What AI project the Plan IA360 will fund, and how to build it

Status as of 22 September 2026. The Plan IA360, the Spanish Government's AI plan, was presented on 21 September 2026. There is no published call for applications, no governing rules and no application window. The plan schedules the general call for the voucher, funded with €600 million, «antes de final de 2027» ("before the end of 2027"), and only after a controlled pilot planned for «primer semestre de 2027» ("in the first half of 2027"). Nothing that follows is a procedure: it's the description of a project, with or without a voucher behind it.

The Government presented the Plan IA360 on 21 September 2026, organised into eight areas of action and fourteen flagship projects; the official communication also summarises it as four broad goals — technological muscle, economic adoption, governance and social contract — which is a different, non-contradictory reading of the same eight areas in the technical document. You can check the full detail in the official plan document; we won't summarise them one by one here, because that isn't what matters to a business that doesn't manage computing infrastructure or talent policy.

What matters is hidden inside flagship project number 11, the bono empresarial de inteligencia artificial (the business AI voucher). The plan doesn't just announce a €600 million allocation: it describes, in more detail than is usual for a document of this kind, what the project receiving that money has to look like. It says, literally, that «la implantación incorporará 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» ("the rollout will include a prior maturity diagnosis, identification of a specific use case and subsequent measurement of the impact on productivity, with an initial phase to validate the model before scaling it up"). It adds that the voucher's design «incluiría pagos por hitos y condiciones acerca de formación en IA de los trabajadores e incorporación de la IA a los procesos internos de los beneficiarios» ("would include milestone payments and conditions on AI training for staff and the incorporation of AI into beneficiaries' internal processes").

Read it without the administrative filter and it says something else: diagnosis, use case, pilot, training, day-to-day integration, measuring the result. That isn't a new bureaucratic requirement: it's, almost word for word, how you build an AI project that doesn't end up abandoned after three months. The plan is the pretext for this article; the subject is your project, whether or not it ever has a voucher behind it.

Phase 1 — Know what to automate before automating anything

The most expensive mistake in an AI project isn't picking the wrong model. It's starting to build before knowing what problem you're solving. A business that types up every supplier invoice by hand because no one has ever stopped to count how many arrive each month, or a team that answers the same type of support query fifteen times a week without anyone ever counting it, doesn't need a model first: it needs a diagnosis first.

A maturity diagnosis looks at three things: what data the business has and what state it's in, which processes could genuinely benefit, and who will maintain them once the project stops being a novelty. And sometimes — more often than a client expects to hear — the honest conclusion is that AI doesn't solve that particular problem; reorganising a process or training a person does. Selling an AI project where none is needed is the fastest way to make the word stop meaning anything inside a business. That's the work of a maturity diagnosis and roadmap.

Phase 2 — The specific use case: what changes and for whom

The plan is explicit about what kind of project has value: services that use foundation models «como insumo» ("as an input"), incorporating «valor añadido propio mediante desarrollo, integración, conocimiento sectorial, tratamiento de datos o rediseño de procesos» ("their own added value through development, integration, sector knowledge, data processing or process redesign"). And it's just as explicit about what doesn't count: «el bono no financiará la mera suscripción a licencias» ("the voucher will not fund the mere subscription to licences").

That line, translated into real processes, separates two very different things. Switching on a generic assistant and hoping the team changes its habits on its own is a licence. Building an agent that reads the supplier invoice, checks it against the original order and only asks for human confirmation when something doesn't match is a process redesign with its own added value. The difference isn't which model sits behind it: it's whether someone actually sat down to redesign the process around it. We go into more detail, with more examples of what kind of project fits and what doesn't under the plan's literal criteria, in which AI projects fit the IA360 voucher.

When the process depends on finding information scattered across contracts, manuals or old emails, the use case is usually different: a document search system that cites the exact source of every answer, instead of forcing someone to reread a sixty-page PDF every time a question comes up.

Phase 3 — Train the people who will use it, not just whoever bought it

The plan provides, in the conditional, for «formación en IA de los trabajadores» ("AI training for staff"), and it makes sense: an AI project that only the person who commissioned it understands dies the moment that person changes role. Training that actually works isn't a generic afternoon talk about what a language model is; it's different depending on who's receiving it: leadership needs to understand risk and governance to make decisions; the technical team needs to understand architecture and operations to maintain it; the team using the tool every day needs to understand what they can ask it and where the limit sits, so they don't trust an answer that hasn't been checked. That's the approach of role-based corporate training, built on the company's own cases, which can be funded through FUNDAE (Spain's foundation for subsidised vocational training).

Phase 4 — Make it part of the process, not sit beside it

This is where more AI projects die than anyone admits in public: the pilot works, everyone applauds at the demo, and three months later no one uses it because it still lives outside the real workflow. The plan calls this «incorporación de la IA a los procesos internos de los beneficiarios» ("the incorporation of AI into beneficiaries' internal processes"), and it's the part that gets neglected most because it isn't the showy part.

Genuinely incorporating it means the system enters through wherever the work already enters today — email, the CRM, the ERP, the warehouse scanner — and comes out wherever it's already reviewed. An automation that connects the tools the business already uses, with the model as just another step in the flow rather than a separate tab someone has to remember to open, is usually what decides whether the project survives its first quarter.

A common example: a document search system that answers accurately but that only the project lead ever opens, because you have to launch a separate chat to ask it anything. The same system connected to email or the CRM, so the answer arrives where people are already working, is the one that actually changes the process. The technology doesn't change; what changes is whether someone bothered to stitch it into the daily flow instead of leaving it as a parallel experiment.

Phase 5 — Measure the impact, not just the launch

The last element the plan lists is «medición posterior del impacto en productividad» ("subsequent measurement of the impact on productivity"). It sounds obvious until you try to do it: measuring the impact requires having recorded, before you start, how long the process took, how much it cost and how many errors it had. Without that baseline, any figure presented afterwards is just an opinion with decimal points.

You don't need a data department to build that baseline. It's enough to record the hours the process consumes today, its cost per hour, and an honest estimate of what share of it is automatable. To estimate upfront whether it's worth it, there's the automation ROI calculator; the measurement comes afterwards, comparing the baseline against the process once it has changed.

Who does each phase, in short

You don't need your own AI department to go through the five phases; you need to know who's responsible for each one. Here it is at a glance:

None of these phases requires waiting for a call for applications: this is simply the order in which it makes sense to do things, whether or not the project has a voucher behind it.

The rest of the plan, in one sentence

The other thirteen flagship projects cover areas such as infrastructure, security, education and public administration: from the computing gigafactory to the AI Safety Institute inside AESIA (Spain's AI supervisory agency). If that part interests you — data centres, talent, post-quantum cybersecurity — it's fully described in the official Plan IA360 document. Here we're sticking with what directly affects your project.

The voucher, in short: what it funds and what it doesn't

The bono empresarial de inteligencia artificial is funded with €600 million and is aimed at services provided by European technology companies, not at buying a standalone licence. The plan schedules a controlled pilot for the first half of 2027; the general call, with those €600 million, «antes de final de 2027» ("before the end of 2027"). The full breakdown of what kind of project fits, with concrete process examples measured against the document's literal criteria, is in which AI projects fit the IA360 voucher (and which don't).

One clarification before we go on: we don't process applications for the IA360 voucher, because there's no procedure to handle today, and we don't promise that a project will get it or how much it would receive. What we do is the work of the five phases above, whether or not a call for applications ever exists.

See AI working in your sector before you decide

Within the same area, the plan provides for a network of centres where you can test it, and it puts it this way: «la empresa ve la inteligencia artificial funcionar en su sector, la prueba con sus datos y la despliega por suscripción» ("the business sees artificial intelligence working in its sector, tests it with its own data, and deploys it by subscription"). That's, in essence, the definition of a well-run proof of concept: first you see it work, then you test it with your own data, and only then do you decide whether to buy it. That order (see, test, decide) is what stops you buying an AI project blind, and you don't have to wait for a physical centre to open to apply it. How to do that today, risking as little as possible, in testing AI with your own data before deploying it.

What you can do now, without waiting for the call

The method the plan describes (diagnosis, use case, pilot, training, measurement) doesn't depend on a call for applications existing. You can start today: the diagnosis and the baseline help you decide better, whether or not a call exists. What the governing rules will require, and whether work done beforehand will count, still isn't known. The quarter-by-quarter roadmap is in your AI project's roadmap, from now to 2027 (with or without a voucher).