AI ROI: How to Tell Whether AI Is Paying Off

AI ROI is what AI earns minus what it costs, divided by what it costs. To measure it, time a task before AI and with AI, checking time included, count every cost, and confirm the saved time turned into money or output.
This guide gives a simple method that uses your own numbers. For help setting it up, see the AI consulting page or my fractional Chief AI Officer service.
What is AI ROI?
AI ROI is the return on what you spend on AI, expressed as a ratio. The spending includes tools, setup, training and the time people spend checking AI output. The return is the value of what changed because of it.
The formula is the same one used for any investment:
AI ROI = (value created − total cost) ÷ total cost
Read the result like this:
| Result | What it means |
|---|---|
| Below zero | The AI cost more than it returned |
| Zero | It broke even |
| Above zero | The value beat the cost, and the number shows by how much for each dollar spent |
ROI sits inside a bigger plan. An AI strategy decides where AI should be used. ROI tells you whether each use is working, so you can keep it, fix it or stop it.
What does research say about AI ROI?
Studies of large companies show that many have not yet seen the return they expected. The figures come from surveys of executives, so they show what leaders report, not audited financial results.
| Finding | Source |
|---|---|
| Only 25% of AI initiatives had delivered the expected ROI over the last few years, and only 16% had scaled enterprise wide. Survey of 2,000 CEOs in 33 countries, run February to April 2025 | IBM Institute for Business Value with Oxford Economics, May 6, 2025 |
| 5% of companies were achieving substantial AI value, and 60% reported minimal revenue and cost gains. Survey of 1,250 senior executives and AI decision makers | BCG, September 30, 2025 |
Two things follow from this. First, “we bought AI” is not a result. Second, the companies that report gains are the ones that picked specific work and tracked it. The same method works for a team of ten.
How do you calculate AI ROI?
Pick one task, measure it before AI, measure it again with AI, and turn the difference into money. Do this for one task at a time. A company-wide ROI number built from guesses is less useful than three tasks measured well.
- Choose one repeated task. Good candidates have a clear start and finish, such as drafting a customer reply, summarizing a call or preparing a weekly report. The AI use cases guide lists ideas by department.
- Measure the baseline before AI. Time the task a few times with the real people who do it. Use a measured average, not an estimate from memory.
- Measure the task with AI. Include everything: writing the prompt, waiting, reading the output, fixing mistakes and the final check. A task is not done until a person would stand behind it.
- Count how often the task happens. Use your records, such as tickets, reports or calendar entries.
- Price the time. Use what an hour of that person’s work costs you, including pay, benefits and overhead. Finance can give you this number.
- Add up the costs (next section) and run the formula.
Fill in this worksheet once per task:
| Line | What to write | Where it comes from |
|---|---|---|
| A | Minutes per task before AI | Timed baseline |
| B | Minutes per task with AI, including checking and fixing | Timed with AI |
| C | Tasks per month | Your records |
| D | Cost of one hour of that person’s time | Finance |
| E | Value of time saved per month | (A − B) × C ÷ 60 × D |
| F | Total AI cost per month | Cost table below |
| G | Monthly AI ROI | (E − F) ÷ F |
If B is larger than A, the AI is slowing that task down, and the worksheet will say so. That is a useful answer.
I built a product-research skill that asks me a few questions, reads the listings and reviews, and recommends the best option. I buy what it recommends about 95% of the time, it saves me about 30 minutes, and it is the skill I use the most.
What counts as a cost?
Count every cost, not just the subscription. Left-out costs are the main reason an AI project looks better on paper than it is. ISACA names technology, talent and infrastructure as the main cost groups, and also points to the upfront cost of data collection and model training.
| Cost | Examples | Easy to forget? |
|---|---|---|
| Tools | Subscriptions, seats, usage fees | No |
| Setup | Connecting tools to your files, writing prompts and templates, building automations | Often |
| Training | The hours people spend learning, plus any trainer or coach | Often |
| Checking and fixing | Time a person spends reviewing output and correcting errors | Very often |
| Ongoing care | Updating prompts, answering questions, handling changes when a tool updates | Very often |
| Rules and risk | Policy writing, access controls, a legal review where needed | Often. See the AI governance guide |
Spread one-time costs, such as setup and training, over the months you expect to use the tool. Say which period you chose, so the number can be checked later.
What counts as a benefit?
A benefit counts when it shows up in a number you already track. ISACA’s framework separates tangible returns, which are financially measurable, from intangible ones, which are harder to price.
| Type | Examples | How to count it |
|---|---|---|
| Time saved | Fewer hours per task | The worksheet above, once the time is used (see below) |
| Money saved | Fewer outside contractors, fewer paid hours | The invoice or payroll line that went down |
| Money earned | More leads handled, faster quotes, higher win rate | Revenue from the work the AI helped with, compared with before |
| Quality | Fewer errors, fewer redos | Error or rework counts, before and after |
| Softer gains | Staff satisfaction, better customer experience, stronger skills | Track with a survey or retention number. Keep it apart from the dollar ROI |
Keep the softer gains in a separate column. Mixing guesses with measured dollars makes the whole number untrustworthy.
Are hours saved the same as money saved?
No. Hours saved are a capacity gain. They become money only when the time is used on something. This is an inference from how business costs work, not a research finding.
Ask what happened to the freed time:
- Fewer paid hours or fewer contractors. This is a cash saving. Count it.
- More output from the same people. This is value if you can sell, use or deliver the extra output. Count the extra output, not the hours.
- Nothing in particular. The time was absorbed. Do not count it as cash. Note it as capacity and decide what to do with it.
Here is a website example. I used to pay thousands of dollars for a logo and a website, and building a site took months. Now Claude builds one in about ten minutes. Those are my numbers, so time your own tasks instead of borrowing mine.
Why can’t you trust how fast AI feels?
People are poor judges of their own time savings, so measure with a clock. A 2025 randomized trial by METR followed 16 experienced open-source developers on 246 real tasks. The developers were 19% slower when AI was allowed. They had expected AI to speed them up by 24%, and afterward still believed it had sped them up by 20%.
The result has narrow limits. METR says it does not show that AI slows down most workers, and it covers one kind of work with early-2025 tools. The lesson for ROI is smaller and safer: a feeling of speed is not a measurement. Time a task with the AI and without, and let the numbers decide.
When should you measure AI ROI?
Take the baseline before anything changes, then compare at fixed points. A schedule like this works for one team, and it is a suggested plan, not a research finding:
| When | What to do |
|---|---|
| Before rollout | Time the task, record the volume, write down the baseline |
| Month 1 | Time the task again once people are using it every week. Early results include the learning curve |
| Day 90 | Run the full worksheet with all costs and decide: keep, fix or stop |
| Every quarter | Re-check. Prompts drift, tools change and people find shortcuts |
Some value takes years to appear, such as a stronger process or a team that knows how to use AI. ISACA describes this as strategic and capability returns, and suggests tracking them over a longer period than the dollar ROI. Report them separately.
If people are not using the tool, ROI will be poor no matter how good the tool is. In that case the fix is usually on the people side, and the AI adoption guide covers rollouts that stall.
What mistakes make AI ROI look better than it is?
Five mistakes account for most inflated numbers:
- No baseline. Without a measured “before,” any improvement is a guess.
- Counting only the subscription. Setup, training, checking and upkeep are real costs.
- Ignoring checking time. If a person must fix the output, that time belongs in line B.
- Counting hours as dollars. Time saved that is not used is not a return.
- Measuring the best week. Use an average over several weeks, across several people, not the fastest person on a good day.
What should you do with the result?
Use the number to decide what happens next:
| Result | Next move |
|---|---|
| Clearly above zero | Keep it, document how it runs, and look for the next similar task |
| Close to zero | Find the biggest cost, often checking time, and fix it. Re-measure in a month |
| Below zero | Check whether the task was a good fit and whether people were trained. Stop or change it if the fix is not clear |
| Cannot be measured | Pick a task with a clearer start and finish |
A result below zero is not a failure of the method. It stops money going to something that is not working. To pick better starting tasks, run an AI readiness assessment first.
Who should own AI ROI?
One named person should own the numbers, and a senior leader should read them. When no one has the time, some companies bring in a fractional Chief AI Officer, who sets the measures and reviews them each quarter. For the role itself, see what a Chief AI Officer does.
What should you do next?
Pick one task this week. Time it five times without AI, write down how often it happens, and fill in lines A, C and D of the worksheet. Then run the same task with AI for a month and finish the sheet. If you want a senior person to set up the measures, train the team and review the results with you, my fractional Chief AI Officer service covers the plan, the policy, the training and the shared systems.
Sources
- IBM, IBM Study: CEOs Double Down on AI While Navigating Enterprise Hurdles, published May 6, 2025, read November 12, 2026.
- BCG, Agentic AI Emerges as the Next Frontier, Driving 17% of Total AI Value Today and Expected to Nearly Double by 2028, published September 30, 2025, read November 12, 2026.
- ISACA, Mary Carmichael, Demonstrating AI ROI: How to Measure and Prove the Value of Your AI Investments, published March 3, 2025, read November 12, 2026.
- METR, Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity, published July 10, 2025, read November 12, 2026. METR has posted newer data for late-2025 tools.


