Why “we only use ChatGPT” is not a compliance strategy

There is a sentence I hear a lot from small teams when the EU AI Act comes up. It goes something like this: "We don't really need to worry about that. We only use ChatGPT." Sometimes it is Copilot, or Gemini, or a translation tool baked into an app they already pay for. The reassurance underneath is always the same: our stack is small, our tools are famous, and famous tools are surely someone else's problem to worry about.

I understand the instinct. When a regulation lands, the natural hope is that you fall outside it. But "we only use ChatGPT" is a description of your tooling, not a compliance position. It answers a question the AI Act never actually asks. Let me explain why, and then, because this is the good news, let me show you how little work it usually takes to put the matter to rest for a low-risk team.

The Act cares what you do, not what you bought

The EU AI Act, formally Regulation (EU) 2024/1689, is built around roles and risk, not brand names. It does not keep a list of approved or exempt products. It asks two questions instead. First, what is your role: are you a provider who builds an AI system and puts it on the market, or a deployer who uses one under your own authority in a professional setting? Almost every small studio, agency or SaaS team is a deployer. Second, what is the risk of the way you are using the system: prohibited, high-risk, limited or minimal?

Notice that neither question has anything to do with how many tools you run or how well known they are. A two-person agency using one chatbot and a fifty-person company running a dozen models are judged by the same framework. The size of your stack changes how much work you have to do. It does not change whether the Act applies to you at all.

Article 4 does not have a small-business exit

Here is the part that catches people out. The AI literacy duty in Article 4 has applied since 2 February 2025. It says that providers and deployers must take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and anyone operating or using AI on their behalf.

Read that again and look for the escape hatch. There is a threshold for company size: there is none. There is a carve-out for minimal-risk tools: there is none. Article 4 applies regardless of how big you are and regardless of which risk tier your tools sit in. If your team uses AI in its work, the literacy duty is live for you today.

The duty is proportionate, which is genuinely reassuring, and I will come back to that. But proportionate is not the same as absent. A duty that scales down to something small is still a duty you have to meet.

"We used common tools" is not evidence of literacy

The most common version of the argument is subtler than "the Act doesn't apply." It is: "our people already know how to use ChatGPT, so surely we are literate enough." And they probably are competent users. But Article 4 is not satisfied by competence that exists only in people's heads. The idea behind it is that your team understands enough to use AI responsibly and to spot its risks, and that you can show a reasonable, role-appropriate effort to get them there.

"We use common tools" tells a regulator, a client or an insurer nothing about whether your staff know that a model can be confidently wrong, that customer data pasted into a prompt may leave your control, or that AI output going to the public sometimes has to be labelled. Familiarity with a product is not the same as understanding its risks. And crucially, none of it is legible from the outside. Literacy you cannot point to is, for practical purposes, literacy you cannot rely on when someone asks.

Minimal-risk tools do not stay minimal by themselves

The other flaw in "we only use ChatGPT" is that it treats the tool as the source of risk. It is not. The use is. The same chatbot can sit in three different risk tiers depending on what you point it at, and a minimal-risk tool can quietly slide into duties the moment the task changes.

A few everyday examples of how a friendly general-purpose assistant stops being minimal-risk:

• Transparency duties (Article 50). If you use AI to run a customer-facing chatbot, people have to be told they are talking to a machine. If you generate or materially alter images, audio or video for the public, that synthetic content has to be disclosed. These transparency obligations attach to the use, not to some special category of tool.
• Feeding in client or personal data. The moment you paste a client's confidential brief or a spreadsheet of personal data into a prompt, you have layered data-protection duties on top, and you have made a decision your staff should understand well enough to make safely.
• Screening or managing people. Use the same assistant to help sift job applicants, rank candidates, allocate tasks or evaluate staff, and you have wandered towards the high-risk employment category in Annex III, which carries real obligations around human oversight, record-keeping and transparency. Some uses, like emotion recognition in the workplace, are prohibited outright under Article 5.

None of this means ChatGPT is dangerous or that you should stop using it. It means the risk tier is not a property of the logo on the tab. It is a property of what your team decides to do with it on a Tuesday afternoon, which is exactly why literacy, and a policy that draws a few lines, matters even for a modest stack.

The timeline you are actually living in

It helps to see where we are. The Act entered into force on 1 August 2024 and applies in phases. The prohibited practices and the Article 4 literacy duty have applied since 2 February 2025. The rules for general-purpose AI models, the governance framework and the penalty provisions applied from 2 August 2025. The transparency duties in Article 50, such as telling people they are talking to a chatbot and labelling AI-generated or synthetic content, apply from 2 August 2026, which is the date most small businesses should have in mind. The heavier high-risk obligations in Annex III, for both deployers and providers, were pushed back from that date to 2 December 2027 by the 2026 Digital Omnibus, so they land later than the original timeline suggested.

Enforcement is national. In the Netherlands the Autoriteit Persoonsgegevens has taken a coordinating role for algorithm and AI supervision, alongside the RDI, and has signalled a focus on transparency and prohibited uses. In Germany there is not yet a single federal AI authority, though the Bundesnetzagentur is widely expected to take a central coordinating role, with data-protection authorities involved in the meantime. The point is not to alarm you with any specific action, there is no need to invent one. The point is that the duties are live, the authorities exist, and "we only use ChatGPT" is not the answer you want to be giving when someone reasonable asks how you handle AI.

So what does a low-risk team actually do

Here is the reassuring turn, and I mean it. If your stack really is low-risk, a few everyday assistants and nothing that screens people or makes automated decisions about them, the work in front of you is small and finite. It is not a legal project. It is an afternoon of tidying, done once and kept current. Concretely:

1. Write down what you use. List the AI tools your team actually touches and, next to each, a one-line note on what it is used for and roughly which risk tier that use sits in. This is a risk-classified inventory, and for most small teams it fits on a single page.
2. Run a short course. Give your people a brief, role-appropriate briefing on responsible use: where models get things wrong, what data must never be pasted in, when AI output has to be labelled, and which uses are off-limits without a conversation first. This is your literacy measure. It does not need a certificate body behind it, Article 4 prescribes no specific course.
3. Set a light policy. One page. What is fine, what needs a second pair of eyes, what is forbidden, and who to ask when someone is unsure. A policy is what stops a minimal-risk tool from wandering into high-risk territory by accident.
4. Keep a dated record. Note when you did the training, who attended, when you last reviewed the inventory and the policy. A date is what turns "we are careful about this" into something you can actually show.

That is the whole shape of it for a genuinely low-risk team. Not a compliance department, four modest artefacts kept honest. This is the model Klaar is built around, turning a small, sensible effort into a dated bundle you can hand to a client, an insurer or a regulator without scrambling. But the tools are secondary. The mindset is the thing: the size of your stack was never the question, and "we only use ChatGPT" was never going to be the answer.

Use whatever tools serve you. Use one if one is all you need. Just do not mistake a short toolset for a finished obligation. The literacy duty is here now, it is proportionate, and for a low-risk team it asks for a course, a policy, an inventory and a date. That is a good afternoon's work, and it is a far better sentence to be able to say than the one we started with.