Governance & law

The EU AI Act: what companies need to anticipate

July 18, 2026 6 min read
Kamel Malek Director of SetRecord.ai

Most compliance conversations about the EU AI Act start and end with a single question: "do we build or deploy an AI system?" That framing misses the risk that actually reaches marketing, legal and communications teams — the Act reshapes what the public expects from any AI system that talks about your organisation, whether you built it or not.

What the Act actually does

The EU AI Act is a risk-based regulation, not a blanket restriction on artificial intelligence. Instead of treating every model the same way, it sorts AI systems by the harm they could plausibly cause and attaches obligations proportional to that harm. A spam filter and a system used to screen job candidates are both "AI", but only one of them carries meaningful legal weight under the Act.

For most companies, the practical relevance isn't building a high-risk system — it's the fact that general-purpose AI systems, the kind that answer questions about your brand, your executives and your track record, now sit inside a formal transparency framework for the first time. That changes the baseline of what "acceptable AI behaviour" means, publicly and legally.

For a communications or legal team, this shift shows up less as a compliance form to fill in and more as a change in public expectation. Journalists covering your sector are increasingly aware that an AI-generated summary of a company is a distinct category of content, not simply "what Google says" — and they are starting to ask sources whether the AI account actually matches the facts. That is a cultural shift the regulation is accelerating, not one it invented from nothing.

The four risk tiers, in plain terms

The Act's structure is easier to remember as a pyramid than as a list of definitions:

  • Unacceptable risk — practices considered incompatible with fundamental rights, such as certain forms of manipulative or exploitative AI, are prohibited outright.
  • High risk — systems used in areas like employment, credit scoring, critical infrastructure or law enforcement carry the heaviest obligations: risk management, documentation, human oversight, and conformity assessment.
  • Limited risk — this is where general-purpose chatbots and content-generating systems typically sit. The core duty is transparency: people need to know they are interacting with AI, and AI-generated content needs to be identifiable as such.
  • Minimal risk — the large majority of everyday AI use cases, with no specific obligations beyond existing law.

Very few organisations reading this will operate a high-risk system. Almost all of them are affected by the limited-risk tier, because it governs the assistants their customers, journalists and investors are already using to ask about them.

Where a specific product falls in this pyramid depends on its intended use, not its underlying technology — the same large language model can power a minimal-risk writing assistant and, embedded differently, a high-risk hiring tool. That use-based classification is worth remembering before assuming any given AI product is automatically low-stakes just because it "only answers questions."

The transparency obligations that matter here

Two transparency principles carry the most weight for information integrity. First, systems designed to converse with people generally must not present themselves as human without disclosure. Second, synthetic content — text, image, audio or video generated or manipulated by AI — is increasingly expected to be identifiable, whether through labelling, disclosure or technical marking.

Neither obligation directly forces a model provider to be factually accurate about your company. Transparency is not the same thing as accuracy — a model can disclose that it's an AI and still assert something false about your CEO in the same sentence. That gap is precisely where reputational exposure lives, and it's why independent verification of what models actually say remains necessary regardless of what the regulation requires upstream.

General-purpose model providers also carry documentation duties that exist upstream of any single output: keeping technical documentation available to regulators and downstream integrators, and publishing a summary of the broad categories of content used to train the model. None of this guarantees a model won't state something false about your company — but it does mean, in principle, that the raw materials behind an answer are somewhat less opaque than they used to be, which is useful context when you are trying to work out where an error came from.

Don't know what AI says about you today?

A SetRecord audit queries supported AI models about your organisation and measures the gap with verifiable facts — within five business days.

Request an audit

Why this creates indirect reputational exposure

Here is the part most compliance checklists skip. The Act formalises a public expectation: AI-generated answers about people and organisations should be identifiable, traceable, and — implicitly — trustworthy enough to be labelled as machine output rather than hidden as if it were neutral fact. As that expectation becomes normal, the gap between what a model says about your brand and what your official record shows becomes more visible, not less. Journalists, regulators and customers increasingly know to ask "is this AI-generated, and is it accurate?"

A company doesn't need to be a "provider" or "deployer" under the Act to feel this. If ten major assistants describe your organisation inconsistently, or attribute a lawsuit, a product or a former executive to you incorrectly, that inconsistency now sits inside a regulatory climate that treats AI transparency as a baseline, not a bonus. Silence on your part reads increasingly like negligence rather than neutrality.

The exposure is often sharper for organisations that are lightly covered online to begin with. A large public company usually has enough verified content in circulation that a model has several accurate sources to draw from; a mid-sized firm, a public institution, or a recently founded company may have only a handful of indexed pages, and a single inaccurate one can end up carrying disproportionate weight in what a model asserts about it.

The enforcement timeline, at a glance

The Act entered into force in 2024 with a staggered applicability schedule rather than a single start date: prohibited practices became enforceable first, obligations for general-purpose AI models followed roughly a year later, and the bulk of high-risk system requirements phase in over a longer horizon, with certain categories given extended runway. The exact dates and scope continue to be clarified through implementing guidance, so treat any specific date as directional rather than a substitute for legal advice on your situation.

What matters operationally is the direction of travel: obligations are tightening in phases, not all at once, which gives most organisations a realistic window to build the internal habits described below before enforcement reaches them in full. Because implementing guidance keeps evolving, the practical approach for most organisations is to treat compliance as an ongoing programme rather than a single deadline to clear — the specific date that eventually applies to your sector then becomes a formality rather than a scramble.

Preparing your organisation now

Three actions are worth taking regardless of how directly the Act applies to you:

  • Inventory your AI exposure — not just tools you deploy, but how external assistants currently describe your organisation when asked.
  • Document your source of truth — a maintained, citable fact sheet reduces the odds a model has to guess, and gives you a reference point when it gets something wrong.
  • Assign ownership — someone in legal, communications or trust & safety should own AI-narrative monitoring the way they already own press monitoring.
  • Rehearse the response — a documented process that has never been tested tends to fail exactly when it is needed; walk through it once against a hypothetical error before a real one appears.

None of this requires you to become an AI Act specialist. It requires treating model outputs about your organisation as a public-facing surface you monitor, the same way you already monitor press coverage and search results — grounded in verifiable evidence rather than assumption, and reviewed with counsel when a specific obligation is genuinely unclear.

See what AI says about you

A SetRecord audit checks what supported AI models say against your verifiable facts — within five business days.