How to calculate AI ROI on a real project

Calculating the ROI of an artificial intelligence project is not hard because of the arithmetic: it is a division. It is hard because the two numbers you are about to divide are almost always the ones nobody measured. The benefit gets estimated from hours that were never timed, and the cost gets estimated from the vendor license, which is usually the cheapest part of the whole thing. What follows is the full procedure, with the formula spelled out and a worked example.

Those four numbers describe the same problem from four angles. Adoption is wide, measurement is thin, the return takes longer than people assume, and part of the saving evaporates before it ever reaches the P&L. None of the four is a reason not to do the project: they are reasons to calculate it properly before signing it.

The first reason is that the expectation gets set before the measurement does. In the *CIO Playbook 2026*, IDC research commissioned by Lenovo covering roughly 500 IT and business decision makers in Argentina, Brazil, Chile, Colombia, Mexico and Peru, 97% of Latin American organizations plan to increase their AI budget over the next twelve months (by 14% on average), 92% anticipate a positive return, and respondents expect to generate around 3 dollars of value for every dollar invested ⁷. Read that carefully: it is a stated expectation in a vendor-sponsored survey, it is a regional average that is not the same as the Mexican figure, and it is not a measured return.

The second reason is that the value is real but highly concentrated. BCG sorted the companies in its global study into three groups: 5% *future-built*, 35% scaling AI and starting to generate value, and 60% laggards getting barely any material value from what they invest; the leading 5% already achieve five times the revenue increase and three times the cost reduction the rest get from AI ⁵. If your business case was built on the market average, you are averaging two populations that look nothing like each other.

The baseline is a measurement of the process as it runs today, taken before the system exists and signed off by whoever will judge the result. If you capture it afterwards it is no longer a measurement: it is a reconstruction, and reconstructions always favor the project that pays for them. This is the most expensive mistake in the whole exercise, and avoiding it costs no money, only two weeks of discipline.

Building a useful baseline means measuring five things about a well defined unit of work (a quote, a case file, a ticket, an invoice), not about an entire department:

• Volume: how many units the team processes per month, with the seasonality of the last twelve months, not of last month.
• Time per unit: real minutes of human work, measured by sampling or from system timestamps, not by asking *how long does this take you?*.
• Loaded cost per hour: gross salary plus benefits, payroll taxes and allocated overhead, divided by the hours actually worked in a year. Take it from your own payroll; the multiplier over nominal salary varies too much between companies to borrow anyone else's.
• Quality: error or rework rate, and the unit cost of each error when it reaches the customer.
• Cycle time: how long the unit takes from entry to exit, waiting time included. It is the metric AI moves the most and the one measured the least.

A warning about the unit of analysis. It is tempting to build the case at the macro level (*AI raises productivity, therefore my company will be more productive*), but those studies measure something else. The OECD estimates that AI could add between 0.2 and 1.3 percentage points to annual labor productivity growth in G7 economies over the next decade, depending on the country and the adoption scenario ⁸. That is a potential aggregate gain, not an observed result, and not something you can drop into your spreadsheet.

The denominator of the formula is the total cost, and the model license is usually its smallest fraction. Gartner, in its analysis of business value and cost for generative AI use cases, offers approximate ranges (it explicitly calls them *ballpark*) that help calibrate the order of magnitude: for defensive use cases such as code assistants, business productivity or marketing content creation, it estimates between 100,000 and 500,000 dollars of up-front investment covering pilot, deployment, integration and training, plus between 220 and 2,100 dollars per user per year in recurring costs, calculated over deployments of 100 to 1,000 users and including SaaS licenses and 10% of the initial implementation cost ³.

For use cases that extend an existing process (generative AI-assisted support, personalized sales content, document search and summarization), the ranges rise to between 750,000 and 1 million dollars up front, with recurring costs of 790 to 11,000 dollars per user per year, assuming a development, data, security and product team working for six months, and application and model maintenance equivalent to 15%-20% of the initial deployment cost ³. These are global 2024 figures, not Mexico specific; they tell you whether your budget is on the right planet, not what to copy.

In projects at mid-sized companies, these are the seven lines that have to be in the denominator. The ones most often forgotten are the fourth and the fifth:

1. 01Diagnosis and design: the work of understanding the process and deciding what gets automated. You pay for it even if the project stops afterwards.
2. 02Build and integration: the system and, above all, the connections to the CRM, the ERP, email and whatever the team already uses. Integrations usually cost more than the model.
3. 03Inference and licenses: the cost of every query, which scales with usage. Gartner expects that by 2028 aggregate inference costs will account for at least 70% of a model's total lifetime cost, far above training ⁴.
4. 04Internal hours: your people's time in interviews, testing, data validation and reviewing outputs. It is a real cost even if it does not come out of the IT budget.
5. 05Change management and training: if the team does not adopt the system, the benefit is zero and you still pay the full cost.
6. 06Maintenance and evolution: retraining, prompt adjustments, vendor API changes, fixing quality degradation. Budget for it from year one.
7. 07Governance and control: observability, permissions, traceability, security and risk reviews. It is the first thing cut and the first thing missed.

There is a concrete reason to be exhaustive here: cost overrun is the most frequent failure mode. Gartner forecasts that by 2028 at least half of generative AI projects will exceed their budget, and attributes it to poor architecture decisions and a lack of operational experience ⁴. An optimistic denominator does not make the project better; it just moves the bad news to next quarter.

The benefit breaks into three families, ordered from most defensible to most arguable. It is worth calculating them separately, because a committee that trusts the first and distrusts the third can still approve the project if you show that it holds up without that line.

Hours recovered. Monthly volume × time per unit × the percentage of that time the system genuinely absorbs × loaded cost per hour. Two cautions: the absorbed percentage is almost never 100% (human review remains) and the saving is only money if those hours get reassigned to something that generates value, or if they avoid a hire. If nobody changes what they work on, you saved time, not cost.

Errors avoided. Error frequency × unit cost × the percentage the system actually prevents. The unit cost includes rework, credit notes, contractual penalties and, when it can be estimated honestly, the lost customer. This line is usually bigger than people think and easier to document than the previous one, because expensive errors leave an accounting trail.

Incremental revenue. More quotes sent, faster response, better conversion rate, less drop-off. It is the most attractive family and the most fragile, because attribution is hard: if the quarter was good, was it the AI or was it the market? Rule of thumb: value incremental revenue at contribution margin, not at billings, apply an explicit attribution discount, and always show the ROI with and without this line.

Every number in this section is invented to illustrate the method: they are round figures from a hypothetical case, not data from any source. Take a mid-sized Mexican company with a team of 10 people who spend 5 hours a week, 45 weeks a year, putting together technical quotes: 2,250 hours a year. The loaded cost per hour is 400 pesos. They implement a system that absorbs 60% of that work.

Now the part almost nobody does: year 1 does not deliver the full benefit, because there is implementation, an adoption curve and adjustments. If we assume 70% of the annual benefit is captured, year 1 contributes 538,000 pesos against 746,000 of cost, which is an ROI of −28%. Year 2 contributes 768,000 against 206,000: +273%. Cumulative over 24 months, ROI is +37% and payback lands around month 17. A healthy project can perfectly well be in the red in its first year; what it cannot do is hide that.

And the stress test: if the committee does not believe the incremental revenue line and we delete it entirely, cumulative ROI at 24 months turns slightly negative and payback slips from month 17 to month 26. The project does not collapse, but it changes category and, probably, changes who has to approve it. That is exactly the conversation that should happen before signing, not in the second-quarter review.

The time horizon deserves its own paragraph because that is where most of the fighting happens. Deloitte found that most surveyed executives reach a satisfactory return on a typical AI use case within two to four years, far beyond the seven to twelve month payback usually expected of a technology investment; even among the most successful projects, only 13% saw returns within the first twelve months ². The sample is Europe and the Middle East, not Latin America, so do not use it as a local benchmark. Use it for something else: to justify evaluating your project over 24 or 36 months instead of 12, which is where good but badly budgeted projects die.

Nobody is backing off because of that: in the same study, 85% of surveyed organizations increased their AI investment over the last twelve months and 91% plan to increase it again, even though only about one in five qualifies as a genuine leader in AI ROI ². Money goes in expecting a return that arrives late. That is only defensible if the calculation says so from the start.

One last criterion, more about governance than arithmetic: decide on day one what result would make you cancel the project. A minimum ROI threshold, a date, a metric that did not move. Projects with no exit condition do not fail: they become permanent, which is more expensive.