If leadership doesn't decide about AI, employees decide for themselves: everyone uses it differently, often on a private account. Leadership then has no idea what data leaves the company, who is accountable for the output, or whether any of it pays off. I teach AI courses for companies and individuals, and I run my own web work through AI agents. This article covers what an owner and leadership team should decide themselves, how to start without a big project, and how you and your team actually learn AI.
Leadership has five things to decide: which workflow to start with, what data may go into AI, who is accountable for the output, how you will know it pays off, and which tools people work in. Start with one workflow for 4 to 6 weeks and measure before and after, including time spent fixing outputs. Since 2 February 2025, the EU AI Act has required companies that use AI at work to take measures for the AI literacy of the people working with it. After the July 2026 amendment you don't have to guarantee a specific level for each person, but you do need the measures. Leadership should learn first (what AI can do, where it fails, how to measure it), then the team, on its own tasks, ideally spread over several weeks.
- What I teach. AI courses for companies and individuals, from a short intro for leadership to longer team programmes. Participants work on their own tasks, not generic demos.
- How I work myself. I manage the work on le-artist.cz through one coordinating chat with an AI agent. It opens parallel sessions for single tasks (an audit, an article, a fix, a review). Each one works in its own copy of the code and reports back. I decide what goes live.
- What this article leaves out. Stories about specific companies from my courses and numbers I can't back up. The legal parts link to the official texts, as of 22 October 2026. This is not legal advice.
Where AI helps a business and where it doesn't
The simplest test I recommend: how expensive is a mistake, and how quickly would you spot it? Where you can check the output in a minute and an error costs nothing, AI is a great assistant. Where an error is expensive and hard to see, AI should at most prepare the groundwork.
| Workflow | What AI does well | What stays with a person |
|---|---|---|
| Emails, quotes, customer replies | a draft, tone variants, a summary of a long thread | prices, deadlines, promises and pressing send |
| Meetings and internal documents | summaries, action lists, searching the text | what actually applies and who owns what |
| Research | an overview of sources, comparing options | checking the source and how current it is |
| Spreadsheets and reports | suggesting a formula, describing a trend, flagging outliers | whether the data is right, and the conclusion |
| Marketing and the website | article outlines, descriptions, image ideas | facts, promises to customers, how the brand looks |
| People, contracts, tax | groundwork and questions to fill in | the decision |
Where AI won't help: a process nobody has written down (AI speeds up chaos too), data that doesn't exist or is scattered across inboxes, a task whose output nobody can check, and a one-off job where the setup takes longer than the work itself.
What leadership has to decide, not IT
Anyone can set up a tool. Nobody else can answer these questions for you, and if you don't answer them, every employee will answer them separately.
| Decision | Why it belongs to leadership | Example answer |
|---|---|---|
| Which workflow we start with | leadership sees where the business loses time and money | replies to enquiries, not accounting |
| What data may go into AI | the company is responsible for personal data and trade secrets | public texts yes, customer personal data only in a tool covered by a data processing agreement |
| Who is accountable for the output | AI can't be accountable, a person can | anything that goes to a customer is read and sent by a person |
| How we will know it pays off | without a measure, AI stays a debate | time to reply, number of corrections, customer satisfaction |
| Which tools people work in | licences, account admin, what happens when someone leaves | company accounts, not private ones |
One page of rules is enough. A long policy nobody reads is worse than a clear traffic light: green (public information and your own texts), amber (internal material without personal data, only in an approved company tool), red (personal data, contracts, passwords, client data under NDA: only with leadership's approval and in an approved tool, or not at all).
How to start: a pilot on one workflow
I don't recommend a big "we're rolling out AI" project. I recommend a pilot you can evaluate.
- Pick a workflow. It repeats every week, the output is text or a table, and mistakes are easy to spot. Ideally the bottom-left field of the matrix above.
- Measure the starting point. For a week or two, log how long the workflow takes and how many mistakes or returns it produces. Otherwise you have nothing to compare against.
- Write the brief and an example. What a good output looks like, what AI must not do (promise prices, invent deadlines) and what data may go in.
- Run it for 4 to 6 weeks in normal operation. People also log the time spent fixing outputs, because that is exactly what enthusiasm tends to overlook.
- Evaluate and decide. Continue, adjust or stop. A stopped pilot isn't a failure, it's an answer you got cheaply.
How to build a budget and work out the return is in my article on what an AI rollout costs. If the pilot should turn into automation, the steps are in AI automation for businesses.
Risks: data, GDPR, confidential information and hallucinations
- Personal data. GDPR applies the moment you paste personal data into an AI tool. If the provider processes it on your behalf, you need a data processing agreement under Article 28 GDPR. An employee's private account doesn't solve that.
- Confidential information. Trade secrets, price lists, contracts and client data under NDA belong in the red category. For every tool, find out whether the provider uses inputs to train its models and how long it keeps them. The defaults differ between personal and business plans. OpenAI, for example, says it does not use data from ChatGPT Business, Enterprise or the API for training by default. For the personal Claude Free, Pro and Max plans, according to Anthropic, it is the user's choice in the privacy settings. Both checked on 22 October 2026.
- Hallucinations. A language model writes confidently even when it's wrong. Always check numbers, names, quotations, legal references and links. That applies to agents that search the web on their own too.
- Reviewing output. Anything that leaves the company (to a customer, onto the website, to an authority) is read by a person before it goes. Anything with a legal or financial impact gets a second reader.
- Private accounts. If the company provides no tool, it risks people finding their own and following no rules at all. Instead of a blanket ban, I recommend one approved tool and clear rules.
The AI Act: AI literacy under Article 4
The AI Act is an EU regulation, so it applies directly in every member state. The AI literacy obligation in Article 4 has applied since 2 February 2025. It covers providers of AI systems and deployers, which includes an ordinary company using an AI tool at work. The only exception is use "in the course of a personal non-professional activity" (Article 3(4)).
In July 2026, Article 4 was amended by Regulation (EU) 2026/1744, the Digital Omnibus on AI, published in the Official Journal on 24 July 2026. Under the new wording, providers and deployers "shall take measures to support the development of AI literacy of their staff and other persons dealing with the operation and use of AI systems on their behalf", taking into account their technical knowledge, experience, education and training, the context of use, and the people the AI is used on. The new part is an explicit reassurance: the obligation "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". The Commission and member states are to support companies, in particular SMEs, and the Commission will publish practical examples. Source: Regulation (EU) 2026/1744, Article 1(5), checked on 22 October 2026.
What I take from it in practice:
- List who uses which AI tool and for what. A table is enough.
- Fit the measures to roles. A salesperson needs different skills from an accountant or a developer.
- Keep a record of what you did. Rules, training, the date and who took part. The regulation doesn't prescribe a format, but without a record you can't show the measures exist.
- Repeat when tools change. A new tool or a new use means updating the rules and the training.
The other AI Act obligations (transparency for chatbots and AI content, high-risk systems, the post-amendment deadlines and supervision in Czechia) are covered in my article the AI Act in practice.
How to learn AI: leadership and the team
When only the team learns AI and leadership stays out, leadership can't judge what the team proposes and ends up deciding on gut feeling. That's why I recommend starting with leadership.
| Who | What they need to know | A format that makes sense |
|---|---|---|
| Owner and leadership | what AI can do and where it fails, the risks, how to measure value, setting the rules | a short intro built on your company's tasks, then short regular reviews of the pilot |
| The team | writing a good brief, checking output, handling data by the rules, their own workflow | a workshop to start, then practice on their own tasks spread over weeks |
| Tool admin (one person) | accounts and permissions, shared prompts, collecting questions and failures | a longer course or one-to-one coaching |
Workshop or course? A one-day workshop gives the team a shared vocabulary, basic rules and first results. Habits change more reliably with a course spread over several weeks: between sessions people try AI on real work, and at the next session we deal with what didn't work. That's why my courses use the participants' own tasks, not generic demos.
Between sessions: a shared library of prompts that worked, 15 minutes in the weekly meeting on what saved whom time and where AI failed, and one person who collects questions. Six areas where a team can learn hands-on are in my article on how to teach your team to use AI.
Choosing tools: what to watch in pricing and licences
One main tool is enough to start with, not five. Which model suits which kind of work is in my comparison of ChatGPT, Claude and Gemini. Before you buy, go through these questions:
| Question | Why it matters |
|---|---|
| Does the provider train on your inputs, and can you switch it off for all users? | it decides what from the amber and red categories may go in |
| Is there a data processing agreement, and where is the data stored? | without one, personal data can't go into the tool |
| Do you pay per user or per usage (credits, limits)? | usage is hard to predict, and work stops when a limit runs out |
| Monthly or annual commitment, minimum number of seats? | business plans often require more seats or annual billing |
| Does the company manage the accounts? | when someone leaves, you need to remove their access and history |
| What are the restrictions on using outputs? | terms for images, video and voice differ between tools |
Plan prices and limits change several times a year, so I don't put them in this article. How credits and limit resets work is explained in AI subscriptions, credits and resets. If the tool will work with company documents, sort out permissions too: who may see which material, see AI over company documents.
What works for me with AI agents, and what doesn't
On my own website, AI agents now do a large share of the work: audits, articles, fixes and reviews. I'm not saying that as an advert but as experience that applies to businesses too.
What works:
- A narrow brief with boundaries. Each session gets one task, a list of files it must not touch, and a description of what "done" means.
- Separate workspaces. Each session works in its own copy of the code, so agents don't overwrite each other.
- Automated checks before publishing. The site build, the type check and the content check must pass, or nothing goes live.
- A second pair of eyes. The review is done by a different session from the one that wrote the text.
- A work log. Each session records what it did, what it didn't verify and what's left. The next one picks up from there.
What doesn't work without a person:
- Agents are confidently wrong. On 20 October 2026, a review of three articles written by an agent found 24 issues, one of them blocking: a security flaw in an example script that meant the article couldn't be published at that point. The agent that wrote the article hadn't caught it.
- Small technical slips. When submitting new URLs to search engines on 22 October 2026, an agent sent one URL mangled by Windows path conversion. Checking the output caught it, and the URL was sent again.
- Facts only a person knows. The number of clients, stories from courses or prices agreed with a client: an agent doesn't know them and must not make them up. So my agents have a rule: whatever they don't know, they send to me.
- More sessions don't mean less responsibility. Parallel work speeds up the preparation, but deciding what goes on the website stays with me.
How I can help
- AI training for companies: a two-hour intro for leadership or the team from €240, a one-day workshop for €1,000 (up to 15 people), a team programme for €2,400 (16 hours over 4 weeks) and a 40-hour transformation for €4,800. Always on your company's tasks and tools.
- Corporate courses: a 20-hour course for €2,400 or 40 hours for €4,800, for individuals or teams, covering AI as well as skills like Photoshop or web building.
- For individuals: open AI courses, with webinars from CZK 990.
- If you want me to put AI into your operations: AI automation, from an audit to a pilot of one process.
FAQ
Does the AI Act apply to a company that only uses ChatGPT?
Yes. Anyone using an AI system in their professional activity is a deployer under the AI Act, and Article 4 on AI literacy has applied to them since 2 February 2025. After the July 2026 amendment, a company has to take measures to support the AI literacy of people working with AI, but it doesn't have to guarantee a specific level for each individual.
Is one training session enough for AI literacy?
The regulation doesn't prescribe the form of the measures or require a certificate. The measures should, however, match people's knowledge, their role and what they use AI for. One training session is a good start. It makes sense to add data rules, a record of the training and refresher training when tools change.
May employees use AI on private accounts?
That's a leadership decision. I recommend company accounts: the company has a contract with the provider, sets whether inputs are used for training, and removes access when someone leaves. A private account covers none of that, and customers' personal data doesn't belong in it.
Where do we start if we have no in-house IT?
With one recurring workflow where the output is text and mistakes are easy to spot, such as drafting replies to enquiries. You don't need IT to work in a chat tool. IT or outside help becomes necessary once you connect company systems or build automations.
Is a one-day workshop or a longer course better?
A workshop gives the team a shared foundation and first results. If you want to change how people work every day, a course spread over several weeks, with tasks on their own work between sessions, works better. A common combination is a short intro for leadership, a workshop for the team and then a pilot.
How will we know AI is worth it?
Measure the workflow before and after the pilot: time per task including fixes, number of errors and the effect on customers. If the time spent fixing equals the time saved, AI isn't worth it for that workflow, and it's fair to say so. How to calculate the return is in my article on what an AI rollout costs.
Sources and links
- Regulation (EU) 2024/1689, the AI Act, EUR-Lex: Article 3(4) (deployer), Article 4 (AI literacy), Article 113 (applies from 2 February 2025)
- Regulation (EU) 2026/1744, Digital Omnibus on AI, EUR-Lex: Article 1(5) (new wording of Article 4), OJ 24 July 2026
- Regulation (EU) 2016/679, GDPR, EUR-Lex: Article 28 (processor)
- OpenAI: Business data privacy and Anthropic: Is my data used for model training?, checked on 22 October 2026
- My articles: the AI Act in practice, how to teach your team to use AI, what an AI rollout costs