WBSO for AI agent and automation development: does your R&D qualify in 2026?

Key takeaways: AI agents have evolved from an experimental concept to the dominant development trend in the software industry in 2026. For Dutch tech founders building their own AI agent software, the WBSO scheme offers a substantial tax advantage: up to 50% wage tax reduction for startups on the first EUR 391,000 in R&D costs. But the boundary between qualifying development work and non-qualifying use of existing tools is sharper than most founders realize.

Building AI agents is R&D, but not always according to RVO

Over the past year, the number of companies developing AI agents has grown explosively. Autonomous systems that independently execute tasks, make decisions, and take actions without continuous human oversight are no longer the exclusive domain of large tech companies. Startups and scale-ups are building agents for customer service, supply chain optimization, financial analysis, legal research, and dozens of other domains.

What many of these founders do not realize: developing AI agent software can qualify for the WBSO (Wet Bevordering Speur- en Ontwikkelingswerk), the primary Dutch tax incentive scheme for R&D in the technology sector. The scheme significantly reduces the wage costs of your R&D staff. Regular companies receive a 36% wage tax reduction on the first EUR 391,000 in R&D wage costs and 16% above that threshold. For startups, the rate on the first bracket is 50%, making it particularly attractive for young AI companies.

But there is a fundamental condition attached that is specifically relevant to AI development: the WBSO exclusively covers the development of proprietary, technically novel software. And that is where the distinction begins that separates an approved application from a rejected one for AI agent companies.

What the WBSO covers in AI development

The WBSO is intended for research and development work (S&O) aimed at developing technically novel software. RVO, the Netherlands Enterprise Agency that assesses WBSO applications, applies three core requirements: you must develop your own software in your own source code, the work must be technically novel compared to existing solutions available to you, and there must be technical challenges that you need to resolve.

For AI agent development, this concretely means that your work qualifies when you write software that solves technically novel problems. It concerns the code you write yourself, not the models you call or the platforms you configure. This distinction is crucial in an era where much AI development consists of orchestrating existing components.

Work that qualifies

The following types of AI agent development typically fall within the WBSO scope, provided you can demonstrate technical novelty.

Custom agent architectures and orchestration mechanisms. When you develop a proprietary framework that determines how an AI agent plans tasks, delegates subtasks, evaluates intermediate results, and self-corrects, that is software development at its core. You write the logic that determines when an agent calls a tool, how parallel tasks are coordinated, and how failure scenarios are handled. This type of work almost always contains technical challenges: how do you prevent infinite loops, how do you optimize the sequence of tool calls, how do you handle conflicting results from multiple sources.

Custom tool-use frameworks. AI agents get their reach from the ability to invoke external tools. When you develop a proprietary system that dynamically determines which tools are available, how tool output is validated, and how the agent learns which tools are effective for specific tasks, you are developing technically novel software. The complexity lies in reliably translating natural language instructions into structured API calls, handling errors, and optimizing tool selection.

Proprietary fine-tuning and training pipelines. Training a neural network in itself does not qualify for the WBSO. But the software you build around the training process does, if it is technically novel. Think of custom data preprocessing pipelines that transform domain-specific data into training formats, evaluation frameworks that measure model quality on metrics specific to your domain, or systems that automatically generate and filter training data based on production usage.

Inference optimization. When you develop proprietary software that improves the speed, cost, or quality of model inference, that typically qualifies. This includes custom caching mechanisms for frequent queries, intelligent routing between different models based on complexity, or proprietary quantization methods that reduce model size without unacceptable quality degradation.

Multi-agent coordination systems. Orchestrating multiple specialized agents that collaborate on complex tasks is a domain full of technical challenges. Proprietary software for task decomposition, inter-agent communication protocols, conflict resolution, and result aggregation typically contains sufficient technical novelty.

What does not qualify

Equally important as knowing what qualifies is understanding which work falls outside the WBSO scope. RVO is strict about this in practice.

Calling existing AI APIs without proprietary software development. When you use the ChatGPT API, Claude API, or another LLM provider and pass the output directly to your user, you are not developing technically novel software. You are integrating an existing service. It does not matter how innovative your application is from a user perspective. The WBSO assesses technical novelty in your own code, not the novelty of your product or business model.

Configuring no-code and low-code platforms. Tools such as Langflow, Flowise, Zapier AI, or similar platforms allow you to assemble AI workflows without writing your own code. No matter how complex your configuration, it remains configuration of existing software. There is no proprietary source code development and therefore no WBSO basis.

Prompt engineering as the primary activity. Writing, testing, and optimizing prompts is not software development. Prompts are instructions to an existing system, not software. Even when you systematically test hundreds of prompt variants with automated evaluation, the core activity is optimizing input for existing software, not developing new software.

Training models without new software. Fine-tuning an existing model with your own data via the standard tools of the model provider (OpenAI fine-tuning API, Hugging Face Trainer with default configuration) does not qualify. You are using existing training software. The resulting model may be new, but the WBSO assesses the software you develop, not the model you produce.

Integrating existing SaaS tools. Connecting your AI agent to Slack, Salesforce, Jira, or other existing services via their standard APIs is integration, not R&D. Even when you combine multiple tools into a workflow that has never existed before, there is no technically novel software if you do not develop your own orchestration mechanisms.

The grey area: when does "using an API" become "developing software"?

In practice, much AI agent work occupies a grey area between pure API usage and full-fledged software development. The boundary is not always sharp, but there are clear indicators.

Scenario 1: you build a customer service agent. You use an LLM API for generating responses and write a proprietary system around it that retrieves customer context from multiple sources, determines when a conversation should be escalated to a human agent, automatically evaluates the quality of generated responses before they are sent, and learns from corrections that staff members make. The LLM API call itself does not qualify, but the proprietary system you build around it potentially does. The technical novelty lies in the orchestration, quality control, and learning mechanism, not in the text generation.

Scenario 2: you develop an autonomous data analysis agent. Your agent receives a natural language question, independently determines which data sources are relevant, writes and executes SQL queries, validates the results, and presents a substantiated answer. The technically novel element is the software you write to enable the agent to reliably and securely generate and execute queries: SQL injection prevention, query optimization, result validation, and the logic that determines when the agent needs additional data. If you develop proprietary algorithms and mechanisms for this that are not directly available in existing frameworks, this typically qualifies.

Scenario 3: you configure an existing agent framework. You use LangChain, CrewAI, or AutoGen and configure agents with specific roles, tools, and instructions. You customize prompts for your domain and connect standard APIs. Although the end result can be impressive, there is no proprietary software development if you remain within the capabilities of the existing framework. This does not qualify.

Scenario 4: you extend an existing framework with custom components. You use LangChain as a foundation but develop proprietary components that do not exist in the framework: a custom memory architecture that retrieves relevant context more efficiently, a proprietary planning algorithm that optimizes task decomposition for your specific domain, or a custom evaluation module that measures agent performance on domain-specific criteria. The proprietary components that are technically novel can qualify; the configuration work cannot.

The rule of thumb: ask yourself whether you are solving technical problems for which no ready-made solution is available. If the answer is yes and you solve those problems by writing your own software, you are likely in WBSO territory.

Structuring your WBSO application for AI agent work

An approved WBSO application for AI agent development requires you to clearly demonstrate what is technically novel about your work. RVO does not assess whether your product is innovative, but whether your software development contains technical challenges that you have not previously solved and for which no existing solution is available.

Describe technical challenges, not product benefits. Do not write: "We are developing an AI agent that autonomously generates financial reports." Instead write: "We are developing an orchestration mechanism that coordinates multiple specialized sub-agents for extracting, validating, and combining financial data from heterogeneous sources, where the technical challenge lies in ensuring data consistency during simultaneous updates and detecting conflicting information from different sources."

Make the distinction explicit between proprietary code and third-party tooling. RVO wants to know which part of your work involves existing tooling and which part is proprietary development. Explicitly state which LLM providers, frameworks, and tools you use, and then describe what you develop on top of them. Transparency about this strengthens your application, because it shows you understand the distinction.

Document your search for existing solutions. The WBSO requires technical novelty relative to what is available to you. Substantiate why existing frameworks, libraries, or services do not solve your technical problem. This does not need to be an exhaustive market study, but you must be able to demonstrate that you investigated whether an existing solution was available before deciding to develop proprietary software.

Maintain an S&O administration. Throughout the project, you are required to maintain an administration of the S&O work performed. Record which technical activities your R&D staff carry out, how many hours they spend on them, and what the results are. This is not only an obligation; it is also your evidence in the event of an audit.

The startup advantage: 50% in your first years

For tech founders starting a new company to develop AI agents, the WBSO startup rate is particularly attractive. Instead of the regular 36%, you receive a 50% wage tax reduction on the first EUR 391,000 in R&D wage costs. With a team of three developers at market-rate salaries, you will not reach that threshold in your first year, which means the full startup rate applies to all your R&D wage costs.

To qualify as a startup, you must not have received an S&O declaration in one or more of the five preceding calendar years. In practice, this means that new companies specifically established for AI agent development almost always qualify.

The financial impact is substantial. At EUR 300,000 in R&D wage costs, the startup rate delivers EUR 150,000 in wage tax reduction, compared to EUR 108,000 at the regular rate. That difference of EUR 42,000 is significant for a startup with limited capital.

Maximize the benefit in year 1. Submit your WBSO application before you begin the R&D work, not afterward. The WBSO operates with application periods: you request an S&O declaration in advance for an upcoming period. Ensure your application is approved before your developers start, otherwise you miss the wage tax reduction for those initially crucial months.

The next step

The WBSO application procedure is not complicated, but it does require you to describe your work in the right way. The difference between an approved and a rejected application often lies not in the work itself, but in how clearly you articulate the technical novelty and challenges.

If you are developing AI agent software and are unsure whether your work qualifies, the first step is to clearly map out which part of your work constitutes proprietary software development and where the technical challenges lie. The boundary between qualifying and non-qualifying work runs straight through many AI projects. That is not a reason to leave the WBSO on the table. It is a reason to structure your application carefully.

Want to learn more about the WBSO scheme and how it applies to your specific situation? Read our comprehensive guide to the WBSO or explore how the scheme relates to other subsidies for AI projects and the MIT scheme.