AI without the chaos

Start by listing the tools and data already in use, then agree how they may be used. The 90-day plan later in the article suggests a sequence; allow for staff time and any specialist support you need.

1. One owner for AI. Not a committee, not "everyone is responsible". One person who has the right to say "this tool is not allowed" and the duty to answer the question "may I use this". In a company of up to 50 people this is often the owner or the operations director, in a larger one the CIO. This person does not need to understand language models. They need to know the company's data and have access to the board.

2. An inventory of what is already in use. Before you ban anything, check what people are using. An anonymous five-question survey (which tools, for which tasks, which data, private or company account, what you would like to be able to do) is enough to start. Without an inventory and a policy you do not know what is in use. The 2025 IBM report shows how common this is among breached organisations: 63% of organisations that had suffered a breach had no AI governance policy or were still developing one. Among those that do have a policy, only 34% regularly check whether employees are using unapproved tools.

3. Three data classes. An employee will not read a twenty-page policy, but will remember a simple rule. I propose this one (it is my proposal, not a result of research):

• green: public content, including published marketing materials, unless it falls into a higher-risk class; approved tools may be used;
• yellow: internal documents without personal data, only company accounts of approved tools;
• red: personal data of customers and employees, trade secrets, source code, customer contracts, financial data. May go into an AI tool only after a decision by the AI owner and a risk assessment.

If material falls into more than one class, use the higher-risk class: red takes precedence over yellow, and yellow over green.

4. A list of approved tools and company accounts. Check whether the company can manage access, revoke it when an employee leaves and establish how the provider handles data. Use the answers to decide which tools and accounts to approve. Choose one or two tools, buy company licences and check in the contract whether the provider uses your data to train models, where it stores the data and for how long. Questions for the provider are in the downloadable materials.

5. A person is responsible for the result. Whoever sends a text, a figure or an answer to a client is responsible for what was sent, regardless of who wrote it. Write this into the policy explicitly. How this looks in practice is shown by the Air Canada case, described in the part on risks.

6. AI literacy: the obligation and how to address it. Article 4 of the AI Act has applied since 2 February 2025 and concerns companies that use AI systems, not only their producers. As amended by Regulation 2026/1744 (described in the section on the law), it requires measures to support the development of AI literacy, without guaranteeing a particular level for each individual. I suggest training based on examples from your company as one such measure; it also reduces the risk of leaks. Whether training alone is enough depends on how widely you use AI, so ask a lawyer. Keep a list of participants.

Before the pilot, establish which obligations apply to your use of AI. The dates and scope are set out later in "Law: what applies and what has been postponed".

How to choose the first task

The criteria below are mine, not from research. A good first task meets four conditions at once: it is performed often (at least a dozen times a week), the data is green or yellow, a person always checks the result before it is sent, and time or quality can be measured before and after the pilot.

Such tasks usually fall in the areas where companies in Eurostat's data use AI most often: sales, marketing and administration. Examples to consider: preparing a draft reply to typical requests for quotation, summaries of meeting notes and email threads, a first version of product descriptions or job advertisements, sorting and categorising tickets.

Do not start with: evaluating candidates or employees (in the AI Act this is a high-risk area, and UODO gives recruitment as an example of a use with a different risk profile from marketing texts), a chatbot answering customers without handover to a person (see Air Canada), and an agent with write permissions in accounting or banking.

How to design a pilot

Again my proposal, but every element follows from the mistakes described above:

• one task, 5–15 users, 6–8 weeks;
• measurement before the start: how long the task takes, how many corrections, how many complaints. Without this, after the pilot you are left with impressions;
• a budget limit set at the beginning, including people's time;
• "go on / stop" criteria written down before the start, not after the results;
• the AI owner reviews the cases in which the tool made a mistake and adds them to the instructions.

Hypothetical example: a company of 40 people, whose sales team prepares 25 sales proposals a week, on average 2 hours each. The pilot involves 5 salespeople and one type of proposal. The criterion set before the start: preparation time falls by at least 25% with an unchanged number of corrections reported by clients. A result below the threshold means the end of the pilot, not "another month". The numbers are illustrative and serve only to show how to set a threshold.

90-day plan

1. Week 1–2

   Appoint the owner, run the survey and inventory the tools

   ResultRegister of AI tools

2. Week 3–4

   Define data classes, choose tools, introduce the policy and a one-hour training session

   ResultSigned policy, list of approved tools

3. Week 5–6

   Choose the pilot task, measure the baseline

   ResultPilot card

4. Week 7–12

   Pilot, weekly review of errors

   ResultPilot data

5. Week 13

   Decision: scale if the criteria are met, stop, or approve a new pilot

   ResultDecision record

A new pilot requires a separate hypothesis, budget and criteria agreed before it starts.

What the data says about results

Two studies are worth knowing, because they show that the effect depends on the task and that people misjudge their own speed-up.

In the version of their study updated on 6 November 2024, Brynjolfsson, Li and Raymond analysed data from 5,172 customer support agents at one company. Access to AI assistance increased resolved issues per hour by 15% on average. Less experienced and lower-skilled workers benefited most. The most experienced and highest-skilled workers saw small gains in speed and small declines in quality.

In July 2025, METR published a trial with 16 experienced open-source developers and 246 tasks. With access to AI they worked 19% slower, although after the study they believed they had been 20% faster. A caveat: the sample is small, one environment, tools from early 2025. In February 2026, METR itself wrote that, with a new design of the experiment, it has an uncertain signal about the effect at that time, because 30–50% of developers did not submit tasks that they prefer to do with AI. At least one thing justifies caution: users' impressions can be a poor measure, so measure.

In McKinsey’s 2025 survey, 88% of organisations used AI in at least one function, while around one third had begun scaling it. These are different stages of adoption. A company’s decision should depend on pilot results, quality and total costs, rather than adoption percentages alone.

Data leaks

In May 2023, Samsung banned its employees from using generative AI. According to Bloomberg's account, the reason was that an employee had uploaded sensitive source code to ChatGPT, and the company feared that the data was reaching external servers, from which it is hard to remove. This is a single, well-documented case. The scale is shown by IBM's Cost of a Data Breach 2025 report, based on 600 organisations that had suffered a breach (March 2024 – February 2025):

• 13% reported a breach of an AI model or application, and 97% of them had no access controls for AI;
• one in five organisations had experienced a breach connected with shadow AI, meaning the use of tools without the organisation's approval;
• breaches involving shadow AI cost on average USD 4.63 million compared with USD 3.96 million for ordinary ones, a difference of USD 670 thousand;
• the average global cost of a breach was USD 4.44 million.

The study covers organisations that had already suffered a breach, and the amounts are global averages. They are not a forecast of the cost of an incident at a particular company. The association between shadow AI and higher costs does not establish that missing controls alone caused the difference.

Wrong answers and liability

In Moffatt v. Air Canada (British Columbia Civil Resolution Tribunal, 14 February 2024), the airline's chatbot told a customer that a discount on a flight because of a death in the family could be obtained by an application made up to 90 days after buying the ticket. The airline's rules provided otherwise. The airline argued that the correct information was elsewhere on its website and that the chatbot was a separate entity responsible for itself. The tribunal did not agree: a company is responsible for all the information on its website, regardless of whether it comes from static text or from a bot. The damages were about 650 Canadian dollars plus interest and costs. The amount is small, the principle is not.

McKinsey, in its State of AI 2025 survey (1,993 respondents, 105 countries, June–July 2025), reports that 51% of organisations using AI experienced at least one negative consequence, and almost one third of all respondents point to consequences of inaccuracies in AI answers.

Personal data

In 2026, UODO (the Polish data protection authority) published lists of control questions for companies deploying AI, in four versions, including a separate one for SMEs that use ready-made systems without training them. The document stresses that the same tool carries a different risk for marketing texts, a different one for customer service and yet another for recruitment. UODO states that the lists are not a binding interpretation and do not replace a full risk analysis.

Integrations and agents

The more permissions you give a tool, the greater the risk. The OWASP Top 10 for applications based on language models (2025 edition) lists, among others, prompt injection (hidden instructions in content that the model reads, making it do something other than intended), disclosure of sensitive information and excessive agency, that is, a situation in which a model has more functions, permissions or autonomy than the task requires. My practical rule: treat the first agent like an intern in their first week. Read access yes, write access to financial systems no.

Gartner forecasts (June 2025) that more than 40% of projects based on agentic AI — systems that can carry out successive steps of a task and use tools within their granted permissions — will be cancelled by the end of 2027 because of rising costs, unclear business value or inadequate risk controls. This is a forecast, not a measurement, and it is based, among other things, on a survey of 3,412 webinar participants. Gartner also estimates that of the thousands of "agent" vendors, only about 130 offer real agentic capabilities, while the rest rebrand chatbots and automations (agent washing). According to McKinsey (survey of June–July 2025), 23% of organisations were scaling agentic systems, and a further 39% were experimenting.

Eurostat, the Ministry of Development and Technology (MRiT) and the Polish Economic Institute (PIE) cover different groups of companies and may define AI use differently. Do not treat their findings as a single comparable series.

For comparison, GUS (Statistics Poland) reported 5.9% for 2024. Eurostat gives the same value for Poland in 2024, so the difference from 8.36% in 2025 is a difference of years (2024 and 2025), not of definition.

The MRiT survey from 2024 (650 micro, small and medium-sized companies) showed 16%. In PIE's September 2025 survey (500 companies, including 100 medium-sized and 50 large), 19% of medium-sized companies and 21% of large companies used generative AI; the percentages refer to the respective size groups. According to the MRiT survey, 77% of companies that do not use AI do not plan to introduce it until it becomes necessary.

Regulation (EU) 2026/1744 was published in the Official Journal on 24 July 2026, entered into force on 27 July 2026 (Article 4 of the regulation) and postponed the deadlines for high-risk systems. The position as of 7 October 2026:

• 2 February 2025: the original prohibited-practice provisions and Article 4 began to apply. The new prohibitions identified in the amended Article 113 have a separate application date: 2 December 2026;
• 2 August 2025: the obligations of providers of general-purpose models began to apply;
• 2 August 2026: Article 50 began to apply. Providers of systems generating synthetic content that were placed on the market before that date have until 2 December 2026 to comply with Article 50(2). This is not a general deferral of disclosure obligations for business users publishing content;
• 2 December 2027: the provisions of Chapter III, Sections 1–3 specified in Article 113 will apply to high-risk systems under Article 6(2) and Annex III (including AI in recruitment and employee evaluation), originally from 2 August 2026;
• 2 August 2028: the same provisions for systems covered by Article 6(1) and Annex I (high-risk systems embedded in regulated products), originally from 2 August 2027. The deferral does not cover Article 6(5).

In its original wording, Article 4 required measures "to ensure, to their best extent, a sufficient level of AI literacy". The amended Article 4 is described under rule 6.

Article 99 sets different fine limits according to the infringement. The limit of EUR 35 million or 7% of annual worldwide turnover concerns prohibited practices, and EUR 15 million or 3% concerns the obligations listed in paragraph 4. For SMEs, including start-ups, paragraph 6 provides for the lower applicable limit; the amendment did not change this rule.

Poland's Act on Artificial Intelligence Systems (Journal of Laws 2026, item 1003) generally entered into force on 11 August 2026. Articles 8–18 and Chapters 3–5, 8 and 9 will enter into force on 28 October 2026; Article 125(4) had a separate, earlier commencement date. The act establishes the Commission for the Development and Safety of Artificial Intelligence (Komisja Rozwoju i Bezpieczeństwa Sztucznej Inteligencji, KRiBSI), which is to conduct inspections, issue opinions and organise regulatory sandboxes.

What follows for most SMEs: your obligations depend on how you use the system, not simply on whether you buy a ready-made tool. Article 50 distinguishes providers from deployers. For publishing, it covers, among other things, deepfakes and text intended to inform the public on matters of public interest; the latter has an exception involving human review or editorial control and editorial responsibility. High risk will appear when AI starts to influence decisions about people (recruitment, employee evaluation).

1. Company AI Use Policy (01-AI-use-policy-template; PDF). A ready template to adapt: rules, data classes, responsibility, incident reporting.
2. AI Tools Register and Risk Assessment (02-AI-tool-register-and-risk-assessment; XLSX). An inventory of tools and uses with automatic calculation of the risk level and a list of approved tools.
3. AI Tool Vendor Assessment Checklist (03-AI-vendor-assessment-checklist; PDF). Questions on data, contract, security and GDPR (RODO), based on UODO's questions and the OWASP list.
4. AI Pilot Card (04-AI-pilot-card; PDF). Hypothesis, baseline measurement, budget, "go on" and "stop" criteria.

Materials are available in Polish and English as PDF or XLSX, as indicated in the links below. The templates are not legal advice; assess your organisation’s requirements before adopting a policy.

Sources

• Microsoft and LinkedIn, 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part (31,000 workers, 31 countries, February–March 2024).
• Eurostat, Use of artificial intelligence in enterprises (data for 2025, table isoc_eb_ai, retrieved in December 2025; barriers among companies that did not use AI but had considered doing so).
• Eurostat, Usage of AI technologies increasing in EU enterprises (23 January 2025; data for 2024, including Poland at 5.9%).
• Polish Economic Institute, PointP 5/2025: AI in Companies (in Polish; original title: PointP 5/2025: AI w firmach) (summary of GUS, MRiT and PIE data; PIE survey: CATI, September 2025, 500 companies, including 100 medium-sized and 50 large).
• IBM, Cost of a Data Breach Report 2025, as reported by HPCwire and Kiteworks.
• Forbes Australia, Samsung bans ChatGPT among employees after sensitive code leak (citing Bloomberg, May 2023).
• American Bar Association, BC Tribunal Confirms Companies Remain Liable for Information Provided by AI Chatbot (Moffatt v. Air Canada, 2024).
• McKinsey, The state of AI in 2025: Agents, innovation, and transformation.
• Gartner, Over 40% of agentic AI projects will be canceled by end of 2027 (25 June 2025).
• Gartner, 30% of generative AI projects will be abandoned after proof of concept by end of 2025 (29 July 2024).
• Brynjolfsson, Li, Raymond, Generative AI at Work, version of 6 November 2024 (5,172 agents, 15% on average).
• METR, Early-2025 AI experienced open-source developer study (10 July 2025) and update of 24 February 2026.
• OWASP, Top 10 for LLM Applications 2025.
• UODO, Initial question lists on personal data protection in AI systems (in Polish; original title: Listy pytań inicjalnych dotyczące ochrony danych osobowych w systemach AI) (2026) and the report by PIT.pl.
• Regulation (EU) 2026/1744 (Digital Omnibus on AI): EUR-Lex, English text and Polish text (Article 1, points 5, 38, 39(b) and 40, that is, Articles 4, 99, 111(4) and 113 of the AI Act as amended, and Article 4 of the regulation on entry into force), National Law Review, Gibson Dunn, Software Improvement Group, artificialintelligenceact.eu.
• AI Act, Art. 4: AI literacy (in Polish; European Commission AI Act Service Desk, original wording).
• AI Act (Regulation (EU) 2024/1689 of 13 June 2024), Arts. 4, 50 and 99: EUR-Lex, English text, Polish text.
• Polish act on artificial intelligence systems (act of 3 July 2026): Journal of Laws 2026, item 1003, text of the act; on KRiBSI also Akademia LTCA, Spider's Web, Bankier.pl.