AI agent development for Dutch businesses

What an agent is, and when it beats a workflow

A workflow follows fixed steps: when this happens, do that. An agent chooses its steps. Given a goal such as find the right spare part for this customer email and prepare a quote, it reads the email, searches your product data, checks stock and drafts the reply. The path can differ each time.

That freedom brings risk. For tasks with a known sequence, a plain automation is cheaper and safer, and we say so. We use agents where the input varies a lot and judgement over text is needed. Our AI agent vs chatbot comparison explains the difference in more detail.

A useful rule is to ask what a new employee would need. If a trainee could do the task with a checklist and access to three systems, an agent probably can too. If it needs deep judgement, personal trust or legal sign-off, keep a person in charge and let the agent prepare the groundwork.

Examples that suit Dutch firms

Each example has a defined set of tools and a clear stopping rule. Where an agent could cost money or send a message, a person confirms first.

Cost per task is worth measuring from day one. An agent that uses several model calls and searches can cost more than the labour it saves if left unchecked. We estimate cost per run in the pilot and set budgets per day, so the business case stays visible.

Safety by design

We apply the least-privilege idea to agents. An agent receives only the tools and data it needs, with read-only access wherever writing is not required. Every tool call is logged with its inputs and results, so you can see why something happened. Spending limits and rate limits stop runaway loops, and a kill switch lets you turn the agent off. Our AI agent security checklist lists the controls.

Agents that read outside content, such as emails or web pages, can meet text written to trick them. We treat that content as data, never as instructions, and we limit what the agent can do after reading it.

AVG, the AI Act and the works council

An agent that touches personal data needs a lawful basis, a processor agreement with each model provider and a retention rule. The Autoriteit Persoonsgegevens enforces the AVG and has published material on AI, so your privacy contact should review the data flows we document. The EU AI Act sets duties that depend on how an AI system is used, with timelines that your lawyer should confirm for your case. Systems that evaluate staff may also need consent of the ondernemingsraad. Raise this before the build, not after.

How we build and test

1. ScopeOne goal, a list of tools, and the cases where the agent must stop.
2. Test setFifty or more real examples from your history, with the answers your staff would give.
3. PrototypeA working agent in a sandbox, scored against the test set.
4. PilotA small group of staff uses it with approvals on, and we fix what they find.
5. Roll outMonitoring, cost alerts and a monthly review of mistakes.

We report error rates from your own examples, not from marketing claims. The LLM evaluation checklist shows how we think about testing. For the wider view, read the Netherlands AI overview.

After roll out we keep a feedback button inside the tool. Staff mark answers as good or wrong, and we use those marks to improve prompts and add test cases. Over a few months the agent improves on the cases your business actually sees.

Pricing

Projects start at $1,500, and our rate is $30 an hour. A narrow agent with two or three tools is near the start. More tools, more approval screens and tighter testing raise the price. Model usage is billed by the provider and depends on volume. See the AI agent development service and the AI agent cost guide. The Netherlands is 3.5 to 4.5 hours behind India, so reviews fit your morning.