3 OCTOBER 2026 · WBSO 2027 · ANNOUNCED
WBSO for AI and software: expansion announced for 2027
The WBSO is intended to align more closely with the development of artificial intelligence and modern software. The coalition agreement announces an expansion of the WBSO for AI and software, and RVO confirms that the government is investigating which changes are needed. This does not mean that expanded WBSO eligibility will automatically apply to every AI or software project from 1 January 2027 — as of 3 October 2026 the concrete details are still under investigation. For software companies, AI startups and businesses developing their own intelligent systems, this is nevertheless an important development: the current rules draw a clear distinction between new functionality and technically new software. Precisely that distinction is becoming ever more relevant with AI.
STATUS · 3 OCTOBER 2026
NOW
AI can already qualify for WBSO — under the current R&D conditions
ANNOUNCED
Expansion of the WBSO for AI/software in the coalition agreement
STILL UNKNOWN
Concrete new criteria — investigation ongoing; current rules apply for now
In addition, RVO is investigating a more consistent approach to prototypes within the existing legislation; further details are expected in the WBSO Manual 2027, due to be published in November 2026.
AI is becoming increasingly important within the WBSO
Artificial intelligence is no longer a niche area within research and development. The WBSO 2025 annual report shows that the number of R&D person-years devoted to AI has increased by 77% over five years. RVO notes that AI is both developed as a technology in its own right and used as a tool in developing other technologies. The WBSO is technology-neutral and has no separate category for AI: anyone developing an AI application must demonstrate, within the same statutory framework, that qualifying research and development is taking place. That is where things sometimes go wrong in practice — modern AI development differs from classic software development, with neural networks and foundation models, machine-learning models, large language models, embeddings, probabilistic rather than fully deterministic systems, model training and fine-tuning, reinforcement learning, generative and multimodal models, specialised inference techniques, GPU and accelerator architectures, and agents and autonomous software chains. The core question is where the application of existing AI technology ends and the development of technically new software begins.
What has the government announced?
In its explanation of the 2027 WBSO changes, RVO states explicitly that the coalition agreement announces an expansion of the WBSO for AI/software. That expansion must remain consistent with the technology-neutral nature of the WBSO; an investigation is underway into whether the scheme still aligns well with rapid developments in the digital domain. No new assessment framework has been published yet stating, for example, which activities will and will not qualify from 2027 — For now, RVO describes this only as an ongoing investigation. The announcement therefore does not mean that all AI development will automatically receive WBSO from now on, nor that using an existing AI service such as ChatGPT, a third-party language model, an ML library or an AI API is automatically considered R&D. Until further rules are published, it remains sensible to test an AI project against the existing conditions for technically new software.
When does software currently qualify for WBSO?
Under the current WBSO, a company must itself develop a technically new product, production process or technically new software, or carry out technical-scientific research. A development project must, among other things, involve technical novelty for the company, concrete technical bottlenecks, uncertainty about technical feasibility, technical approaches and development work undertaken by the company itself. RVO describes software as the non-physical logical subsystem of an information system that determines the structure of data and processing operations, insofar as laid down in a formal programming language. A key starting point:
New software is not automatically technically new software
Almost every new application contains code that did not exist before — that alone does not make the development a WBSO project. Technically new software requires a new information-technology principle in which the company itself solves technical bottlenecks. The development process is usually iterative and involves genuine technical risks or uncertainties. The difference between functionally new and technically new is particularly important in AI projects.
An AI feature is not the same as AI R&D
Suppose a company extends an existing webshop with a chatbot based on an existing LLM provider via an API. The developers build a user interface, a link to the product database, a prompt, a standard vector database, a retrieval layer and an API connection to an existing language model. For the company this may be an entirely new feature — but that does not automatically mean technically new software is being developed. If existing technology is combined, configured or implemented using known methods, the project is primarily application or systems development rather than R&D. RVO explicitly names applying, assembling or implementing existing software as activities that do not in themselves qualify as R&D.
When might an AI project qualify for WBSO?
The situation changes when genuine IT problems arise during development that cannot be solved with available techniques — for example because existing technology underperforms in terms of computing time, memory usage, latency, scalability, synchronisation, fault tolerance, model compression, inference speed, resource allocation, data throughput, deterministic behaviour, parallel processing, model architecture or the technical interaction between self-developed components. It is not primarily about what the AI must be able to do, but which information-technology problem the developers themselves must solve to make it possible.
EXAMPLE: MAKING AN AI MODEL RESPOND FASTER
Functional requirement: "The system must respond within 100 milliseconds." That is not in itself a technical bottleneck. A bottleneck arises, for example, when a self-developed model cannot function within that limit due to its data structures and computational methods, and existing optimisation techniques prove insufficient. The developers may then investigate new data structures, algorithms, computational strategies, partitioning, caching methods, quantisation techniques, model representations or parallel execution methods.
If those solutions are technically new to the company, are programmed in-house and there is advance uncertainty about whether the solution will work, it becomes much easier to substantiate why R&D is involved. RVO itself uses the comparable principle of the new algorithm: if an existing algorithm cannot meet the specifications and the company therefore develops a new, more efficient algorithm, that can qualify — provided the principle is actually laid down and developed in a formal programming language. Merely conceiving or describing it is not enough.
Machine learning, generative AI and LLMs
The same distinction between use and own development applies to machine learning. Training with existing libraries, collecting and labelling data, standard preprocessing, setting up or comparing existing models, regular hyperparameter optimisation, AutoML, or applying an existing model to a new market — none of this is automatically R&D; commercial or functional novelty is not decisive. With generative AI the distinction is even sharper: using an existing foundation model via an API, prompting, prompt chaining, retrieval-augmented generation, standard embeddings and vector search, function calling, or agents built on existing frameworks does not in itself produce a technically new software principle. The central question remains: which concrete information-technology problems can the company not solve with existing knowledge and techniques, and which new technical solution principles does it programme itself? An application must therefore go beyond terms like LLM, machine learning, neural network, RAG or agentic AI — those describe the technology, but not yet the technical R&D.
What could the announced expansion change?
That is not yet known as of 3 October 2026. RVO states that it must be investigated where the current WBSO insufficiently aligns with the digital domain; the expansion must stay within the scheme's generic character. It cannot be concluded from this that certain activities will automatically become eligible from 2027. What is clear is why the topic is on the agenda: modern software development increasingly consists of building on open-source frameworks, cloud platforms, foundation models, external libraries, accelerators, APIs, existing model architectures and complex ecosystems — technical innovation takes place in a different layer than in classic development. The question is whether the boundary between application and technically new software still matches where R&D actually happens.
What does this mean for a 2027 WBSO application?
Do not anticipate an expansion that has not yet been worked out; base your preparation on the existing assessment criteria. Five questions help make the application technical:
1. What is being developed technically? Describe the software component or technical operating principle your own developers actually programme. Not: "we are developing an AI platform for automated document analysis", but for example: "we are developing a new processing layer that enables concurrent inference over highly variable document structures without the current linear growth in memory load".
2. What are the technical bottlenecks? RVO explicitly asks for concrete IT bottlenecks — not customer wishes, features, user problems, planning, privacy requirements, commercial risks or quality objectives. It must become clear why the desired behaviour cannot technically be achieved with standard techniques.
3. What will you develop yourself? Link a concrete technical approach to every bottleneck: algorithms, data structures, processing logic, model components, protocols, computational methods, architecture components or new programming techniques your team will actually investigate and develop.
4. Why is the outcome technically uncertain? If it is already clear in advance how the technical problem will be solved, a key characteristic of R&D is missing. There must be a genuine risk that the chosen approach will not work technically.
5. Which activities are development and which are routine engineering? R&D: new algorithmic behaviour, programming and testing a technically new operating principle, new data structures, solving fundamental performance problems, technical experimentation. Not automatically R&D: UI development, routine API integration, implementing standard software, data entry, standard database work, configuration, documentation, market validation, functional design, routine testing, deployment and maintenance. Clearly delineate the R&D part within a larger project.
Using AI to develop software ≠ developing AI
Developers increasingly use AI as a development tool: code generation, copilots, automated debugging, AI-assisted testing, code review, documentation. That does not in itself make a project technically new. The WBSO assesses what the company itself develops technically and which own R&D activities are carried out; the fact that AI is used during those activities does not automatically determine whether a project qualifies.
RVO is also looking at prototypes
Separate from the announced AI/software expansion, RVO is investigating whether a revised, more uniform line on prototypes is possible within the existing legislation; further details are expected from November 2026 in the WBSO Manual 2027. This can also be relevant for software and AI projects, because a development project under the WBSO in principle ends once the new technical operating principle has been demonstrated. How the new interpretation will be phrased exactly is not yet known; this page will be updated as soon as RVO publishes the Manual.
WBSO and AI in 2027: what do we know now?
AI projects can receive WBSO — yes, if they meet the current R&D conditions
New separate AI scheme — no
WBSO expansion for AI/software — announced
Concrete new AI criteria — not yet published; investigation ongoing
Current rules on technically new software — leading for now
New prototype line — under investigation; details expected in the WBSO Manual 2027 from November 2026
Should you wait with an AI application?
No. AI and software projects can already qualify under the current WBSO if they meet the conditions for technically new software or technical-scientific research. If you want to start on 1 January 2027, it makes sense to inventory your projects now — for employers, the special deadline of 20 December 2026 applies to a 1 January start. The further elaboration of the expansion can be incorporated later, once RVO publishes more information. See also WBSO 2027 – all announced changes.
Have an AI project assessed for the WBSO?
With AI and software projects the technical delineation is often decisive. An application that mainly describes features — automated classification, intelligent search, chatbot, predictive model, generative AI, recommendation engine — does not yet make clear why R&D is involved. Klaren Subsidieadvies assesses AI and software projects in advance on these technical aspects and translates the development into a WBSO project description that matches RVO's assessment framework. What is not technically possible yet, why existing techniques do not solve it, which own software-technical solution is being investigated — and why it is uncertain in advance whether it will work?
Frequently asked questions about WBSO and AI
Does artificial intelligence fall under the WBSO?
Will the WBSO be expanded for AI in 2027?
Is building an AI app automatically WBSO?
Does using ChatGPT or an LLM API qualify?
Is prompt engineering WBSO?
Can developing a new algorithm be WBSO?
Can machine learning fall under the WBSO?
Does WBSO also apply to generative AI?
When will the new rules be known?
CHECKED BY PETER KLAREN · WBSO SPECIALIST SINCE 2004
LAST SUBSTANTIVELY UPDATED: 3 OCTOBER 2026
Sources: RVO — WBSO in 2027: announced changes (AI/software expansion, investigation) · RVO — WBSO Manual 2026 (criteria for technically new software, technical bottlenecks, operating principles) · RVO — WBSO software development project form · RVO — Focus, WBSO Annual Report 2025 (AI-related R&D person-years +77% in five years).