Research allowance application: the BSFZ example, good and poor
The BSFZ uses its own fictional project, an AI-controlled feeding station for squirrels, to show what a good and a poor description look like: field by field, with character limits and with its reasoning. We show both versions in full and, alongside, calculate what turns a description into an assessment basis (Bemessungsgrundlage).
- Five fields, two versions each, in the BSFZ's own words
- Character limits per field: 200, 500, 1,000 and 1,500 characters
- Work plan with eleven work packages and person-months
Written by Max Nodes, Founder and Managing Director · 25 of 25 supported BSFZ applications certified (as of 09/2026) · Last updated: · Law as of:
At a glance
Turning an idea into an application
Figures and details
The project was invented by the BSFZ; the texts below are its original.
Key facts in brief
The BSFZ has published its own example: a fictional project, an AI-controlled feeding station for squirrels, with every field described once well and once poorly, together with the assessors' reasoning. The difference is never in the product, always in the description: the good version names the knowledge gap, the state of the art and its limit, the technical risk of the project and the work steps in order; the poor version names market goals, superlatives and requirements. Both versions are quoted in full below, plus the tabular work plan.
The essentials of the example
How does the BSFZ tell a strong application from a weak one?
For every field, in the BSFZ's own words: first the version that convinces the assessors, below it the one that fails, each with the BSFZ's explanation. The sample project, an AI-controlled feeding station for squirrels, is fictional.
Title of the project
Good description (96 of 200 characters in the German original)
AI-based feeding station for recognising and feeding individual squirrels
BSFZ's assessment: Meaningful title
AI-based feeding station for recognising and feeding individual squirrels
BSFZ's assessment: Meaningful title
Poor description (36 of 200 characters in the German original)
Development of the TX108A feeding station
BSFZ's assessment: Title that says nothing, using an internal product name without any further explanation
Development of the TX108A feeding station
BSFZ's assessment: Title that says nothing, using an internal product name without any further explanation
Objective of the project
Question in the application portal: What is the objective of your project? Which challenge or problem is the project meant to solve, or which knowledge gap is it meant to close?
Good description (1461 of 1500 characters in the German original)
An AI-based squirrel feeding station is to be developed that autonomously recognises individuals within a population and dispenses special feed with additives (e.g. vitamins) to individual animals. For the AI-based station, both the set-up including motion sensors, cameras and control and a new AI solution are newly designed. Although first AI approaches for recognising different animal species in images exist, there is no solution for identifying individual squirrels, which differ only minimally in their external features. To achieve fast recognition despite these small differences, model elements with new functions and adapted weightings and links are to be developed for the single-stage and therefore fast YOLO (You Only Look Once) object-detection algorithm. The feeding station has both a standard wildlife camera (for daytime images) and an IR camera (for images at dusk) with motion sensors. After motion is detected, images are captured, transmitted one after another to a server and evaluated using deep-learning algorithms. As soon as the individual is correctly identified with 90 % certainty, the signal is given to open the container with the special feed. At an accuracy below 90 %, the animals would suffer severe malnutrition.
BSFZ's assessment: The intended product and how it works are described comprehensibly. The available knowledge and the knowledge gap are presented. The essential properties of the project objective are named and technically substantiated. Characteristics of the intended objective are presented. A solution/implementation strategy is presented. Abbreviations used are explained.
An AI-based squirrel feeding station is to be developed that autonomously recognises individuals within a population and dispenses special feed with additives (e.g. vitamins) to individual animals. For the AI-based station, both the set-up including motion sensors, cameras and control and a new AI solution are newly designed. Although first AI approaches for recognising different animal species in images exist, there is no solution for identifying individual squirrels, which differ only minimally in their external features. To achieve fast recognition despite these small differences, model elements with new functions and adapted weightings and links are to be developed for the single-stage and therefore fast YOLO (You Only Look Once) object-detection algorithm. The feeding station has both a standard wildlife camera (for daytime images) and an IR camera (for images at dusk) with motion sensors. After motion is detected, images are captured, transmitted one after another to a server and evaluated using deep-learning algorithms. As soon as the individual is correctly identified with 90 % certainty, the signal is given to open the container with the special feed. At an accuracy below 90 %, the animals would suffer severe malnutrition.
BSFZ's assessment: The intended product and how it works are described comprehensibly. The available knowledge and the knowledge gap are presented. The essential properties of the project objective are named and technically substantiated. Characteristics of the intended objective are presented. A solution/implementation strategy is presented. Abbreviations used are explained.
Poor description (809 of 1500 characters in the German original)
Squirrels enjoy ever-growing popularity in Germany. Squirrels live mainly in forests with old tree stands but, as followers of human settlement, are increasingly found near people, e.g. in gardens and parks, as long as there is enough food and shelter. To ensure correct feeding, this project will develop a feeding station for squirrels that will revolutionise the feeding of wild animals. With the help of artificial intelligence (AI), the animals are to be identified with an accuracy of 99.99 per cent and within 3 seconds and fed specifically. With the new product we hope to increase our market share in feeding stations in Germany by 50 per cent.
BSFZ's assessment: Only requirements for the intended product are listed; no challenges or problems to be solved by the project and no knowledge gap to be closed are set out. Requirements for the product and targeted parameters are not comprehensible or justified. Presenting the commercial exploitation prospects is unnecessary, since marketing follows the R&D project. Only the buzzword AI is mentioned, without describing the AI to be developed.
Squirrels enjoy ever-growing popularity in Germany. Squirrels live mainly in forests with old tree stands but, as followers of human settlement, are increasingly found near people, e.g. in gardens and parks, as long as there is enough food and shelter. To ensure correct feeding, this project will develop a feeding station for squirrels that will revolutionise the feeding of wild animals. With the help of artificial intelligence (AI), the animals are to be identified with an accuracy of 99.99 per cent and within 3 seconds and fed specifically. With the new product we hope to increase our market share in feeding stations in Germany by 50 per cent.
BSFZ's assessment: Only requirements for the intended product are listed; no challenges or problems to be solved by the project and no knowledge gap to be closed are set out. Requirements for the product and targeted parameters are not comprehensible or justified. Presenting the commercial exploitation prospects is unnecessary, since marketing follows the R&D project. Only the buzzword AI is mentioned, without describing the AI to be developed.
Distinction from the state of the art
Question in the application portal: To what extent does the intended product, process or service go beyond the state of the art?
Good description (474 of 500 characters in the German original)
Existing feeding stations for squirrels work mechanically, i.e. the animals open the feed flap themselves when needed. For other animal species there are first approaches, but these cannot distinguish individual animals from one another, only different species. The combination of motion sensors, the type of feeding station and the AI-based control also differs from previous solutions.
BSFZ's assessment: A comprehensible distinction from the state of the art is given.
Existing feeding stations for squirrels work mechanically, i.e. the animals open the feed flap themselves when needed. For other animal species there are first approaches, but these cannot distinguish individual animals from one another, only different species. The combination of motion sensors, the type of feeding station and the AI-based control also differs from previous solutions.
BSFZ's assessment: A comprehensible distinction from the state of the art is given.
Poor description (351 of 500 characters in the German original)
The TX108A squirrel feeding station is highly innovative and sets new standards for everyone, because it not only greatly improves our product TX107 but also puts all other providers in the shade. Feeding squirrels is to be maximally automated and optimised for a feeding experience never seen before.
BSFZ's assessment: A distinction from the state of the art is not described concretely. There is no substantive distinction from the company's own products or from routine development work, because neither the predecessor product nor the intended new product is described concretely.
The TX108A squirrel feeding station is highly innovative and sets new standards for everyone, because it not only greatly improves our product TX107 but also puts all other providers in the shade. Feeding squirrels is to be maximally automated and optimised for a feeding experience never seen before.
BSFZ's assessment: A distinction from the state of the art is not described concretely. There is no substantive distinction from the company's own products or from routine development work, because neither the predecessor product nor the intended new product is described concretely.
Description of all work
Question in the application portal: Which work is carried out to address the challenge or problem named above, or to close the knowledge gap?
Good description (998 of 1000 characters in the German original)
After the feeding station has been designed and built, first test images of squirrels are taken with the wildlife camera and the IR camera. On this basis the sensors and relevant parameters (exposure time, resolution) are optimised. Then around 1,000 images of about 50 squirrels are taken and annotated. In parallel, YOLO model elements for recognising the squirrels are developed so that the model can be trained and validated once the test data has been captured. The animals are to be identified by features of the head (ears, nose, mouth, fur colour, whiskers). After validation is complete, the algorithm is tested with live images from both cameras and adjusted with regard to the weighting of individual head features. Finally, the whole station is tested under different weather and lighting conditions, with the station (sensor alignment) and the algorithm (feature weighting) adapted to the different conditions.
BSFZ's assessment: R&D activities are broken down in detail and systematically. The work steps build on one another. The methodology and the technical background are described concretely. Test parameters are itemised. Completion of the project is described concretely.
After the feeding station has been designed and built, first test images of squirrels are taken with the wildlife camera and the IR camera. On this basis the sensors and relevant parameters (exposure time, resolution) are optimised. Then around 1,000 images of about 50 squirrels are taken and annotated. In parallel, YOLO model elements for recognising the squirrels are developed so that the model can be trained and validated once the test data has been captured. The animals are to be identified by features of the head (ears, nose, mouth, fur colour, whiskers). After validation is complete, the algorithm is tested with live images from both cameras and adjusted with regard to the weighting of individual head features. Finally, the whole station is tested under different weather and lighting conditions, with the station (sensor alignment) and the algorithm (feature weighting) adapted to the different conditions.
BSFZ's assessment: R&D activities are broken down in detail and systematically. The work steps build on one another. The methodology and the technical background are described concretely. Test parameters are itemised. Completion of the project is described concretely.
Poor description (648 of 1000 characters in the German original)
In the course of the project, the requirements for the feeding station are first researched. On this basis the feeding station is then developed, tested and revised, using agile methods. The station is equipped with a multitude of brand-new, ultra-modern sensors that capture data about squirrels and transmit it to the forward-looking AI software. The AI software then adapts the feeding to the squirrels. In this way the operators of the station can focus their time and resources on other important tasks. This technology takes feeding to a whole new level.
BSFZ's assessment: The intended product, its characteristics and the solution approaches are not described in adequate technical and scientific depth. Numerous marketing phrases and adjectives that are superfluous for the assessment are used.
In the course of the project, the requirements for the feeding station are first researched. On this basis the feeding station is then developed, tested and revised, using agile methods. The station is equipped with a multitude of brand-new, ultra-modern sensors that capture data about squirrels and transmit it to the forward-looking AI software. The AI software then adapts the feeding to the squirrels. In this way the operators of the station can focus their time and resources on other important tasks. This technology takes feeding to a whole new level.
BSFZ's assessment: The intended product, its characteristics and the solution approaches are not described in adequate technical and scientific depth. Numerous marketing phrases and adjectives that are superfluous for the assessment are used.
Scientific, technical and methodological risks
Question in the application portal: Which scientific, technical and/or methodological uncertainties exist in addressing the challenge or problem, or what could cause the approach to closing the knowledge gap to fail?
Good description (819 of 1000 characters in the German original)
The biggest risk is that the AI, based on only a few images, cannot identify the squirrels with the required speed and accuracy. The reasons are that the individual animals differ only minimally in their external features, that decisive features (e.g. fur and ear tufts) change with the season and over the course of life, and that a squirrel does not always position itself frontally in front of the camera. It is therefore not foreseeable whether the image quality will be sufficient to distinguish the individual animals unambiguously. Furthermore, the night-time infrared images lack important colour information for distinguishing them. The time factor is also relevant, since wild animals stay in the same place only very briefly, so the system has to work in under 5 s.
BSFZ's assessment: Challenges are described concretely and justified. Possible problems during implementation are presented concretely and are comprehensible on the basis of the project content.
The biggest risk is that the AI, based on only a few images, cannot identify the squirrels with the required speed and accuracy. The reasons are that the individual animals differ only minimally in their external features, that decisive features (e.g. fur and ear tufts) change with the season and over the course of life, and that a squirrel does not always position itself frontally in front of the camera. It is therefore not foreseeable whether the image quality will be sufficient to distinguish the individual animals unambiguously. Furthermore, the night-time infrared images lack important colour information for distinguishing them. The time factor is also relevant, since wild animals stay in the same place only very briefly, so the system has to work in under 5 s.
BSFZ's assessment: Challenges are described concretely and justified. Possible problems during implementation are presented concretely and are comprehensible on the basis of the project content.
Poor description (799 of 1000 characters in the German original)
A potential risk is that the AI does not meet the requirements (accuracy 99 % and speed < 3 seconds) and may not assign the squirrel correctly, so that the animal receives the wrong feed. Furthermore, training on a CPU takes a long time and the data quality and quantity may not be sufficient. If zero-error tolerance of the AI cannot be achieved for BLEU score, perplexity and area over curve, the project has failed. It could also be that the feeding station does not work because the cameras are not maintained and calibrated regularly, so that they do not deliver reliable data. In addition, the operators of the station must be able to react quickly to problems and, if necessary, control the system manually.
BSFZ's assessment: Merely listing requirements that may not be achievable, without a recognisable connection to the problem addressed, the components used or the solution approaches, is not sufficient for the assessment. The challenges show no relation to the concrete project content. There are technically incorrect statements. Risks that may occur after completion of the R&D project are not taken into account in the risk assessment.
A potential risk is that the AI does not meet the requirements (accuracy 99 % and speed < 3 seconds) and may not assign the squirrel correctly, so that the animal receives the wrong feed. Furthermore, training on a CPU takes a long time and the data quality and quantity may not be sufficient. If zero-error tolerance of the AI cannot be achieved for BLEU score, perplexity and area over curve, the project has failed. It could also be that the feeding station does not work because the cameras are not maintained and calibrated regularly, so that they do not deliver reliable data. In addition, the operators of the station must be able to react quickly to problems and, if necessary, control the system manually.
BSFZ's assessment: Merely listing requirements that may not be achievable, without a recognisable connection to the problem addressed, the components used or the solution approaches, is not sufficient for the assessment. The challenges show no relation to the concrete project content. There are technically incorrect statements. Risks that may occur after completion of the R&D project are not taken into account in the risk assessment.
The difference is the same in all five fields: the strong version describes the route and the open problem, the weak one the finished product. Under § 2 FZulG the BSFZ assesses novelty, technical risk and systematic approach: all three are properties of the route, not of the result.
The example comes from the BSFZ itself, including the character limits per field. We write applications in the same order: the knowledge gap first, then the route, then the risk.
How does the BSFZ recognise a convincing project description?
By how concrete it is, not by its length. The table sets weak and strong wording for the three criteria side by side; the BSFZ example above shows it on a whole application.
| Criterion | Weak wording | Strong wording |
|---|---|---|
| Novelty | no “We are developing innovative software.” No state of the art, no distinction. | yes Existing solutions need labeled training data. The example company detects anomalies without labeled data, an approach beyond the state of the art. |
| Technical Risk | no “Implementation is demanding.” No concrete risk named. | yes It is open whether the model generalises across different machines and sensor types. That is an unsolved problem. |
| Systematic Approach | no “We work in a structured way.” No steps named. | yes Four steps: literature review, model development, evaluation with three pilot customers, iterative optimisation, documented in a research log. |
Weak wording is not wrong, only incomplete: the BSFZ cannot assess novelty, risk or systematic approach from such sentences and raises a query. If the deadline for that is missed, the application can be rejected without a substantive review.
Why the strong version carries the application
- The state of the art is named, not just claimed: existing solutions and their limit.
- The technical risk is concrete: where the project could have failed.
- The research plan shows steps instead of a label like “structured”.
- The result stays open; the description does not end with the finished product.
What does the tabular work plan look like in the BSFZ example?
Eleven work packages, each with period, person-months and who carries it out. Contractors are listed on every package they contribute to.
Tabular work plan
For every project, a tabular work plan stating the personnel effort must be submitted. If the project includes contracted work, the work plan must state in which work packages the contractor(s) are involved. The template in the application portal must be used for this.
| Work package | Period | Employed staff (person-months) | Carried out by |
|---|---|---|---|
| Design and construction of a feeding station with cameras and motion sensors | 04.2024 – 07.2024 | 4 PM | Sample Company GmbH; contractor: Sample Services GbR |
| Capturing test images of squirrels with an IR and a standard wildlife camera | 07.2024 – 08.2024 | 2 PM | Sample Company GmbH |
| Adjusting the alignment of the sensors and the parameters (exposure time, resolution) for the images | 08.2024 – 10.2024 | 2 PM | Sample Company GmbH |
| Capturing 1,000 images of 50 different squirrels with the IR and the standard wildlife camera | 10.2024 – 12.2024 | 3 PM | Sample Company GmbH |
| Image annotation (assigning the individual squirrels in the images) | 10.2024 – 12.2024 | 3 PM | Sample Company GmbH |
| Development of YOLO model elements for squirrel recognition, implementation: PyTorch & YOLO package | 07.2024 – 12.2024 | 4 PM | Sample Company GmbH; contractor: Sample Server GmbH |
| Data preparation, training and validation of the model for identifying squirrels | 01.2025 – 04.2025 | 3 PM | Sample Company GmbH; contractor: Sample Server GmbH |
| Testing the algorithm with live images from the IR and wildlife camera | 05.2025 – 07.2025 | 2 PM | Sample Company GmbH; contractor: Sample Server GmbH |
| Adjusting the algorithm with regard to the weighting of individual head features | 05.2025 – 07.2025 | 2 PM | Sample Company GmbH |
| Testing the entire AI-based feeding station under different conditions (weather, day/night) | 07.2025 – 08.2025 | 2 PM | Sample Company GmbH; contractor: Sample Services GbR |
| Adapting the station and the algorithm to different weather and lighting conditions | 08.2025 – 12.2025 | 2 PM | Sample Company GmbH; contractor: Sample Services GbR |
- The work packages are meaningful, build on one another and refer directly to the description of the work.
- Only common or previously introduced abbreviations are used.
- A systematic approach with a comprehensible methodology and solution strategy is recognisable.
- The personnel effort is stated for each work package and is comprehensible and plausible in the context of the work carried out.
- The contractors are listed in every work package in whose implementation they are involved.
How does the project description turn into an assessment basis?
From personnel costs, the managing director's allocated salary and contract research, not from an estimate. The example calculates with illustrative line items.
The example company employs four developers at EUR 65,000 employer gross each with a 70 % R&D share. The employed managing director also contributes 46 % of their time to the project; their salary counts as personnel costs under Section 3(1) FZulG, not as an own contribution. Own contributions at EUR 100 per hour are reserved for sole proprietors and co-entrepreneurs.
| Cost type | Calculation | Amount |
|---|---|---|
| Personnel costs, developers | 4 × EUR 65,000 employer gross, 70 % R&D share | EUR 182,000 |
| Personnel costs, managing director | EUR 200,000 employer gross, 46 % R&D share | EUR 92,000 |
| Contract research | EUR 40,000 contract volume, 70 % eligible | EUR 28,000 |
| Assessment basis | Total | EUR 302,000 |
| Research allowance | 35 % SME rate | EUR 105,700 |
Result: EUR 105,700 tax credit. The tax office first offsets it against corporate tax; any surplus is paid out. With an expected corporate tax of EUR 80,000, EUR 25,700 remains for direct payout.
Which mistakes turn the strong example into a rejected application?
It is rarely the project that fails, usually the presentation or a deadline. These five mistakes are the ones we see most often, and the BSFZ's poor version above shows most of them.
- Vague description: “We are developing new software” replaces neither the state of the art nor the technical risk.
- Result instead of route: the description names only the finished product. The BSFZ assesses the route there, not the result.
- Personnel costs misallocated: the managing director's salary counts as personnel costs under Section 3(1) FZulG, not as an own contribution; entering it wrongly risks a cut.
- Contract research overstated: only 70 % of the contract volume is eligible. Claiming 100 % results in a cut.
- Missing time tracking: without time records, the tax office can reduce the R&D share. Time tracking starts before the project, not after.
BSFZ queries come with short deadlines. A late or vague answer can lead to rejection before the substantive review is complete.
Which questions come up most often about this application example?
No. The BSFZ designed the AI-controlled feeding station for squirrels as a fictional project to show, on an example, what separates a good description from a poor one. The texts on this page are its original, field by field.
The strong version names the state of the art concretely, describes an actual technical risk and shows the work steps. The weak version only claims innovation without letting the BSFZ assess novelty, risk or systematic approach.
No. Section 3(3) FZulG reserves own contributions at EUR 100 per hour exclusively for sole proprietors and co-entrepreneurs of partnerships. The salary of an employed managing director counts instead as personnel costs under Section 3(1) FZulG.
The assessment basis is capped at EUR 12 million per fiscal year. At the 35 % SME rate that yields at most EUR 4.2 million allowance; without SME status the standard rate of 25 % applies.
A rejection can be appealed with a revised description. Often it is enough to name the state of the art or the technical risk more concretely, rather than reinventing the project.
The five fields of the project description allow around 4,200 characters combined. What matters is not the length but whether novelty, technical risk and systematic approach are answered concretely.
Have your project description reviewed before filing
In the first consultation we check whether your project description carries the three criteria of § 2 FZulG and show where it is still too vague. Success-based, no upfront costs.
Book a first consultationRead on
- Applying for the research allowance: step by stepThe BSFZ application and research allowance application (ELSTER) in sequence, with deadlines and documents.
- Calculating the assessment basisHow personnel costs, own contributions and contract research add up to the assessment basis.
- Own contributionsWho may claim EUR 100 per hour and who may not.
- Contract researchWhich 70 % of a contract volume is eligible.
- BSFZ portal: guideRegistration, form fields and submission step by step.
- Research allowance calculatorCalculate the assessment basis and allowance from your personnel costs.
Max Nodes
Founder and Managing Director of NOVARIS Consulting. Specialised in the German R&D tax credit (FZulG) since 2021; 25 of 25 accompanied BSFZ applications approved (as of 09/2026). Max Nodes