By Rodrigo Cornejo, Fixer at Mamífero
On page 10 of the AI Maturity Index Chile 2026, published by Forbes Chile with support from Kyndryl, there’s a sentence that sums up the whole problem. It says that the 34% of executives who identify data quality as the main barrier to scaling artificial intelligence is “double” the 30% who mention legacy systems.
Thirty-four isn’t double thirty. It’s four points.
I checked it against the original file because I assumed it was an extraction error. It isn’t: the sentence is intact and says exactly that. And it’s not just any typo. It’s the sentence that establishes the report’s thesis (the brake on AI in Chile is data) and it establishes it by inflating a difference that, as I’ll show, doesn’t even exist.
This piece isn’t about that report. It’s about why we’ve spent two years repeating AI adoption figures without ever asking where they come from.
What the report does well
I’ll start there, because otherwise this reads like settling scores, and it isn’t.
The report’s arithmetic is impeccable. Its four tables add up to 100. The data maturity chart publishes the bars in absolute values (2, 14, 22, 7 and 5) that add up to exactly the panel’s 50 responses, and the 2.98 average the text reports can be reproduced to the decimal from those bars. The derived percentages also check out: 44% at the middle level, 32% at the low levels, 10% at the top level. Nobody cooked anything.
The five interviews are the best part of the document and survive everything that follows. The distinction Gonzalo Oyanedel of Walmart Chile draws between flashy projects that barely pay for themselves and those that get buried in operations and move the bottom line is a useful way to organize a portfolio. Andrés Kipreos of Buk’s thesis, that the problem isn’t a lack of pilots but a lack of someone to take them into operation, precisely describes something anyone who has tried this will recognize. Reinaldo Rodríguez of Banco Itaú’s distinction between the corporate tool under risk control and the personal license outside the perimeter is exactly the conversation that needs to happen. None of that depends on the sample size. They’re five people telling what they did, and that’s where their value lies.
The problem appears when the document stops recounting what five people told it and starts describing Chile’s business landscape.
The panel is fifty
Fifty executive responses, collected between March and April 2026, spread across eleven sectors: banking, insurance, utilities, mining, logistics, retail, manufacturing, technology, agriculture, education and services. Four point five responses per sector.
On that base, each percentage point is half a company. It’s worth translating the figures that are going to circulate:
The 10% that report fully mature data are five companies. The 4% that have a formal AI ethics committee are two. The 8% that describe themselves as “AI First” are four. And in the return-by-area chart, the 2% for Finance and the 2% for Human Resources are one response each.
On that 2% (one executive) the report builds the suggestion that the little-explored back office could be where medium-term competitive advantage is decided. And from there it moves, in the next paragraph, to the infrastructure modernization agenda of the company sponsoring the study.
The rankings that aren’t rankings
Here’s the core finding, and it’s what made me write this.
The report publishes two charts in ranking format: the barriers to scaling AI and the perceived return by functional area. It presents them as an order, interprets them as an order, and builds two of its five final recommendations to the board on that order.
I ran the appropriate test for comparing two categories within a single-choice question. The results, using the report’s own 50 responses:
- Data quality (34%) versus legacy systems (30%): p = 0.86.
- First place in the return chart (48%) versus second (24%): p = 0.065.
- Second versus third: p = 0.36. Third versus fourth: p = 0.77.
- Legacy systems (30%) versus cybersecurity (16%): p = 0.21.
No consecutive position in either ranking separates from chance.
The first line is the most uncomfortable. Between the 34% and the 30% the report describes as a twofold difference, the probability that the order flips when the survey is repeated with another 50 people is practically that of a coin toss. It’s not that the gap is small: there’s no gap to measure.
And the report’s fifth recommendation says, verbatim, that you should choose one or two functional areas to build mature AI products, and points to Operations and Customer Service because that’s where the return is. That recommendation rests entirely on a ranking the sample doesn’t support.
The intervals nobody published
In twenty-three pages there isn’t a single margin of error. Not one.
With 50 cases, a 10% ranges between 4.3% and 21.4% at 95% confidence. A 4% ranges between 1.1% and 13.5%. The 10% that gives the report’s thesis its title could be double. The 4% that supports its regulatory chapter could be more than triple.
The gap between ambition and readiness exists all the same, and it’s probably real. But citing “10%” as a fact is citing the midpoint of a range that reaches double, and presenting it with two significant figures suggests a precision the sample doesn’t have.
There’s more. The report claims that large traditional companies tend to report more resistance to change than SMEs or younger companies. That’s a subgroup claim. To make it you have to split 50 responses by size and by age, which leaves cells of ten or fifteen cases. The report doesn’t publish the number of cases in those subgroups. Nor does it publish the ten-question questionnaire, so we don’t know how “fully mature” data was defined for an executive.
A minor and telling technical detail: the maturity scale is ordinal, from 1 to 5, and the report calculates an arithmetic mean for it. That assumes the distance between level 1 and 2 equals the distance between 4 and 5, which nobody established. The panel’s median, which would be the correct statistic, is 3. A 2.98 looks more precise and says less.
The problem the report itself documents
All of the index’s measurements are self-reported. Nobody audited a single data repository. What the report measures isn’t data maturity: it’s what fifty executives believe about the maturity of their data.
And here’s the cleanest contradiction in the document, because it brings it up itself.
On page 12, citing a global study by LeBow College, the report notes that 87% of global leaders say their data is AI-ready while 43% simultaneously cite data readiness as their main barrier. The report uses that figure as proof that executives’ self-perception of their data doesn’t match reality.
That’s evidence that self-reporting, on this specific question, isn’t reliable. And the report’s entire instrument is self-reporting on that question.
There’s a second problem in the same family. The 80% who “perceive themselves as leaders or on par with their competitors” is a composite category. Declaring yourself on par is declaring yourself average, which isn’t the same as declaring yourself a leader. The report never publishes the breakdown. The paradox that gives it its title (high ambition, weak foundations) is built by contrasting that 80% with the 10% of mature data, and simply opening the category changes the size of the paradox. It may still be large. We don’t know, and the document had the figure.
Which way the error leans
A small sample isn’t necessarily a biased sample. This one probably is, and in a specific direction.
The methodological note states that the panel is geared toward companies on Forbes Chile’s media and advisory radar. That’s self-selection twice over: who is in that orbit, and who in that orbit agreed to answer. The company furthest behind on AI is also the least likely to answer a survey about AI maturity.
If so, the panel is more mature than the country. Which means the real 10% could be even lower, and the gap the report denounces would be worse than it reports. The methodological critique doesn’t favor optimism: it pulls the floor out from under the figure in both directions.
The bibliography
Three of the report’s seven references are from 2024, in a document published in 2026 about a field that moves every month. One, from Entel Digital and CENIA, has no year.
The LeBow College study is cited in the body as “2026 State of Data Integrity and AI Readiness” and in the bibliography as “Global AI Readiness Study (2024).” Different title, different year. The body attributes three different figures to it (87%, 51% and 78%) without distinguishing which of the two each belongs to.
The MIT figure that 95% of AI projects never get past the pilot phase, which appears in the Walmart Chile case, isn’t in the bibliography. It’s a figure that circulated widely last year and was thoroughly debated when it was published. It appears here with no traceable source.
And a minor one, which I leave with its nuance: the introduction refers to “the first AI Maturity Index Chile,” while the bibliography cites two previous Chilean measurements of AI maturity, by MAS Analytics and by Entel with CENIA. You could argue that an index isn’t a barometer, or that “first” refers to the first edition of this particular index. That’s a reasonable reading. It’s also the kind of ambiguity that later gets cited as the first AI maturity study in Chile.
None of this is hidden
It’s worth saying plainly, because it changes what we’re talking about.
The document identifies itself on every one of its pages as sponsored content. The methodological note warns that the study isn’t equivalent to a nationally representative probability measurement and that the results should be read in the context of a targeted sample. Forbes Chile and Kyndryl hid nothing.
It’s just that the note is on page 22. After the conclusions. After the five recommendations to the board. After the reader has already taken away the 92%, the 80% and the 10%.
And the figures don’t circulate with page 22 attached. They circulate on their own, on an executive committee slide, in the second paragraph of a proposal, in a LinkedIn post. By then only 10% of Chilean companies have mature data is already common knowledge, and nobody remembers that it’s five companies from a panel of fifty with an interval that reaches 21%.
The question we don’t ask
This is where this stops being about one report.
Almost every AI adoption figure cited in Chile comes from someone with something to sell: a consultancy, an infrastructure provider, a platform, a media outlet with a sponsorship. That’s not an accusation; it’s the structure of the market: nobody funds an expensive measurement of a subject that doesn’t matter to them commercially. The bias isn’t in bad faith; it’s in who decides what’s worth measuring and which question gets asked.
What is our responsibility is having stopped asking. We repeated percentages for two years without ever finding out how many people answered.
There are three questions, and they take thirty seconds:
How many responses are behind it? If the study doesn’t say on the first page, you already know something. If there are fewer than a hundred and the headline talks about a country, the figure is good for conversation, not for decisions.
Who answered, and how did they get there? A convenience panel describes the panel. If the participants came from the publisher’s network of contacts, the result says more about that network than about the market.
Do they publish a margin of error? Without intervals, there’s no way of knowing whether a ranking is a finding or an accident. And if there are no intervals but there are rankings, the burden of proof has been reversed.
I didn’t ask any of the three for quite a while, and I cited figures of this kind in presentations. This column exists because this time I did ask them, and the result was worse than I expected.
What’s left
The AI Maturity Index Chile 2026 documents one thing well: five people carrying out this transition told how they’re doing it, and that stands on its own. Everything else is a snapshot of fifty executives presented with the syntax of a national measurement.
The gap between what Chilean companies say about AI and what their foundations allow probably exists. It’s a reasonable intuition and it matches what you see working in the field. But a reasonable intuition with a 34% stuck in front of it isn’t truer: it’s just harder to argue with.
And the irony is that the report is titled The Readiness Paradox. The paradox it documents most precisely is its own: a document that denounces organizations for trusting their data more than that data can bear, built entirely on data that trusts itself.
And a personal note: 👏Let’s👏stop👏believing👏everything👏that👏says👏AI. Please.