EIC Grant Writing in the Age of AI: Quality, Confidentiality and Speed

This talk is about the EIC Accelerator and about AI, two topics that are now hard to separate. I am a grant writer myself. I started out writing everything by hand and saw the whole transition when grant writing moved to AI. AI is more than a tool here: it is becoming the centre of how grant writing works.

The EIC Accelerator in Brief

The EIC Accelerator is a funding programme of the European Commission under Horizon Europe. It is meant for later-stage companies that already have a prototype and have developed their technology to a certain degree, usually at technology readiness level (TRL) 5 or 6. The formal rule is TRL 5 completed: you have something in your hands that you have already validated. The goal is to bring it to TRL 8 or 9 and to scale.

The biggest draw is the grant, but the programme also offers equity financing. In total a company can receive up to €12.5 million: €2.5 million as a grant and up to €10 million in equity. The programme is very well known and very well funded, with a budget of about one billion euros a year that changes from year to year.

Who Gets the Funding: DeepTech

Most of the programme's marketing is about DeepTech: technologies that are capital-intensive and highly technical, often come from science, and took a lot of time and resources to develop. They are unique and not broadly available on the market. I add trendy technologies to that list, because the funded companies show clear trends, such as hydrogen in transport or quantum computing. Science is not a requirement. Many funded software and engineering companies do not come from a university and are not spin-offs, but the programme as a whole is very technical.

I analysed every company selected since 2019, over a thousand of them, and grouped them by sector. Health leads with 36%, followed by digital and computing (11%), energy (10%) and agrifood (8%). Health covers software, hardware, implants, medical devices and pharma. I believe it leads because health innovation is a natural fit for DeepTech: it is patent-heavy, often comes out of universities and needs a lot of money for the clinical process.

After that comes a very long tail: climate and environment (7%), electronics and semiconductors (7%), industrial technology (6%), mobility and transport (5%), materials and chemicals (5%), then space (3%), construction (2%) and consumer goods (1%). Small percentages of over a thousand companies are still a lot of companies.

n = 1: Do Not Copy Past Winners

The way I like to put it is that the programme is n = 1. Innovation means funding something unique, and most companies are unique and interdisciplinary. If you look through the list of winners and find a company exactly like yours, that is more of a bad sign: what you are doing may not be unique, and it was already around a year or two ago when that company was funded.

I still recommend looking at the list of funded companies, looking them up and reading what their projects do, because it shows what the Accelerator is looking for. What I do not recommend is trying to match an industry or copy past winners. Finding a similar company also does not exclude you. Quantum computing is a good example: many quantum companies with competing architectures have been funded. Unlike a venture capital fund, the EIC has no conflict of interest and does not sit on boards, so it does fund companies that compete with each other.

17.5% of Winners Build on AI

Looking at the technology rather than the industry, the most common one is no surprise: 17.5% of the selected companies use artificial intelligence, ahead of computer vision (4.4%), robotics (4.4%), cloud software (3.8%) and semiconductor design (2.9%). AI can be added to almost any technology, and it is one of the biggest technological revolutions in a long time because it disrupts knowledge work: anything where a human needs to think.

There is a caveat. A pure wrapper around an existing AI model, built on someone else's API, is usually a tough sell. Wrappers still get funded, far less often, and the Scale Up Europe Fund, whose public money comes through the EIC, recently invested in a company built around existing models. Most of the selected companies that use AI do not develop their own models either: a €2.5 million grant is not enough to build a frontier or foundation model.

Three Steps and the Deadlines

The process has three steps. Step 1 is a short proposal: a 12-page PDF, a read deck of about ten slides and a video. Step 2 is the long proposal: everything from Step 1 extended to 20 pages plus 10 pages on the project, with annexes such as letters of intent, financials and the budget, more financial annexes and another video. Step 2 now also includes an online interview with an evaluator. Step 3 is the interview in Brussels: a 10-minute pitch followed by 35 minutes of questions. You have to pass all three steps.

Step 1 has twelve cut-offs a year, on the first Tuesday of every month. In 2027 they are 5 January, 2 February, 2 March, 6 April, 4 May, 1 June, 6 July, 3 August, 7 September, 5 October, 2 November and 7 December. Step 2 has four cut-offs in 2027: 20 January (recently moved from 21 January), 17 March, 1 September and 4 November. Step 3 interviews happen twice a year, after every second Step 2 batch: after 17 March and after 4 November.

That calls for planning. Even a Step 1 evaluation takes about 50 days or longer, and with only four Step 2 cut-offs a Step 1 approval that arrives just after a deadline means a long wait for the next one.

The rule on how often you can apply is three strikes and you're out. If you are rejected three times at any stage, you are blocked until the end of the framework programme in 2027. The block used to apply per project; it now applies per legal entity.

Should You Apply? Most Companies Should Not

The EIC advertises the programme very broadly, and anyone can apply. But it takes a lot of effort, so most companies should not. The reasons not to apply:

• Thematic and technology fit: the programme looks for a very particular fit, technology of strategic importance for Europe. Not everything has to be core strategic, as a lot of funded MedTech shows, but a cashew farm in Sri Lanka will not work even if the company operates in the EU.
• A long process: without the commitment to stick with it, especially after a rejection, it will not work. I estimate that most winners were rejected at least once somewhere in the process.
• Interviews: you have to defend your company, project and plan at prototype level, including commercialisation. Evaluators do not expect a flawless performance, but they do expect flawless answers.
• Randomness: four evaluators in Step 1, four more in Step 2 and a jury of five, sometimes with a European Investment Bank representative, in Step 3. With that many people involved, the outcome can appear random.
• Expertise: it requires a lot of it.

The reasons to apply: you are a good thematic and technological fit; the funding is €12.5 million, €2.5 million of it as a grant; much of the writing and submission can be outsourced, although you have to understand and stand behind the project because you will face two interviews; interviews and a trip to Brussels are no issue for you; and, most importantly, the Accelerator is part of a larger funding strategy. You should be committed to the project, talking to investors and already raising money, with a plan that works even without the grant. If you need money now to avoid bankruptcy, it will not work.

A Long-Term Engagement, Not a Short-Term Grant

The application process is long, the grant project runs for two years with reporting, and the equity comes through a multi-year process in tranches. Equity takes time. Of the companies selected in each year, about 50% (2021), 56% (2022), 32% (2023), 13% (2024) and 1.4% (2025) had received at least a partial equity payment. The EIC and the agencies around it are not investment bankers or venture capitalists and had to work out how to handle equity, which created a backlog.

The grant is straightforward, because the Commission has paid out grants efficiently for a very long time. For the equity, plan for one to three years. The data shows some companies receiving their first equity injection soon after the deadline while others are still waiting, and the amount can change in due diligence: some companies had €15 million approved and received about €29 million to €30 million.

How AI Has Changed Grant Writing

I started grant writing at the end of 2016, completely manually. It used to be a quiet, solitary job: a few days of research, reading everything by hand, then one or more client calls to understand the technology, then weeks or a month of writing without the client hearing from you, then the first draft and one or two weeks of editing. It was almost like writing a book, and I enjoyed it.

Today the direction is the same, but the work is very different. I now work mostly with AI. When I start a project I get a large data dump from the company: pitch decks, financials, technical data and research. I spend the first few days, rather than weeks, discussing that data with AI along with my notes and my idea of the narrative. Client calls then only cover the questions that are still open, because I already understand the project. The first draft comes quickly, and there is more time for editing.

The transition took me a while, even though I had built my own AI writer for the EIC Accelerator, ChatEIC. I believed the highest quality came from writing everything myself, and it took time to see that the quality comes from doing the thinking myself, not the writing. Over a month of writing, your subconscious keeps producing ideas, such as a strong angle for the commercial strategy, and those moments of serendipity matter. More editing time also suits the new process: with more interviews, clients have to be able to defend every part of the proposal, so there is more room to align it with them.

AI is not only speeding up lengthy steps but changing how the work is done. I do not think it makes proposals better or worse, but it makes them different: proposals written with AI include decisions, sources and research I might not have found on my own.

Disruption: Consultancies, Grant Writers and Grant Agencies

AI is disrupting consultancies, grant writers and grant agencies. Applications rise, grant agencies are overwhelmed, and consultancies are cut out. A consultant's daily work is talking, emails, analysis, writing and advice, which is exactly what AI is good at, and some consultancies are in real trouble with tanking sales and revenues.

Over the last few weeks the change has accelerated, and I believe the cause is the spread of personal agents with their own knowledge base. Clients are now so used to them that the agent comes first, before they reach out to a consultant. I would not be surprised if 2027 forces many consultancies to become much leaner and to change radically how they work, because a client who is agent-first may no longer talk to you but through Claude or ChatGPT.

AI is a wave, and the only way to survive is to surf it. Whether you are a consultant or an applicant, use AI as much as possible, less for an immediate benefit than to understand what the technology can do. I use it every day, including for this presentation: every word comes from me, but AI produced the design in about ten minutes.

Grant Agencies Are No Better Off

Individuals adopt AI quickly, companies more slowly, and grant agencies slowest of all: they are bureaucratic institutions without a profit incentive. They are flooded with applications that keep prolonging evaluation times. AI in submissions is far more widespread than AI in evaluations, and no one can control which AI tools evaluators use on their own.

Through Subsdy, my platform for EU grants with a calendar of the calls due in the next two months, I have come across many grants that comment on AI. Agencies react in four ways: prohibiting AI, requiring prompts or labelling, shortening proposals, and switching to interviews instead of written proposals. Some examples:

Banning AI in a grant for software development shows a lack of understanding of how the technology works: telling a developer in 2026 not to use AI to write code is ridiculous. The same is true for grant writing. Asking applicants to attach every prompt they used is just as unrealistic. The EIC is actually good here: it neither restricts AI nor asks for anything of the kind.

The EIC Accelerator did add a second interview in Step 2. I believe it was not introduced specifically against AI proposals, but against a backlog caused by the sheer number of proposals, which AI has made worse. The effect is a more human evaluation: if AI wrote everything, you still have to explain the proposal to someone who read it and asks you questions. More interviews also catch applicants who reach the Step 3 interview without knowing their own project. While most companies scale with AI, the EIC scales with people, paying for an hour of evaluator time for each of the thousands of Step 2 submissions every year.

AI Has Made Proposals Neither Worse Nor Better

There have been two phases. In the first, there was a quality gap: proposals written with early ChatGPT were little more than lists of bullet points, what I call AI slop grenades. In the second, high-quality agents with context and a knowledge base produce much better work.

Even so, if you ask Claude, which I use most and consider very good, to write an EIC Accelerator proposal from scratch, with all your company data and the official templates, the result will be a very bad proposal. That is not because the model is bad but because it does not know what the proposal has to look like. The official template does not help: it is very general and does not include all the evaluation criteria, and some required content has no section to go in. I use detailed instructions that take the official template and extend it with many more segments. Even the smartest models may never write a great EIC proposal on their own, unless a lot of public data shows exactly what winning proposals look like and the model can transfer it from, say, a MedTech proposal to an engineering or robotics company.

You could always write a good or a bad proposal. AI just does it faster: if you want a bad proposal, you can now have it very quickly, and the same goes for a good one. Most AI proposals are still bad, and good proposals are still rare.

The Most Effective Tools: Work Backwards from the Deliverables

There are many good tools, and whatever you use now is probably fine. I prefer to work backwards from the deliverables. For the EIC you need PDF files that look good, spreadsheets in XLSX and in the old macro format XLSM that the EIC still uses, and MP4 videos, since Step 1 and Step 2 each include a video. Together with the two interviews, the videos are a strong human component and a natural antidote to applicants who want AI to do everything without understanding the subject.

My hot take is that the best tool is an agent in the terminal. I use Claude Code for about 95% of my AI work, together with Markdown and text files. From the terminal one agent handles every format: it builds scripts, accesses Google Docs, Sheets and Drive through MCP, edits XLSM and XLSX files locally, turns Markdown into PDFs, merges and compresses PDFs, edits videos, transcribes and animates visuals, and designs pitch decks and graphics.

A small case study: the EIC budget spreadsheet is an XLSM file where you click buttons to copy work packages around, and it feels at least two years out of date. After resetting my computer I no longer had Office installed, so I asked the agent to convert the XLSM file to XLSX and run all the scripts that fill in the work packages. I was sure it would fail. It took about two minutes and the result was perfect.

You no longer need Office or Adobe; free tools now do a better job. For video I use DaVinci Resolve, which is free, with local Whisper models for transcript editing and voice enhancement, and you can do the same for the EIC Accelerator video. The agent writes scripts on the fly for whatever you need, from formatting financial charts to exporting PDFs. Working in a team is no obstacle: you can draft everything locally and let the agent put the proposal into Google Docs, even as suggestions, and email the team to review it. In my experience Gemini cannot do the formatting in Google Docs, but Claude can. The result is a high-quality proposal with minimal friction, for almost free. Claude Cowork, Codex and Cursor are good environments too.

How to Use AI for the EIC Accelerator: Seven Steps

My method has seven steps:

1. Research and data analysis
2. Understand the industry, technology and market
3. Develop the narrative
4. Draft and structure
5. Read and monitor everything
6. Fact check
7. Collaborate with your team

The most important rule is not to let the AI think for you. I noticed that as I used AI more, I stopped reading the responses properly and just wanted it to finish the task, so I was looking away from what I should have been looking at: the proposal itself. You have to understand the topic better than the AI does. The AI should either do the task you gave it or explain things to you like a good teacher. If it thinks for you, you lose control and stop building expertise, and you replace yourself with a tool that costs about twenty dollars a month, the last thing you want to do as knowledge work changes in 2027.

So the first step is deep research with AI as a teacher. I spend days talking to it and letting it teach me, because I have to understand both the high level and the detail. I create the structure and the narrative myself; the AI can add things and comment. Then I draft with detailed instructions and let the AI fill in the content. Early on there is more freedom, but the closer the submission, the more control I take: in the last two weeks no change goes in without me seeing it. The writer must understand the project. AI is a great servant but a poor master, or perhaps better, a great teacher but a poor master.

The Agent Is Still Bad at the Narrative

The narrative starts with a problem and offers the solution, but it is also the selling point of the proposal: why anyone should care, and why the project is strategically important for Europe. Everything has to feed into it: the problem, the technology, the customers, the competitors, the market, the products and the revenue streams. I give the problem description a lot of attention, research and references, because it sets up the narrative. Policy and problem are where you must create the narrative yourself.

Not every business allows a perfect narrative. But I recently had a difficult Step 1 case where, after digging for an angle that was not obvious, the evaluation came back with no negative comments at all. Even with the maximum score of four out of four, evaluators usually add a note that they want to see more on IP, financials or commercialisation in Step 2. This time there was none, and one evaluator thanked us for a great proposal. That is a narrative hitting hard.

You Build the Outline, the Agent Fills It

I segment the proposal and keep general instructions alongside very specific ones, then place all the segments into a template. I work with a voice mode built on a local model, which lets the agent take as long as it needs rather than answering immediately. I walk around, ask for the next module, say module 3.2 on the commercial strategy, tell the agent what to include because I know the project inside out, and have it read the instructions and give me a first draft. It reads the draft back to me, I ask for a couple of changes, and it goes into the proposal. The whole draft is built on my understanding and my input, with the AI doing the filling.

Create strict guidelines and grant writing skills for your agents, then an outline to fill section by section. Stay in charge: use standardised research instructions to screen large datasets, read and monitor everything, and keep version control against last-minute hallucinations and agent mishaps. Agents hallucinate in fact checks too, so you fix the actual issues, and they can suggest edits in Google Docs through MCP for your team.

Tools Are Useless If You Don't Know What the Outcome Should Look Like

This brings me back to why even the best model cannot write a great proposal on its own: it does not know what a great proposal looks like. It knows the general structure but not what belongs inside it, and it does not know which data to ask for. Take an engineering company with interesting hardware, a patent application and one million euros raised that thinks it is at TRL 5 or 6. The agent will not ask for its unique selling points in relative or absolute terms, such as a certain percentage better or faster. It takes what you give it and follows the official template. What you need is domain knowledge: a good template, good section-by-section instructions and review.

The Gap Agents Can't Fill

Agents lack the grant know-how for a very specific programme. I have not used any general grant writing app, and I do not think they are a good business: grants differ so much that a general grant tool is little more than a general agent, and today's language models already are the perfect general tool. Some consultancies have turned to these apps out of desperation, and I think that is the blind leading the blind. If you work at a consultancy, use AI aggressively instead: Claude, Codex or whatever you prefer.

Agents also cannot handle the interviews and negotiations, the grant agreements and equity, the project and its reporting with the project officers, or a narrative without red flags.

Is Using AI Safe?

AI is a tool for efficiency, not an autopilot. Subscribing to ChatGPT and Claude and using as much of your allowance as possible is almost an insurance policy: it keeps you on the ball, shows you how others and your customers use the technology, and prepares you for whatever comes next.

The usual objection is: "But what if I can't trust them?" Cloud storage caused the same debate. People asked why they would hand sensitive files to Google, and now everyone uses the cloud. The same will be true for AI, if it isn't already. The real fear is training data, so check the policy of the tool you use: most paid tools exclude your data from training, unlike some free AI tools in the browser.

My main argument is that AI already touched it. If you upload a file to SharePoint or Google Drive, or open a PDF in Adobe Acrobat and see its summary, a language model has already read the file, probably on a server in the US. Cloud storage and email already hold your most sensitive files, and email is one of the least secure services there is.

If you are serious about confidentiality, go local: use encrypted email, encrypt files before uploading them to the cloud, and run open-source models from Hugging Face on your own GPU. It is feasible, but overkill for most people. I run local models for voice, transcripts and video editing, but for writing I use Claude, because it is far more efficient and my GPU would not be fast enough for a large model anyway.

Learn More

I have created a number of guides and tools on the EIC Accelerator, including ChatEIC, a simple AI writer that gives a good first draft. When someone sends me a proposal to review that is clearly written by AI and really bad, a quick draft with ChatEIC is about ten times better and gives me a solid base to edit from. For grants beyond the EIC, Subsdy shows the calls due in the next two months, and you can find my consulting at seglerconsulting.com.

Q&A: Voice Mode, Version Control and Local Data

I use AI extremely heavily, and I think doing so is an insurance for 2027 for everyone, whether you are a consultant, run a business or work at a university. I built my own terminal app to have everything the way I like it. My voice mode is a script: I ask the agent in the terminal to open a voice channel, speak, and a local model, free from Hugging Face, turns my voice into text. When I stop, the agent reads what is new and answers through a text-to-speech output. A screen display next to it shows a slide that updates with every answer. That way I can walk around while the agent researches, for example a company like Nvidia for a stock analysis video. I do not know how to code; Claude does, and it built exactly what I wanted.

I also use version control for everything, through Git in VS Code. A Git repository tracks every file in a folder, so every proposal I create shows up in my list of changes. After I commit a first draft and ask Claude to fact check and fill in the placeholders, I open the difference and see exactly what it changed. Claude can never change anything behind my back, and the same protects a proposal from five team members editing a number in one place and forgetting it in another. Dropbox keeps files safe too, but it does not show line-by-line changes, so I use both.

I do not use personal agents such as OpenClaw. Claude already has access to everything on my computer, can run sub-agents and even drive other tools, so I see no difference worth paying for, and I expect many of these products to be short-lived. For research I use two approaches. For a stock analysis I have standardised research instructions that pull SEC filings and run long web searches, then I discuss the data with the agent by voice until every document is covered and the follow-up questions snowball. For a grant proposal the research starts with the company's own data, which I discuss with the agent document by document, with web search only for policy, market and competitors. I keep data in local folders rather than databases, because the agent searches local files with its own scripts far more easily than through API access, and it searches all the client files, filings and transcripts on every answer.