How to measure AI ROI: a method a board can trust

Boards that ask how to measure AI ROI are often shown hours saved, and hours are the wrong unit. Workday found that nearly 40% of AI time savings are lost to rework, as staff correct errors, rewrite drafts and check output. Even the hours that survive count only when someone uses them to cut a cost or grow revenue.

The fix is to measure money, not activity. Take a baseline before launch. Compare a team that uses the system with a team that does not. Count the full running cost, including the people who check the output. Count a benefit only when a cost line falls, revenue rises or a planned hire is no longer needed.

Few firms do all four: MIT Sloan Management Review research found few apply the same financial discipline to AI as to a new factory. Results are a separate worry. In 2025, CEOs told IBM that only 25% of AI initiatives had delivered the expected return in recent years.

What is AI ROI, and what should it measure?

AI ROI is the money an AI system earns or saves, minus everything it costs, divided by what it costs, over a set period. ROI stands for return on investment. A system that costs one million over three years and brings back two million has returned its cost and the same again.

Two more measures matter as much to a finance director:

• Payback: the number of months until the money the system has saved or earned covers what it has cost so far.
• Net present value: total benefit minus total cost, with later years discounted, because money in year three is worth less than money today.

A large ROI that arrives in year four can lose to a smaller one that arrives this year. That is why a business case needs all three measures, not ROI alone.

The cost side is the easier half. The benefit side is where most cases fail. Research in MIT Sloan Management Review, based on interviews with CEOs and senior leaders, found that few companies apply the same financial discipline to AI as they would to a new factory or piece of machinery. The method below closes that gap.

How to measure ROI of generative AI, analytics and AI agents

The kind of AI changes what you can measure.

• Analytical AI, such as demand forecasting or fraud scoring, predicts or optimises one decision you already track. Its return is usually the easiest to trace.
• Generative AI, which writes, summarises and drafts, mostly makes work faster, better or larger in volume. MIT Sloan notes that this kind of gain requires deliberate translation into financial impact.
• AI agents, software that plans and carries out multi-step tasks inside company systems, can remove whole steps of a process. Measure them by the finished task, not the hour saved. They also cost more to run and to check: of the firms in Deloitte's survey that already use agentic AI, just ten per cent said they were realising significant ROI from it.

For AI aimed at sales, the unit of work is a lead or a deal. The comparison group is a set of sales staff, regions or campaigns without the tool over the same period, so a strong market is not credited to AI. The benefit is the gross margin on the extra sales, not the revenue itself, because those sales still carry their own costs. Anthropic found its sales staff closed deals with more clients after adopting agents, and warned that reps with more sales opportunities may have adopted first.

Why is AI ROI so hard to measure?

AI ROI is hard to measure for three reasons: AI rarely arrives alone, its costs move with use, and its output is easy to count but hard to value.

AI arrives with other change. A new system usually comes with cleaner data, a redesigned process or a new team structure. Any of them can move the numbers. Attribution, crediting a gain to its true cause, is the hardest step. An executive at an energy and industrials company told Deloitte:

"We only managed to get a ballpark estimate of the benefits because it was hard to separate the gains from AI initiatives from those of other initiatives, like operational excellence, team reorganisation or changing roles."

AI costs move with use. Use through an API, the direct link between your software and an AI model, or through an AI agent bills by the token, a unit of text of about three quarters of a word, so the bill rises with every task. Seat licences for staff assistants bill a fixed fee per user instead, so the bill rises with every person added. The business case must price both at full volume, not pilot volume.

Vendors have noticed. On IBM's third-quarter call on 21 October, chief executive Arvind Krishna said its Apptio software now connects AI spending with business results, providing visibility into usage, token consumption and operating costs. His wider point: "Deploying more models or building more agents will not be enough."

AI output is easy to count and hard to value. Anthropic studied its own staff and found output per person rose sharply as people adopted AI agents. Its researchers still warn that more code does not necessarily mean better code or business impact. The same study found that meta-work, the time staff spend instructing, checking and monitoring agents, rose from 10% to 22% of their AI sessions between autumn 2025 and spring 2026. That work is a real cost, and most business cases leave it out.

Patience is also running out. Deloitte's survey of 1,854 executives across Europe and the Middle East found that most firms see a satisfactory return on a typical AI use case only within two to four years. They expect seven to 12 months from other technology. Boards that funded early AI on faith now want the number.

How to calculate AI ROI, step by step

Calculate AI ROI one workflow at a time, on a unit of work, against a measured baseline and a comparison group. This is the measurement framework in eight steps.

1. Pick one workflow and its unit of work. An invoice paid, a claim settled, a contract reviewed, a support case closed. Cost per unit lets you compare before and after even when volume changes.
2. Measure the baseline before launch. Record time per unit, cost per unit, error rate and volume over at least one full business cycle, usually a quarter. The baseline is your measured before. Without it, any later number is a guess.
3. Set up a comparison group. Give the system to some teams or sites and not others, over the same period. One practitioner quoted by CIO puts it plainly: "We baseline the pre-AI process, then run controlled rollouts so every metric has a clean counterfactual." A counterfactual is what would have happened without the AI.
4. Price the full cost. Include build, integration, licences, usage fees, hosting, monitoring, security review, training, the staff who check output and the cost of measuring. Our guide to AI cost management covers the cost side in detail.
5. Subtract rework and review. Gross time saved, minus the time spent correcting and checking output, gives net time saved.
6. Turn time into money only where a decision uses it. Released hours are worth money when they avoid a planned hire, end a contractor or outsourcing contract, cut overtime, absorb growth without hiring, or win revenue. Name the decision and the budget line before launch.
7. Adjust for risk. Reduce the benefit for errors that reach customers, for incidents and for a high override rate: how often staff reject or change the system's output.
8. Calculate ROI, payback and net present value over three years. Let the benefit ramp up. It rarely starts at full value on day one.

Anthropic's study shows the comparison group at work. Its business was growing fast, and as the researchers note, "It is easier to close deals when you sell a strong model." So they compared 192 engineers who adopted coding agents with 74 who did not, over the same months, rather than credit the tools with all the growth. They also flagged that adoption was not random: staff about to start a big project may adopt first, which would flatter the result. Your measurement plan should name the same risk.

AI ROI metrics to track at each stage

Track six groups of metrics, in this order, because each one only matters once the one before it holds.

Use and quality are leading indicators: they move within weeks and warn you early. Unit cost and money are lagging indicators: they move over quarters and prove the case. The CIO guide quotes one practitioner who tracks adoption by active users "because an unused model has zero ROI."

A recent AWS case study on Bayer Crop Science shows what good operating metrics look like. Bayer's troubleshooting agent cut a diagnosis that once took over an hour of manual log-hunting to under 10 minutes. In its first phase, the platform resolved about 80 percent of user questions without an engineer in the loop. These are strong operating results. They become ROI only when finance asks what the engineers now do with the time, and whether a budget line moved.

What people get wrong about AI ROI

Four mistakes recur. Each one inflates or hides the real return.

Hours saved are not money saved. Time saved is an input. It becomes a return only when it changes a cost, a hire or a sale. Workday found that only 14% of employees consistently get clear, positive net outcomes from AI. Many of the rest lose much of the time they save to fixing output. Count net hours, then count only the hours a decision uses.

The famous failure rate measures something narrower than it sounds. The MIT NANDA study reported by Fortune found that about 5% of AI pilot programmes achieve rapid revenue acceleration, and that for 95% of companies in its dataset generative AI was falling short. That measures fast, visible profit impact, not whether the tools work. The same research found the biggest ROI in back-office automation, such as replacing outsourced processing and agency work. That fits the method above: those costs already sit on a budget line, so the saving is easy to prove.

A good pilot does not prove the case. Pilots run on clean data, keen users and close attention. Deloitte warns that proofs of concept run on unrealistic dummy data are often a source of optimism, and that problems emerge once they meet real data. A proof of concept is a small trial build. The CIO guide puts it more bluntly: "AI ROI almost never scales cleanly." Check every assumption, and the baseline, again before you scale.

Fear of falling behind is not a business case. IBM's survey found that 64% of CEOs say the risk of falling behind drives investment in some technologies before they have a clear understanding of the value. Speed is a fair reason to test quickly and cheaply. It is not a reason to skip the baseline, because without one you cannot tell a winner from a loser later.

How to decide which AI projects to fund

Fund AI in two lanes, and hold each lane to a different test.

Cisco's senior vice president of corporate finance, Andrew Ashton, told EY that AI spend is additive rather than a substitute for existing budget. So every AI project competes with something else the business could fund. Cisco approves some pilots "based on a very finite kind of business case" and runs others as experiments, to learn where AI delivers value before it commits more.

• Business-case lane. The project names a budget line, a baseline, a comparison group and a payback date. Fund it when its risk-adjusted payback clears your hurdle: the minimum payback or return the business demands before it funds a project. Hold projects that remove an existing cost to your normal technology hurdle. For the rest, write down a longer limit before you fund them, because most firms in Deloitte's survey report two to four years to a return on a typical AI use case.
• Learning lane. The project tests an idea whose value is unclear. Cap its budget at an amount you are willing to write off, give it a fixed end date, and judge it on what it teaches. It moves to the first lane only with a measured result.

Before you fund either lane, ask the team or the vendor these questions:

1. What is the unit of work, and what does one cost today?
2. Which budget line moves, and who in finance signs for it?
3. What is the comparison group?
4. What does the system cost at full volume, not pilot volume?
5. Who checks the output, and how many hours a week does that take?
6. What error rate or override rate stops the project?
7. When does it pay back, and is it still worth doing if that slips by a year?
8. Who runs it after launch, and from which budget?

A team that cannot answer the first two questions is not ready for the business-case lane. Start where cost already sits on a budget line: outsourcing, agency fees, overtime, contractor spend and planned hires. Back-office work like this gives the cleanest proof.

How to run a pilot that measures AI ROI

A pilot that measures ROI is a controlled test with a written plan, not a demonstration. Run it in this order.

1. Write the measurement plan first. Name the unit of work, the baseline, the comparison group, the decision that will bank the benefit and the stop rules. Finance signs it before any build starts.
2. Collect the baseline. Measure the current process for a full cycle. Where system records exist, use them, and check them against a few weeks of direct observation.
3. Launch to a small, gated group. Bayer launched its platform to a gated group first, rather than to everyone at once. A small group keeps the comparison clean and the risk contained.
4. Measure every week. Track use, rework, override rate and cost per unit. Count failed tasks, not only system errors: as the AWS write-up notes, an answer can be wrong even when every technical step succeeded.
5. Check the result against the comparison group at the end. Calculate the net benefit with the eight steps above, and have finance check the numbers.
6. Decide: scale, fix or stop. Scale only if the result holds and the decision that banks the benefit has been made, such as a revised hiring plan or a contract notice.
7. Keep measuring after launch. Usage costs grow with adoption, and model updates change output quality. Review cost per unit and quality every quarter.

Risks to an honest ROI figure, and the control for each

Most bad AI ROI figures are not lies. They are honest numbers built on a weak method. These are the common weak points.

• Attribution creep. Gains from other changes get credited to AI. Control: a comparison group, and a record of which steps the machine did, which a person checked and which a person improved.
• Self-reported gains. Staff surveys flatter new tools. CIO describes one firm where 73% of employees said a new tool improved their productivity, though the gain had not been objectively measured. Control: treat surveys as an early signal, never as a benefit line.
• Selection bias. Your keenest people adopt first, so early results flatter the tool. Control: compare like with like, and expect results to fall as use spreads.
• Cost drift. Usage bills grow faster than plans. Control: track cost per unit of work every month, with a spending cap per workflow.
• Hidden human work. Instructing, checking and monitoring AI takes staff time. At Anthropic, written standing instructions for agents grew from a few hundred to 8,500 in a year. Control: record that time and count it as cost.
• Vendor-reported results. A vendor's case study uses its own baseline. Control: use your own baseline, measured by your own team, and ask for the vendor's method in writing.
• Quality drift. Model updates change output without warning. Control: a fixed test set, run again after every change, with a quality floor that pauses the system.

Where Thriiv fits

Thriiv works from the same principle: a quantified saving before budget is committed. We find where the operating cost sits, measure it, and price our work against the value it returns, not the hours it takes. A working pilot follows in weeks, not quarters. We are paid against that value, and we host and run what we ship, so the cost and the benefit stay measured after launch.

For Anglo, a custom design automation system cut design time from ten hours to ten minutes per design, a reduction of 98%. To test an AI business case against this method, tell us where the cost sits.

What to do this quarter

• List every live AI project with the budget line it moves. Any project without one gets a measurement plan or an end date.
• Baseline one workflow now, before any new build, starting with a cost that already sits on a budget line.
• Make finance the owner of the ROI figure, and have it sign each measurement plan before launch.
• Cut licences nobody uses. An unused system returns nothing.
• Report AI to the board in money: net benefit, payback and risk-adjusted ROI, with hours saved shown only as a leading indicator.

The firms that get a return from AI will not be the ones that spend the most. They will be the ones that can show, line by line, where the money came from.