The four risk tiers, explained with everyday tools

The EU AI Act sorts every AI system into one of four risk tiers, and the tier decides what you have to do. That sounds abstract until you realise the tools in question are the ones already open in your browser tabs. A chatbot on your website, a CV screener your recruiter tried last month, the image generator your designer uses, the ChatGPT window drafting your emails. Each one sits in a tier, and each tier carries a different set of duties, from none at all to a full compliance programme.

This guide walks through the four tiers using tools a small team actually recognises. For each one you get the plain meaning, a concrete everyday example, and what the Act asks of you. By the end you should be able to look at your own stack and place each tool with reasonable confidence.

Tier one: prohibited, the practices you simply cannot use

At the top sits the banned list. Article 5 of the Act names practices that are considered an unacceptable risk to people's rights, so they are not allowed on the EU market at all. There is no paperwork that makes them fine. You either avoid them or you break the law.

For a small team the relevant bans are narrower than the headlines suggest, but two are worth knowing well:

• Emotion recognition on staff or customers. Software that claims to read the emotional state of your employees or of people in a shop or classroom is prohibited in workplace and education settings. If a vendor pitches you a tool that scores how happy, stressed or engaged your team looks on camera, that is a hard no.
• Most real-time remote biometric identification in public spaces. Live facial recognition that scans crowds to identify people as they pass is banned for general use, with only narrow law enforcement exceptions that do not apply to a studio or agency.
• Social scoring and manipulative techniques. Systems that rank people by unrelated behaviour to disadvantage them, or that use subliminal techniques to distort decisions, are also out.

The prohibited practices have applied since 2 February 2025, earlier than most of the Act. They also carry the heaviest penalty: up to 35 million euro or 7% of worldwide annual turnover, whichever is lower for a smaller company. What you must do here is simple to state and important to follow. Do not buy, build or switch on anything on this list, and if a supplier's demo edges into it, walk away.

Tier two: high-risk, the tools that pull you into real duties

The second tier is where a small business is most likely to trip without noticing. High-risk systems are listed in Annex III of the Act. They are allowed, but only if a serious set of obligations is met, because they make or shape decisions that affect people's lives.

The category most relevant to small teams is employment and worker management. That covers AI used for recruitment, for filtering or ranking job applicants, for allocating tasks, and for evaluating performance or deciding who to promote or let go.

The everyday example is a CV screener or automated hiring tool. Say your recruiter starts using a system that ranks applicants and surfaces the top ten. The moment that tool influences who gets an interview, you are using a high-risk system, and as the deployer you inherit real duties:

• Human oversight. A person must be able to review, question and override the system's output. The tool advises, a human decides, and that human needs to understand the tool well enough to catch a bad call.
• Record-keeping. You need to keep logs and documentation of how the system is used, so that a decision can be traced and explained later.
• Transparency. People affected, in this case candidates, should be informed that an AI system is involved in the process.
• Following the provider's instructions and monitoring the system so it keeps working as intended and you flag problems.

Other Annex III areas include education, access to essential private and public services such as credit or insurance decisions, certain uses of biometrics, and critical infrastructure. Most small studios and agencies will only meet this tier through hiring or evaluation software. The heavier high-risk obligations for Annex III systems, for both deployers and providers, were moved by the 2026 Digital Omnibus from 2 August 2026 to 2 December 2027, so that is now the date to plan around. Note that 2 August 2026 is when the transparency duties apply, not the high-risk duties. Getting the wrong answer here is still expensive, with penalties of up to 15 million euro or 3% of worldwide annual turnover.

Tier three: limited risk, where you just have to be honest

The third tier is lighter. These systems are not dangerous, but they can mislead people about what they are dealing with, so Article 50 asks for one thing: transparency. Tell people when they are interacting with AI or looking at AI-generated content. That is essentially the whole duty.

Two everyday examples cover most small teams:

• A customer-facing chatbot. If you put a support or sales chatbot on your site, it has to make clear that people are talking to an AI, not a human colleague. A short line at the start of the conversation does the job.
• AI-generated images or synthetic media. If you publish an AI-generated image, voice or video, or you meaningfully manipulate media so it looks real, that has to be disclosed. Deepfakes and synthetic content need a label so nobody is fooled.

None of this stops you using these tools. It just asks you to be upfront. The cost of compliance is a sentence, a label or a small notice. The transparency obligations become fully relevant from 2 August 2026, and they sit under the same penalty ceiling as other obligations, up to 15 million euro or 3% of turnover, though in practice a missing disclosure is far easier to fix than a banned practice.

Tier four: minimal risk, almost everything else

The last tier is where most of your stack actually lives. Minimal risk covers everyday tools that do not make consequential decisions about people. The Act sets no specific legal obligations for these systems.

The everyday examples are the ones you use constantly:

• ChatGPT or similar assistants for drafting emails, copy and first drafts.
• Canva and design tools with AI features for layouts and images.
• Spam filters, translation tools, transcription, most productivity software with a bit of AI baked in.

There is no register to file, no oversight process to document, no disclosure required for the tool itself. That said, minimal risk is not the same as no responsibility. Two things still apply. First, if you use one of these tools to produce synthetic media, the transparency duty from tier three still bites on that output. Second, and this catches everyone, the AI literacy duty applies here too.

The one duty that applies to every tier

Article 4 of the Act, the AI literacy duty, does not care which tier your tools sit in or how small your company is. It has applied since 2 February 2025, and it asks providers and deployers to ensure a sufficient level of AI literacy among the staff and others who use AI on their behalf.

It is deliberately proportionate. It does not prescribe a particular course or certificate. It takes into account your people's technical knowledge, their experience, the context they use AI in, and who the AI is used on. The core idea is plain: the people using these tools should understand enough to use them responsibly and to spot the risks, and you should be able to show you made a reasonable, role-appropriate effort. So even a team whose entire stack is minimal risk still owes its people basic AI literacy.

How to classify your own tools

You do not need a lawyer to make a first pass. Work through your tools one at a time and ask a short chain of questions:

1. Does it do anything on the banned list? Emotion recognition on staff or customers, live public biometric ID, social scoring. If yes, it is prohibited. Stop using it.
2. Does it make or shape decisions about people in hiring, evaluation, education, credit, insurance or similar? If yes, treat it as high-risk and plan for oversight, record-keeping and transparency.
3. Does it talk to people as if human, or generate realistic content? If yes, it is limited risk. Add a clear disclosure.
4. None of the above? It is minimal risk. No specific duties, but the literacy duty still applies.

Classification is not a one-off. Tools change, features get added, and a minimal-risk app can grow an AI hiring feature overnight. Klaar keeps this inventory current for small teams and pairs it with the training, policy and dated compliance record that back it up, so when a client, insurer or regulator asks, the answer is a folder rather than a shrug. Start with the four questions above, and you will already know where you stand well before the transparency duties apply on 2 August 2026 and the heavier high-risk obligations follow on 2 December 2027.