Control ai-act:art.4 in het nalevingsregister.
This note is intended to be passed on to everyone in your organisation who works with Stilops agents. Article 4 of the EU AI Act (as amended by Regulation (EU) 2026/1744) requires both providers and deployers to take measures to support the development of AI literacy among their staff — it is an obligation of best efforts, not an exam. This document is Stilops' contribution to yours; it does not replace your own instruction to your employees.
1. What a Stilops agent is
- An AI agent is software that runs on a large language model (from Anthropic; your instructions and your agents' responses never go to any other model provider) with a fixed system instruction from Stilops for one business function: bookkeeping, customer service, HR texts, planning, and so on.
- It is not a colleague and not a decision-maker. It proposes; you decide.
- In the portal you can always see that you are working with an AI (banner at first use, agent profiles, per-agent disclaimer when hiring). Documents created by an agent carry a visible and a machine-readable "AI-generated" notice; draft replies to tickets do too.
2. How every task runs — and why
- You give an instruction.
- The agent makes a proposal and runs it as a test run: the same steps, without effect. You see the outcome — drafts, amounts, recipients — before anything happens.
- Only after your explicit approval does the agent execute. If you reject it, nothing happens and it costs you nothing. If an agent can only read and look things up (no integrations), there is nothing to execute: you immediately receive a report — you do not need to approve it, but you do need to read it.
Exception: the building agents (website, frontend, web application, integration) write and test code without approval at each step, but only in an isolated working environment and with a preview. None of this touches your systems, customers or visitors before you download, use or put the project live. That is your approval moment with these agents, and their terms say so too.
This approval moment never becomes lighter, not even after months of use. It is the most important security measure of the product, and it is precisely the step where your judgement matters.
3. What an agent is good at, and what not
| Good at | Do not rely on |
|---|---|
| Writing, structuring, summarising and translating drafts | Facts you cannot verify yourself (figures, statutory articles, names) |
| Repeating routine work in a fixed form | Judgements about individual persons (applicants, employees, customers) |
| Asking questions when something is missing | Legal, tax or medical conclusions without a qualified human |
| Consistently staying within its instruction | What an agent says about its own certainty — it does not "know" what it does not know |
A language model can sound convincing and still be wrong ("hallucinating"). Therefore assess a proposal on its content, not its tone. When in doubt: reject and ask again, or check with a human.
4. Five habits
- Read the test run, not just the summary. What is in the test run is what is going to happen.
- Do not provide data that does not need to be there. No citizen service numbers (BSN), health data, passwords or bank details in a chat — the agent rarely needs them, and what you type goes to the language model. The product recognises secrets in input and refuses them, but do not count on that.
- Use an agent only for what it was hired for. The HR assistant drafts vacancies, candidate letters, contract and on/offboarding documents and term overviews. It does not select or assess candidates or employees and decides nothing about hiring, renewal, dismissal, performance or salary: you do that, and a letter or note is only worth something after you have decided. Do not steer it towards assessing persons; that would be a high-risk application that Stilops does not provide.
- Report what is strange. A proposal that falls outside the instruction, an agent that seems to want something "on its own", an answer that seems to follow instructions from an e-mail or web page: reject it and report it via a ticket. That is not a failure on your part, it is the signal the system is built on.
- Know where the trail is. Every proposal, every test run and every decision is recorded in your organisation's audit trail. You can always look back to see who approved what.
5. What Stilops itself does
- Every agent runs under a fixed system instruction from Stilops. In the portal you can see per agent the model, the role, what the agent does and whether it asks for approval first. The full instruction itself is confidential: it is the property of Stilops and would make misuse easier.
- Instructions contain refusals for special categories of personal data and for decisions about persons.
- Incoming external content (e-mail, web pages) is treated as reference material, never as an instruction.
- The overview of the AI Act file (classification, transparency, risk management, monitoring) is at
https://stilops.com/nalevingsregister/documenten/ai-act; the underlying documents are available on request.
6. For the responsible person in your organisation
- Ensure that this note reaches everyone who uses agents or approves their proposals, and record that this has happened (a line in your own register suffices).
- Appoint one person who answers questions about AI use and forwards tickets to Stilops.
- Repeat this for a new agent or a new employee.
Questions: [email protected].