Digital transformation with AI: what the reports leave out and where I would start

According to the IndesIA Barometer 2026, only 4.1% of Spanish SMEs show identifiable use of artificial intelligence. That same year, a Wolters Kluwer report says 76% of Spanish SMEs use AI solutions at least once a week. Both figures were published in 2026 and, in theory, describe the same thing.

There’s no need to decide which one is lying. They measure different things, and understanding how they differ is more useful to anyone trying to digitalize a company than either number. All the figures below refer to Spain.

Four figures, four different questions

The 4.1% figure comes from the barometer of the IndesIA association, produced with Informa. I haven’t been able to find the original report, so I’m citing it through press coverage (in Spanish). According to that coverage, the study analyzes 131,554 SMEs and looks for a detectable presence of AI in each company, without asking anyone. It’s up from 2.1% in 2024 and 2.9% in 2025, and ranges from 3.6% in micro-businesses to 6.5% in medium-sized companies. A micro-business whose manager uses ChatGPT to write emails leaves no detectable footprint, so it’s left out.

The 76% comes from Wolters Kluwer’s Future Ready Business 2026 report (Spanish edition), a survey of more than 1,000 managers of SMEs with fewer than 250 employees in eight European countries, Spain among them. I haven’t found the size of the Spanish subsample. It’s a self-reported survey by a business software vendor, and the figure is the percentage who say they use AI solutions at least once a week, which anyone who opens a generative assistant every Monday satisfies.

In between there are two more sources. The Gigas Digitalization Barometer, with a sample of 1,300 companies, puts the share of SMEs using some form of AI at 36%, of which only 15% pay for it. And the Bank of Spain business survey (EBAE) (in Spanish) reported in 2025 that almost 20% of companies had AI systems, most of them at an experimental stage.

Lined up, the figures draw a fairly consistent staircase: many people use an AI chat often, quite a few fewer have put it into some process even on a trial basis, and very few companies have it integrated in a way you can see from outside. What headlines call “adoption” is usually the first step.

The gap between what companies believe and what they have

The Gigas barometer has another figure that I find more revealing than any usage percentage. 44% of companies rate their level of digitalization between 8 and 10. Only 9% make advanced use of technology when their technical foundations are analyzed: cloud, cybersecurity, AI and talent.

The details explain the gap. Two thirds use some cloud service, mainly for email and backups; only 30% run some critical process there, barely 9% have moved their ERP, and more than half keep physical servers in the office.

For AI, this matters more than any usage figure. A language model doesn’t fix data that lives in an Excel file on someone’s desktop, or processes that only exist in the head of whoever does them. If the data can’t be reached through an API or a query, the most capable agent on the market ends up as a chat someone copies and pastes things into. That helps whoever uses it, but it doesn’t digitalize the company.

My position is that most companies asking about AI agents should be asking about something more boring: where their data is, who owns each process and which part of that process is written down anywhere.

Technology isn’t what holds them back

The Bank of Spain also asks about obstacles, and the first one in its analysis is the lack of skilled staff, ahead of implementation costs and the lack of available data. Gigas points the same way: half of SMEs say they struggle to move forward with digitalization for lack of qualified staff, and 67% neither ran nor planned any technology training for their team in the past year.

That’s an awkward finding for much of the sales pitch, which presents AI as a replacement for the skilled people you don’t have. The surveys say that to adopt it you need precisely those people. Someone has to understand the process, decide which part to automate, review what the model does, maintain it when the provider’s API changes and respond when it gets things wrong. AI greatly reduces the cost of writing the software that automates a process, but it doesn’t remove the need for someone who knows which software needs writing.

That’s why I’m wary of digitalization proposals that start with the tool. “We’re going to roll out agents” is not a goal. “The monthly close takes four days because three exports have to be cross-checked by hand” is, and it’s probably solved by a script before an agent.

What has actually changed: the cost of writing the script

The most useful change AI has brought to digitalization strikes me as far less flashy than autonomous agents: writing a small, correct program now costs very little.

I see it in the internal digitalization of the company I work for, which doesn’t offer that service to others, and in the projects I do on my own. The advantage of tools like n8n for integrating two systems or moving data was that building a visual flow cost less than writing and deploying a service, and at work we use them less and less. With an AI assistant, a script that does the same thing takes minutes, and once dockerized it’s deployed like any other service, with tests, versioning and a diff you can read in a review. I already explained in another article why n8n shouldn’t replace your backend; what has changed since then is that low-code’s speed advantage has shrunk a lot, while its testing and versioning drawbacks are still there.

For a company that wants to digitalize, the bottleneck moves from programming the automation to specifying it well and operating it afterwards, which leads back to the skilled people that, according to the Bank of Spain, are missing.

Where I would start

If someone asked me to digitalize a company with AI tomorrow, I wouldn’t start with any agent. I’d follow roughly this order.

I’d pick a process with an owner, repetitive, with clear inputs and outputs, and whose current cost someone can measure, even roughly. Better a small one that gets finished than a strategic one that stays a pilot.

I’d check that the data the process needs is reachable from code. If it isn’t, that’s the first project, and there’s nothing AI about it: scheduled exports, a database instead of a shared spreadsheet, an API instead of an email with an attachment.

I’d write the automation as normal software, using AI to help write it, but deployed as a service with its tests, its logs and its repository. The language model would only come in at the step that needs it (classifying a document, extracting fields from free text, drafting something) and with human review until there’s data to justify removing it. I explain the difference between this and a demo in how to tell an AI PoC from a maintainable system.

I’d put in writing the knowledge the automation needs and that today lives in someone’s head: rules, exceptions, formats. If it later makes sense for an agent to work with that knowledge, the natural format is skills. One example is a skill we’ve built at eLearning Media that generates SQL over Blackboard Data (Illuminate): it packages the current schema, the valid fields, the candidate relationships between tables and query examples, and it requires every table and column to be validated against that schema before returning anything. The model already knew SQL; what the skill adds is the domain knowledge that only people who had written those queries many times used to have.

Only after that, with digitalized processes and accessible data, would I consider agents with more autonomy and minimal permissions; an agent with access to all the email and the ERP is a risk that deserves its own analysis.

None of this promises a specific saving, and I’d be wary of anyone who promises one before looking at the processes. What it offers is a way forward in which each step leaves something useful behind even if the next one never comes.

The question I would ask before hiring anyone

Adoption figures will keep going up, and each report will pick the metric that best fits its message. For a specific company, the useful question is which of its processes a program could run today if someone specified it properly.

If the answer is none because the data isn’t accessible, the first project is clear and involves no AI. If there are some, start with one of those, using normal software and a language model only where it’s needed. If whoever offers to digitalize your company doesn’t ask you that question before talking about agents, you know which step of the staircase their proposal is stuck on.