Before you continue. The Government presented the Plan IA360 on 21 September 2026: it's an announcement, not an open call, and today there are no governing rules or deadline to apply for anything. What follows doesn't depend on that funding to be true — it's the explanation of why so many AI projects get stuck at the pilot stage, with a public voucher or without one.
A business tries out an assistant that answers emails, an invoice classifier or a customer service chatbot. It works well in the demo, with the examples someone prepared to show it off. Three months later, nobody uses it: the team went back to the manual process, or the pilot is still "in testing" with nobody remembering why it never made it to production. It isn't a problem with the model. It's almost always a problem with how the project was set up before the first line of code was written.
It's also, in part, the explanation behind a figure the Plan IA360 itself sets out as its starting diagnosis: only 21.1% of Spanish businesses with ten or more employees were using artificial intelligence in the first quarter of 2025, according to the ICT survey by the INE (Spain's National Statistics Institute) (Plan IA360, flagship project 10). The plan's target is to reach 55% by 2030. Many of those businesses have already tried AI at some point; what drags the figure down is that a good part of those pilots never turned into real, sustained use.
The use case nobody asked for
The most common mistake isn't in the technology, it's in the starting point. Someone in management decides "let's put AI into something" and picks the process that sounds most impressive in a presentation, not the one that actually takes hours off the team's workload. A flashy content generator loses out, in hours of work saved, to a boring classifier that sorts a hundred emails a day. The pilot that survives isn't the flashiest one: it's the one that solves a bottleneck someone can actually point to.
There's no baseline to measure against
If nobody wrote down, before starting, how long the manual process took and how many errors it had, there's no honest way to say afterwards whether the pilot improved anything. The comparison becomes subjective — "it seems to be going better" — and a subjective impression doesn't convince whoever has to decide if the project continues or gets shut down. The baseline is measured before the first deployment, not reconstructed from memory when someone asks whether it worked.
The pilot has no owner after the demo
A pilot that depends on the person who pushed it dies the moment that person's priorities change. Nobody is left to review the exceptions, to decide whether the model got it wrong or the process itself changed, to approve moving from ten test cases to the hundred that arrive each week. Without an operational owner, the pilot doesn't fail loudly: it fades out on its own, and three months later nobody would be able to explain when it stopped being used.
The data wasn't ready, and that's discovered too late
The document that was going to feed the system turns out to have three different formats depending on who filled it in. The history of customer conversations is scattered across a CRM, an email inbox and the notes of someone who no longer works at the company. None of this shows up in the demo, because the demo is built with the cleanest data available. It shows up in week three of real production, when the first case that doesn't fit any expected pattern comes in.
The method the plan itself describes for the voucher, with no need to wait for it
The Plan IA360 sets out how it plans to implement the business AI voucher (bono empresarial de IA360): «diagnóstico previo de madurez, identificación de un caso de uso concreto y medición posterior del impacto en productividad, con una primera fase que permita validar el modelo antes de su escalado» (an upfront maturity diagnosis, identification of a specific use case and subsequent measurement of the impact on productivity, with a first phase that allows the model to be validated before it is scaled up; Plan IA360, flagship project 11). You don't need a call to exist to copy that order: first you measure where you're starting from, then you pick a single specific process, and only then do you deploy it, with the way of checking whether it worked already decided in advance. This is the sequence that separates a pilot that reaches production from one that stays in the folder of pending projects — the plan intends to require it in order to grant public funding through the voucher; a business can apply it without waiting for a call to exist.
The operational owner isn't an administrative detail
In a well-planned agents deployment, the person with authority over the process signs off on the action that costs money or reaches the customer directly, until the model demonstrates stability above a threshold set in advance — that threshold is, literally, the baseline turned into a decision criterion. When the project is simpler — connecting the CRM, email and a spreadsheet with a language model in between, without building a full agent — the cheapest way in is usually automation with n8n: the flow and its credentials stay on the business's own infrastructure, though whether the text itself sent to the language model leaves that infrastructure or not depends on where that model runs — the same question covered in where your data lives when you use AI. If it works, the automation itself reveals whether it's worth moving up to something more sophisticated.
Four signs your pilot isn't going to scale
- Nobody can say, with a figure, what improved compared with the previous process.
- The project has an enthusiastic champion but no operational owner assigned in writing.
- The use case was chosen because it "looked good", not because it was the real bottleneck.
- The data feeding the pilot is a curated sample, not the real, messy flow that comes in each week.
Frequently asked questions
What do you do with a pilot that's been "in testing" for months with nobody closing it?
Give it a review date and an owner with the authority to decide, even if that comes late. A pilot without those two things doesn't die visibly: it stays open indefinitely, consuming attention without anyone noticing until someone asks why it's still there.
Who should own the pilot within the business?
Someone with operational authority over the process being touched, not just whoever had the idea. That person reviews the exceptions, decides whether a failure is the model's fault or the process's, and authorises moving from test volume to real volume.
Does a pilot that doesn't scale mean AI isn't right for my business?
Not necessarily. Most of the time it means the use case, the metric or the starting data weren't well defined before you started, not that the technology failed. Repeating the same mistake with another process gives you the same result.
Do I need a complex project to get started, or can I try something small?
The second option usually works better. A scoped flow using automation with a language model as just one more node is, for many SMEs, the cheapest way to check whether the process scales before building something more ambitious.
Further reading
This method of validating before scaling is the same one set out in the complete guide to the Plan IA360 for the business voucher. And it's also the logic behind the Red NEURONA, designed so an SME can test with its own data before committing.