Skip to content

Odduel

We put the odds in your favour

Insights

AI ROI: how to measure the return on an AI project

How to measure AI ROI step by step: a baseline, the benefits to count, the costs people forget, how long the return takes, EU AI Act effects and the most common mistakes.

Two people at a meeting table reviewing a printed worksheet of costs and benefits beside a laptop, measuring the return on an AI project

AI ROI is the value an AI project returns compared with everything it costs to build, run and maintain, measured against a baseline recorded before the project started. You measure it by choosing one process, recording how it performs today, estimating the gain the AI should bring, then tracking the real gain and the real costs after launch. This guide gives a worksheet you can use step by step, explains how long the return usually takes to show and lists the mistakes that make AI projects look better or worse than they are.

Last updated: 21 November 2026.

What is AI ROI?

Return on investment for AI follows the same logic as any investment: the net benefit divided by the total cost. What makes AI different is that both sides of the calculation are easy to get wrong. The benefits are often spread across many small tasks, and the costs continue long after launch, because models, prompts and data need care.

A useful definition: AI ROI is the measured change in a business outcome caused by an AI system, minus the full cost of owning that system, over a set period. Every word matters. “Measured” means you need a baseline. “Caused by” means you must separate the AI’s effect from other changes. “Full cost” means more than the build.

At Odduel, a creative and technology agency in Barcelona and São Paulo, we start every project in our AI development service by writing down the baseline and the target, before any model is chosen.

In short: pick one process, measure it now, estimate the gain, list every cost including running and review time, set a time window, then compare. An AI project without a baseline cannot show a return, however good it is.

What are the steps in measuring AI ROI?

The worksheet below works for most business uses of AI: answering customer questions, extracting data from documents, drafting content, classifying tickets, forecasting demand.

  1. Name the process and its owner. For example: “first replies to customer emails, owned by the support lead.”
  2. Record the baseline. Volume per month, time per task, error rate, cost per task, and the business outcome it affects, such as response time or conversion.
  3. State the expected effect. Which part of the process will the AI do, and which part stays with people? Be specific: “drafts replies for routine questions; a person reviews and sends.”
  4. Estimate the gain range. A cautious and an optimistic estimate of time saved, errors avoided or revenue added. Avoid a single precise number before you have data.
  5. List every cost. Build, integration, model usage, hosting, monitoring, human review, training staff, and maintenance.
  6. Run a pilot. Use the AI on a defined share of the work, keeping the rest as a comparison group where possible.
  7. Measure the real effect. Compare the pilot with the baseline and with the comparison group over the same period.
  8. Calculate and decide. Net benefit over total cost for the period; then scale, adjust or stop.

Which benefits should you count?

Count benefits you can observe in your own data. They usually fall into four groups.

Benefit typeWhat to measureExample
Time savedHours per task before and after, times volumeInvoices processed per hour by the finance team
QualityError rate, rework, complaintsFewer wrongly routed support tickets
RevenueConversion, average order value, retentionMore enquiries answered outside office hours
Risk and complianceMissed deadlines, audit findingsContract clauses flagged before signature
Benefit types to count in an AI ROI calculation

Time saved only becomes a financial return when it is used for something: more volume handled with the same staff, faster service, or work that was previously left undone. Say which of these applies, or the “saved hours” stay on paper.

Which costs are usually forgotten?

The build is the visible cost. These are the ones that make returns look better than they are when left out:

  • Model usage. Most AI systems pay per request or per amount of text processed; volume growth raises the bill.
  • Human review. Where a person checks AI output, their time is a running cost.
  • Data preparation. Cleaning, labelling and connecting data is often the largest share of the first phase.
  • Monitoring and evaluation. Someone must check quality regularly, because model behaviour and inputs change.
  • Maintenance. Prompts, integrations and models need updates as providers release new versions.
  • Compliance work. Documentation, risk assessment and data protection checks, especially for higher-risk uses.
  • Change management. Training staff and adjusting processes so the tool is actually used.

Our article on AI agent development cost breaks down these lines for agent-style systems.

How long does it usually take to see a return?

It depends on the type of use. Narrow tasks with high volume, such as classifying documents or drafting routine replies, can show measurable time savings within the first weeks of a pilot. Projects that depend on behaviour change, new data or customer adoption, such as a recommendation engine or a sales assistant, usually need several months before the effect is clear.

Three factors shorten the time to a visible return:

  1. High volume. A small gain on a task done thousands of times a month shows up fast.
  2. Clean, available data. Projects that must first build their data foundation take longer.
  3. A clear owner. Someone responsible for adoption and measurement keeps the project from drifting.

Set the measurement window before the pilot starts, and resist judging early: the first weeks often include learning effects in both directions.

How do you prove the AI caused the result?

This is where many ROI claims fall apart. If sales rose during the month a chatbot launched, the chatbot may or may not be the reason.

  • Use a comparison group. Run the AI on part of the work, such as one region, one product line or alternate days, and compare.
  • Hold other changes steady. Avoid launching a price change or a campaign at the same time.
  • Look at the mechanism. If the AI drafts replies, check response times and resolution, not only overall sales.
  • Track quality, not only speed. Faster but less accurate work can cost more later.

When a comparison group is not possible, compare with the same period last year and with the weeks before launch, and say clearly in your report that the effect is estimated.

What rules affect AI ROI in Europe?

Regulation changes the cost side. The EU AI Act sorts AI systems by risk: most business uses, such as drafting text or sorting tickets, carry light obligations such as transparency, while uses in areas like recruitment, credit or access to essential services count as high-risk and require documentation, human oversight and risk management. Where personal data is involved, the GDPR applies as well.

For the ROI worksheet this means:

  • Check early whether your use could be high-risk, because the compliance work is a real cost line.
  • Budget for transparency, such as telling people when they talk to an AI system.
  • Keep records of data sources and evaluation results; they help with compliance and with measuring quality.

Companies across Spain, France, Germany, Italy, Portugal and the rest of the EU work under the same act, which makes the cost of compliance easier to plan across markets.

What are the most common AI ROI mistakes?

MistakeWhy it distorts the returnWhat to do instead
No baselineNothing to compare againstMeasure the process before starting
Counting hours saved as moneyHours are only money if reusedSay how the time is used
Ignoring running costsReturn looks larger than it isInclude model usage, review, maintenance
Starting with the technologySolutions look for problemsStart with a costly process
Measuring too earlyLearning effects distort resultsAgree the window in advance
Scaling a pilot that was never measuredSpreads an unknown effectProve one use, then extend
Common AI ROI mistakes and how to avoid them

How should a small business approach AI ROI?

A small business has less data and less room for long experiments, so focus matters more:

  1. Pick one painful, repetitive task, such as answering the same questions, preparing quotes or entering data from documents.
  2. Try existing tools first, before commissioning custom work; the cheapest experiment is often a configured service.
  3. Commission custom work when the task is specific to your business or must connect to your own systems.
  4. Keep the worksheet on one page so it is reviewed every month.

Our guides on how to start an AI project and on AI integration for business explain the first steps and the ways AI connects to existing software.

Questions people ask

How do you calculate the ROI of an AI project?

Subtract the full cost of owning the AI system from the measured benefit over a set period, then divide by that cost. The benefit comes from comparing the process with a baseline recorded before launch. The cost includes build, integration, model usage, hosting, human review, monitoring and maintenance, not only the initial development.

How long does it take for AI to pay back?

Narrow, high-volume tasks such as sorting documents or drafting routine replies can show time savings within weeks of a pilot. Projects that depend on new data or customer adoption usually take several months. Agreeing the measurement window before launch prevents both early celebration and early abandonment of a project that needs time.

What is a good baseline for measuring AI ROI?

A good baseline records the process as it is today: monthly volume, time per task, error or rework rate, cost per task and the business outcome it affects. Measure it over several weeks to smooth out unusual periods. Without this record, any later improvement is an opinion rather than a measured return.

What are the most common AI ROI mistakes?

The most common mistakes are starting without a baseline, counting saved hours as money without saying how the time is used, leaving out running costs such as model usage and human review, judging results too early, and scaling a pilot whose effect was never measured against a comparison group.

Does the EU AI Act change the cost of an AI project?

It can. Most business uses carry light duties such as transparency, but systems used in areas such as recruitment or credit decisions count as high-risk and need documentation, oversight and risk management. That work is a real cost line, so check the risk category before calculating the expected return.

Choose one process this week, write down how it performs today and what it costs: that baseline is the start of any honest AI ROI. When you want an AI system scoped, built and measured against it, see our AI development service or tell us about the process.

More articles

More from our Insights.

Explore all insights

Get in touch

Let’s change your odds.

Odds do not change on their own. If you have something to build, launch or fix, tell us about it and we will show you how we would approach it.

    Need an NDA before you share details? Start a project

    Chat with us