In this article7 sections
An EU AI Act regulator inspection does not start with a 500-page questionnaire. It is likely to start with a handful of targeted questions about whether your organisation has done the basic work. If you cannot answer them clearly, you will be on the back foot for everything that follows.
The EU AI Act entered into force on 1 August 2024. AI literacy (Article 4) and the prohibitions (Article 5) have applied since 2 February 2025, and national market surveillance authorities supervise the Act from 2 August 2026. The obligations for deployers of high-risk AI systems apply from 2 December 2027, after the Digital Omnibus on AI (Regulation (EU) 2026/1744) moved them. In the Netherlands, the Autoriteit Persoonsgegevens (AP) is the intended coordinating supervisor but has no AI Act powers until the Dutch implementing law is in force; in Belgium, no supervisor had been formally designated at the time of writing. The time to prepare your answers is now, not when you receive a letter.
Q1: What AI systems do you currently deploy?
This is the inventory question. Every inspection starts here because regulators need a baseline. They want a documented list: which tools you use, what they do, who supplies them, and in which business processes they operate.
Without a system inventory, you cannot classify risk, assign responsibilities, or demonstrate compliance with any other obligation. The AI Act does not require an inventory as such, but you cannot meet Article 4 for the tools your staff use, or Article 26 for high-risk systems, without knowing what you run. Writing it down is the simplest way to show that.
For an SME with 20 to 200 employees, a practical inventory is a simple register: system name, supplier, version or API endpoint, business function, data inputs, outputs, and the staff roles that interact with it. Include everything: the CV-screening plugin in your ATS, the credit-scoring module your finance team uses, the chatbot on your customer portal. Regulators look for completeness. A partial list signals that your governance is partial too.
Action: Build or update your AI system inventory this quarter. Assign one person as owner.
Q2: What is the highest risk classification in your portfolio?
Once the regulator has your inventory, they immediately look for high-risk systems as defined in Annex III of the AI Act. Annex III lists eight areas where AI systems are presumed high-risk: biometric identification, critical infrastructure, education, employment, essential private and public services, law enforcement, migration, and administration of justice.
For most SMEs, the employment category is the one that bites. AI tools used in recruitment, CV filtering, performance monitoring, or deciding who gets promoted or terminated fall directly under Annex III, point 4. If you use an AI-powered ATS that ranks candidates or a platform that evaluates employee performance, you are likely deploying a high-risk system.
From 2 December 2027, high-risk classification triggers deployer obligations under Article 26. You must use the system according to the provider's instructions for use, which in practice means checking that the supplier has delivered them along with its CE marking and EU declaration of conformity. You must assign human oversight to competent, trained staff. You must keep the logs the system generates for at least six months. You must inform workers before using the system at work. A Fundamental Rights Impact Assessment (FRIA, Article 27) is required only of public bodies, private providers of public services, and deployers of credit-scoring or life and health insurance AI; registration in the EU database is mainly for providers, and for deployers that are public authorities.
Regulators know the Annex III categories by heart. If you misclassify a high-risk system as limited-risk or minimal-risk, that error is both a compliance failure and a credibility problem. Get your classification documented, with the reasoning written down.
Action: Map each system in your inventory to an Annex III category or record explicitly why it does not fall within one.
Q3: What AI literacy training have your staff received?
This question has a legal basis in Article 4 of the AI Act, which, as amended by Regulation (EU) 2026/1744, requires providers and deployers to "take measures to support the development of AI literacy" of their staff and others using AI on their behalf. It does not require you to guarantee any specific level for any individual. The obligation applies now: it has applied since 2 February 2025, together with the prohibitions, the earliest compliance date in the regulation.
The law does not require training records, but they are how you show the measures you took. Expect to be asked what you did, for whom and when, and a training plan or good intentions are a weak answer. Records showing which staff completed what training, when, and what it covered are a strong one. For high-risk systems, the bar will be higher from 2 December 2027. Staff who operate, oversee, or interpret outputs from a high-risk AI system need role-specific training on how that system works, its known limitations, and the human oversight procedures.
A one-hour generic "AI awareness" module is a weak answer if your HR team is using an AI system to screen candidates. That team needs training tied specifically to the tool and the obligations attached to it.
Action: Create a training log with employee names, training content, completion dates, and the AI systems covered. Store it somewhere you can retrieve it in 30 minutes.
Q4: What does your AI policy say?
A written AI policy is the governance anchor the regulator looks for to confirm that AI use in your organisation is managed, not accidental. The policy does not need to be long. It needs to be specific.
At minimum, your AI policy should state: which AI systems are approved for use, who has authority to approve new systems, how risk classification decisions are made and recorded, what the human oversight requirements are, and how staff report concerns or anomalies. For deployers of high-risk systems, the policy should also describe how you will meet Article 26 obligations, and, if you are in the Article 27 group, your FRIA process.
AVG/GDPR interaction matters here too. If your AI systems process personal data, and most HR and customer-facing tools do, your AI policy should cross-reference your DPIA process and your data protection obligations under the AVG. In the Netherlands the AP is both the data protection authority and the intended AI supervisor, so one unified policy is the practical choice.
If you currently have no AI policy, draft one. A two-page document that directly addresses the points above is far better than nothing. Nobody expects perfection. The question is whether you have thought about this at all.
Action: Draft or review your AI policy before the end of this quarter. Make it version-controlled and date-stamped.
Q5: How do you handle AI-related incidents?
Something will go wrong. An AI-powered tool will produce a biased output. A system will flag a candidate for the wrong reason. A decision will be made on the basis of an AI recommendation that later proves incorrect. The question is not whether incidents happen. It is whether you have a process for catching and responding to them.
From 2 December 2027, Article 26(5) requires deployers of high-risk AI systems that identify a serious incident to immediately inform the provider first, and then the importer or distributor and the relevant market surveillance authority. If a deployer believes a system presents a risk, it must inform the provider or distributor and the authority, and suspend use. You need a process ready before the incident, not invented during it.
Your incident handling process should cover four steps: detection (how does a staff member flag a suspected AI-related incident?), assessment (who determines severity and whether it is a serious incident?), notification (whom do you inform, in what order, and how fast?), and remediation (what do you do to prevent recurrence?). Document this process. Test it at least once a year with a tabletop exercise.
For lower-risk systems, incident handling is still good practice even where it is not legally mandated. Regulators view a working incident process as evidence of a mature compliance posture.
Action: Write a one-page AI incident response procedure and assign a named owner for each of the four steps.
Being Ready for All Five
The pattern across all five questions is the same: document everything, assign owners, and store records where you can find them quickly. Regulators are not trying to catch you on technicalities. They are testing whether AI governance in your organisation is real or cosmetic.
For SMEs, the practical roadmap runs in this order. First, build the inventory. Second, classify risk against Annex III. Third, check your training records against Article 4. Fourth, produce or update your AI policy. Fifth, confirm your incident process is documented and tested. None of these steps requires a large team or an external consultant to get started, but each one requires dedicated time.
The Article 4 and Article 5 questions can be asked today. The high-risk questions apply from 2 December 2027. Starting with these five questions now gives you a calm, documented answer whenever an authority asks.
Run a free 2-minute compliance check at comply.khairos.ai to see where your organisation stands on each of these five areas before a regulator asks first.
References
Part of our guide to Audit, Evidence & Governance
Not sure where your company stands?
Answer 15 questions about how your team uses AI. You get a compliance score and the gaps to close first, as a PDF.