The EIC Accelerator funding (grant and equity, with blended financing option) by the European Commission (EC) and European Innovation Council (EIC) awards up to €2.5 million in grant and €10 million in equity financing per project (€12.5 million total) and is designed for startups and Small- and Medium-Sized Enterprises (SME), often supported by professional writers, freelancers or consultants.

This article is Part 4 of the series and explores the use of Artificial Intelligence (AI) in the evaluation of EIC Accelerator grant proposals (see ChatEIC). Part 1, Part 2, and Part 3 can be found here.

The Open Sea

After restructuring the entire EIC Accelerator evaluation criteria, there is an opportunity to develop a new system for AI assessments that has the potential to be far superior to human evaluators. As described in the previous parts of the series, it also allows the EIC to improve its most important metric, the Step 3 interview success rate, since it shows how well the current process filters applicants in earlier stages.

While the EIC has so far failed to improve this central metric, it has the opportunity to turn things around and fix its shortcomings.

Creating an AI evaluation system is not straightforward, and this article might not reflect the best architecture, but it elaborates on a concept that has the potential to be effective.

Round and Round

One key downside of generative AI technologies is that they can hallucinate and can add a lot of bloat to the content they produce. This can be challenging if one wants to use it as a scalpel instead of a broad sword, since it is important to get minuscule details right.

Of course, to mitigate such risks, one can always introduce redundancies such as multiple iterations of AI evaluations that are then averaged to gain an enhanced result. This is actually an easy and highly recommended approach since the cost of repeating an individual evaluation dozens of times will still be far below any fees paid to remote evaluators or consultants. This asymmetry allows the EIC to not only enhance the AI instructions quickly but also to eradicate randomness to an unprecedented level.

The Good, The Bad, And The Judge

Leading questions will always be a problem for AI since it will likely aim to create a conclusive answer to any question instead of presenting a nuanced case. That means, asking if the technology is novel might lead to a result that is hard to rank against other companies since each AI might just highlight the innovative character.

A better approach could be adding a step that only extracts data before it is being used, similar to the process of extracting data from companies during the AI grant writing process (i.e., ChatEIC).

This extraction step would utilize multiple AIs that each have their own focus, all working in unison to generate a final score regarding each criterion.

Good Cop, Bad Cop: Extracting Data

For AI evaluations, this extraction step could be implemented through a positive AI (the good cop) and a negative AI (the bad cop) that only try to find reasons for, and reasons against, a certain criterion.

This way, it can be assured that all negative and all positive aspects of the project are extracted before a judgment is made. This avoids the risk of a single AI glossing over all arguments because it has already made up its mind halfway through the document.

It also simplifies the debugging by the EIC’s AI team since it allows them to fully trace why decisions were made and if data has been missed. The more steps are included, the more oversight the EIC will have.

Judge Draft: Final Scoring

After the good and the bad cop AIs have added their input to the respective section, the final results can be handed over to the judge AI, which provides a final score based on the pros and cons. While the good and bad cops are purely trying to find praise or fault, the judge will use a weighted assessment system to identify which project ranks high and which ranks low.

This is the aspect where the EIC can integrate its own expectations and agendas into the scoring since it can decide what is more or less important overall. For example, an inexperienced team will be highlighted by the bad cop AI, while the good cop will highlight the domain expertise and advisors of the company. The judge can take that data and have specific instructions to favor domain expertise over industry experience since this is a natural feature of DeepTech founders, as described above.

The EIC can likewise introduce similar weights in other aspects, such as customer traction over existing revenues, past team accomplishments over small team size, or the scientific nature of the technology and IP over the maturity of the revenue stream development.

This way, the EIC can color the entire evaluation process to be DeepTech-aligned while having more power over the evaluation process than ever before. The EIC Board can easily obtain a briefing regarding the exact evaluation process and can make direct suggestions for improvements as well as gain quantified feedback on the results.

Spam Filters

On the flip side, any proposal that reaches the Step 3 interview stage but should have never made it this far can act as a new filter to ensure this mistake is not repeated. The EIC can therefore set up its AI system to act as a spam filter to avoid low-quality or simply unfitting projects from making it to Step 2 and especially Step 3. If the AI system missed critical reasons as to why the company was a poor fit for the EIC, changes can be implemented globally.

For the EIC, such a system would be a game changer and it will be a breath of fresh air for the applicants since it has the potential to become so specific and accurate over the coming years that it will reject most unfitting projects already in Step 1, while Step 2 and Step 3 should have success rates exceeding 60%. Not only will the EIC save millions in consultant fees, but the applicants will save thousands of hours in wasted grant application efforts.

All this time and money can be used to build great companies instead of fueling the hamster wheel of grant applications.

Hive Minds: Multiple Evaluator Disorder

To finalize the AI evaluation system, it simply needs to be scaled across all general and sub-criteria, whereas a good cop, bad cop, and judge trio is set up for each. Instructions for each set can be easily kept separately and updated according to overall feedback from the EIC and other stakeholders.

This system is not only comprehensive, but it is also very easy to incorporate. For ChatEIC, this would be extremely easy to integrate since the architecture is already present and it uses a system where all instructions for the AI models are separated to allow easy updates and additions of new AI modules.

Albeit ChatEIC will likely not adopt an AI evaluator in the near future since it is not a product that applicants are interested in as of today. It is something only the grant agency itself, namely the EIC, will need in the future.

Confidentiality: Will European LLMs Please Stand Up?

Confidentiality is an important consideration for an AI evaluation process since the proposals will contain some of the most important IP in the entire EU. This means that using non-EU-based AI models is a risk. It is a minimal risk since blocking data from being used in training activities is usually the default, but one cannot be careful enough in the current geopolitical climate.

The only way for the EIC to reasonably integrate AI would be through using a European LLM or at least an open-source LLM that is operated by a European company. While the former reduces the pool of options down to just a handful of companies, the latter would allow for a significant breadth of options.

There are likely many companies that the EIC could contract to integrate such an architecture since the concept and integration are very straightforward. ChatEIC uses API to connect to AIs and uses server-side instructions to create the correct responses. Such a system could easily be run locally on the EIC’s servers using an open-source LLM. Additionally, integrating web search to help the AI is an easy function to include (i.e., ChatEIC has web-search functionality).

This concludes the article series on AI evaluations for the EIC Accelerator as well as specific concepts on how to integrate them.

These tips are not only useful for European startups, professional writers, consultants and Small and Medium-Sized Enterprises (SME) but are generally recommended when writing a business plan or investor documents.

Deadlines: Post-Horizon 2020, the EIC Accelerator accepts Step 1 submissions now while the deadlines for the full applications (Step 2) under Horizon Europe are:

The Step 1 applications must be submitted weeks in advance of Step 2. The next EIC Accelerator cut-off for Step 2 (full proposal) can be found here. After Brexit, UK companies can still apply to the EIC Accelerator under Horizon Europe albeit with non-dilutive grant applications only - thereby excluding equity-financing.

Contact: You can reach out to us via this contact form to work with a professional consultant.

AI Grant Writer: ChatEIC is a fully automated EIC Accelerator grant proposal writer: Get it here.

EU, UK & US Startups: Alternative financing options for EU, UK and US innovation startups are the EIC Pathfinder (combining Future and Emerging Technologies - FET Open & FET Proactive) with €4M per project, Thematic Priorities, European Innovation Partnerships (EIP), Innovate UK with £3M (for UK-companies only) as well as the Small Business Innovation Research (SBIR) and Small Business Technology Transfer (STTR) grants with $1M (for US-companies only).

Any more questions? View the Frequently Asked Questions (FAQ) section.

Want to see all articles? They can be found here.

For Updates: Join this Newsletter!