Where the data comes from
Sources are grouped by the question they answer, not by their number, because a count ages on its own. What firms do: INEGI's 2024 Economic Census in its definitive series, and Centro México Digital for the size breakdown on those same microdata. What the Plan says: the ATDT document of April 2026. What it is compared against: the Eurostat and OECD series, which measure the same class of firm. What moving is worth: the Ministry of Economy's communiqué of 5 June 2026, which circulates a study by the same centre on manufacturing. What instrument exists: the Plan México incentive decree in the Official Gazette of 21 January 2025, and the annual reports of the Ministry of Economy, which runs it. And what computing exists: the TOP500 lists and Secihti's communiqués on Coatlicue. The full list, with each one's trap, is on the sources page.
The census's four bases
The census publishes three universes and confusing them changes the result by a factor of four. The census universe is 5,468,180 economic units. Of those, 1,444,950 used the internet. Of the latter, 1,266,352 used some digital tool, and it is on that base that INEGI calculates the 2.1 percent for artificial intelligence. The fourth base is not a universe but a cut: among firms with ten or more people, Centro México Digital computes 8.0 percent from the same census's microdata. It is the figure that circulates in the press, and the one this report uses to compare with the OECD, which measures with that same cut. It is marked as reported, not official: the microdata are not freely downloadable and the calculation cannot be reproduced from outside. A note on vocabulary: an economic unit is an establishment, not a company, and the report writes firm where the sentence calls for it, its own headline included; what is counted is always establishments. And that comparison has a limit: the size cut matches, the instrument does not. Eurostat asks about several artificial intelligence techniques and counts anyone using any of them; the Mexican census asks a single yes or no inside a list of nine tools. How much that moves the figure is unknown, and the report does not estimate it. The report always uses INEGI's base and declares the conversion when taking it to the full universe.
What is our own calculation
Seven things, and the most important goes first. The band of where Mexico would be today, between 8.0 and 19.8 percent, which comes from applying to the Mexican figure the growth Eurostat did measure in Europe: it is arithmetic on someone else's series, not a Mexican measurement, and the report publishes no point inside that band. The size ratios against Europe, 0.94 for small firms and 0.56 for large ones, which are the division of two measured figures. The count of units using artificial intelligence, which comes from applying the percentage to the base and is published with a rounding band, between 25,960 and 27,227. Coatlicue's multiples against the TOP500. And its two delivery dates, 2027 and 2028, which come from adding the declared 24 month horizon to each of the two starts the official documents declare: planning in one and construction in the other. And the two ratios, which also come with a band because they derive from rounded figures: the share of the whole census, one in every 206, falling between 201 and 211; and the distance between the first and last state, 5.4 times, falling between 5.1 and 5.8. All seven are labelled as own calculation and none is presented as official.
How the Plan was counted
Over the document's body, 6,746 words, trimming the cover, table of contents and bibliography: counting over the whole PDF repeats section titles and adds the words of the nineteen bibliographic entries. The comparison is done without accents and in lower case. Each term keeps its literal context and those contexts are published in the data dashboard, up to ten per term: above ten the occurrence is counted but not transcribed, and the chart declares it. Ten covers in full any term the report makes a claim about. In every quotation in the report, the original's long dash is written as a comma; the rest is left as it is.
An access problem
The Plan's canonical URL on ATDT's portal returns 403 to any automated client, on both domains. The file that was measured was downloaded from a public mirror, and that is why the repository stores its SHA-256: a word count over a document that cannot be re-downloaded from the source does not hold without the fingerprint of the exact file that was counted. The citation keeps the official URL.
That was not the only wall, and all four have a similar shape: the page opens in a browser and the document will not come down to a program, or there is no way to find its path. The PPEF 2027 index on the Finance Ministry's site serves HTML but exposes no links to the PDFs, whose paths carry a session token: without the path, 404; with the exact path, which came from a person, the Draft Decree came down with a 200, and the same pair of tokens later opened the analytics and the volumes. The open data mirror hosted by ATDT returned an Akamai 403 on 15 September. The OECD's press page also returns 403, though its SDMX API answers without trouble: the official figure was one endpoint away from a release that looked unreachable.
From that comes a rule the report applies throughout: the wall gets recorded, not smoothed over. When a document that is not published yet cannot be told apart from one that is blocked to clients that are not browsers, the report claims neither.
What the report does not claim
It does not claim causality between data centres and adoption: the census has a single measurement and a single measurement cannot sustain a causal claim. It does not claim the Plan is lying about Coatlicue, but that the comparison cannot be verified without the unit. It does not turn the count of AI-using units into an exact figure, because the percentage it comes from is rounded. And above all it does not claim where Mexico stands today: the census measures 2023 and there is no second Mexican measurement, so the report publishes a band and refuses to pick a number inside it. Nor does it claim that Coatlicue's construction began or that the foundation stone was laid: the Plan declares both dates and no later official document confirms them, so what the report says is that there is no confirmation, not that it did not happen. And it does not credit Crece tu MIPYME with the 841,156 digitalised firms: the report counts from three and a half months before the programme existed and does not separate the two.
Why other documents are counted, and how
Comparing two documents measured differently is not comparing. The Ministry of Economy's annual report is counted with the same method as the Plan: word roots, declared exceptions, and the literal context of each occurrence stored in the dataset. The difference is scope. For the Plan only the body is counted, trimming cover, index and bibliography, because counting the index would inflate its own terms. For the report everything is counted, cover, index and annexes included, and that is deliberate too: what holds up the chapter is an absence, and giving the document more text than it is due can only make a zero stop being one. The third is the 2027 budget's Draft Decree, also counted whole and without cuts, with two rules of its own: every whole-word term is counted twice, by substring and tolerating the spaces the extractor inserts inside words, and if they differ the count aborts; and no amount is attributed to annex rows, because the extractor separates the amounts column from the text. The only figure taken from there is the Ramo 55 total, read from the table of totals by ramo.
The volumes and the analytics from the same package come in by two roads. The programme strategies of Ramos 55 and 38 are text and are counted with the decree's rules, whole. The analytics are not counted, they are added up: the Analítico de Claves carries every budget line of the project with its amount, and the package's catalogues give each programme, unit and ministry its name. The guard is one of coincidence: the Ramo 55 total from adding up its 399 lines has to be the one in the decree's table, or curation aborts. Two documents from the same package saying the same thing by different roads is what counts here as verified. And the analytic only lists lines with an amount, so a programme without a line is published as such, not as zero pesos.
The trap that does change the result is SME. Counted as a substring it comes out 51 times, because every MSME contains it. On its own it comes out 5. The discount travels in the dataset, in the excluded mentions column, not hidden in the code: comparing an inflated 51 against the Plan's one would have been counting the same word twice to win an argument.
How the regional index is used
The 2026 Latin American Artificial Intelligence Index is by ECLAC and Chile's National Centre for Artificial Intelligence. Its scores carry the academic label: they are a composite index of 126 sub-indicators, normalised and weighted by its authors, not a measurement. Positions are published with the distance to the country above and the one below because the index carries no margin of error, and none is compared with the 2025 edition, which its authors declare not fully comparable.
Business adoption by country is marked as estimated and, with the IDB's projection, these are the only estimated figures in the report. The report does not estimate them: the adoption ones are by Raúl Katz and Juan Jung, in a study for ECLAC's Digital Development Observatory, and come from a model trained on a 2019 to 2025 panel of the countries that do measure, with eight variables, among them 5G coverage, digital skills, specialised talent and the presence of firms selling AI solutions. They are used for what they are good for, ranking twenty countries, and never beside the census figure as if they measured the same thing. The OECD reference the index places next to its estimates is marked as reported, because it is taken from the index and not from the OECD database.
The figures come from the index's PDF and, where the PDF does not print them, from the data workbook of its viewer. The two do not always agree: on enabling factors they differ by up to 0.3 points, and the PDF prevails. The rates by size for the twenty countries sit in a diagram the PDF carries as an image: they were transcribed by hand and the transcription is checked against the averages and rates the text does print. The PDF is not kept in the repository, because ECLAC asks that copies not be uploaded to external platforms: its fingerprint is kept and the program downloads it again.
How the IDB figures are used
The Inter-American Development Bank's figures come from its report From Digitalization to Artificial Intelligence: Turning Promises into Productivity, in the Development in the Americas series, presented on 16 November 2026. On 17 November the bank's publications site answered an automated client with an anti-bot filter, and these figures were published from the press release. On the 18th the document was read in its Spanish edition, downloaded by hand, and they were redone from it: the overview chapter and the one on firms in full, and from the chapters on productivity, digital adoption and government procedures, the sections the figures come from; in the rest of the book every mention of Mexico was searched. The PDF is kept in the repository with its fingerprint, because its licence allows it.
None is marked as official. The projection is marked as estimated: it comes from a model whose high scenario assumes AI adopted in 80 percent of exposed tasks and 40 percent more productivity in them, and whose low one assumes 11.5 and 13.5. These are effects accumulated over ten years, not annual growth, for thirteen countries, with parameters taken mainly from evidence for the United States; the report asks that they be read as projections under common assumptions and not as estimates of the region's behaviour, and warns that they may overstate the gains in the most informal economies. In its own table, with data for 2018 to 2022, 66.3 percent of Mexican workers are not enrolled in social security, against 55.5 in the average of the countries it lists. The surveys and the catalogue of procedures are marked as reported: the report does not have their microdata.
The figures on people are from the 2025-26 AmericasBarometer: a face-to-face household survey of adults aged 18 to 64, with about a thousand interviews per country, fielded between October 2025 and May 2026 and weighted to represent each country. There are sixteen countries, plus Barbados from another survey. The report publishes no margin of error, which is why Mexico's place, which comes from ordering the values the report prints, is given as fragile and the distance from the regional average is not. The comparison groups come from other sources: European OECD countries from Eurostat; the United States from the Harvard Project on Workforce; and the OECD on procedures from Eurostat and the International Telecommunication Union. They count people using a tool, not firms: none is placed beside the census, the OECD series or the regional index's estimate as if they measured the same thing.
The firm survey is that of ConnectAmericas, the bank's business platform: voluntary, with 1,549 firms in 25 countries, 70 percent of them with ten employees or fewer. The report says it is not representative and probably overstates adoption, and gives no rates by country. The 82 percent is calculated over the firms that already use generative AI, not over all those surveyed: with the press release this could not be known, and that is how it was published on 17 November. The catalogue of procedures comes from another study by the bank, from 2025, which was not read, and the report does not say what date it corresponds to.
Two discrepancies are left in view. The press release gave a 4.2 percent wage increase over a decade; the book does not print that average but the range across countries: from 2.3 to 5.3 percent if workers can change task, job or sector, and falls of 13.9 to 20.9 if they cannot. The range is what is published. And within the book, generative AI use in European OECD countries is 43 percent in the charts and in the adoption chapter, and 44 in the text of the overview chapter: 43 is used, which is also the press release's figure.