The Fund Is Not Buying Your Technology

Read the first paragraph of the funding logic and the shape of the whole exercise appears. KMU-innovativ: Kunstliche Intelligenz is funded by the BMBF and administered by DLR PT as an economic policy instrument, not as a research grant for artificial intelligence. The numbers it exists to move are published. German small and medium-sized enterprises are 99.4 per cent of all businesses, 55 per cent of gross value added and 60 per cent of employment. AI adoption in core processes sits well below that footprint, and the documented figure used in the programme framing is that 74 per cent of German SMEs have not adopted AI in core processes at all. That is the target. The research is the method. So the money is not buying a clever model. It is buying R and D that private capital will not fund because the risk is too high, and that is nonetheless close enough to market to show measurable economic benefit within three to five years of the project ending. Too far from market and the programme has no instrument to use. Too close and the applicant did not need it. The window is the point. What the instrument explicitly does not fund is now predictable: pure science with no commercial path, consumer AI with no industrial or public-sector use case, and projects whose main output is a dataset or a benchmark rather than something deployable in a factory or a clinic or a registry.

The scoring sheet is public enough to write against, and it is weighted. Scientific and technical quality carries roughly 40 per cent. Commercial exploitability, the Verwertungsplan, carries roughly 30 per cent. Qualification of the applicants carries roughly 20 per cent. Foerdernotwendigkeit, the additionality argument, carries roughly 10 per cent. Every paragraph in a submission serves one of those four boxes or it is filler. That gives a read on where a strong technical team usually loses: a project with brilliant architecture and no pilot partner caps out around 40 per cent of the sheet, because the 30 per cent box was left empty on purpose. The container is fixed too. Twelve pages. Arial 11 point. Minimum 1.5 cm margins on all sides. 1.15 line spacing. Anything past twelve pages is truncated, and there are no appendices, because anything not inside those twelve pages does not exist for the reader. Tables count as pages. Figures count as pages. And the title has to match the easy-Online portal entry character for character, and the applicant name has to be the Handelsregister name, not the brand name. Both of those are small, both are checkable in a minute, and both have ended applications. The competition is real: 50 to 300 submissions in a cycle, 30 to 60 minutes of reviewer attention each. The probability that a technically strong project loses on the commercial box is far higher than the probability it loses on the technical box, and the weights say so before anybody reads a paragraph.

A submission passes three people before a funding decision exists, and each one is doing a different job. The first is the DLR PT programme officer running the completeness check. Five to ten minutes, no content judgement at all. They are checking whether the mandatory fields are populated, whether the format is compliant, whether the entity data matches easy-Online, and whether the file landed before the portal closed. One missing mandatory field returns the submission as inadmissible before a reviewer ever sees a word of the argument, and the probability of that happening rises sharply in the last 24 hours, which is why the target filing window is 24 to 48 hours early. The second is the assessor, and there are typically two independent Gutachter per submission. These are senior academics in machine learning or senior industry R and D managers. They carry a structured sheet with the four weighted criteria, they read start to finish, they assign sub-scores, and they have almost certainly read 50 to 200 other submissions in the same cycle. They are not hostile. They are immune to hype, which is a different thing and a harder one to write against. The third is the project manager, who appears only after a positive recommendation, when the full application is invited. They read the long version, check the budget, and stay for the multi-year reporting. A Skizze written at application-level detail — real costs, named subcontractors, timelines that survive contact — is what shortens that stage. The practical consequence is the shortest rule in the whole file. The reviewer reads the first 200 words and forms a judgement, and while they will revise it, they rarely reverse it. The first 200 words should carry the novel technical claim, the risk, and the reason this team can hold it. Background is not the opening.

The playbook contains its own example of a sentence that passes and one that fails, and the difference is entirely countable. "The global AI market is 500 billion euros" is worthless. The version that scores reads: AI-supported process automation in mid-size German manufacturing, Destatis WZ08-28, around 12,400 firms, average IT budget 180,000 euros a year, AI adoption under 5 per cent as of 2024, giving a serviceable addressable market near 200 million euros growing at 18 per cent a year against the BITKOM 2024 report. Same subject. One of them can be checked. The business model follows the same standard. "We will license the software" is not a business model. An annual licence at 24,000 euros per seat for manufacturing firms of 50 to 250 employees, targeting 12 customers by the end of year three for 288,000 euros of recurring revenue and break-even at 8 customers, is a business model. It can be argued with. Unrealistic beats vague, because unrealistic at least proves somebody did the multiplication. Intellectual property gets a topic and a month: a utility patent for a named training method, filed at the DPMA in month 20. Or an explicit statement of why there is no filing and what carries the protection instead — trade secret, copyright over the training data, an exclusive commercial licence from the research partner. Silence reads as IP weakness. The audit that holds it together is two columns. Claim on the left, source on the right. Every quantitative claim appears in the list with a primary reference. Any row with an empty source column comes out of the document or gets rewritten so it makes no numeric claim. Twelve sourced figures beat twenty unsourced ones, every time.

The programme is an economic instrument with a research budget attached.

None of the compliance floor rules make a submission better. Every one of them can make it void. Personnel stays at or below 20 per cent of the total budget, with external services carrying 80 per cent or more. That is a deliberate signal as much as a rule: a budget weighted to wages reads as an attempt to subsidise staff, and a budget weighted to external research reads as an R and D project. No research partner gets named without written confirmation — an LoI or an email from an authorised representative — and no new partnership is realistically buildable inside a three-week deadline. Every subcontract needs three independent arms-length offers, documented before award. The entity data has its own traps. The VR or HRB number of every entity is a mandatory portal field, so an entity whose number is unknown cannot be included. The KMU self-declaration uses the last completed fiscal year only, never projections. Connected enterprises at 25 per cent or more cross-ownership have to be consolidated under EU Recommendation 2003/361/EC. Duplicate funding across every group entity has to be checked and documented before submission. The registered managing director signs, or a notarised power of attorney does. The clock is the part people misjudge. Target submission 24 to 48 hours before the deadline, because Easy-Online degrades on the final day. Qualified electronic signature coordination needs 48 hours with the managing director. Every project invoice carries the Foerderkennzeichen from the first day of the funding period. The whole application file is retained for ten years after the project ends. And two which are about character rather than administration: never present an estimate as a confirmed figure, never present a draft as ready to file, and when in doubt, ask. The cost of a question is zero. The cost of a false statement is full repayment.

Most submissions end on confidence. The winning version ends on a named failure mode and what happens next. The template is short. The primary technical risk is the convergence failure of a reward function in sparse-feedback environments. Mitigation: a curriculum learning schedule inside work package two, with an explicit fallback to supervised pre-training if the reinforcement route has not converged within six months. Go or no-go decision point: month eight. That reads as intellectual honesty and it also proves the milestones are real. A milestone that cannot fail is not a milestone, it is a date. The additionality argument is the counterfactual, written plainly. Without funding, the work continues at the current pace and reaches a stated point in a stated number of years. With funding, it reaches technical readiness level six and commercial deployment in three. It cannot be that nothing would happen without the grant, because that implies nobody wants the product. One more thing about framing, and it is the most common unforced error. A grant is never described as income. It is described as risk coverage: the portion of the work that produces no directly attributable commercial output during the project period. Same money, opposite reading. Sixteen years of field work, from rural Nigeria to post-conflict Serbia to Kazan to Fukushima to Silicon Valley, and a industrial group restructured along the way, taught the same lesson that this playbook states in its own words. The reviewer is the only person who matters. Not the hotline, not the guidelines PDF, not the programme manager. One person, one sheet, four weighted boxes, twelve pages, and an economic gap they have been asked to close. Write for the person holding the sheet, and the programme's purpose stops being background and starts being the argument.