AL+

RO EN

Email us

EU AI Act obligations for small businesses: start with an inventory

Start an AI Act review with the systems you use, your role and their purpose. Chatbot providers must assess output marking as well as disclosure to users.

. By . Software business

Citiți articolul în română

On 21 May 2024 the Council of the EU gave its final approval to the AI Act, and the coverage that week was about risk tiers, bans and fines measured in percentages of global turnover. Two and a half years on, most of the rules apply, the high-risk deadlines have moved, and searches for the law return page after page of compliance offers.

My position is that most small and medium businesses need far less of that compliance work than they are sold. A company that uses AI tools, or adds a chatbot or an AI feature to its product, and does not build high-risk systems should start with three practical checks under the Act: know which AI systems it uses and in what role, take measures so the people using them understand them, and assess the transparency duties for its systems. For chatbot providers, disclosure is only one of those duties. A company that buys a heavy compliance programme before it has done that inventory is paying for fear, because it does not yet know which rules apply to it.

What was approved, and what changed since

The Council's press release described a "risk-based" law: banned practices such as social scoring, high-risk systems allowed but subject to requirements before reaching the market, and "very light transparency obligations" for systems of limited risk. TechCrunch's report the same day put chatbots in that last category. The regulation was published as Regulation (EU) 2024/1689 on 12 July 2024 and entered into force on 1 August 2024, with its rules phased in.

The phasing is where most of the later confusion came from. In July 2026 the EU adopted the Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026. Its recitals say the delayed standards and national frameworks had made compliance "heavier than expected". Read against both texts, the calendar as of November 2026 is:

Rule Applies from
Banned practices (Article 5) and the AI literacy duty (Article 4) 2 February 2025
Two new bans on generated sexual imagery without consent and child sexual abuse material 2 December 2026
Transparency duties for chatbots, deepfakes and generated content (Article 50) 2 August 2026
Machine-readable marking for generators already on the market before 2 August 2026 2 December 2026
High-risk systems listed in Annex III (hiring, credit, education and others) 2 December 2027
High-risk AI inside regulated products (Annex I) 2 August 2028

Much of what was published about the Act before July 2026 gives the old high-risk dates. Check the date of any advice before paying for it.

The question that decides everything: provider or deployer

The Act assigns obligations by role. A provider (Article 3(3)) is whoever develops an AI system, or has one developed, and places it on the market or puts it into service "under its own name or trademark". A deployer (Article 3(4)) is anyone "using an AI system under its authority", outside a purely personal, non-professional activity.

A business with a ChatGPT or Copilot subscription for its staff is a deployer. The vendor is the provider and carries the heavy obligations.

The case that deserves more attention is the small company that builds a customer chatbot on a model API and puts it on its own website under its own name. In my reading of Article 3(3), that company is the provider of the chatbot as an AI system, even though someone else trained the model underneath. That matters for more than a visible notice. Article 50(1) places the chatbot disclosure on providers: people must be informed that they are interacting with an AI system, unless it is obvious to a reasonably well-informed, observant and circumspect person in the circumstances. Article 50(5) says this must happen "at the latest at the time of the first interaction". Article 50(1) also excludes systems legally authorised for detecting, preventing, investigating or prosecuting criminal offences, with safeguards for third parties, unless available to the public for reporting an offence. For an ordinary customer chatbot, the practical starting point is one clear sentence in the chat widget and its first message, a visible notice that still needs accessibility and interaction checks.

Article 50(2) also requires providers of systems generating synthetic text, audio, images or video to ensure machine-readable marking and detectability. Technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, considering content limitations, cost and the state of the art. It excludes systems to the extent they assist standard editing or do not substantially alter the deployer’s input or its semantics, and legally authorised criminal-law uses. Ordinary open-ended chatbot generation is not automatically standard editing. Record which model or platform supplies markings, whether the application preserves them through display and export, how detectability is verified, and the reason for any exception. A chat label alone does not answer this check.

What the inventory looks like

The inventory is a short document, kept current, with one line per AI system: the tool or feature, who uses it, for what, whether you are provider or deployer, and whether anything about the use touches the categories the Act names. For most small companies it fits on a page.

Article 4, as rewritten by the Omnibus, requires providers and deployers to "take measures to support the development of AI literacy" of their staff, taking into account their knowledge and the context of use, and states that this "does not require providers or deployers to guarantee any specific level of AI literacy of any individual". The original text had asked for measures to ensure, "to their best extent, a sufficient level". The inventory tells you whom to train and on what: the person answering customers with a drafting assistant needs to know what it gets wrong. Article 4 is also not among the obligations for which Article 99 sets a fine ceiling. A short internal guide and a record of who read it is a proportionate answer.

The inventory also shows the cases where Article 50 reaches deployers: emotion recognition or biometric categorisation (Article 50(3)), deepfake image, audio or video (Article 50(4)), and AI-generated text published to inform the public on matters of public interest, unless a person has reviewed it and someone holds editorial responsibility.

Where this conclusion does not hold

For some companies the inventory reveals real work.

Hiring, credit and other Annex III uses. Annex III lists, among others, AI used "to analyse and filter job applications, and to evaluate candidates". A system that ranks or rejects candidates on substantive assessments is the relevant high-risk example. A tool that only reformats CVs is not automatically equivalent. Under Article 6(3), Annex III systems can be excluded where they pose no significant risk to health, safety or fundamental rights, including by not materially influencing decisions, and perform a narrow procedural task, improve a completed human activity, detect decision patterns without replacing or influencing a completed human assessment without proper review, or perform a preparatory task. Annex III systems that profile people remain high-risk. Providers claiming the exclusion must document the assessment and register under Article 6(4). For deployers of systems that remain high-risk, Article 26 applies from 2 December 2027 and requires them to use the system according to its instructions, assign human oversight to people with the competence and authority to exercise it, monitor it, keep its logs for at least six months where it controls them, and inform workers' representatives and affected workers before using it at the workplace. That is a programme, though a bounded one.

Becoming a provider by accident. Under Article 25, a deployer who puts its own name on a high-risk system, substantially modifies it so that it remains high-risk, or changes the purpose of a general system so that it becomes high-risk takes on the provider's obligations. A software company that adds AI candidate ranking to its HR product is in that position. The inventory has to record purpose, because purpose decides classification.

Banned practices. They are rarely relevant to ordinary businesses, but they already apply and carry the highest fines: up to €35 million or 7% of worldwide turnover.

Data protection. Article 2(7) states that the Act does not affect the GDPR. A chatbot that receives customer data is a data protection question first, and the inventory belongs next to whatever records you already keep for the GDPR.

The strongest counterargument

The case for spending early is serious. Fines for breaching Article 50 or deployer obligations can reach €15 million or 3% of worldwide turnover, whichever is higher. Enforcement is uneven and nobody can say which first cases will set the tone. Large customers send procurement questionnaires that ask for AI governance policies, and a small supplier without answers loses the contract.

Each point is fair, and none changes the order of work. For SMEs, Article 99(6) uses the lower of the percentage and fixed maximum: at €2 million turnover, 3% is €60,000. Actual fines depend on the infringement and statutory factors. The point is to identify and implement applicable duties, not buy a generic programme first. A procurement questionnaire is best answered with a concrete inventory, role assessment, training measures and transparency controls.

Where a compliance budget is justified, the inventory makes it efficient, because it shows the consultant which systems need attention.

The position

Write the list first, then assess obligations system by system. A deployer of ordinary tools and a provider of a generative chatbot have different work to do. The chatbot provider must check disclosure and machine-readable marking, including technical implementation and any applicable exception. Hiring, credit and education uses need a documented classification assessment; the inventory is the starting point, not proof of complete compliance.

Our AI agent service proposes written data flows and human approval boundaries before development. A project that interacts with people also needs its role, classification and applicable transparency controls agreed explicitly.

Sources