Self-correcting AI agents as a development approach in the Dutch startup scene

Self-correcting AI agents are not a distant prospect: several Dutch startups have been building products with them for years. From pricing software that monitors margins to sensor systems that signal machine failures weeks in advance, the underlying approach is the same. A system observes, adjusts and learns from the result, in a continuous cycle.

Reinforcement learning, feedback loops and self-correction appear in marketing presentations from companies around the world, but behind those terms lie divergent technical choices. At a number of Dutch startups, those choices were made years ago and have since been tested in practice. A comparison of five companies shows what this approach delivers in practice and where the differences lie.

Pricing and forecasting as a feedback machine

Rotterdam-based SYMSON, founded in 2011 and the recipient of a Series B round of nearly €60 million in 2022, builds AI-driven pricing software for companies looking to protect their margins. The core of the system is a feedback loop: the software proposes a price, measures what subsequently happens in sales and margins, and adjusts its next recommendation accordingly. That is, in essence, what reinforcement learning means in a commercial context. The system does not learn once on historical data, but continuously corrects itself based on what the market returns.

A comparable logic underlies the work of Visualfabriq from Maarssen, active since 2013. The company focuses on revenue forecasting and promotion management for consumer goods companies. Promotions are inherently cyclical and measurable: you launch a campaign, measure the response and adjust the next cycle accordingly. Visualfabriq formalised that cycle in software that helps planners make better estimates. With €1.8 million in private equity financing (2021), the company is smaller than SYMSON, but it operates on the same principle of iterative adjustment based on measured outcomes.

Detecting failures before they become visible

In the industrial sector, the feedback loop takes a different form. At Samotics, founded in 2015 in Leiden and financed with €20 million, the focus is on electrical signal measurement on industrial machines. The system continuously reads current and voltage patterns and compares them against a baseline profile it has previously built up. Deviations are flagged, the model adjusts its reference based on new data and in doing so refines its sense of what is normal. The result, according to the company, is that failures are detected up to five months earlier than would be the case with conventional monitoring.

Jungle AI from Baarn, founded in 2017 and financed with €10 million in a Series A round in 2022, works along similar lines but targets a broader range of industrial assets, from wind turbines to production lines. The system detects underperformance and anomalous behaviour by continuously comparing against what the machine did previously under similar conditions. That self-correcting reference model is the core of the approach: the machine learns what healthy behaviour looks like and signals when that pattern shifts.

Both companies thereby solve the same problem as SYMSON and Visualfabriq, but in an entirely different context. Where pricing software draws feedback from sales figures, industrial AI draws it from electrical signals and sensor data. The architecture is analogous; the application is fundamentally different.

Oversight of the system itself

A distinct category is formed by Deeploy, founded in 2020 in Rotterdam and financed with €8.5 million in a Series A round in 2024. Deeploy does not build an AI agent that adjusts itself based on domain data, but a platform that monitors AI models in production. The system measures whether models are still performing as expected, signals drift and gives teams the tools to intervene before a model goes off course.

This is a different interpretation of the feedback loop. It is not the model that corrects itself, but an external system that monitors model quality and informs the people responsible for it. In the context of the AI Act and increasing compliance requirements, that is a relevant position: organisations are required to demonstrably maintain control over their AI systems, and Deeploy provides the infrastructure to do so.

The comparison with the other four companies reveals something important. SYMSON, Visualfabriq, Samotics and Jungle AI build systems that learn and self-correct. Deeploy builds the oversight system that monitors whether that self-correction is still proceeding responsibly. Both layers are necessary in an environment where AI models autonomously influence decisions.

What the comparison reveals

The five companies show that self-correcting systems are already live in various sectors, with concrete customers and measurable outcomes. The shared development approach is a cycle of measuring, comparing and adjusting, but the execution varies considerably. Some systems learn from transaction data, others from sensor signals, and others still monitor models from the outside.

What stands out is that most of these startups have built their product around a specific, well-defined decision problem. This is a pattern that recurs in successful applications of machine learning: the system has a clear objective, a measurable outcome and a mechanism to evaluate its own performance. Without those three elements, self-correction is difficult to achieve.

For investors and other founders in the Dutch ecosystem, it is notable that this approach is not confined to one sector or one type of data. Pricing, revenue forecasting, machine behaviour and model oversight are all areas where feedback loops add value. The underlying technical principles are transferable; the key lies in choosing the right domain and defining the right objective function.