VAL.ER.AI: What AI-Assisted Evaluation Means for Those Who Write Erasmus+ Projects

At the end of June, INDIRE, Microsoft and Hevolus Innovations presented VAL.ER.AI (Valutazione Erasmus+ Assistita da Intelligenza Artificiale): a platform, built in just eight weeks, that uses agentic AI to support the evaluators of the Italian National Agency in analysing Erasmus+ applications.

It is one of the first applications of agentic AI in an Italian public-administration process of this scale.

For those who write these applications, the news deserves careful reading, not because everything changes overnight, but because something quietly shifts in how our work will be read on the other side.

The numbers behind the decision

The context helps understand why VAL.ER.AI exists.

In 2026 alone, and only for KA220 Cooperation Partnerships, INDIRE received 667 applications, each up to 80 pages long, reviewed by 102 evaluators. Multiply that by the other actions of the programme, and the scale becomes clear.

The platform is designed to reduce evaluation time by at least 30%, automating documentary checks and pre-analysis while, and this is the point INDIRE has been very explicit about, leaving the final decision to the human expert.

Director General Flaminio Galli framed it as a way to free evaluators from the most repetitive tasks so they can focus on higher-value analysis.

Sara Pagliai, coordinator of the INDIRE Erasmus+ Agency, added a more ambitious goal: over time, comparing human and AI evaluations to build a shared training base, with the aim of achieving greater equity across contexts, including school, university and adult education, and across territories.

The project has been shared with the European Commission and operates under GDPR.

What actually changes for those writing proposals

Officially, nothing. The evaluation criteria are the same, the expert is still the one who decides, and the scoring architecture does not change.

In practice, though, three things are worth keeping in mind.

First: the readability of the text becomes even more strategic.

If a first layer of analysis is now handled by an agent that scans structure, coverage and consistency with the call’s criteria, then proposals written with a clear architecture, explicit sections, keywords used with the meaning the call gives them, verifiable claims, will be recognised as complete more quickly.

Confusing text does not just tire the evaluator: it also gives the machine less to work with.

Second: consistency becomes more visible.

An agent that cross-checks documents can identify internal contradictions, between objectives and activities, between budget and workplan, between impact and target, with a systematic quality that a human under pressure cannot always guarantee.

This is good news for those who build carefully. Less so for those who rely on the fact that a single reader, running low on time, might miss things.

Third: standardised language works better up to a point, and then works against you.

An agent trained on the criteria will easily “recognise” the vocabulary of the programme. But if a proposal is nothing more than a well-composed mosaic of Erasmus+ keywords, it also has less of the distinctive substance that convinces a human evaluator in the final decision.

The lesson, if anything, is one many good writers already know: clarity of language, distinctive content.

The open questions

VAL.ER.AI is a well-communicated initiative, with clear governance and a transparent approach. But some questions remain open, and are worth watching in the coming months.

How will the “shared training base” between human and AI evaluations really be built? Pagliai spoke of it as a tool for equity across contexts. This is the most ambitious element and also the most delicate: an evaluation memory reflects choices, some explicit, others less so, that will shape how future proposals are read.

Will there be transparency for applicants on where AI intervenes and where it does not? In an evaluation process, the ability to understand why a proposal received a certain score is a right that also protects the quality of the system.

Will other National Agencies follow? INDIRE moved first, but the volume problem is European.

A precedent that others will read closely

INDIRE is the first National Agency to formalise an AI-assisted evaluation process at this scale, but the pressure that led it there is European, not Italian.

National Agencies across Europe are dealing with a similar equation: growing application volumes, dense multi-page proposals, and a pool of evaluators that cannot expand at the same rate.

When one Agency finds a way to reduce 30% of the operational load, it is reasonable to expect others to look at that solution very carefully.

There are good reasons to expect the model to spread. The European Commission was already involved in the VAL.ER.AI design, particularly on the data-protection side, creating a regulatory foundation that could facilitate future developments. Microsoft and Hevolus now have a working reference case. And the broader European political focus on industrial AI and innovation in public administration points in the same direction.

What is less clear is how coordinated this diffusion will be.

If each National Agency develops its own tool with different training bases and different logics, there is a risk of further fragmenting the evaluation landscape, the opposite of the equity ambition VAL.ER.AI was built around.

If, on the other hand, the Commission uses this as an opportunity to define shared standards for AI-assisted evaluation across the network, we could see one of the more interesting governance experiments of this programming cycle.

Ripples towards the next MFF

This is where the story connects to the negotiation that will shape the next seven years of European programming.

The next Multiannual Financial Framework (MFF) will be discussed under strong pressure to demonstrate efficiency, reduce administrative burden and show measurable results faster.

Tools like VAL.ER.AI fit that narrative closely, and it is possible to imagine at least three potential directions of change.

Application architecture may change. If AI-assisted pre-analysis becomes standard, application forms may evolve to make them easier to analyse, more structured fields, more explicit identification of impact indicators and less free-form narrative. Proposals may become shorter and more schematic, rather than longer.

Evaluation criteria may become more granular. An agent works best when criteria are broken down into checkable elements. In the next generation of programme guides, we may therefore see a move towards more explicit and verifiable evaluation dimensions.

The competitive edge may shift. In a system where machine-readable clarity is rewarded at the first stage, organisations that begin structuring their proposals for both audiences, human and AI, may find themselves in a stronger position. Not because they learn to “game” the tool, but because they are pushed to write clearer and more consistent projects.

None of this is guaranteed.

The direction, however, is consistent with several signals coming from the European Commission: simplification, efficiency, the development of industrial AI and the use of new tools to improve public administration.

VAL.ER.AI is one piece of this emerging picture. It will be worth watching whether it remains a limited experiment or anticipates a broader change in how European projects are evaluated.