AI ROI: how to measure it, and what to do when the numbers say stop
Measure AI ROI one workflow at a time, against a baseline, with the full cost on the table.
BY STEFAN MARITZ··7 MIN READ
TL;DR
- AI ROI is the business value an AI initiative creates relative to its full cost, including people's time, data work, and change. Measure it per workflow against a baseline, because a licence count tells you nothing about return.
- Most organisations take two to four years to see a satisfactory return on a typical AI use case, against the seven to 12 months usually expected of technology spend. Only 6% report payback inside a year.
- The picture is improving. Nearly half of companies now capture meaningful value from AI, and AI spend has roughly doubled to 3.3% of revenue, with more than 80% of it sitting outside the IT budget.
- Revenue growth is still mostly a hope: 20% of organisations already grow revenue through AI, while 74% expect to.
- Time saved is an input. Decide before you start whether each use case should move cost, capacity, revenue, or quality, and measure that.
Most leadership teams we speak to can tell us how many AI licences they pay for. Far fewer can tell us which workflow now costs less, runs faster, or produces better work because of them. That gap is where most AI ROI conversations go wrong.
This guide covers what AI ROI means, why it is hard to measure, how to set up a measurement you can defend to a board, and when to stop.
What is AI ROI?
AI ROI (return on investment) is the value an AI initiative produces compared with everything it costs. The value side covers lower cost, more capacity, higher revenue, or better quality. The cost side covers licences and build work, plus the data clean-up, review time, training, and management attention that make the tool useful.
We measure it at the level of a workflow, meaning a recurring piece of work with a start, an end, and an owner, such as month-end reconciliation or first-line customer queries. A company-wide AI ROI figure blends successes and failures until neither is visible. A workflow-level figure tells you what to scale and what to switch off.
Why is AI ROI so hard to measure?
AI rarely arrives alone. It usually lands alongside a data clean-up, a team reorganisation, or a process change, which makes it difficult to say how much of the improvement belongs to the model. Many executives end up with only a ballpark estimate of AI's contribution for exactly this reason.
Other things make it harder:
- Some benefits are hard to price. Better customer conversations or less frustrated staff matter, and they rarely appear as a line in the management accounts.
- Proofs of concept run on clean demo data. Optimism built on a proof of concept with dummy data tends to fade once real, messy data arrives.
- The technology moves faster than the metrics. A use case scoped in January may look different by June because the models changed.
- Adoption decides the outcome. A tool people do not use returns nothing, however good it is.
- Financial discipline is unusual. Few companies apply the same financial rigour to AI that they would to a new factory, and companies making similar investments often define success differently.
The fix is to design the measurement before the work starts, while the old process is still running and can be timed.
How to measure AI ROI
A plant manager judges a new machine on a production line by what happens to throughput and scrap. AI deserves the same test, and the steps below apply it.
Step one: pick the workflow and its owner
Choose one recurring workflow with a named owner who can change how the team works. Without an owner, even a good result fades, which is one of the main reasons AI pilots stall before they pay back.
Step two: record a baseline
A baseline is the before measurement you compare against. Record cycle time, cost per transaction, error or rework rate, and volume handled per person for at least four weeks before anything changes.
Step three: decide what kind of return you expect
Different use cases return value in different ways, and each one has its own trap.
| Type of return | What to measure | Example | Watch out for |
|---|---|---|---|
| Cost | Cost per transaction, external spend | Fewer agency hours on first drafts | Savings that reappear as spend elsewhere |
| Capacity | Volume per person, backlog size | Same team handles more queries | Freed hours with no plan for them |
| Revenue | Conversion, win rate, speed to quote | Proposals out in hours instead of days | Crediting AI for a lift something else caused |
| Quality and risk | Error rate, rework, compliance findings | Fewer reconciliation errors at month end | Errors moving to a later step |
Capacity is the one leaders most often misread. If AI frees 10 hours a week and nobody decides what those hours are for, they disappear into email.
Revenue is the one most teams overlook, because they start from efficiency. One of us saw it from the customer side recently. Buying car insurance on a Sunday afternoon, they got a quote and confirmed cover on the spot through King Price's AI assistant. Another insurer called back just after eight the next morning, by which point the business had already gone elsewhere. That was new business going to whoever could answer first.
Step four: count the full cost
Include licences or usage fees, integration and build work, data preparation, the time people spend reviewing AI output, training, and ongoing monitoring. Review time is the cost most often left out, and it can quietly cancel the saving.
Step five: measure at 90 days and decide
Compare against the baseline after 90 days. Keep what worked, fix what nearly worked, and stop what did not. Stopping is a legitimate result. It frees budget and gives you evidence for the next decision.
When a different answer beats AI
Some of the best ROI decisions we make with clients involve no AI at all. If a workflow fails because the process is unclear or the data is wrong, a model will make the same mistakes faster. Often a rule-based fix is cheaper and more reliable, which is why the difference between automation and AI matters when you are paying for one and needed the other.
Some AI investment is capability building, such as teaching a team to work with AI at all. That work has value, and it should be judged by whether behaviour changed. The same logic sits behind measuring AI training effectiveness and judging real AI adoption by the workflows that changed.
An AI ROI checklist for this quarter
- List your live AI initiatives and the workflow each one is meant to change. Anything without a workflow is an experiment, so label it as one.
- Name an owner and record a baseline for each workflow you plan to touch next quarter.
- Write down the expected type of return for each use case: cost, capacity, revenue, or quality.
- Add review time to the cost line. Sample it for two weeks if you do not know it.
- Decide in advance what happens to freed capacity, whether that is more volume, new work, or faster turnaround.
- Book a 90-day review with the authority to scale, fix, or stop.
- Put the change in the work itself. Update templates, checklists, and handovers so the new way is the default, because a result that depends on enthusiasm will not survive the next busy month.
Frequently asked questions
How do you measure the ROI of AI?
Pick a single workflow, record a baseline for time, cost, and errors, and count the full cost of the AI change, including review time and training. Compare the same measures after 90 days and decide whether to scale, fix, or stop.
How long does it take to see a return on AI?
Most organisations report two to four years for a typical use case, far longer than the seven to 12 months usually expected of technology. Narrow, well-owned workflows can show cost or capacity gains much sooner, which is why we recommend reviewing each one at 90 days.
Why do so many AI projects fail to deliver ROI?
Usually because nobody owns the workflow, nobody recorded a baseline, or the tool was added to a process that stayed the same. The model is rarely the main problem.
Is time saved a good AI ROI metric?
It is a useful leading indicator, and it becomes a return only once the freed time turns into lower cost, more output, or better quality. Decide what the time is for before you count it as value.
What costs should be included in AI ROI?
Licences or usage fees, integration and build work, data preparation, review time, training, and ongoing monitoring. Leaving out review time is the most common way to overstate a return.
If you would like a scored view of where value is likely to sit in your organisation before you commit budget, the AI Readiness Diagnostic takes about five minutes and ends with a report you can take to your exec team.
WORK WITH PRAXES
Ready to move from reading about AI to getting your organisation ready for it? Start with the diagnostic.
Take the AI Readiness Diagnostic