AI implementation cost: what an AI project really costs a company

The question always comes up on the same call, almost always around minute fifteen: *what is this going to cost me?*. The honest answer is that any number said before seeing your process would be made up. Since that is no use to you, let's do the opposite. Instead of a price, here is the full cost structure: what you pay once, what you pay forever, and which line is the one that blows up budgets.

Those four numbers mark out the terrain. Reference ranges do exist, but they are global and measured on large companies. The dominant cost is not building, it is operating. And in Mexico the digital starting base is lower than international reports assume, which changes the order of the investments. Every hard figure carries a source; where there is none, we say so.

Because the data is not produced locally, and that has a measurable root. Stanford HAI's 2026 AI Index reports that the United States captured USD 285.9 billion of private investment in AI during 2025, 23 times China's USD 12.4 billion, and warns that the Chinese figure probably understates real spending because of government guidance funds ¹³. That report does not break out Mexico or Latin America. ECLAC does make the uncomfortable comparison: the region is 6.6% of world GDP and receives barely 1.12% of global AI investment ¹¹. Without volume, there are no analysts measuring local prices.

And it is worth taking apart one misunderstanding before you budget. Gartner forecasts that worldwide AI spending will grow 47% in 2026, to USD 2.59 trillion, and specifies that infrastructure is more than 45% of that total and that the spending is driven above all by technology providers and hyperscalers ⁵. Those trillions are not what companies like yours spend: to a large extent they are what the industry spends building the capacity you are going to rent.

The most detailed public reference we have found on generative AI costs was published by Gartner in February 2024 (ID G00805323): it sorts use cases into three tiers and estimates, for each one, upfront investment and recurring cost. Two warnings from Gartner itself before you read the table: these are ballpark estimates derived from clients in pilot or early deployment, and they are global figures in 2024 dollars, not Mexican prices.

What matters is not the numbers, it is the jump between rows: moving from improving a process to building a new one multiplies the cost by an order of magnitude, and that decision is taken in the first week. That is where an AI consultancy either saves its client's budget or burns it, before anyone writes a line of code.

Every serious proposal breaks down into these eight lines. If the one you received is missing some of them, it is not that they are free: it is that they are not budgeted and will show up later.

There is one more cost, the most expensive of all: the project that gets abandoned. In 2024 Gartner forecast that by the end of 2025, 30% of generative AI projects would be abandoned after proof of concept because of poor data quality, inadequate risk controls, escalating costs or unclear business value ¹. It is a forecast, not a measurement, but it describes the pattern: the money is not lost in the pilot, it is lost in the months when nobody dares to cancel it.

The digital starting point is lower than international reports assume, and that reorders the budget. According to the preliminary results of INEGI's 2024 Economic Censuses, with data for 2023, only 26.2% of the country's economic units used the internet, 25.3% used computer equipment and 5.6% had a website as a communication tool ¹⁰. In many Mexican projects the first real line item is not AI: it is digitizing the process you want to automate.

The official figure closest to AI adoption has to be read carefully. Of the 1,255,625 economic units that reported using digital technology in 2023, 5.5% flagged "other technologies", a category that, according to INEGI's own footnote, includes 3D printing, artificial intelligence systems and advanced robotics ¹⁰. That is not AI adoption: the share corresponding to AI alone is smaller, the denominator is only the companies already using digital technology, and each economic unit could report more than one option.

The gap by company size is persistent. Across the OECD as a whole, 40% of companies with 250 or more employees reported using AI in 2024 (or in the latest year available), against 20.4% of those with 50 to 249 employees and 11.9% of those with 10 to 49 ⁹. These are averages centered on the G7, built on self-reported use in surveys and limited to companies with 10 or more employees, so they exclude micro-enterprises: they do not portray Mexico, but they do portray the asymmetry a mid-sized company budgets against. The regional bottleneck is well known: in the IBM survey fielded in November 2023 among IT professionals in six Latin American countries, at organizations with more than 1,000 employees that were already deploying or exploring AI, the most cited barrier was limited skills and expertise, at 32% ⁸.

Talent is the genuinely scarce input. The 2025 Latin American Artificial Intelligence Index, from CENIA and ECLAC covering 19 countries in Latin America and the Caribbean, reports that Brazil and Mexico account for 68% of the region's "active researchers" in AI (a bibliometric measure: distinct authors who published on AI, not a census of job positions) ¹². Read the other way round: Mexico is one of the two countries with the most talent in a region that receives 1.12% of global investment ¹¹, and that pushes senior hours up. Demand does not help either: in the CIO Playbook 2026 (IDC research commissioned by Lenovo among some 500 decision makers across six countries), 97% of Latin American organizations plan to increase their AI budget over the next 12 months, with an average rise of 14% ⁷. These are regional figures from a vendor-sponsored study, not Mexican ones, but implementation prices are not going to ease off any time soon.

None of the steps below requires knowing the final price before you start. All of them require deciding something in writing.

1. 1

   Define the number that has to move

   Hours recovered, time to first response, cost per case file, error rate. Without a metric agreed before kickoff, the project ends up at the mercy of whoever spoke last in the steering committee.

2. 2

   Buy the discovery separately

   One or two weeks paid on their own, with their own deliverable: process map, real state of the data and an estimate of the recurring cost. It is the only purchase in this article whose best possible outcome can be "don't do it".

3. 3

   Model the cost per query before you build

   Suppose, with round and purely illustrative numbers, that the process receives 200 requests a day and each one involves three model calls: that is 18,000 calls a month. That volume defines your bill, not the license. The exercise exists so you find out in the spreadsheet, and not in production, whether the architecture scales.

4. 4

   Reserve maintenance from day one

   Gartner models it as a percentage of the initial deployment cost: on the order of 10% for simple cases and 15% to 20% for those that extend an existing process ¹. A budget that only covers the build is financing an asset that starts degrading on delivery day.

5. 5

   Set a date to kill the project

   Decide in advance when and on what criteria it gets cancelled if the metric does not move. Remember that forecast of 30% of projects abandoned after proof of concept ¹: the expensive part is not cancelling, it is taking six months to accept that it already cancelled itself.

Realism helps here. Deloitte surveyed 1,854 senior executives across 14 countries in Europe and the Middle East between August and September 2025: 85% of organizations increased their AI investment over the previous 12 months and 91% plan to increase it again, but only about one in five qualifies as a genuine leader on AI return ⁶. The study does not include Latin America. What does travel is the substance: investing more and capturing return are different things, and the report accepts that returns take years and are not always financial ⁶.

On the optimistic side, in IDC and Lenovo's CIO Playbook 2026, 92% of Latin American organizations anticipate a positive return and expect, on average, around three dollars of value for every dollar invested ⁷. That is a stated expectation from the people doing the buying, in a vendor-sponsored study, not a return measured in the books.

There is one variable working in your favor that almost nobody puts into the model: the unit price of capability is falling fast. According to Stanford HAI's 2025 AI Index, querying a model with performance equivalent to GPT-3.5 (64.8 points on MMLU) went from 20 dollars per million tokens in November 2022 to 0.07 in October 2024: more than 280 times less in under two years ³. That is the price floor for that specific capability, not a general fall in the cost of AI. Even so, the conclusion holds: a budget written two years ago describes a world that no longer exists.

And one recommendation that saves more than any negotiation: do not train your own large model from scratch. Gartner forecasts that by 2028 more than half of the companies that did so will abandon the effort because of cost, complexity and technical debt ¹. Gartner does not segment that forecast by company size (that reading is ours), but the operational conclusion is direct: your advantage is in your data and your process, not in your training weights.