What the AI Act requires of a company using generative AI at work
If your staff already use generative AI assistants, two duties apply now. Article 4 AI literacy has applied since 2 February 2025 to every provider and deployer, whether or not the AI system is high-risk. The transparency duties for deepfakes and AI-generated public-interest text apply from the Act's general application date of 2 August 2026, while high-risk duties depend on the use case. Under the 2026 amendment, Annex III high-risk systems move to 2 December 2027. The answer for your organization depends on three things: the use case, whether you act as a deployer or a provider, and whether you modify or rebrand the tool.
Start with the use. Asking an assistant to draft internal notes is a different case from using it to screen job applications or rank workers for promotion. The Annex III point 4 category covers AI used to recruit or select people, including filtering applications, and AI used for promotion, termination, task allocation and monitoring of worker performance. Publishing AI-generated text or deepfake images to the public triggers disclosure under Article 50(4), unless the text has undergone human review or editorial control and someone holds editorial responsibility. Confirm each use against the Annex III list and the Commission's draft classification guidelines before you rely on a conclusion.
Your role can also change. Under Article 25(1), a deployer becomes the provider of a high-risk system if it puts its own name or trademark on the system, makes a substantial modification that keeps it high-risk, or changes the intended purpose of a general-purpose tool so that it becomes high-risk. Your organization owns its controls, policies, training records and evidence of AI literacy. DeployOpen can help plan and prepare systems, controls and documentation for these duties and can test them. Only independent market surveillance authorities, assessors or certification bodies issue findings, reports or certificates.
Which deployer duties apply to general-purpose assistants today
Organizations that roll out ChatGPT-style assistants or other generative AI are deployers under the AI Act, and several duties already apply whether or not any use is high-risk. The four cards below cover the obligations that bite first, who supervises each one, and the records that show you meet them.
AI literacy under Article 4
Article 4 requires providers and deployers of all AI systems to take measures so that staff and others operating AI on their behalf have a sufficient level of AI literacy, judged against their role and the context of use. It has applied since 2 February 2025, and national market surveillance authorities, not the AI Office, supervise it from 2 August 2026. Check that you can show who uses which assistant, what training they received, and when. Keep internal training records, since the Commission notes that relying only on vendor instructions for use is often not enough.
Deepfake disclosure under Article 50(4)
If your staff use generative AI to create or manipulate image, audio or video that resembles real people, places or events, you must disclose that the content is artificially generated or manipulated. For evidently artistic, creative, satirical or fictional works, the duty narrows to disclosing that such content exists, in a way that does not hamper the work. Check your marketing, training and internal communication templates for synthetic media. Record where each synthetic asset was published and what label it carried.
Disclosing AI-generated public-interest text
A deployer that publishes AI-generated or manipulated text to inform the public on matters of public interest must disclose that origin. The disclosure can be skipped only where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it. Review the channels where assistants draft press notices, public guidance or policy explainers. Keep a named editor, the review log and the date of publication for each piece you rely on that exception for.
Fines for operator breaches
Article 99 sets the maximum fine for most operator obligations at EUR 15 million or 3% of worldwide annual turnover, whichever is higher. Prohibited practices carry a higher ceiling, and misleading information given to authorities carries a lower one. Treat the EUR 15 million or 3% figure as the exposure for a missed literacy, disclosure or record-keeping duty. Ask your counsel which national authority would supervise your sector, and keep a dated file that shows each duty was assessed.
How to evidence AI literacy for staff
The key point is that the AI Act does not require a certificate, and it does not prescribe a training format. The Commission's AI literacy questions and answers say so directly, and they add that national market surveillance authorities, not the AI Office, supervise Article 4 from 2 August 2026. Your job as a deployer is to show that the people operating and overseeing your generative AI tools have a sufficient level of understanding for their role. The evidence you keep is what proves that, so the question to answer before anything else is what you can show an inspector, not which course to buy. Confirm the supervisory details against the Commission's AI Act overview and your national authority's guidance.
Instructions for use from the tool vendor are a starting point, but the Commission's Q&A notes that relying on them alone is often not enough. Build a short evidence set instead: internal records of training and other literacy activities (who attended, what was covered, when, and the version of the material), role-based guidance that explains what a marketing writer, a recruiter, a developer or a customer service lead should and should not put into an assistant, and a written acceptable-use rule that names approved tools, prohibited data types and the duty to flag AI-generated content that goes to the public. Record who approved each document and when it was last reviewed. A training log with no link to these rules shows attendance, not understanding.
Literacy should follow exposure. A staff member who only drafts internal meeting notes needs a different level of understanding from someone who uses an assistant to screen job applications, which may touch Annex III point 4 and, if it becomes high-risk, much heavier duties. Map each role to the tools it uses, the decisions it influences and the disclosure duties it triggers, then set a minimum competence for each group. Ask your vendors and your internal owners to confirm which roles touch recruitment, customer-facing content or public-interest publishing, because those roles need the deepest coverage and the most frequent review.
When does a generative AI use become high-risk?
Annex III point 4 is the trigger that matters most here. It covers AI used to recruit or select people, including targeted job adverts and filtering or ranking applications, and AI used to make or inform decisions on promotion, termination, task allocation, and monitoring of workers' performance and behavior. A general-purpose assistant that ranks CVs or scores candidates falls within that wording even though the vendor never designed it for hiring.
- Recruitment screening: using a chatbot or assistant to filter, rank or score applications is the clearest case of an Annex III use.
- Targeted job adverts: AI that selects who sees a job advertisement is covered by the same point.
- Workforce decisions: AI that informs promotion, termination or task allocation sits within the trigger.
- Performance monitoring: AI that monitors worker performance or behavior is covered, even if the output is only advisory.
- Classification: the draft Commission guidelines on classifying high-risk systems, dated 19 May 2026, are the place to check how these uses are read. They are draft, so confirm the final position before relying on them.
The 2 December 2027 date applies to Annex III systems under the 2026 amendment. Ask your legal or HR counsel to confirm which internal tools fall within point 4 before that date.
Drafting emails, summarizing meeting notes, translating documents or producing first versions of reports is ordinary office work. These uses are usually not high-risk, because they do not fall within an Annex III category or an Annex I product context. The main AI Act duty that applies to them is AI literacy under Article 4, which has applied since 2 February 2025 to deployers of all AI systems.
- Low-risk default: routine drafting and summarizing usually falls outside Annex III.
- AI literacy: staff who use these tools need sufficient understanding of them, and records of training help show this.
- Scope creep: if a summary is later used to score people or make employment decisions, the use may change and trigger Annex III.
- Public text: if AI-generated text is published to inform the public on matters of public interest, Article 50(4) disclosure may apply unless human review or editorial control applies.
- Purpose control: a written list of approved uses helps show that office tools stay outside high-risk categories.
Confirm with counsel whether any drafting use touches a regulated decision in your sector before treating it as out of scope.
A customer-facing chatbot is usually not high-risk on its own. Its main AI Act duty is transparency: people interacting with an AI system must be told they are doing so, unless this is obvious from the context. The chatbot becomes high-risk only if it is used for an Annex III purpose, such as assessing eligibility for an essential service or credit, or if it is part of a product covered by Annex I.
- Disclosure: users should know they are talking to an AI system, which is an Article 50 duty.
- Purpose test: a chatbot answering product questions is usually a transparency case, while one scoring creditworthiness may fall within Annex III.
- Synthetic content: generated images, audio or video shown to customers may need deepfake disclosure under Article 50(4).
- Changing purpose: under Article 25(1), changing the intended purpose of a non-high-risk system so it becomes high-risk makes the deployer a provider.
- Documentation: keep the intended purpose statement and the disclosure wording as evidence.
Check the intended purpose in writing and confirm the disclosure text with your legal team before launch.
How to get from inventory to evidence-ready deployer compliance
This sequence turns an informal list of AI tools into a dated, role-specific file that shows which uses fall under Article 5, Annex III and Article 50, what staff were trained on, and which obligations depend on vendors. Each step closes with a record you can hand to an assessor or a regulator.
Inventory every generative AI use and its purpose
Start by listing each place generative AI touches your business: chat assistants, coding tools, document drafting, image and video generation, customer-facing chatbots, and AI features built into software you already buy. For each entry, record the owning team, the vendor or model, whether staff or customers interact with it, and the specific purpose it serves. Ask department heads for entries they know about, and check expense reports and software approvals for tools added without central review. The output is a single AI system inventory with one row per use. The step is done when every department has signed off on its rows.
Classify each use against Article 5, Annex III and Article 50
Test each inventory row in a fixed order. First, confirm the use does not fall under a prohibited practice in Article 5. Second, check whether the purpose matches an Annex III category, such as AI used to recruit or select people or to make decisions on promotion, termination, task allocation or worker monitoring. Third, note whether the output is a deepfake or AI-generated text published to inform the public, which triggers Article 50 disclosure duties. Record the reasoning for each row, because an assessor will ask why a use was not classed as high-risk. Confirm borderline cases with legal counsel and check the Commission's draft classification guidelines dated 19 May 2026. The output is a classification register with a risk label and a reason for every use.
Set AI literacy records and disclosure rules
Article 4 applies to deployers of every AI system, regardless of risk level, so each team that operates a tool needs a documented level of AI literacy suited to its role and the systems it uses. The Commission states that no certificate is required and that no training format is mandatory, so choose a format your organization can repeat and evidence. Keep internal records of who was trained, on which tools, when, and what they covered. Pair the training with written disclosure rules that tell staff when AI-generated content must be labeled, including deepfakes and public-interest text under Article 50(4). The output is a training log, a role-based literacy outline and a one-page disclosure standard approved by a named owner.
Confirm whether any use changes your provider role
Under Article 25(1), a deployer becomes the provider of a high-risk system if it puts its own name or trademark on the system, makes a substantial modification that keeps it high-risk, or changes the intended purpose of a non-high-risk system, including a general-purpose AI system, so that it becomes high-risk. Review each row flagged as high-risk or heading that way, and each case where a team has configured a model for a new business purpose, rebranded it for customers or built an internal product on top of it. Where a role change is possible, stop further changes and escalate. The output is a short role memo for each affected use stating whether you remain a deployer or have taken on provider duties.
Collect vendor documentation and identify gaps
For each vendor or model in the inventory, request the documentation it supplies: instructions for use, intended purpose statements, known limitations, disclosure features, data handling terms and any technical information the vendor provides to downstream users. Note which documents you have, which are missing and which vendor statements are general marketing claims rather than usable instructions. The Commission's AI literacy guidance notes that relying only on instructions for use is often not enough, so pair vendor material with your own records of how the tool is used. The output is a vendor evidence file per supplier with a gap list that names the owner who will chase each missing item.
Keep internal records that show ongoing compliance
Assemble the evidence that proves the earlier steps happened and stay current. Bundle the inventory, classification register, literacy log, disclosure standard, role memos and vendor files into one controlled repository with version history and a named owner. Set a review trigger so that new tools, changed purposes or vendor updates send the affected rows back through classification. Plan around the dates that matter: the general application date of 2 August 2026 and the 2 December 2027 date for Annex III high-risk rules, as amended in 2026. The output is a dated evidence pack that an internal auditor or market surveillance authority can read without a briefing from the team that built it.
Who owns each deployer duty and who can help
This table maps each deployer duty under the AI Act to the person inside the company who should own it, what an implementation partner can prepare, and what only an independent authority or assessor decides. Read it as a split of work: the organization owns the duty and its decisions, a partner prepares systems and records, and regulators or assessors make any formal finding. DeployOpen prepares evidence but does not issue reports or certificates, and Article 4 requires no certification.
| Deployer duty | Owner inside the company | What an implementation partner can prepare | What only an authority or assessor decides |
|---|---|---|---|
| AI literacy (Article 4) | HR or learning lead with the AI governance owner | Role-based training outline, attendance and completion records, refresher schedule | Whether training is sufficient, enforced by national market surveillance authorities |
| AI system inventory | Head of IT or procurement, with business unit owners | Discovery of AI tools in use, including shadow tools, and a register template | Nothing formal; the inventory is the organization's own record used in any inquiry |
| Deepfake and synthetic content disclosure (Article 50(4)) | Communications or marketing lead | Disclosure wording templates, labelling workflow, content tagging steps | Whether a given item qualifies as a deepfake or a regulator's view of a disclosure |
| Public-interest text disclosure (Article 50(4)) | Editorial or publications lead | Review and sign-off workflow showing human review or editorial responsibility | Whether the editorial control applied is sufficient in a particular case |
| High-risk classification of a use (Annex III point 4 for HR) | Legal or compliance lead, with HR as process owner | Use-case assessment questionnaire, decision record, logging design for any high-risk use | The final classification in a disputed case, by national authorities or courts |
| Role change to provider (Article 25(1)) | Legal counsel and product owner | Map of rebranding, substantial modification or purpose changes that trigger provider duties | Whether a modification meets the legal test, decided by authorities if challenged |
| Conformity assessment and certificates | Quality or compliance owner of the high-risk system | Quality management documentation and technical file preparation | Conformity assessment and any certificate, by an assessor or notified body where required |
Questions buyers ask about AI Act deployer duties
Do we need a certificate for AI literacy under the AI Act?
No. Article 4 requires providers and deployers to take measures to ensure a sufficient level of AI literacy among their staff, but the Commission's AI literacy Q&A states that no certificate is required and there is no mandatory training format. What matters is that you can show what you did. Organizations can keep internal records of training and other AI literacy initiatives, and the Commission notes that relying only on a vendor's instructions for use is often not enough. Your records should reflect the roles, the tools in use and the risks each group faces. Ask your legal or compliance team to confirm how the duty applies to your specific mix of tools and staff, since the answer depends on your context.
Who enforces Article 4 on AI literacy?
National market surveillance authorities enforce Article 4, not the Commission's AI Office. The Commission's AI literacy Q&A states that these national authorities supervise the duty as of 2 August 2026. Article 4 has applied since 2 February 2025 and covers providers and deployers of all AI systems, not only high-risk ones. Because enforcement sits with national authorities, the authority in each Member State where you operate or where your output is used is the one to watch. Confirm which authority applies to your organization with your counsel, as national arrangements can differ.
Does using ChatGPT or a similar assistant make us a provider?
Usually no. Under Article 3(4), a deployer is anyone using an AI system under its authority, except in a personal non-professional activity, so using a commercial assistant at work makes you a deployer. Article 25(1) can change that role. If you put your name or trademark on a high-risk system, make a substantial modification that keeps it high-risk, or change the intended purpose of a general-purpose AI system so it becomes high-risk, you become its provider. For example, an illustrative internal tool that screens job applications built on a general-purpose assistant could cross that line. Review each new use case against that test before rollout.
When must we label AI-generated text?
You must disclose AI-generated or manipulated text only when you publish it to inform the public on matters of public interest, and the text has not undergone human review or editorial control with someone holding editorial responsibility (Article 50(4)). Deepfake images, audio and video carry their own disclosure rule under the same paragraph. For evidently artistic, creative, satirical or fictional works, disclosure is limited to the existence of such content without hampering enjoyment. Internal drafts that never reach the public are generally outside this duty. Document who reviews and approves public-facing content so you can show the editorial control exists.
What do fines look like under Article 99?
Article 99 sets the maximum fines by type of breach, and each cap is the higher of a fixed amount or a share of worldwide annual turnover. Prohibited practices carry up to EUR 35 million or 7%. Most operator obligations, which include the deployer duties discussed here, carry up to EUR 15 million or 3%. Misleading information given to authorities carries up to EUR 7.5 million or 1%. These are maximums, and actual penalties depend on the authority, the severity and the cooperation shown. The figures come from the Act as researched, so confirm the current consolidated text before relying on them in a board paper.
Can we do this ourselves, and what does outside help cost?
Many organizations handle deployer duties in-house, starting with an inventory of AI tools, a training record and a review step for public content. Outside help is most useful when a use may become high-risk, when you might become a provider, or when you need the evidence set organized for an assessment. Any external party's cost follows from a written scope that lists the systems, the number of use cases, the deliverables and the evidence required. Ask for the scope, the roles each party owns, and which items recur, such as training refreshes. DeployOpen plans, implements and documents the technical side of self-hosted systems and prepares evidence, but it does not issue compliance reports or certifications, and the quote follows from the written scope.
What to ask for in a quote for AI Act deployer readiness
Ask for a written, fixed scope before anything else. DeployOpen does not publish a price list. A quote follows from the scope agreed in writing, so the quality of that scope decides how comparable your quotes will be. Ask every provider to name the generative AI systems covered: each assistant, API integration, embedded feature in SaaS tools, and any self-hosted model your staff use at work. Ask whether the scope includes personal accounts used for work tasks, and how the provider will find out about tools staff adopt without approval. A quote that does not list the systems is not yet a quote you can compare.
Ask for the deliverables by name and by format. For deployer readiness, the core items are an AI system inventory showing each tool, its owner and its business use; a record of AI literacy initiatives for the staff who operate these tools, since the Commission's Q&A says organizations can keep internal training records and that relying only on instructions for use is often not enough; and a disclosure procedure covering deepfakes and AI-generated text published to the public under Article 50(4). Also ask whether the provider will test whether a use case could fall into Annex III, such as AI that filters job applications, and what the deliverable says if it does. Confirm that each item is delivered as an editable document or record your team keeps.
Ask whether ongoing support is a separate retainer or included, and what it covers: inventory updates when a new tool is adopted, refresher training, and review when a vendor changes its terms or a use changes. Then ask what is excluded. Legal opinions on how the Act applies to your sector, fees charged by any independent assessor or certification body, and the internal time your own staff spend are normally outside an implementation partner's work, so confirm each one explicitly. Only independent auditors and certification bodies issue reports or certificates. Your legal counsel interprets the law. The partner prepares systems and evidence, and your organization owns the decisions.
How DeployOpen prepares your AI Act deployer evidence
DeployOpen works on the deployer side of generative AI use: the AI literacy records, staff usage rules, disclosure workflows and high-risk screening that a company needs to show. We prepare systems and evidence inside your own infrastructure, while your organization owns its compliance decisions and an independent assessor or authority judges them.
How we scope
- Engagements are fixed-scope, with scope, deliverables and timelines agreed in writing before work starts.
- The written scope lists which generative AI tools and self-hosted models are in use, and which roles and departments they touch.
- We confirm with your legal and compliance leads which uses could fall under Annex III or Article 50 before any build begins.
How we build
- We set up a register of AI systems in use, with owner, purpose, data types and deployment location for each entry.
- We draft AI literacy training records, usage policies and acceptable-use rules that match the roles of your staff.
- We configure disclosure workflows for AI-generated images, audio, video and public-interest text, with a named human reviewer step.
How we hand over
- You receive the inventory, policies, training records and disclosure procedures as documents your team can maintain.
- Logs and test evidence from self-hosted tools are documented with retention settings, so an assessor can trace how they were produced.
- We walk your compliance and IT owners through each artifact, and list the items that only your legal counsel or an auditor can confirm.
How we support
- Ongoing support, upgrades and periodic reviews are available as a retainer after the initial handover.
- We help you reassess the register when a new tool is adopted or a use changes, including changes that could alter a deployer's role.
- Confirm with DeployOpen during scoping which dates and obligations apply to your organization, since the 2026 amendments changed several timelines.
