How our AI works

We use a multi-provider architecture: the system routes each task to the most suitable model and can switch providers without you noticing. No product feature depends on a single vendor, which is also why we do not publish a list of specific model versions — it would change every few months and leave this page wrong. What does not change is the deterministic workflow around them, which is where the value is.

No. enisa.ai is a private, independent service with no relationship to, affiliation with or endorsement from ENISA (Empresa Nacional de Innovación, S.A.) or the Spanish government. We use the term "ENISA" only to describe the loans we help you apply for. You submit the application yourself, through ENISA's own client portal.

No — and be wary of anyone who does. The decision is ENISA's alone and depends on your project, your accounts and the available budget. What we do is tell you early and honestly whether applying is worth it: that is what the €3 analysis is for, and sometimes its conclusion is that today you are not a candidate.

The system is built so that it cannot. Every generated claim has to rest on a document you uploaded, an integration you connected, or an assumption explicitly flagged as one, and it is cited against its source. When a required fact is missing, the workflow stops and asks you instead of inferring it. An unsupported draft is flagged in review and cannot be exported.

We keep a versioned rule registry. Each requirement stores its source, the literal quote it came from and the date a human verified it, and every report states which version answered it. An automated watcher monitors the official pages and opens a review when something changes — but it never publishes a rule on its own. A person approves that.

No. The reconciliations are arithmetic and deterministic code does them: that assets equal equity plus liabilities, that projections start where the last financial year ended, that the investment plan totals the same as the financing sources, and that own funds reach the amount requested. The AI explains and drafts from numbers that have already been checked; it neither invents nor recomputes them.

It is an indicative estimate, not a calibrated probability: there is not yet enough decision history to calibrate one, and claiming otherwise would be false. It is computed from explicit rules and rubrics you can read in the report itself, it never reaches 100%, and it drops below 20% when a critical signal is present. It always comes with the factors that moved it and the version date that produced it.

Yes, all of it. Every answer is editable, every change is recorded with its author and date, and you can restore a previous version. You approve the final text before anything is submitted.

You and the people you invite to your account. Each account is isolated from every other at the database and storage level, files live in a private encrypted bucket, and they are only served through short-lived signed links. When you buy the full application, the assigned consultant gets access to that specific application, and that access is logged.

No. Your documents, answers and conversations are not used to train models — ours or anyone else's. We work with providers under contractual no-training terms, and the sub-processor list is published in our privacy policy.

On infrastructure hosted in the European Union. The specific providers, regions and retention periods are in the privacy policy, and we will sign a data-processing agreement if you need one.

You can export everything that is yours — original documents, extracted data, reports and applications — at any time, and you can request deletion. Deletion removes the derived material too: indexed chunks, transcripts and stored files, not just the database row.

They are transcribed so they can join your company's knowledge, and we always show you the transcript to confirm or correct before it is used. The audio and its transcript are deleted together when you remove the note.

In the full application, yes. Before anything is published, a reviewer checks that each answer rests on real documentation, that the financial statements reconcile, that the form's length limits are respected and that the tone is right. If something is missing, we return it to you with the reason. The €3 analysis, by contrast, is automatic: it is a fast diagnosis, not a signed report.

That your application has a named person assigned from the moment you buy it, and that you can write to them while ENISA evaluates it: questions about the process, additional documentation they request, or how to read a communication. We cannot influence or accelerate the decision, and we are not your legal or tax advisers.

Tell us. Everything generated is traced back to the document and the prompt version that produced it, so we can reconstruct what happened, correct it and — if the fault is ours — fix the cause for everyone. If an analysis fails technically, we re-run it without charging you again.