The EU AI Act for Mid-Sized Companies: What Deployers Must Do
Most mid-sized companies do not build AI. They buy it and use it: a chat assistant for office work, an agent that drafts quotes, a scoring feature inside the HR software. The EU AI Act calls this role the "deployer". This guide sorts deployer obligations by role and risk class, reflects the 2026 Digital Omnibus amendments (status: October 2026) and is orientation, not legal advice.
What is the EU AI Act and who counts as a deployer?
The AI Act is Regulation (EU) 2024/1689. It entered into force on 1 August 2024 and applies in stages. It is not only aimed at vendors: any organisation that uses an AI system under its own authority in a professional context is a deployer. Purely personal, non-professional use is the only exception.
Article 3 separates two main roles:
- A provider develops an AI system, or has one developed, and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge.
- A deployer uses an AI system under its authority.
For a company with 50 to 500 employees this usually means: for bought-in tools such as Microsoft Copilot, ChatGPT Enterprise or an AI feature in your ERP, you are the deployer. Technical documentation, conformity assessment and similar duties sit with the vendor. Your own duties depend on what you use the system for.
That is the key point of this guide. The technology does not determine your obligations; the use case does. The same language model can trigger almost nothing when it drafts a sales quote and become a high-risk system when it shortlists job applicants.
How is the AI Act structured? The four risk levels
The Act sorts AI uses into four levels: prohibited practices, high-risk systems, systems with transparency duties, and minimal-risk systems. The higher the level, the more obligations apply. On top of that, deployers have one cross-cutting duty: supporting AI literacy among staff.
The European Commission summarises the levels as follows:
- Unacceptable risk: banned, for example social scoring or untargeted scraping of facial images to build recognition databases.
- High risk: uses that can seriously affect health, safety or fundamental rights, such as recruitment tools or credit scoring.
- Transparency risk: systems where people need to know AI is involved, such as chatbots and deepfakes.
- Minimal or no risk: no additional AI Act rules, for example spam filters.
In practice a deployer works through the levels in order. Is the use prohibited? If not, does it fall into one of the high-risk areas? If not, do transparency duties apply? Anything left is minimal risk. A separate guide on risk classes covers that classification step by step.
Which obligations apply by role and risk level?
Deployer duties fall into four groups. Prohibitions and AI literacy already apply. Transparency duties apply from 2 August 2026. High-risk duties under Article 26 apply from 2 December 2027. And if you build or substantially change a system yourself, you may take on provider obligations as well.
Prohibited practices (since February 2025)
Since 2 February 2025, Article 5 bans a short list of practices, including manipulative techniques that cause significant harm, social scoring and inferring the emotions of people at work or in education, unless the purpose is medical or safety-related. The Omnibus adds a ban on AI systems that generate non-consensual intimate imagery or child sexual abuse material, applying from 2 December 2026 according to the Omnibus text. Few ordinary businesses come near these lines, but the workplace emotion ban deserves a check if you run call-centre or video-meeting analytics.
AI literacy (since February 2025, softened)
Article 4 has applied to providers and deployers since 2 February 2025. The Omnibus reworded it: rather than ensuring a "sufficient level" of AI literacy, organisations must now take measures that support the development of AI literacy among their staff, as Gibson Dunn explains. No specific level per individual is required any more. What that looks like in practice is covered in AI literacy under Article 4.
Transparency duties (from 2 August 2026)
Article 50 splits transparency duties between providers and deployers. As a deployer you must:
- tell people when they are exposed to an emotion recognition or biometric categorisation system;
- disclose that image, audio or video content has been artificially generated or manipulated when it would otherwise appear authentic (deepfakes), with a lighter touch for evidently artistic, satirical or fictional work;
- disclose that text is AI-generated when it is published to inform the public on matters of public interest, unless a human has reviewed it and a person or organisation holds editorial responsibility.
Making sure users know they are talking to an AI is primarily the provider's job under paragraph 1. If you build your own website chatbot, you are normally that provider yourself. The disclosure has to be clear at the latest at the first interaction.
High-risk deployer duties (from 2 December 2027)
If you deploy a high-risk system listed in Annex III, Article 26 applies. For most mid-sized firms the relevant area is employment: AI for recruiting and candidate selection, and AI that informs decisions on promotion, termination, task allocation based on personal traits, or performance monitoring. The core duties:
- Follow the instructions for use, backed by appropriate technical and organisational measures.
- Assign human oversight to people with the competence, training, authority and support to do it.
- Input data you control must be relevant and sufficiently representative for the purpose.
- Monitor operation; if you see a risk, inform the provider and the market surveillance authority and suspend use; report serious incidents.
- Keep logs under your control for at least six months.
- Inform workers and their representatives before using a high-risk system in the workplace.
- Inform affected people when an Annex III system makes, or helps make, decisions about them.
- Run a data protection impact assessment under Article 35 GDPR where required, using the provider's information.
A fundamental rights impact assessment under Article 27 applies only to certain deployers: public bodies, private entities providing public services, and deployers of systems for credit scoring of individuals or for risk assessment and pricing in life and health insurance. After the Omnibus it can build on an existing data protection impact assessment.
When you become a provider
Under Article 25, a deployer takes on the provider obligations for a high-risk system if it puts its own name or trademark on one, makes a substantial modification to one, or changes the intended purpose of a system so that it becomes high-risk. That last case is easy to trigger: a general-purpose assistant that someone repurposes with a prompt to rank job applicants. A separate guide covers the provider-versus-deployer question for in-house AI agents.
What did the Digital Omnibus change?
The Digital Omnibus, Regulation (EU) 2026/1744, entered into force on 27 July 2026. It postpones the high-risk deadlines, softens the AI literacy duty, adds a new prohibition and simplifies documentation for smaller firms. The risk-based architecture of the AI Act is unchanged.
Changes that matter to deployers, based on the Commission, the Omnibus text and the Cloud Security Alliance's note on the original dates:
| Topic | Before | After the Omnibus |
|---|---|---|
| High-risk, Annex III | 2 August 2026 | 2 December 2027 |
| High-risk in products (Annex I) | August 2027 | 2 August 2028 |
| AI literacy (Art. 4) | ensure a sufficient level | take supporting measures |
| Transparency (Art. 50) | 2 August 2026 | 2 August 2026; marking grace period for existing generative systems until 2 December 2026 |
| New ban on intimate deepfakes | – | from 2 December 2026 |
The Article 50 grace period only covers the machine-readable marking that providers of generative systems must build in. It does not delay deployer disclosure duties. The Omnibus also extends simplified documentation and a lower-of-the-two fine cap to small mid-cap companies, mostly relevant to providers. A separate guide sets out the full AI Act timeline.
What do you need in place?
Three things: an overview of which AI systems are in use, a named person who owns the classification, and a working line to data protection and employee representatives. Mid-sized firms rarely need a dedicated compliance team for this, but they do need clear ownership.
An inventory. Many AI features arrive quietly through software updates, such as a summary button in the CRM or a candidate ranking in the applicant tracking system. Without a list, classification is guesswork.
Ownership. Someone has to decide whether a use falls under Annex III and record why. In many companies this sits with the IT lead, the data protection officer or an AI lead working with senior management. How such a role can be set up part-time is described on the page about a fractional Head of AI.
Interfaces with other law. The AI Act sits alongside the GDPR rather than replacing it. Where personal data is processed, your data protection officer stays involved; the guide on GDPR and AI agents covers what that means in practice. Where works councils or similar employee bodies exist, national rules on workplace monitoring may apply anyway; Article 26 adds an information duty for high-risk systems on top.
Vendor contracts. For high-risk systems you need the provider's instructions for use and the information the Act requires them to supply. Ask for it during procurement, not after go-live.
How do you approach it in five steps?
A sensible sequence starts with what you already have and ends with a routine: list, classify, clarify roles, define measures, review. For most companies the effort stays modest as long as nothing falls into a high-risk area.
- List your AI systems. Record every tool with an AI feature: name, vendor, purpose, department, data processed, users. Ask departments directly instead of relying on procurement records alone.
- Classify each use. Prohibited, high-risk under Annex III, transparency duty under Article 50, or minimal risk. Write two or three sentences explaining the decision.
- Clarify your role. Are you deploying a bought-in system, or have you built, rebranded, substantially modified or repurposed something? If so, check Article 25 and the provider obligations.
- Define measures. For every system: AI literacy measures and an acceptable-use policy. For transparency cases: labelling rules for deepfakes and published text. For high-risk systems: name the oversight person, set log retention, inform staff, check whether a data protection impact assessment is needed.
- Review regularly. Require that new AI features pass the same check before use, and review the inventory at least once a year.
In our own operations, starting the inventory as a plain spreadsheet that grows with every new tool has worked better than starting with a legal document. The legal assessment builds on top of it.
What are the limits and risks?
The biggest risk is not an exotic edge case but misclassification: an HR tool assumed to be harmless, or an in-house agent that quietly turns into a high-risk system. On top of that, guidance and harmonised standards are still incomplete, so some questions remain open.
Fines. Under Article 99, breaches of deployer duties in Article 26 and transparency duties in Article 50 can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher; for prohibited practices up to EUR 35 million or 7%. For SMEs, the lower of the two figures applies.
Supervision. Who supervises you depends on the member state you operate in. In Germany, the Federal Network Agency (Bundesnetzagentur) becomes the central market surveillance authority and complaints point; according to Otto Schmidt, the Bundesrat approved the implementing act on 10 July 2026, and the agency must also run at least one regulatory sandbox aimed mainly at smaller firms.
Open questions of interpretation. When a change counts as "substantial", where task allocation based on personal traits begins, and how much human review exempts text from labelling all depend on the case. Commission guidelines and case law will sharpen these lines over time.
What this guide does not do. It is no substitute for legal advice or for your data protection officer's review. In HR, credit and insurance decisions, and anything touching critical infrastructure, get specialist input.
False comfort from the delay. High-risk duties starting in December 2027 does not mean nothing applies until then. Data protection law, employment law and the parts of the AI Act already in force apply today.
How do you get started?
Start with the inventory and one question: does any system evaluate, select or monitor people? If the answer is no for every system, focus on AI literacy and transparency. If it is yes for any of them, plan for the high-risk duties now rather than in autumn 2027.
A pragmatic four-week plan:
- Week 1: inventory of all AI tools through a short survey of departments.
- Week 2: classify each entry into the four risk levels, with a short rationale.
- Week 3: agree an acceptable-use policy and AI literacy measures; add labelling rules for published content.
- Week 4: for possible high-risk cases, consult legal counsel, data protection and employee representatives.
If you plan to build your own agents rather than only use them, settle the role question early and let it shape your architecture. For running AI under your own control, see sovereign AI; for an overview of agents in business processes, see AI agents.
Frequently asked questions
Does the AI Act apply if staff only use ChatGPT, Copilot or Claude for office work?
Yes. Any professional use of an AI system under your authority makes you a deployer. For ordinary office use the duties are modest: support AI literacy, stay clear of prohibited practices and label deepfakes or AI-written public-interest text where required.
Did the Digital Omnibus remove obligations for deployers?
No. It postponed the high-risk deadlines and softened the AI literacy duty, but kept the risk-based structure of the Act. Transparency duties still apply from 2 August 2026.
Who supervises the AI Act?
Each member state designates market surveillance authorities. In Germany the Federal Network Agency (Bundesnetzagentur) takes the central role under the national implementation act approved by the Bundesrat in July 2026.
When does building our own AI agent make us a provider?
If you develop an AI system, or have one developed, and put it into service under your own name, you are its provider by definition. For high-risk systems you also become the provider if you substantially modify someone else's system or repurpose it for a high-risk use.
Is this article legal advice?
No. It is an orientation guide. For classifying specific systems, particularly in HR, involve your legal adviser or data protection officer.
Sources
- European Commission: AI Act – Regulatory framework, timeline and risk levels digital-strategy.ec.europa.eu
- Digital Omnibus on AI, Regulation (EU) 2026/1744 (text and recitals) artificialintelligenceact.eu
- AI Act, Article 3: Definitions (provider, deployer) artificialintelligenceact.eu
- AI Act, Article 5: Prohibited AI practices artificialintelligenceact.eu
- AI Act, Article 26: Obligations of deployers of high-risk AI systems artificialintelligenceact.eu
- AI Act, Article 27: Fundamental rights impact assessment artificialintelligenceact.eu
- AI Act, Article 25: Responsibilities along the AI value chain artificialintelligenceact.eu
- AI Act, Article 50: Transparency obligations artificialintelligenceact.eu
- AI Act, Article 99: Penalties artificialintelligenceact.eu
- AI Act, Annex III: High-risk AI systems artificialintelligenceact.eu
- Gibson Dunn: EU AI Act Omnibus Agreement – Postponed High-Risk Deadlines and Other Key Changes gibsondunn.com
- Cloud Security Alliance: EU AI Act High-Risk Deadline – Deferred, Not Cancelled labs.cloudsecurityalliance.org
- Otto Schmidt: Bundesrat approves German AI supervision act (13 July 2026) otto-schmidt.de