Over 2,000 programs, and nobody in the company has time to read them.
The German federal funding database alone lists over 2,000 programs. On top of that come sixteen state-level worlds, EU calls, and the programs of KfW and BAFA. A mid-sized manufacturer or trading company qualifies for several of them at any given moment. Almost nobody knows which. The reason is mundane: nobody owns the topic. Research only happens when someone hears about a program by chance, usually from the house bank or the chamber of commerce, and usually too late for the current call.
Once a program is found, the real work begins. A project description in the language the funding body expects. Forms with fields that assume prior knowledge. Evidence from four departments. A cost and financing plan. That costs several person-days that nobody has free alongside daily business.
And it does not stop after approval: drawdowns, interim reports, and the proof of use have their own deadlines that nobody remembers a year later. This is exactly why many companies hand the topic to external funding consultants, who usually take 10 to 15 percent success commission on the approved grant.
We start with the research, the drafting work comes later.
The most common objection in our conversations is honest and justified: “We gave up on funding applications, the effort eats the grant.” That is exactly the calculation the AI employee flips, and only that one. It takes over the parts that are effort without judgment. What it does not do is decide for you. And what nobody promises is an approval.
- Company profile before the program search: Industry, headcount, location, legal form, planned projects, and the history of previous funding get recorded properly once. Without that profile, program matching is guesswork.
- Monitoring and pre-checks as the first scope: The AI employee first only watches the funding databases and reports hits with a check of the hard criteria. That delivers value in week one and needs no access to sensitive documents.
- Drafts from your own material: Only then do we connect previous applications, requirement specs, and quotes. The draft is written in your wording, not in generic funding prose. Every passage that is derived rather than found gets flagged.
- Deadlines and proof of use last: Submission windows, drawdowns, interim reports, and proof-of-use dates move into a monitored calendar. This part pays off after approval, when the initial euphoria is long gone.
80 percent is done by the AI employee. The human decides at two points.
The AI employee searches the public funding databases and matches every hit against your stored company profile. Then every hit runs through the hard eligibility criteria: size class, industry, type of project, minimum investment, own contribution, and the limits for cumulation and de minimis. These are fixed rules, not matters of discretion. Every single one is reported as met, not met, or unclear. Criteria that need interpretation are never resolved silently.
For the programs you pursue, it drafts the project description and the form answers from your own documents, assembles the list of required evidence, and keeps the deadline calendar current. Two stations belong to your team: the decision whether to apply at all, and the expert review including submission. Nothing is ever submitted automatically.
A manufacturing company wants to connect its machines to a production data system and train the shop-floor team. Investment in the low six figures, start planned for next quarter.
One field is flagged for review: whether the production data system meets the program’s innovation criterion is a matter of interpretation. The AI employee names it instead of assuming an answer, and a human clarifies it with the funding body.
That was before, this is today.
This is how funding work ran before, and this is how it runs today with the AI employee. Conservatively calculated, the effort on your side down to a finished application draft drops by about 70 percent. Approval still lies with the funding body.
| Before | Today | |
|---|---|---|
| Program research | Happens when someone hears about it by chance | Continuous matching against the public databases |
| Eligibility | Surfaces after weeks of work, or never | Hard criteria checked before work goes in |
| Application draft | Blank page, several person-days | Draft from your documents, human sharpens it |
| Collecting evidence | Chased through the departments by email | Document list with owners and reminders |
| Deadlines and proof of use | Excel list, often noticed too late | Monitored calendar across the whole project |
* A typical scenario, not a measurement from a single customer project. Savings conservatively calculated: research, pre-checks, project description, and collecting evidence cost 3 to 5 person-days per application before. Today the decision, the expert review of the draft, and the release remain, together about 1 to 1.5 person-days. That is about 70 percent less effort on your side down to a finished draft. We deliberately say nothing about approval rates, those lie with the funding body.
How do you build an AI employee like this, technically?
The knowledge from this use case to take away, whether you build with us or on your own:
Matching: programs against the company profile
The AI employee reads the program pages of the German federal funding database, the state portals, the EU calls, and those of KfW and BAFA, and matches them against a maintained company profile: industry, headcount, revenue class, state, legal form, planned project, and previous funding. Without that profile you get plausible-sounding but useless suggestions. That is a data problem, not a model problem.
Drafts: language models fed with your documents
For the project description and the form answers, GPT from OpenAI or Claude from Anthropic work well, but only with your own material as the basis: previous applications, requirement specs, annual accounts, quotes. A draft without those sources sounds good and carries nothing. Every passage that is derived rather than found gets flagged so the review stays targeted.
Hard eligibility is a rule set, not an AI question
Size class, industry, type of project, minimum investment, own contribution, and the cumulation and de-minimis limits are decidable from data. They belong in a fixed rule set, not in a language model. Soft criteria, for instance whether a project counts as innovation, always go to a human as an open point. A plausible invention devalues the application later.
Deadlines and proof of use need their own logic
Approval brings obligations: drawdown windows, interim reports, documentation of hours and receipts, and the proof of use within a fixed deadline. These dates live in a monitored calendar, and receipts get collected continuously during the project. That way the proof of use grows along instead of being reconstructed from unlabeled folders a year later.
UX: the interface decides adoption
Every program found gets an effort-versus-benefit template: expected grant, criteria met and open, realistic remaining effort. The drafts sit there as editable documents ready for expert review, nothing goes out unseen. And a deadline board shows every upcoming date. Exactly these three views turn a list of programs into a decision.
What does it cost in comparison?
Superkind charges per use case, independent of whether a single application gets approved. Here is the honest comparison:
| Doing the research yourself | Funding consultant | Superkind AI employee | |
|---|---|---|---|
| Cost | Several person-days per program, never booked internally | 10 to 15% success commission on the approved grant | Price per use case, a fraction of a full-time position |
| What is included | Whatever one person manages on the side | Judgment and argument, usually per mandate | Research, pre-checks, draft, evidence list, and deadlines |
| Scales with | Spare time, which is rare | The size of the grant | The number of programs, without new positions |
| Exceptions | Left unhandled | Cost extra consulting hours | Flagged and sent to a human |
| Rollout | Immediate, but without structure | Contract and mandate | 2 to 3 weeks to the first productive version |
The honest comparison is not the one against zero cost. It is the one against the programs you do not apply for at all today. None of the three paths promises an approval, and neither does this one.
What we learned about automating funding work.
The scarce resource in German funding is not the money, it is attention. Companies do not miss programs because they would not qualify, but because reading over 2,000 program descriptions is nobody’s job. Once that reading runs automatically, the topic changes character: from the occasional heroic act of whoever happens to have time, to a routine decision with a number behind it.
The second thing: only honesty about outcomes makes this use case credible. Everyone in this market knows someone who promised funding and delivered invoices. So we draw the line clearly. The AI employee prepares, your team decides, the funding body approves. Everything on our side of that line can be measured and improved. Nothing on the other side should be promised.
What it is not suited for: If you file a single application every three years, no setup pays off. Highly specialized research applications live on consortia and personal contact with the project agency and belong to an experienced expert consultant. And if nobody has the capacity to actually deliver and document an approved project, more applications make things worse.
