Approved, scored 61 of 100, never drawn: what an EU grant evaluator actually buys

On 6 October 2025 a project whose technical proposal I had written cleared its evaluation. The evaluator gave it 61 points out of 100 and recommended it for support under OP TAK, the Czech operational programme for business competitiveness co-funded by the EU. Total costs 48 million koruna, of which 44.3 million eligible; 15.5 million of grant, 28.8 million of own funds. Eight people in R&D, 25 months of implementation, the wider Central and Eastern European region as the target market. The subject was a large enterprise application with a long history.

Not one crown of that grant was ever drawn. The project never started.

It was not a technical failure and the business case was never disproved. It was a strategic decision at owner level, where a three-to-five-year payback horizon did not hold up next to short-term results. I left the company where I ran development some months later. Today my company is one person, and whenever somebody asks me whether to fund AI development with a grant, I go back to that scoring sheet.

What the evaluator actually scored

The score came broken down:

The full marks did not go to the technology. They went to technical uncertainty and to the plan for testing. AI and machine learning itself landed on four out of five, among the weakest parts of the proposal, alongside big data and cybersecurity.

I had written the uncertainty chapter as an admission: which parts we did not know would work, why, and what happens if they do not. The testing plan followed — for each uncertainty, which experiment would settle it and in what order. I expected that to be the weakest section, because it holds the most “we don’t know”. It was one of only two with a perfect score.

I compress that into a sentence I now use well outside the grant world: an evaluator does not buy your model, they buy your uncertainty and your method for shrinking it. The logic is cold: research money goes to things whose outcome is not known in advance, and anyone claiming to know it is either not doing research or not telling the truth. That holds regardless of company size, and regardless of grants: when I scope my own projects now, I start from the list of what I do not know.

The number nobody looks at

Grant: 15.5 million koruna. Own funds: 28.8 million. Everyone reads the first number, because that is what the company receives. The second one decides.

A grant is not money you receive. It is money that unlocks roughly twice as much of your own — and commits it for 25 months, inside a budget you can no longer rearrange around what you learn. A grant therefore stops being a funding question and becomes a horizon question. Not “can we afford it?” but “will we still be doing this in two years, if year one produces nothing?”.

That is exactly where the project ended. The money was there; the horizon was not. What remains is worth saying out loud: the proposal was good enough for the state and still did not survive its own company. I have seen far more projects die that way than on engineering. Engineering can be fixed. A horizon cannot, because it does not live in the documentation — it lives in the head of whoever decides.

What the odds actually are

That was one project. Here is the measurable part. The Czech Ministry of Industry and Trade runs TWIST, which supports research and development in artificial intelligence: up to 30 million koruna per project, up to 70 % of eligible costs, open to enterprises of any size anywhere in the country, with 5 billion koruna across 2025 to 2031.

The ministry announced the second call’s results on 15 July 2026. 318 proposals were submitted, asking for 4.27 billion koruna. 630 million went to the 40 best-rated; more than thirty were held in reserve depending on the 2027 state budget.

Forty out of 318 is 12.6 % of projects. 630 out of 4,270 is 14.8 % of the money requested. Roughly one application in eight got funded.

Now the number that appears in no announcement. A proposal like that takes several weeks of the most expensive person in the company — somebody who understands the technology and the budget and can make the two agree. In a large enterprise that cost disappears into overhead. In a firm of one to three people it is weeks of not invoicing and not shipping. The expected value nobody mentions in a webinar is one in eight, paid with the scarcest capacity you have.

The programmes small firms get pointed at, checked on 5 October 2026

A small Czech company asking how to pay for AI training gets three answers.

POVEZ II, the employee-training programme, is recommended most often. It ran from 1 December 2015 to 31 December 2023. It is over. While it ran it was genuinely employer-facing: a contribution to training costs plus reimbursed wages for time spent training, at a rate set by state-aid rules rather than a flat number, with employers seated in Prague excluded. A sensible instrument that has not existed for nearly three years.

DIGI pro firmu, its digital-skills successor funded from the national recovery plan, ran from April 2024 to 28 February 2026. Also over. Its official page, last updated 25 June 2026, says that “during the year a new project will be launched under the Operational Programme Employment Plus (OPZ+)”. As of 5 October 2026 I could not find it announced. I am not claiming it will never come; I am recording when I looked.

Jsem v kurzu, the labour ministry and employment office portal (page last updated 25 September 2026), does exist and is open. But it is not an employer instrument: to use it you must be registered with the employment office as a jobseeker or an interested person. A chosen retraining course is paid up to 50,000 koruna, the ceiling across three years; digital-skills courses up to 35,000 koruna with the participant co-paying 23.27 %. In 2026 approval was tightened: priority to professions with real labour-market demand, support “as a rule, one-off”.

The honest conclusion is unflattering. Much of the advice circulating about “subsidies for AI training” points at programmes that are closed, or at programmes never addressed to the employer side.

What is left and realistic

Innovation vouchers, call IV under OP TAK. 75 % support for a small enterprise, a grant of 100,000 to 2 million koruna, for SMEs in the EU block-exemption sense. Applications run from 31 March 2025 to 30 April 2027, and on 24 September 2026 the allocation was raised from 250 to 350 million koruna because of high demand. Eligible costs explicitly include software and hardware development, modelling, simulation, testing and certification.

The catch is not in the small print but in the design: the voucher buys services from a research organization or an accredited subject — a testing laboratory or certification body with a valid accreditation certificate. It does not pay your own team. For a firm of one that is decisive, because your own time is the entire cost structure. The voucher pays for somebody else to do the work, not for you to do it.

TREND, subprogramme 2 (“Newcomers”) from the Technology Agency is the one call explicitly designed for companies with no prior R&D track record — the closest fit a micro firm has. Its announcement of 1 July 2026 states that no public call in subprogramme 2 will open this year. Subprogramme 1, for technology leaders, which does require an existing R&D record, has its thirteenth call postponed on 15 September 2026 with reference to state-budget negotiations.

None of this is a conspiracy against small companies. These instruments were designed for firms with a payroll and the capacity to co-fund. A micro firm is not their addressee — it simply fails the assumption the construction rests on.

The advantage a firm of one has over a 44-million project

Here is the conclusion I did not expect, and I am not going to soften it. For a firm of one to three people the arithmetic almost always favours a small self-funded project with a one- to two-year payback over a grant-shaped one. Not because grants are bad, but because their real price is not the co-funding share: it is the proposal, the administration, and spending your own co-funding share on somebody else’s calendar — a quarter to a third of the budget in today’s open calls, two thirds in the project I just described. A micro firm pays all three in the currency it earns in — the time of one person.

And now the turn. The 44-million project was not killed by engineering, competition or missing money. It was killed by nobody being able to hold a three-to-five-year horizon. That, though, is a corporate disease: one person holds the horizon, another decides, and a quarterly result stands between them.

Where the owner is the person who decides, that disease does not exist. So the real advantage of a small firm over that 44-million project is not agility and not low cost. It is that the horizon and the decision sit in the same head. If I say I will work on something for two years, nobody revisits it after six months. The uncomfortable half belongs here too: when a project fails there, the reason is in that same head. No owners to point at, no bad quarter.

When a grant does make sense

None of which rejects the category. A grant does make sense for a small firm under three conditions met at once: when something already works and the job is to scale it rather than find out whether it works; when there is real capacity to co-fund without endangering operations; and when the owner can hold the horizon the application commits them to, the one line nobody fills in for you. A grant is a good second step; as a first one it is the most expensive way to find out you did not believe in the project enough.

These parameters are verified as of 5 October 2026 and this world changes by the month, so re-check at the source if you read this later. One thing is checkable immediately. The DIGI pro firmu page has said since June that the OPZ+ successor will launch “during the year”. When I looked on 5 October it had not been announced. Find it there before the year ends and I was too sceptical. If not, the instrument most often recommended here for digital-skills training has been missing for over half a year — long enough to have built and paid for a whole small project out of your own money.