AI ROI for one job is the hours AI gives back each week, minus the hours your team still spends checking it and looking after it, set against the hours it took to set up. Work it out from four numbers you already have: how many times the job happens, how long each one takes, the share AI really handles in a test, and the time a person still spends on it. Where it stops: an AI ROI calculator gives you a forecast. Someone still has to measure the real hours and keep the job working so the number holds.
What an AI ROI calculator should measure
Many AI ROI calculators online are made by vendors. They ask for a few inputs, assume AI handles almost all of the work, and show a big percentage. The percentage is the least useful number on the screen.
A useful AI ROI calculation counts hours. How many hours does this job take today? How many will AI take off your team's plate, knowing it will hand some cases back? How many hours will people still spend checking its work and looking after it? And how many weeks until the hours saved cover the hours it took to set up?
Hours are honest because you can check them. Your team knows how long a job takes. A month after AI starts, you can count again and see if the forecast was right.
A vendor AI ROI calculator vs working it out yourself
| Vendor calculator | Working it out in hours | |
|---|---|---|
| Where the inputs come from | Typed in from memory | Last month's real numbers for one job |
| Share of cases AI handles | Often set high by default | What AI handled in your own test |
| Hours still needed from people | Often left out | Checking, odd cases and upkeep, counted |
| The answer you get | A large percentage | Hours back each week, and weeks to pay back |
| Can you check it later? | Hard to | Yes, count the hours again after a month |
How to work out AI ROI in four steps
Step one: get the starting numbers before you open any calculator. For one job, write down how many times it happens each week, how long each one takes, how often it goes wrong, and how long a mistake takes to put right. Use last month's real numbers, not a guess.
Step two: set the share of cases AI handles from a test, not a slider. Run AI on real work for two to four weeks and count how many cases it finished without help. Use that number. If you haven't tested yet, start low and raise it only when the test proves it.
Step three: count every hour it still takes. Setting it up, cleaning the data it needs, checking its work each week, handling the cases it hands back, and fixing it when a program changes. If a calculator has no place for these, write them next to it.
Step four: read the weeks to pay back, not the percentage. Divide the setup hours by the hours saved each week. That tells you how many weeks until the job has paid for its own setup. A first job should pay back within a few months. If it takes a year, test a different job.
An AI ROI example, in hours
Say your office enters 120 order emails a week, and each takes about 6 minutes. That is 12 hours a week.
In a four-week test, AI finished 70% of the orders on its own and handed the rest to a person with the facts attached. That takes about 8.4 hours a week off your team. Checking a sample of its work takes about an hour a week, and upkeep about half an hour. So the real gain is close to 7 hours a week.
If setting it up took 40 hours of your team's time, the job pays back in about six weeks. Now move the 70% up and down by a fifth and see how the weeks change. If the answer only works at the top of the range, it is a weak first job for AI.
Ways to work out AI ROI
Use a vendor's AI ROI calculator
- How it works
- Many AI vendors offer a free calculator on their site. You type in your volumes and times, and it shows a return.
- Good for
- A quick first look at whether a job is worth testing at all.
- Where it breaks
- The defaults are set by the seller. Many leave out the hours people still spend checking and looking after the AI.
Work it out in a spreadsheet
- How it works
- Put the four steps above in a sheet: starting numbers, the share AI handles from your test, every hour still needed, and weeks to pay back.
- Good for
- Owners and managers who want a number they can check a month later.
- Where it breaks
- The sheet is only as good as the test behind it. Without a real test, the share AI handles is still a guess.
Arkeo agents, set up and run for you
- How it works
- We go through your weekly jobs with you, count the hours each one takes, and set up an agent in your own accounts for the first one. The agent checks its own work, and its results show what it actually handled each week.
- Good for
- Owners who want the hours counted from real work, not a forecast.
- Where it breaks
- You give us the starting numbers for the job: how often it happens and how long it takes. We help you count them.
What an AI ROI calculator leaves out
A calculator forecasts the gain. Most AI projects don't reach it, and the reason is rarely the AI.
In IBM's 2025 study of 2,000 CEOs, only 25% of AI efforts had delivered the return expected over the last few years, and only 16% had been scaled across the company. MIT's 2025 study found that about 5% of AI pilots reached fast revenue growth, while the large majority showed little or no measurable effect on profit. The researchers pointed to poor fit with how work is actually done.
Where AI is put to work properly, the gains are real. In PwC's 2025 survey, 66% of companies using AI agents said they were getting more done. BCG found that the companies getting the most from AI chase about half as many projects as the rest, and expect more than twice the return.
The gap between the two is hand work no calculator counts. Someone pulls the starting numbers. Someone checks the AI's work each week. Someone handles the cases it hands back, and fixes it when a program changes. The Small Business Administration's advice is the same: test AI tools to see if they add value, and have another person review what AI produces. Leave those hours out and the forecast looks better than the year will.
How to measure AI ROI once AI is working
The forecast is a guess. These steps turn it into a number you can trust.
- 1
Write the starting numbers down
Before AI starts, record how often the job happens, how long it takes, and how often it goes wrong. Date the page.
- 2
Test on real work first
Run AI next to the person for two to four weeks. Count what it finished alone and what it handed back.
- 3
Log the hours people still spend
Checking, odd cases and fixes. A rough weekly count from the person who owns the job is enough.
- 4
Count again after a month
Compare the real hours to the forecast. If they are close, give AI more of the job. If not, find out why before you add more.
- 5
Review it each quarter
Hours back, mistakes, and how often a person stepped in. Retire the jobs that stopped paying. NIST's framework has ideas on what to measure.
AI agents that report their own hours
Each of these does a job that comes back each week, so the hours it saves are easy to count.
The Monday meeting starts from the same page.
- Pulls your numbers from five toolsMondays at 7 a.m.
- Calls out anything that moved past your set percentageeach week
- Posts a one-page summary to Slackevery Monday
Bills go out on time, and late payers get caught.
- Checks finished work against each client's planat month end
- Builds the invoiceper client
- Flags unpaid invoicesevery week
Your CRM stays right without anyone typing.
- Logs emails and callsas they happen
- Checks off the task when the reply landsright away
- Moves the deal to the right stageper deal
Questions
Is my customers' information safe?
Yes. Agents work in your own accounts and see only what each job needs. Customer details never go into public AI. Agents connect to AI through business-grade connections audited under SOC 2.
How do you calculate AI ROI?
Take the hours a job takes each week, multiply by the share AI really handles in a test, subtract the hours people still spend checking and looking after it, and compare that weekly gain to the hours it took to set up.
How long does AI take to pay back?
For a single, well-chosen job, a few months is a fair target. If your numbers say a year or more, test a different job first.
What is the most common AI ROI mistake?
Trusting the share of cases a vendor's calculator assumes AI will handle, and leaving out the hours people still spend checking and fixing. Use your own test numbers instead.
What does it cost?
You get the price on the call, along with the hours it saves your team.
Sources
- newsroom.ibm.com: 06 ibm study ceos double down on ai while navigating enterprise hurdle
- fortune.com: mit report 95 percent generative ai pilots at companies failing cfo
- pwc.com: ai agent survey.html
- bcg.com: ai adoption in 2024 74 of companies struggle to achieve and scale valu
- sba.gov: #ai for small business
- nist.gov: ai risk management framework