The short answer: The July amendment to the AI Act pushed back the date from which the rules for high-risk systems apply — to 2 December 2027 for Annex III and to 2 August 2028 for Annex I. The obligation to take measures supporting the development of AI literacy under Article 4 was not postponed and has applied since 2 February 2025. It does not concern technology companies alone: it affects providers and any company whose people use AI systems at work. The regulation does not ask you for a certificate or for any particular document — it asks for measures adapted to roles and to the context of use. Rules for data, review of outputs, and role-based training are our recommended practical framework for putting such measures into shape and documenting them.
Regulation (EU) 2024/1689, commonly referred to as the AI Act, entered into force on 1 August 2024 and its obligations are being phased in. In July 2026 the timetable changed — some of the deadlines moved by one to two years. Inside companies, that news was translated as „so we can relax until 2027." That is a misconception worth unpacking, because the AI literacy measures did not move by a single day.
This text describes the situation as of 31 July 2026.
What was postponed in July 2026 and what was not
The postponement came with Regulation (EU) 2026/1744, which the European Commission put forward in November 2025 as part of the „digital omnibus." It entered into force on 27 July 2026 and, alongside the deadlines, it also amended the wording of Article 4.
| Date | What starts (or has already started) to apply | Typical example |
|---|---|---|
| 2 February 2025 | Prohibitions under Article 5 and AI literacy under Article 4 | Employees writing quotations in ChatGPT |
| 2 August 2025 | Rules for general-purpose models (GPAI), governance and the penalty framework | Providers of large language models |
| 2 August 2026 | The remainder of the regulation, including transparency under Article 50 | Labelling AI content, a chatbot that admits what it is |
| 2 December 2027 | High-risk systems under Annex III | Recruitment, performance evaluation, access to education |
| 2 August 2028 | High-risk as a safety component of a product (Annex I) | AI in a machine, a medical device, a lift |
The postponement therefore does not apply to the whole of the AI Act. It did not touch Article 4, it did not touch the prohibitions under Article 5 (social scoring or emotion inference in the workplace, for example), and it has no bearing whatsoever on legislation that applies independently of the AI Act — the GDPR, employment law, protection of trade secrets, and contractual commitments towards customers.
In practice, that last group is the most expensive one. When an employee pastes a client list into a public AI service, there is no need to wait for the AI Act — the GDPR has a penalty ceiling of 20 million euros or 4 % of worldwide turnover and has been in effect since 2018.
Does Article 4 also apply to a company of five people
Yes. The AI Act does not distinguish by company size, but by role. Most ordinary companies are a „deployer" of an AI system — in other words, whoever uses someone else’s tool in the course of their activity. An accounting practice, an e-shop, a design office, and a service company all fall into the same category as a corporation.
Article 4 is also not aimed solely at permanent employees. It speaks of staff and other persons dealing with the operation and use of AI systems on the company’s behalf — which includes contractors, temporary workers, and suppliers working in your name.
The obligation typically arises when people use AI for:
- writing quotations, contractual documents, and marketing copy,
- translating and summarising documents,
- analysing spreadsheets and customer data,
- responding to customers and transcribing meetings,
- screening CVs,
- generating source code.
In its interpretative Q&A on AI literacy from the spring of 2025, the European Commission expressly states that Article 4 also applies to a company whose people use generative AI for, say, producing advertising copy or a translation. Human review at the end does not relieve you of the obligation — the person doing the checking has to know exactly what to verify and against what.
The rest of this article focuses on the timetable and the impact of the July amendment. A separate, practical guide on how to describe and record the measures can be found in the article AI literacy under Article 4 of the AI Act in practice.
What Article 4 asks of you: measures, not a certificate
The AI Act does not prescribe a uniform training course, a final test, or a certificate. Nor does it prescribe a number of hours. Following the July amendment, Article 4 requires measures supporting the development of AI literacy, taking into account people’s knowledge, experience, education and training as well as the context of use. It expressly does not require the company to guarantee a specific or „sufficient" level of AI literacy for every individual.
That soft wording is deliberate. The regulation determines neither the form of evidence nor any particular documents — the following list is therefore not an enumeration of statutory requirements, but a recommended practical framework for describing, maintaining and, where needed, documenting the measures adopted. Its purpose is to demonstrate a system, not an individual’s knowledge:
- A list of the AI tools in use. Name, purpose, who uses it, what data flows into it, who is responsible for it.
- Rules for entering data. An employee has to know in advance whether they may enter personal data, a contract, a non-public price list, source code, or medical records. They must not be working it out at the moment they have the chat window open.
- Rules for reviewing outputs. Which texts, calculations, and decisions must pass through a human before use, and exactly what they are verified against.
- Role-based training. Marketing, HR, a technician, and a managing director each face different risks. One shared webinar for everyone is a starting point, not the goal.
- A record. The content of the training, an attendance sheet, the internal instruction, the date it was last updated.
- An incident procedure. Who is notified about a data leak, a faulty output, or the use of an unapproved tool.
None of those documents is a mandatory requirement of the AI Act, and none of them on its own guarantees compliance. Together, however, they show something rather different from a „use AI carefully" instruction in an email from last year. The Commission also maintains a public repository of AI literacy practices under the AI Pact — useful inspiration, not a standard.
What an employee should actually be able to do after training
AI literacy is not a course in writing clever prompts. It is knowledge of the specific risk of a specific tool. Three examples that people remember better than definitions:
The model makes things up, confidently. In Mata v. Avianca (New York, 2023), lawyers filed a submission with the court citing judgments that ChatGPT had generated and that did not exist. The court fined them 5,000 dollars. Verifying figures, citations, and legal assertions is not a formality — a hallucination looks exactly as trustworthy as the truth.
The company is liable for what a chatbot says. In Moffatt v. Air Canada (2024), the airline’s chatbot gave a customer incorrect information about a discount. The airline argued that the chatbot was a separate entity responsible for its own statements. The tribunal rejected that and awarded compensation.
Deepfakes are no longer theory. In 2024, an employee at the Hong Kong branch of the engineering firm Arup transferred roughly 25 million dollars following a video conference attended by deepfake colleagues, including the „chief financial officer." Training also has to cover the fact that video and voice have ceased to be proof of identity and that a payment process needs verification through a second channel.
Alongside that comes the duller but equally important part — configuring the tools:
- ChatGPT in a personal workspace (Free, Plus, and Pro alike) may use new conversations to improve the models until the user turns off the „Improve the model for everyone" option. With Business, Enterprise, Edu, and the API, inputs and outputs are not used for training by default. An employee with their own personal account is a channel that is hard for a company to govern: the company has no central account settings, no audit trail, and no assurance that the person has configured the data options correctly.
- Microsoft 365 Copilot keeps data inside the tenant, but makes available to the user everything they have permission to see. The typical problem on deployment is therefore not the model, but years of accumulated oversharing in SharePoint — Copilot will expose it within a week. What belongs before deployment is a permissions review, not training.
- A paid plan is not a permission slip. Plenty of people automatically equate „we pay for it„ with „we may put anything in there." In 2023, Samsung banned generative AI on company devices after employees pasted internal source code into ChatGPT.
- The origin of a tool matters. In 2025 NÚKIB, the Czech national cyber security agency, issued a warning about DeepSeek tools; what that means for Czech companies is discussed in the article NÚKIB has issued a warning about DeepSeek. For workloads that must not leave your network at all, the answer is on-premises deployment — see AI without the cloud.
How to tell whether you have a high-risk system
What decides it is neither the brand nor the fact that the supplier uses a large language model under the hood. What decides it is the purpose and the way it is deployed. Drafting an email in Copilot will not be high-risk. The same technology used to score job applicants may well fall under Annex III.
Pay attention if AI:
- selects, ranks, or filters out applicants,
- evaluates work performance, allocates tasks, or influences the termination of employment,
- decides on access to education or to an essential service,
- works with biometric data,
- is a safety component of a regulated product.
Amazon demonstrated that this is no academic worry: according to a 2018 Reuters report, an internal CV-screening tool systematically penalised CVs containing the word „women’s," because it had learned from ten-year-old data from a field that hired predominantly men. The tool ended up in the bin. That is exactly the type of system Annex III describes.
The postponement to 2027 and 2028 is time to prepare, not confirmation that you may use such a system without rules in the meantime. Discriminatory impact is already dealt with today by anti-discrimination law and employment law.
What the preparation costs and what to check with a supplier
The law sets no minimum length or price for training. The cost follows the number of job roles and the riskiness of the use, not the number of heads — a twenty-person company with three types of AI usage is a cheaper case than an eight-person company that simultaneously does recruitment, marketing, and development.
As a working framework (not a statutory minimum), a small company is well served by a shared block of 2–4 hours plus a shorter follow-up part for the roles that work with personal data, customers, recruitment, or automated decision-making.
Before ordering, ask whether the price includes:
- an analysis of the tools the company actually uses (including the unapproved ones),
- training on your own work situations, not generic examples,
- rules for data and for verifying outputs, in writing,
- material for new joiners,
- a record of the content and of attendance,
- a recommendation on when to update the training.
A generic webinar with no knowledge of your environment will raise awareness, but it will not produce rules you can lean on in practice. If you do not have the capacity to put this together yourselves, we do it as a service — an inventory of tools, the rules, and AI training for companies based on your real cases, with a record that stands up as evidence. The funding options are summarised in our overview of subsidies for AI training in 2026.
What actually happens if you do nothing
This calls for precision, because a great deal of scaremongering circulates around it. The catalogue of fines in Article 99 of the AI Act does not lay down a separate automatic rate in millions of euros for a breach of Article 4. That does not mean it is a voluntary recommendation, however. According to the current interpretative Q&A of the European Commission, national market surveillance authorities may impose proportionate penalties and other corrective measures under national law; the supervisory and enforcement rules apply from August 2026. The specific Czech authority, the procedure, and the penalty therefore need to be verified against the rules in effect at the time of the inspection.
The practical risk, though, arises earlier and elsewhere:
- an employee enters non-public data into an unauthorised service — that is a matter of the GDPR and trade secrets, not of the AI Act,
- the company sends a customer invented information and is liable for it (see Air Canada),
- an AI output becomes the basis for a personnel decision and the company cannot justify it,
- in any dispute or inspection, the position of a company that has not trained anyone and has no rules is markedly weaker than that of a company holding a dated internal instruction and demonstrable attendance at training.
A certificate reading „AI Act compliant" from the tool’s supplier changes none of that. Responsibility for how it is used within your own company cannot be transferred to the manufacturer.
A checklist for the business owner
This is not an enumeration of statutory requirements — it is a practical route to arriving at proportionate measures:
- Write down the AI tools the company has approved.
- Find out what people are using without approval — as a rule, it differs.
- For each tool, describe the purpose, the data, and the person responsible.
- Separate ordinary work support from recruitment, evaluation of people, and other sensitive decisions.
- Set out in writing what must not be entered into public AI services.
- Train people by role and keep both the content and the attendance.
- Review permissions in SharePoint and on shared drives before you let Copilot loose.
- Repeat the review whenever a new tool is introduced or a process changes.
Frequently asked questions
Does every company have to provide mandatory AI training?
If a company provides or uses AI systems in the course of its activity, it must adopt proportionate AI literacy measures. The regulation does not prescribe the format of a course, its length, or a final examination — it prescribes that the measures must match the risks and the context of use.
Does the AI Act also cover the use of ChatGPT or Copilot?
Yes. If employees use the tool on the company’s behalf — for marketing, translations, data analysis, or communicating with customers — the company must adopt proportionate measures for their AI literacy. It chooses the specific form itself, such as rules for data, review of outputs, and role-based training; the regulation does not prescribe it.
Was the AI literacy obligation postponed to 2027?
No. What was postponed were selected obligations for high-risk systems. Article 4 has applied since 2 February 2025.
Do we need a certificate or a company AI officer?
No. To satisfy Article 4, the AI Act requires neither a certificate nor any particular organisational function. The company should, however, be able to document what measures it adopted and when.
When do the rules for high-risk AI start to apply?
For the areas under Annex III, such as certain systems in recruitment and employment, from 2 December 2027. For AI as a safety component of regulated products under Annex I, from 2 August 2028.
Primary sources on the current timetable
- Regulation (EU) 2024/1689 — the AI Act
- Regulation (EU) 2026/1744 — the AI Omnibus
- European Commission: the AI Omnibus enters into force
- European Commission: questions and answers on AI literacy
Where to find us
ITHOPE is a company based in Brno, Czechia: we do data recovery from disks, RAID and NAS, PC and laptop service and IT for business, all in our own lab at Viniční 3587/216, 615 00 Brno.
Phone +420 775 556 063 (24/7 line). You can bring the device in person, or send it by courier (Zásilkovna, PPL).