EU AI Act Compliance Checklist for Enterprise Deployers (2026)
Team Sonomos · Sonomos
The Sonomos research team writes about AI privacy, data protection, and how to use generative AI safely at work.
Short answer, checked 18 November 2026: most of the EU AI Act now applies, but the part that weighs most on deployers does not yet. Regulation (EU) 2026/1744, the "Digital Omnibus on AI", was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. It moved the high-risk rules (Chapter III, Sections 1 to 3, which include the Article 26 deployer duties and the Article 27 fundamental rights impact assessment) to 2 December 2027 for stand-alone systems listed in Annex III, such as hiring and credit scoring tools, and to 2 August 2028 for AI built into products covered by the Annex I product-safety legislation.
The rest is live. The Article 5 prohibitions and the Article 4 AI literacy duty have applied since 2 February 2025, the obligations on providers of general-purpose AI (GPAI) models since 2 August 2025, and everything not dated separately since 2 August 2026, including the Article 50 transparency duties.
This checklist is for deployers: organisations that use AI systems built by others in a professional context. Providers carry most of the design and documentation duties; the steps below are what falls to you. Every date and article reference comes from the consolidated text of the AI Act as amended on EUR-Lex.
The timeline after the AI Omnibus
| Date | What applies | Where it is set |
|---|---|---|
| 2 February 2025 | Prohibited practices (Article 5) and AI literacy (Article 4) | Article 113(a) |
| 2 August 2025 | GPAI model obligations (Chapter V), governance, penalties | Article 113(b) |
| 2 August 2026 | General date of application, including Article 50 transparency | Article 113 |
| 2 December 2026 | Article 50(2) marking for generative systems placed on the market before 2 August 2026; new prohibitions on AI-generated non-consensual intimate imagery and child sexual abuse material | Articles 111(4) and 113(a) |
| 2 August 2027 | GPAI models placed on the market before 2 August 2025 must comply | Article 111(3) |
| 2 December 2027 | High-risk rules for Annex III systems | Article 113(c)(i) |
| 2 August 2028 | High-risk rules for AI in Annex I products | Article 113(c)(ii) |
| 2 August 2030 | High-risk systems already in service that are intended for use by public authorities | Article 111(2) |
The European Commission's AI Act overview gives the same Annex III and Annex I dates and the 27 July 2026 entry into force.
What the Omnibus changed for deployers
The dates moved; the duties did not. The Omnibus left Article 26 and the Annex III list of high-risk uses unamended. Its recital 40 gives the reason for the delay: standards and national authorities were not ready.
Systems already in use are caught only if they change. Under Article 111(2), a high-risk system placed on the market or put into service before the Chapter III date falls under the Act only if its design changes significantly from that date. Recital 39 says the test runs per type and model, so later units of an unchanged design already on the market keep the same status. Systems intended for use by public authorities are the exception: they must comply by 2 August 2030. Earlier versions of this guide described a one-year transition to August 2027 for legacy systems; neither the original Act nor the amended text works that way.
AI literacy was softened, not removed. Article 4 now requires providers and deployers to "take measures to support the development of AI literacy" of their staff, without guaranteeing any particular level in any individual.
Two prohibitions were added. From 2 December 2026, Article 5 bans AI systems that generate realistic intimate images of identifiable people without their explicit consent, and systems that generate child sexual abuse material. For a deployer, the ban bites when a system is used for that purpose.
Some paperwork can be shared. Amended Article 27(4) lets the fundamental rights impact assessment cross-reference your GDPR data protection impact assessment. New Article 4a lets deployers of high-risk systems process special categories of personal data where strictly necessary to detect and correct bias, under strict safeguards.
Phase 1: prohibited practices and AI literacy (since February 2025)
There is no compliant version of a prohibited practice, and the Act's highest fines attach to them.
- No social scoring. Article 5(1)(c) bans scoring people on social behaviour or personal characteristics where the score leads to unjustified or out-of-context detrimental treatment. It is not limited to public authorities: behavioural scores from employee monitoring, used against staff in unrelated contexts, are in scope.
- No untargeted facial scraping from the internet or CCTV footage to build recognition databases (Article 5(1)(e)).
- No emotion recognition at work or school, except for medical or safety reasons (Article 5(1)(f)). Affect detection in video interviews and emotion-inferring proctoring tools are the usual cases.
- No biometric categorisation that infers race, political opinions, trade union membership, religion, sex life or sexual orientation (Article 5(1)(g)).
- No manipulation or exploitation of age, disability or social or economic circumstances that materially distorts behaviour and causes significant harm (Article 5(1)(a) and (b)).
- Know the limits of the biometric ban. Article 5(1)(h) bans real-time remote biometric identification in public spaces for law enforcement. A retailer's live facial recognition is not caught by it, but it is an Annex III high-risk use (point 1(a)) and needs an exception under Article 9 of the GDPR.
- AI literacy. Document training matched to each role, the system and the people it affects. An AI acceptable use policy is a reasonable place to record it.
If anything in use falls under Article 5, stop it and record the decision.
Phase 2: general-purpose AI (since August 2025)
GPAI obligations run to model providers. Your job is vendor due diligence, which goes better when you know what you are owed.
- Know which documents reach you. The Article 53(1)(a) technical documentation goes to the AI Office and national authorities on request, not to customers. Article 53(1)(b) information goes to providers who build the model into their own AI systems; if you build on a model API, that is you.
- File what is public. Every GPAI provider must publish a summary of its training content (Article 53(1)(d)) and keep a copyright policy (Article 53(1)(c)).
- Ask about systemic risk. A model trained with more than 10^25 floating-point operations is presumed to carry systemic risk (Article 51(2)), and its provider must evaluate and adversarially test it, mitigate systemic risks, report serious incidents to the AI Office and secure it (Article 55).
- Note the legacy date. Models on the market before 2 August 2025 have until 2 August 2027 (Article 111(3)).
Phase 3: Article 50 transparency (since August 2026)
Article 50 applies whether or not a system is high-risk, so it is the duty most organisations using generative AI actually meet.
- Chatbots you build are probably yours to label. Article 50(1) puts the duty to tell people they are dealing with AI on the provider. A customer-facing assistant you build on a model API and run under your own name likely makes you its provider.
- Deep fakes. Disclose AI-generated or manipulated image, audio or video that is a deep fake (Article 50(4)); for evidently artistic or satirical work, a light-touch disclosure is enough.
- Public-interest text. Disclose AI-generated text published to inform the public on matters of public interest, unless it went through human review and someone holds editorial responsibility (Article 50(4)).
- Emotion recognition and biometric categorisation. Where lawful at all, tell the people exposed (Article 50(3)).
- Marking. Providers of generative systems must mark outputs as machine-readable synthetic content (Article 50(2)); systems on the market before 2 August 2026 have until 2 December 2026.
Phase 4: high-risk systems (from 2 December 2027)
This is where deployers carry the heaviest direct obligations, and the classification and contracting work is worth doing now.
Step 1: decide whether a use is high-risk
Annex III covers biometrics, critical infrastructure, education, employment, essential private and public services, law enforcement, migration and border control, and the administration of justice and democratic processes. The uses most enterprises meet:
- Employment (point 4): recruitment, filtering applications and evaluating candidates; decisions on promotion, termination and terms of work; allocating tasks by behaviour or traits; monitoring and evaluating performance.
- Education (point 3): admissions, grading and monitoring students during tests.
- Essential services (point 5): credit scoring (fraud detection is excluded), risk assessment and pricing for life and health insurance, eligibility for public benefits, emergency call triage.
Article 6(3) excludes a listed system that poses no significant risk, for example one doing a narrow procedural task, but never a system that profiles people.
- Inventory the AI systems in use, including ones adopted without approval.
- Record for each whether it serves an Annex III purpose and whether Article 6(3) applies, with reasons.
- Re-check when the use changes. Drafting a job advert with ChatGPT is not high-risk; ranking applicants with it is.
Step 2: check that you are still only a deployer
Under Article 25(1), you take on the provider's obligations if you put your name or trademark on a high-risk system, substantially modify it, or repurpose a general-purpose system so that it becomes high-risk. A home-built CV-screening workflow on a general chatbot is the usual way deployers cross that line.
Step 3: meet the Article 26 deployer duties
- Use it as instructed by the provider's instructions for use (Article 26(1)).
- Assign human oversight to people with the competence, training, authority and support to do it (Article 26(2)). Rubber-stamping outputs does not meet the standard.
- Check your inputs are relevant and sufficiently representative where you control them (Article 26(4)).
- Monitor and report. If the system presents a risk, inform the provider and the market surveillance authority and suspend use; report serious incidents to the provider first (Article 26(5)). Article 73 allows 15 days from awareness, or two days for a widespread infringement.
- Keep the logs under your control for at least six months unless other law says otherwise (Article 26(6)).
- Tell workers first: inform workers' representatives and affected workers before using a high-risk system at work (Article 26(7)).
- Tell the people affected that a high-risk Annex III system makes or assists decisions about them (Article 26(11)).
- Register if you are a public body, and do not use a system its provider has not registered (Articles 26(8) and 49(3)).
- Feed the DPIA with the provider's Article 13 information (Article 26(9)).
Step 4: fundamental rights impact assessment (Article 27)
Bodies governed by public law, private entities providing public services, and any deployer using credit scoring or life and health insurance pricing systems (Annex III points 5(b) and 5(c)) must assess the impact on fundamental rights before first use and notify the market surveillance authority of the result.
Step 5: get the provider's side in writing
- The instructions for use and Article 13 information on capabilities, limitations, accuracy and oversight.
- The EU declaration of conformity (Article 47) and CE marking (Article 48).
- Confirmation of registration in the EU database (Article 49(1)); critical infrastructure systems are registered nationally instead.
- A commitment to tell you when a release is a significant change in design, since that decides whether a system you already use falls under Article 111(2).
Where the AI Act meets the GDPR
- DPIA. High-risk uses of personal data will usually meet the Article 35 GDPR threshold, which names systematic and extensive automated evaluation of people, including profiling, with significant effects.
- Automated decisions. AI Act oversight sits alongside the limits on solely automated decisions in Article 22 GDPR; see GDPR data subject rights and AI.
- Notices. Article 50 and Article 26(11) notices sit next to the GDPR's Articles 13 and 14 privacy notices.
- Transfers. The AI Act says nothing about where a model runs. That remains a GDPR question, covered in our GDPR and AI guide and, for one provider, Claude EU Data Residency in 2026.
Frequently asked questions
Does the EU AI Act apply to AI tools we use but didn't build?
Yes. The Act puts separate obligations on deployers, meaning anyone using an AI system under its authority in a professional context. AI literacy and the prohibitions apply now, Article 50 applies where you publish deep fakes or public-interest text, and the Article 26 duties apply to high-risk uses from 2 December 2027. Internal use counts: a tool that evaluates your own employees is an Annex III employment use.
We use a third-party HR screening tool. Is that high-risk?
Almost certainly. Filtering applications and evaluating candidates are listed in Annex III point 4, and screening tools usually profile candidates, which rules out the Article 6(3) exception. From 2 December 2027 you will need human oversight, log retention, worker and candidate notices and incident reporting. Our guide to AI in hiring covers the US rules that apply alongside.
When do high-risk obligations apply after the AI Omnibus?
On 2 December 2027 for systems listed in Annex III, and on 2 August 2028 for AI that is, or is a safety component of, a product covered by Annex I legislation and needing third-party conformity assessment. Both dates come from Article 113 as amended by Regulation (EU) 2026/1744; the original Act had set 2 August 2026 and 2 August 2027.
What happens to high-risk systems already in use before December 2027?
Under Article 111(2), they fall under the Act only if their design changes significantly from that date (2 August 2028 for Annex I systems). Systems intended for use by public authorities must comply by 2 August 2030 regardless. There is no general one-year transition for legacy systems.
What are the penalties for non-compliance?
Under Article 99, breaching a prohibition can cost up to €35 million or 7% of worldwide annual turnover, whichever is higher; most other breaches, including of Article 26 and Article 50, up to €15 million or 3%; and misleading information to authorities up to €7.5 million or 1%. For SMEs, and for small mid-caps on the lower two tiers, the cap is whichever is lower. National market surveillance authorities enforce most of the Act; the AI Office supervises GPAI providers.
The bottom line
The Omnibus gave deployers sixteen more months on high-risk systems, not an exemption. Between now and 2 December 2026, confirm nothing in use is prohibited, document AI literacy and review Article 50 duties. Through 2027, classify every Annex III use, check whether Article 25 makes you a provider, collect provider documentation and design oversight, logging and notices. Our EU AI Act glossary entry and NIST AI RMF guide cover the governance frameworks around this work.
Much of what flows into these systems is personal data typed by staff into AI tools. Locke runs on the device and replaces names, account numbers and other sensitive values with placeholders before a prompt leaves it, in the AI tools it screens; it is one technical control in a deployer's programme, not a conformity assessment or a registration tool.
Sonomos publishes this blog and makes Locke, which detects and masks sensitive information on your device before it reaches an AI model. How Locke works
This article is general information about privacy and AI, not legal advice, and reading it does not create a solicitor–client or attorney–client relationship. Regulatory requirements turn on your specific facts and jurisdiction and change over time. Consult qualified counsel before relying on anything here, and check the publication and last-reviewed dates above. We may revise posts, but we do not undertake to keep any post current as the law changes.
Reference pages for this post
Related Articles
Best Privacy-Focused AI in 2026: Chatbots That Keep Your Data Private
The most private AI runs locally, via Ollama or LM Studio. Lumo, Brave Leo and Venice say they keep no chat logs; Duck.ai hides your IP. Compared, Nov 2026.
Can My Employer See My AI Chats? ChatGPT, Claude, Gemini and Copilot
Yes, if your employer controls the account, the computer or the network. Work accounts on ChatGPT, Claude, Gemini and Copilot let it read or keep chats. A side-by-side table of what each vendor documents, as of November 2026.
Can My Employer See My Copilot Chats? Work vs Personal Accounts
Copilot chats on a work account sit in your mailbox, where your employer can search, hold and export them with Purview. A personal account sits outside the tenant, but a work laptop can still capture it. What Microsoft documents, as of November 2026.