RPA and AI automation comply with the AI Act when you correctly classify the systems you use based on risk, comply with the associated obligations, and clearly assign responsibilities between the supplier and the user. For most AI-driven process automation in business environments, the systems fall into the “minimal” or “limited” risk categories, for which the requirements are significantly less stringent than for high-risk applications. In this article, we answer the most frequently asked questions about the AI Act, specifically focused on organizations that work with RPA, Agentic AI, and process automation.
Which AI systems are covered by the AI Act?
The AI Act applies to any AI system offered or deployed in the European Union, regardless of where the provider is established. Specifically, a system falls under the scope of the law if it contains a machine learning component, rule-based reasoning, or a statistical model that generates output, such as predictions, recommendations, decisions, or generated content.
For process automation, this means that traditional rule-based RPA bots fall into a gray area. Purely deterministic scripts that execute fixed instructions step by step without any inference or learning component are generally not considered AI systems under the law. However, as soon as an automation solution uses machine learning, natural language processing, computer vision, or self-learning models to make decisions, the system does fall under the AI Act.
Agentic AI systems—which independently plan and execute tasks based on context and objectives—almost always fall under the definition. The same applies to chatbots that respond based on language models, AI-driven routing systems in contact centers, and computer vision applications that interpret images. In short: the more a system reasons autonomously, the more certain it is that the AI Act applies.
How does the AI Act’s risk classification work?
The AI Act classifies AI systems into four risk levels: unacceptable risk (prohibited), high risk (strictly regulated), limited risk (minimal transparency requirements), and minimal risk (unregulated). The vast majority of business AI applications fall into the minimal or limited-risk categories.
The four levels work as follows:
- Unacceptable risk: Prohibited practices such as manipulative techniques, social scoring by governments, real-time biometric identification in public spaces, and emotion recognition in the workplace. These prohibitions will take effect as of February 2, 2025.
- High risk: Systems used in eight specific areas, such as biometrics, critical infrastructure, education, employment, access to essential services, law enforcement, migration, and the administration of justice. Systems that create profiles of individuals are also always considered high risk.
- Limited risk: Systems such as chatbots and deepfakes where users must be aware that they are interacting with AI. The primary requirement is transparency.
- Minimal risk: Spam filters, AI in video games, and similar applications. No specific obligations.
With regard to process automation in business environments, most applications fall into the “minimal” or “limited” risk categories, unless they are used in one of the eight high-risk domains or perform personal profiling.
What are the requirements for high-risk AI systems?
High-risk AI systems are subject to a comprehensive set of obligations that apply to both providers and users. The core obligations revolve around documentation, transparency, human oversight, and quality assurance of data and processes.
The main obligations for providers of high-risk AI are:
- Establishing and maintaining a robust quality management system
- Prepare technical documentation demonstrating that the system meets the requirements
- Implement automatic activity logging to ensure the system is traceable
- Provide transparent user information about capabilities and limitations
- Enable human oversight so that an employee can adjust or stop the system
- Ensuring robustness, accuracy, and cybersecurity
- Conduct a conformity assessment and affix the CE marking
- Register the system in the EU database for high-risk AI
Users (deployers) of high-risk AI are subject to additional obligations: they must use the system in accordance with the provider’s instructions, organize human oversight, and, when using the system in HR decisions or in the provision of services to citizens, conduct a fundamental rights impact assessment. Users are also required to report incidents and train employees in AI literacy, a requirement that will take effect as of February 2, 2025.
How do you determine whether an RPA process is high-risk or not?
To determine whether an RPA process is high-risk, ask yourself two questions: Does the system fall under Annex I (safety component in regulated products) or Annex III (one of the eight high-risk usage scenarios)? If the answer to both questions is no, the system is not a high-risk system in most cases.
For each automation process, follow these steps:
- Determine whether the system incorporates AI: Does it use machine learning, NLP, or computer vision? If not, then it falls outside the scope of the AI Act.
- Check the Annex III areas: Is the system used for HR decisions (hiring, promotion, termination), creditworthiness assessments, access to government benefits, or other areas within the eight categories?
- Check for personal profiling: Does the system create profiles of individuals based on personal data? If so, it is always considered high risk.
- Assess the impact on fundamental rights: Can the system make decisions that have significant consequences for people? This is an indication of high risk.
- Document your conclusion: Even if you conclude that the system is not high-risk, you should document your reasoning in writing. This will protect you in the event of an audit.
A practical example: An RPA bot that processes invoices and prepares payments for approval is, in most cases, not a high-risk system. An AI system that evaluates resumes and ranks candidates for hiring processes does fall into the high-risk category, however, because it falls under employment and human resources management.
Who is responsible for compliance with the AI Act: the supplier or the user?
Both the provider and the user (deployer) bear responsibility, but for different obligations. The provider is primarily responsible for the system’s technical compliance. The user is responsible for the proper deployment of the system within their own organization and context.
The division of responsibilities is as follows:
- Provider: Conformity assessment, technical documentation, CE marking, registration in the EU database, post-market monitoring, and reporting of incidents to authorities.
- User: Use the system in accordance with the instructions, ensure human supervision, train staff, conduct a fundamental rights impact assessment where required, and report incidents to the provider.
There is one important point to note: a user can become a provider themselves, with all the associated obligations. This occurs when you put your own name on a system, make a substantial change to an existing system, or modify the intended purpose in such a way that a system falls into the high-risk category. Organizations that modify or combine AI systems from suppliers must pay close attention to this.
In addition, suppliers outside the EU must appoint an authorized representative in the EU. Importers and distributors also have their own verification obligations before they place a system on the market or distribute it.
When will the AI Act take effect, and what are the deadlines?
The AI Act will take effect in phases. The first requirements took effect on February 2, 2025. Most requirements for high-risk AI systems will take effect on August 2, 2026, which is the most relevant deadline for many organizations in terms of their process automation.
The complete timeline at a glance:
- February 2, 2025: Prohibited practices (Article 5) are in effect. The requirement for employees to be AI-literate applies.
- August 2, 2025: Requirements for GPAI models (such as large language models), the governance structure, national supervisory authorities, and the provisions on fines take effect.
- August 2, 2026: Most requirements for high-risk Annex III systems will be fully in effect. This is the critical deadline for most business AI applications.
- August 2, 2027: Requirements for high-risk AI used as a safety component in regulated products (Annex I) take effect. GPAI models that were already on the market before August 2025 must be compliant by that date at the latest.
The fines are substantial: violations of the prohibited practices can result in fines of up to 35 million euros or 7% of global annual revenue. Non-compliance with other obligations can result in fines of up to 15 million euros or 3%. For SMEs, the lower of the percentage or the fixed amount applies in each case.
Given the August 2026 deadline, it is wise to start right away by taking stock of your AI systems, conducting a risk classification, and identifying the necessary documentation.
How Pegamento Helps with AI Act-Compliant Process Automation
If you work with process automation and AI, you want to be sure that your solutions comply with the AI Act without compromising your organization’s efficiency or flexibility. We help organizations use AI and automation responsibly, from risk classification to implementation.
What we specifically offer:
- Agentic AI assistants that don’t just follow instructions, but take the initiative and act on their own. This is what we mean by Agentic AI: an evolution from executive bots to self-thinking assistants that actively drive processes. Learn more about our Agentic AI for customer service.
- Customized solutions using standard building blocks—not costly custom work, but a smart combination of proven modules that are perfectly tailored to your situation and industry.
- Everything under one roof, from development and implementation to management and support, without complex supplier management or silos.
- ISO 27001-certified information security, supplemented by ISO 9001 and ISO 26000, ensuring that compliance and quality are structurally guaranteed.
- Guidance on preparing for the AI Act, including assistance with risk classification, documentation, and establishing human oversight of automated processes.
Would you like to know how your current or planned automation solutions align with the AI Act? Contact us, and we’ll work with you to determine the best approach for your organization.
Frequently Asked Questions
Moet ik voor elke RPA-bot afzonderlijk een risicoclassificatie uitvoeren, of kan ik dit per proces of afdeling doen?
Het is verstandig om de risicoclassificatie op systeemniveau uit te voeren, dus per individuele automatiseringsoplossing die een afzonderlijke AI-component bevat. Bots die dezelfde technologie gebruiken voor vergelijkbare taken kun je wel groeperen in een classificatietemplate, maar zodra de gebruikscontext of het domein verschilt, moet je de classificatie herhalen. Een factuurverwerkingsbot en een HR-screeningsbot vallen namelijk in totaal verschillende risicocategorieën, ook al draaien ze op hetzelfde platform.
Wat moet ik concreet doen om te voldoen aan de AI-geletterdheidsplicht die al geldt sinds 2 februari 2025?
De AI-geletterdheidsplicht vereist dat je als gebruikersorganisatie zorgt dat medewerkers die met AI-systemen werken voldoende kennis hebben om deze systemen verantwoord te bedienen en te beoordelen. Concreet betekent dit: breng in kaart welke medewerkers AI-systemen gebruiken, stel een basistraining samen over de werking, beperkingen en risico’s van de betreffende systemen, en documenteer wie wanneer welke training heeft gevolgd. Je hoeft geen AI-experts te maken van je medewerkers, maar je moet aantoonbaar kunnen maken dat ze bewust en bekwaam omgaan met de AI-tools die ze dagelijks inzetten.
Wat gebeurt er als ik een AI-systeem van een externe leverancier aanpas voor mijn eigen gebruik — word ik dan zelf aanbieder?
Ja, in bepaalde situaties wel. Je wordt juridisch gezien aanbieder als je een substantiële wijziging aanbrengt in een bestaand AI-systeem, je eigen naam of merk op het systeem plaatst, of het beoogde doel van het systeem zodanig aanpast dat het in een hogere risicocategorie terechtkomt. Praktisch voorbeeld: als je een standaard taalmodel van een leverancier herconfigureert om zelfstandig HR-beslissingen te ondersteunen, ben je niet langer alleen gebruiker maar ook aanbieder met alle bijbehorende verplichtingen zoals conformiteitsbeoordeling en technische documentatie. Leg contractueel altijd vast wie welke verantwoordelijkheid draagt.
Hoe richt ik 'menselijk toezicht' op een geautomatiseerd AI-proces in de praktijk in?
Menselijk toezicht betekent niet dat een medewerker elke beslissing van een AI-systeem handmatig moet controleren, maar wel dat er een werkende mechanisme bestaat om in te grijpen wanneer dat nodig is. Concreet kun je denken aan: een goedkeuringsflow voor beslissingen boven een bepaalde drempelwaarde, een dashboard waarop afwijkingen of uitzonderingen zichtbaar zijn voor een verantwoordelijke medewerker, en een duidelijk gedocumenteerde procedure voor het pauzeren of stopzetten van het systeem. Zorg er ook voor dat de medewerker die toezicht houdt voldoende context en bevoegdheid heeft om daadwerkelijk bij te sturen — toezicht op papier zonder echte handelingsruimte telt niet.
Geldt de AI Act ook voor AI-systemen die we intern gebruiken en niet aan klanten aanbieden?
Ja, de AI Act maakt geen onderscheid tussen intern gebruik en extern aanbod. Zodra je een AI-systeem inzet binnen de EU — ook puur voor interne processen zoals HR, financiën of operations — ben je als gebruikersorganisatie gebonden aan de verplichtingen die voor deployers gelden. Als je het systeem zelf hebt ontwikkeld én intern inzet, ben je tegelijkertijd aanbieder én gebruiker en gelden beide sets verplichtingen. Interne AI-tools die hoog-risico taken uitvoeren, zoals geautomatiseerde beoordelingen van medewerkers, vallen dus volledig onder de wet.
Wat is het grootste praktische risico als ik nu nog niets doe ter voorbereiding op de AI Act?
Het grootste risico is niet zozeer de directe boete, maar de operationele verstoring als je vlak voor de deadline van augustus 2026 ontdekt dat een of meerdere van je AI-systemen hoog-risico zijn en nog niet voldoen aan de documentatie- en toezichtsvereisten. Het achteraf inrichten van een kwaliteitsmanagementsysteem, het opstellen van technische documentatie en het organiseren van menselijk toezicht kost aanzienlijk meer tijd en geld dan wanneer je dit meeneemt in de reguliere ontwikkel- en implementatiecyclus. Begin daarom nu met een inventarisatie van je AI-systemen, ook als je vermoedt dat de meeste in de lage risicocategorieën vallen.
Zijn er specifieke sectoren of use cases binnen procesautomatisering waarbij de kans op hoog-risico classificatie extra groot is?
Ja, de kans op een hoog-risico classificatie is aanzienlijk groter in sectoren en use cases die direct raken aan mensenrechten, toegang tot diensten of werkgelegenheid. Denk aan: geautomatiseerde kredietbeoordelingen of fraudedetectie bij financiële instellingen, AI-gestuurde werving- en selectietools in HR, geautomatiseerde triage of prioritering in de zorg, en systemen die bepalen of iemand in aanmerking komt voor overheidsuitkeringen of -diensten. Ook contactcenterautomatisering waarbij AI bepaalt hoe klanten worden gerouted of beoordeeld, kan in de hoog-risico sfeer komen als het gaat om toegang tot essentiële diensten. Een sectorcheck is daarom altijd onderdeel van een goede risicoclassificatie.
