How to Measure AI ROI (2026): Method and Calculator

AutomateWisely Editorial Team · Last reviewed November 1, 2026 · How we research

The CFO asks: “What did the AI pilot actually return?”

“People love it.”Satisfaction is not a return. It may predict one, but it is not money.

“We saved 2,000 hours.”Only if those hours turned into more output, faster service or a hire you did not make.

“Usage is up 300%.”Usage measures cost as much as value, especially with pay per use pricing.

The answer that survives the meeting: “Against a four week baseline, cost per invoice fell from $4.00 to $1.50. We absorbed 15% more volume without the planned hire. Net of all costs, that is $21,600 in year one, payback in under three months.”

In short: AI ROI is the usual formula, (benefits minus costs) divided by costs. What makes it credible is the work around it: a baseline measured before launch, every cost counted (including the time people spend checking AI output), and an honest view of how much saved time really turns into money. The calculator below lets you model that last part, which most ROI estimates skip.

What the evidence says about AI returns

~6%

are AI high performers88% of organizations use AI, but only about 6% attribute more than 5% of EBIT to it.McKinsey, The State of AI, November 2025 (1,993 respondents, 105 countries)

95%

of GenAI pilots stalledPilots with little or no measurable P&L impact. Specialized vendor tools succeeded about two thirds of the time, internal builds about one third.MIT NANDA, The GenAI Divide, reported August 2025 (150 interviews, 350 employee survey, 300 deployments)

31%

saved one to two hoursOf US workers who used AI in the previous week, 31% said it saved them one to two hours; 10% saved none and 3% lost time.US Census Bureau pulse survey, fielded March 2026, published August 2026

Time saved by AI in the previous week, US workers who used it

1 to 2 h3 to 4 h4 h+None

1 to 2 hours: 31%3 to 4 hours: 15%More than 4: 15%No time: 10%Added time: 3%Other or not sure

The MIT figure is often quoted as “95% of AI projects fail”. A fairer reading: most pilots never set a financial baseline or never fit the workflow, so their impact could not be shown. A project you cannot measure is a project you cannot defend.

The method in four moves

01BaselineTwo to four weeks of volume, time, cost, quality and speed before anything changes.

02BenefitsPick one or two that turn into money and agree how to value them in advance.

03Full costSetup, licences, usage fees, owner time and the hours spent reviewing output.

04ROI and paybackYear one and year three separately, against the baseline.

ROI = (benefits − costs) ÷ costs × 100 · Payback = one time costs ÷ monthly net benefit

If you are still choosing where to start, pick a use case with a clean baseline; how to implement AI in your business covers how to choose.

Where saved hours leak away

This is the step most estimates skip. Hours saved on paper only become money if the time is redeployed: more volume with the same team, faster service that keeps customers, or a hire you no longer need.

Hours saved

100 hours a month

Hours redeployed

60 hours

Money realised

$1,800 at $30 an hour

Illustrative. The share of saved time that becomes value is what the calculator calls the realisation rate.

What to measure, by use case

Use case	Core metric	How it becomes money	Watch for
Customer service	Cost per resolved ticket, resolution rate	Fewer agent hours or more tickets per agent	Repeat contacts and satisfaction dropping
Invoice processing	Minutes and cost per invoice, error rate	Volume growth without new hires, fewer late fees	Review time eating the savings
Sales prospecting	Meetings booked per rep, conversion	Extra closed deals × margin	Activity up, pipeline quality down
Marketing content	Cost and time per published asset	Lower agency spend, more campaigns	Output that nobody reads
Software development	Cycle time, changes shipped	Features delivered sooner, avoided contractor cost	Defects and rework rising

A worked example: invoices

A company handles 1,200 supplier invoices a month. An AI tool reads each invoice and pre-fills the accounting system; a person checks every entry. Illustrative numbers:

Before

Time per invoice 8 min
Hours per month 160
Team cost per hour $30
Labour cost per invoice $4.00

After, including review

Time per invoice 3 min
Hours per month 60
Setup, one time $6,000
Tool and owner time per month $700

Value of 100 h a month$36,000/yr

Year one cost$14,400

ROI, year one150%

Payback2.6 months

Those numbers assume every saved hour is used. If only 60% is, the year one return drops to 50%. That gap is the realisation rate, and the simulator below shows it month by month.

ROI simulator

AI project ROI calculatorYearly benefit, cost, net value and payback, with a realisation rate

Hours saved per month

Loaded cost per hour ($)

Other yearly benefits ($)

One time setup ($)

Running cost per month ($)

Horizon

Realisation rate 100%Share of saved hours that turns into value. 100% only if every hour goes into more output or an avoided cost.

Defaults reproduce the invoice example. Horizon is in years (1 to 3) and changes ROI and net value; the chart always shows 24 months for three realisation rates. Nothing you enter is stored or sent.

Five ways AI ROI gets overstated

Counting every saved hour as cashApply a realisation rate and say where the time went
Leaving out review and correction timeTime a sample of real tasks including checks
Measuring against memory, not a baselineRecord two to four weeks before launch
Ignoring usage growth in pay per use toolsModel cost at next year’s volume, not today’s
Reporting the best team onlyReport the median team, or compare with a control group

Common questions

How long before measuring?

Three to six months is typical. The first weeks include a learning curve that understates the steady state.

What if benefits are hard to price?

Track them separately, such as fewer errors or faster replies, and do not force a dollar value onto them.

Should we include licences we already pay for?

Include the share used for this project, or the uplift if a plan was upgraded for AI features.

Is a negative year one ROI a failure?

Not always. Setup costs land in year one. Check payback and the year three view before deciding.

For small teams looking for first projects with a measurable return, see AI for small business and AI agent use cases.

A one page ROI memo

What finance usually wants to see, in this order:

The process and its baseline (volume, time, cost, quality)
What changed and since when
Results against baseline (same metrics, same period length)
Full cost (one time and monthly, including review time)
Realisation: where the saved time went
ROI, payback and the year three view
Risks and what you will measure next

Sources

McKinsey, The State of AI (November 2025)
Tom’s Hardware, reporting on MIT NANDA’s The GenAI Divide (August 2025)
US Census Bureau, About a third of workers who used AI said they completed tasks one to two hours faster (August 2026)
Nextgov, AI is already on the job, and Census is keeping count (August 19, 2026)

General information, not financial advice. Spotted an error? Tell us.

How to Measure AI ROI: A Method Finance Will Accept