turning a technician's notes into an invoice, answering a supplier's delivery question, or preparing a monthly close package.

For that workflow, find the records that show how the work runs today. Job management systems timestamp when a job closes and when the invoice goes out, payroll shows overtime by person and week, accounting shows credits and write-offs, and the email inbox shows how long customers wait for an answer. Records like these beat recollection for two reasons: they existed before anyone had an opinion about the AI, and they'll keep being recorded the same way afterward. Two weeks of records is enough to size the opportunity, but for an ROI reading, collect four to eight weeks before launch so ordinary variation shows up, and if a number you need isn't recorded anywhere, start recording it now, before the AI arrives. The four basics (time per task, volume, rework, and loaded labor cost) are laid out in our consulting ROI guide.

Pick one primary measure the workflow is judged on, plus two or three supporting ones. For an invoicing workflow, the primary measure might be days from job completion to invoice sent, supported by staff minutes per invoice and the share of invoices disputed. Writing these down before launch matters more than choosing the perfect set, because measures chosen after launch tend to be the ones that happened to improve.

Step 2: Count Net Time, Not Gross Time

Count the time an AI workflow saves after subtracting the time people spend reviewing, correcting, and redoing its output, because that checking time is real labor and it can erase much of the saving. Gross time saved is the number that shows up in vendor dashboards and in staff estimates. Net time saved is the number that belongs in an ROI calculation.

The gap is large. In a survey Workday commissioned from Hanover Research in November 2025, covering 3,200 full-time employees at organizations with $100 million or more in revenue, 85% of employees said AI saved them one to seven hours a week. Yet nearly 40% of AI time savings were lost to rework, including correcting errors, rewriting content, and verifying outputs, and only 14% of employees consistently got clear, positive net outcomes from AI (Workday, Beyond Productivity: Measuring the Real Value of AI, January 14, 2026). The figures are employees' own estimates, but the direction matches what we see when a workflow's minutes are actually counted.

The practical way to get net time is to measure the whole workflow, end to end, across everyone who touches it. If AI cuts a technician's report-writing time by eleven minutes but the billing coordinator spends four extra minutes fixing the drafts, the workflow saved seven minutes on that piece, and the coordinator's minutes belong in the count even though they happen at a different desk on a different day.

Step 3: Compare Against Work Done Without AI

Compare the AI workflow against the same work done without AI over the same weeks, so the result reflects the AI and not the season, the economy, or a new hire. A simple before-and-after reading can't separate those. If invoices go out faster in April than they did in January, the AI might be the reason, or the January backlog from holiday closures might have cleared on its own.

A smaller business can't run a laboratory experiment, but it usually has a natural comparison sitting in its own operation. A second branch, a second crew, a second shift, or a different job type can keep working the old way for a few weeks while the first one switches. Comparing the change in each group, rather than the level of each, cancels out whatever affected both, a method economists call a difference-in-differences. If both branches got 7% busier and only one got the AI, the AI can claim only what the second branch didn't also gain.

Keep the comparison running long enough to cover at least one ordinary cycle of the work, often four to eight weeks, and write down anything else that changed during that time, like a price increase, a new hire, or a software update. When the comparison group finally gets the AI, its own before-and-after gives you a second reading of the effect, which is a useful check on the first.

Step 4: Count Freed Time Only Where It Turns Into Money

Freed time counts as a return only when it shows up as lower cost or higher output in your books. Saving a technician eight minutes a job is worth nothing on paper if the technician's day ends at the same time, overtime stays the same, and the number of jobs per day doesn't move. This is the single most common way AI ROI gets overstated, because "hours saved times hourly wage" assumes every saved hour was converted into money, and most of them aren't converted unless someone deliberately decides where they go.

Large-scale evidence comes from Denmark. Anders Humlum and Emilie Vestergaard linked surveys of AI chatbot adoption to administrative records on earnings and hours across exposed occupations. Workers reported productivity benefits and new AI-related tasks were widespread, yet the researchers estimated "precise null effects on earnings and recorded hours" at both the worker and the workplace level. Employers absorbed AI through task reorganization instead, adding new work like content generation, AI oversight, and AI integration (Humlum and Vestergaard, "Still Waters, Rapid Currents", NBER Working Paper 33777, revised 2026). By default, time saved by AI gets spent on other work, whether or not that work is worth anything.

There are only a handful of ways freed time becomes money in an operations business, and each one leaves a trace in a record you already keep.

Time that doesn't land in any of these rows counts as zero in the ROI, even if everyone agrees the work feels easier. Record it as spare capacity, and make a decision about where it should go, because unassigned capacity is the easiest return to lose. Our page on the AI skills gap shows what happens when nobody owns a workflow, and our guide to whether hiring an AI consultant is worth it shows how to haircut a time-saved estimate before you buy.

Two of those rows need extra care. Cash collected sooner is a one-time improvement in working capital, so the annual return is what that cash would otherwise cost you (the interest on a credit line, say), not the full amount. And revenue from added capacity only counts once customers have paid for the extra work, because until then it's a forecast of what the freed hours could sell.

Step 5: Read the Numbers at Fixed Checkpoints

Set the dates you'll read the numbers before launch, usually at about 30, 90, and 180 days, and decide in advance what each reading needs to show. Fixed checkpoints stop two opposite mistakes: declaring victory on the first good week and giving up during the first bad month.

Expect the early reads to look worse than the later ones. Economists Erik Brynjolfsson, Daniel Rock, and Chad Syverson describe a "productivity J-curve" for general-purpose technologies like AI. The complementary investments a business makes, such as redesigning the work, writing down its rules, and training people, are costly and poorly measured at first, so measured productivity looks low early and improves once those investments pay off (Brynjolfsson, Rock, and Syverson, "The Productivity J-Curve", American Economic Journal: Macroeconomics, 2021). A single workflow often shows a smaller version of the same shape, with the first month carries the setup time and the correction load at their highest.

The decision rules matter more than the exact dates. If a workflow still isn't covering its running cost at six months, and you can't name the reason it would start to, stopping is a legitimate result, and our guide to recovering a failing AI project covers how to choose between fixing, narrowing, and stopping.

An 85-Person Door Service Company Measures Its First Workflow

Consider an 85-person commercial overhead door and loading dock service company with two branches, North and South (an illustrative composite). Its technicians repair and maintain dock levelers, high-speed doors, and fire doors for warehouses and distribution centers. After each job, the technician wrote a service report, often at home at the end of the day, and a billing coordinator turned the report into an invoice, calling the technician whenever the parts or labor weren't clear. North handled about 250 jobs a week. The median job took seven business days to invoice, about 8% of invoices were disputed, and technicians were logging a combined 60 hours a week of overtime at North, mostly on paperwork.

The company set up an AI workflow at North that turned a technician's dictated notes and photos into a draft service report and invoice lines, which the billing coordinator reviewed before sending. Before launch, the service manager pulled six weeks of job-close and invoice timestamps, overtime by technician, and dispute credits for both branches, and the owner agreed that South would keep working the old way for 16 weeks. The primary measure was days from job completion to invoice sent, with overtime and dispute credits as the supporting measures.

The first surprise was the size of the saving. Technicians said the new process saved them about 20 minutes a job. The timestamps showed report writing falling by 11 minutes and invoice building by 4, but the coordinators were spending about 4 minutes a job fixing drafts that misread part numbers or missed a second trip, and callbacks fell by 2 minutes on the technicians' side. The net saving across everyone was 13 minutes a job, not 20, and none of it was the coordinators' (their faster invoice building was used up by the corrections).

The second surprise came at the 30-day read: overtime at North hadn't moved. Technicians were still finishing the day's paperwork at home out of habit, just faster. Two weeks later, the service manager changed the routine so each report was dictated and submitted from the truck before leaving the site, and North's paperwork overtime fell over the following four weeks while South's stayed flat.

By the 180-day read, the company could value three of the conversion rows and had decided to leave the fourth at zero. Here's the arithmetic, all of it illustrative:

• Overtime. North's paperwork overtime ran about 18 hours a week below South's. At a loaded overtime rate of $50 an hour over 48 working weeks, that's about $43,200 a year.
• Disputes. North's disputed invoices fell from about 8% to 5% while South's stayed near 8%, cutting credits by about $400 a week relative to South, or about $19,200 a year.
• Faster invoicing. The median time to invoice fell from seven business days to three at North, pulling about $120,000 of receivables forward once (four days of North's roughly $150,000 in weekly billing). At the 8% rate on the company's credit line, that's worth about $9,600 a year, not $120,000.
• More jobs per technician. Jobs per technician-day rose 9% at North, but they also rose 7% at South with no AI, and two points over a single season was too small to separate from noise. It went into the ROI at zero, and onto the owner's list as capacity to sell.

Annualized at the 180-day run rate, that's about $72,000 a year in measured value. Year-one cost came to about $35,400: $24,000 to build and connect the workflow to the job management and accounting systems, $7,800 a year in software, and about 80 hours of staff time during setup, valued at $3,600. ROI on those numbers is roughly 100%, a little lower for the first calendar year because the overtime saving didn't start until week 10. Counted the usual way, with the technicians' 20-minute estimate multiplied across all 250 jobs a week at the $50 rate, the same workflow would have shown about $200,000 a year, close to three times the measured value and an ROI above 400%, and the owner would have had no way to tell which of the two numbers was real. When South switched over in week 17, its overtime fell by a similar amount, which gave the owner a second reading pointing the same way.

Common Mistakes When Measuring AI ROI

The most common mistakes in measuring AI ROI all inflate the return, which is why projects can look successful in a quarterly update and still get cut when the budget gets tight. Each one has a simple fix.

1. Counting adoption as return. Usage tells you people are trying the tool. It says nothing about whether the work got cheaper or better, so treat it as an early checkpoint, not a result.
2. Taking the vendor's dashboard at face value. "Hours saved" in a dashboard is usually gross time, calculated from an assumed minutes-per-task figure. Rebuild it from your own timestamps and include the checking time.
3. Annualizing the best week. A strong week multiplied by 52 ignores the slow months, the correction load, and the weeks the tool was down. Use the average across a full checkpoint period.
4. Counting cash pulled forward as profit. Faster invoicing improves cash flow once. Its annual return is the financing cost you avoid, which is a much smaller number.
5. Averaging everything into one company number. A blended "AI ROI" lets one strong workflow hide two that should be stopped. Report each workflow separately, and add them up only after each has its own number.

Citation-ready summary: To measure AI ROI, measure one workflow at a time from operating records, subtract the time spent checking and fixing AI output, compare the change against work done without AI over the same weeks, and value freed time only where it shows up as lower overtime, avoided hiring, faster cash, fewer errors, or billed added work. ROI is that measured value minus total cost, divided by total cost, read at roughly 30, 90, and 180 days. In Workday's survey (fielded November 2025), nearly 40% of AI time savings were lost to rework, and Danish register data found precise null effects of chatbot adoption on earnings and recorded hours, so gross time-saved figures routinely overstate the return.

Where Avolis Fits

Every engagement starts with our diagnostic, and measurement starts there too. We work through each candidate workflow with the people who run it, measure volumes, time per task, and handoffs, and rank where AI would actually pay off, so the starting numbers for whichever workflow gets built are in hand before anything is built. You keep that ranked result whatever you decide next.

When we build, we agree the primary measure, the comparison group, and the checkpoint dates with you before launch, and we pull the numbers from your own systems rather than from a tool's dashboard. When we stay on after launch, we read the 30-, 90-, and 180-day numbers with you, and when a workflow isn't converting its freed time into anything, fixing that is part of our work, whether the answer is a changed routine, a narrower scope, or stopping. Our guide to AI readiness assessment services explains what a good diagnostic should produce, and our guide to which AI consulting company to choose covers how to tell a partner who will be measured on the result from one who reports activity.

Frequently Asked Questions

How do you measure the ROI of AI?

Pick one workflow, measure it from your own records before and after the AI goes in, and compare the change against similar work done without AI over the same weeks. Subtract the time spent checking and fixing AI output, value freed time only where it lowers cost or raises billed output, then divide the measured value minus total cost by total cost.

What is a good ROI for an AI project?

There's no reliable benchmark for a single workflow, because results depend on the work and how much of the freed time gets used. A practical test is whether the measured value covers the running cost by about six months, with a credible path to covering the build cost. Broad results are modest so far: executives in a 2026 survey of nearly 9,000 firms reported AI had added 0.2% a year to productivity growth.

Why do AI time savings not show up in profits?

Usually because the saved time is never converted into lower cost or more billed work. Danish research linking AI adoption to payroll records found "precise null effects on earnings and recorded hours," with employers absorbing AI through task reorganization instead.

How long does it take to see ROI from AI?

Plan on about 90 days to see a clear signal against a comparison group and around six months for a dollar figure you can defend. Early readings tend to look worse than later ones, because setup work, rule writing, and corrections are heaviest in the first month, a small-scale version of what economists call the productivity J-curve.

Can a small business measure AI ROI without a data team?

Yes. Most of the records needed already exist in the job management system, payroll, and accounting software. The work is choosing one primary measure before launch, keeping a branch, crew, or job type on the old process for a few weeks as a comparison, and reading the numbers on fixed dates.

Continue Learning

Measuring AI ROI comes down to three subtractions that most reported numbers skip: the checking time that eats into gross savings, the change that would have happened anyway, and the freed time that never turned into money. A smaller business can handle all three with records it already keeps, one comparison group, and a few dates on the calendar, and the result is a number that holds up when someone asks how you know.

If you're about to launch an AI workflow, write down its primary measure and pick its comparison group this week, before the first draft goes out.

Why projects fail, in detail:

• Why do AI projects fail? Common reasons and statistics
• Why are AI projects failing?
• The AI skills gap: what small businesses are actually missing
• How to recover a failing AI project

Starting the right way:

• AI readiness assessment services
• Which AI consulting company should I choose?
• How to choose an AI consulting service for business ROI
• Is it worth hiring an AI consultant for a small business?

---

Sources

• Ivan Yotzov, Jose Maria Barrero, Nicholas Bloom, Philip Bunn, Steven J. Davis, Brent H. Meyer, Paul Mizen, Gregory Thwaites, and others, "Firm Data on AI," NBER Working Paper 34836, 2026, doi:10.3386/w34836, retrieved 2026-10-30: https://www.nber.org/papers/w34836
• Workday, "New Workday Research: Companies Are Leaving AI Gains on the Table," press release on the report Beyond Productivity: Measuring the Real Value of AI, 14 January 2026, retrieved 2026-10-30: https://newsroom.workday.com/2026-01-14-New-Workday-Research-Companies-Are-Leaving-AI-Gains-on-the-Table
• Anders Humlum and Emilie Vestergaard, "Still Waters, Rapid Currents: Early Labor Market Transformation under Generative AI," NBER Working Paper 33777 (previously circulated as "Large Language Models, Small Labor Market Effects"), doi:10.3386/w33777, retrieved 2026-10-30: https://www.nber.org/papers/w33777
• Erik Brynjolfsson, Daniel Rock, and Chad Syverson, "The Productivity J-Curve: How Intangibles Complement General Purpose Technologies," American Economic Journal: Macroeconomics 13(1), 2021, retrieved 2026-10-30: https://www.aeaweb.org/articles?id=10.1257/mac.20180386

On the NBER firm survey. The figures are from the abstract of the current version, which covers nearly 9,000 CFOs, CEOs, and senior executives in seven countries surveyed in summer 2026. An earlier version of the same paper (February 2026) covered about 6,000 executives in four countries and reported different figures, so check the version before quoting it. The 0.2% and 0.6% figures are executives' own estimates of AI's effect on annual labor productivity growth.

On the Workday figures. The survey was fielded by Hanover Research in November 2025 among 3,200 full-time employees at organizations with $100 million or more in annual revenue in North America, Asia-Pacific, and Europe, the Middle East, and Africa. The time-saved and rework figures are respondents' own estimates. Workday sells AI-enabled software.

On the Danish study. The "precise null effects" quotation and the task-reorganization finding are from the abstract of the current NBER version, which estimates effects using difference-in-differences and rules out effects larger than 2% two years after the launch of ChatGPT.

On the door service company. It's an illustrative composite, not a client. Its figures (85 people, two branches, 31 and 18 minutes per job, 20-minute estimate, 60 and 42 overtime hours, 8% and 5% disputes, seven and three business days, $150,000 weekly billing, $50 overtime rate, 8% credit line, $72,000 in value, $35,400 in cost) are representative examples, not measurements or benchmarks.

---

About Avolis Research Group

Avolis Research Group is Avolis's in-house research practice, focused on how operations-heavy small and mid-sized businesses actually adopt AI. It synthesizes primary economic research, government survey data, and results from real implementations into practical, vendor-neutral guidance.

More about Avolis and how we work · Get in touch