EU AI Act Transparency Rules: What Small Teams Actually Have to Do
Most coverage is written for companies with a compliance department. The list that reaches a five-person team is much shorter, and specific.
On this page
Most coverage of the AI Act is aimed at companies with a compliance department. If you're 5 people shipping a product with an AI feature in it, the list that reaches you is shorter.
#The parts that reach a small team
Tell people they're talking to a machine. If users interact with an AI system directly, that has to be obvious. A chat widget that implies a human is on the other end is the clearest failure case.
Label synthetic media. Generated or materially manipulated image, audio and video content needs marking as such, in a machine-readable way where feasible. Know whether you're a provider or a deployer. Building a system and putting your name on it makes you a provider, with meaningfully more obligations than using someone else's. This is where small teams most often misjudge their position.
#What this looks like in a product
- A one-line disclosure at the top of any AI chat surface
- Metadata on generated images, plus a visible marker where context doesn't make it obvious
- A written note of which third-party models you call and for what, kept current
That last one takes an afternoon and is the thing people skip.
#What's not required of most small teams
The heavy obligations — conformity assessment, registration, post-market monitoring — attach to high-risk categories: employment decisions, credit scoring, biometrics, essential services. A writing assistant or a support triage bot is usually not in that set.
Usually is doing real work in that sentence. Check your own use case rather than assuming.
#Timelines
Obligations phase in over several stages rather than all at once, and the dates have moved before. Track the official text rather than reporting on it — including this article.
#What you should write down this week
Spend an hour on an inventory before you worry about anything else. List every model you call, what it does, and whose users see the output — 5 or 6 rows is typical for a small product, and you'll probably find 1 you'd forgotten about. That document is what makes every later question answerable, and if you're ever asked whether you're a provider or a deployer, it's the difference between a 10-minute answer and a week of archaeology.
Frequently asked questions
Does this apply if we are not in the EU?
It can. Scope follows the users, not the company. If your output reaches people in the EU, assume it's in scope and get advice.
Does using an API make us a provider?
Usually it makes you a deployer (you run someone else's system rather than publishing your own), which carries fewer obligations. Fine-tuning and branding the result as your own can change that.
Do we need to label AI-assisted writing?
The labelling rules target generated and manipulated media. Text assistance sits differently, but client contracts increasingly require disclosure regardless of the law.
Written by
Vemuri Abhiram
Founder & Editor
I write every post here. The tools I review are ones I pay for and run on real client work, not demo accounts, and every number you see came off my own invoice or dashboard.
You might also like
More from Underread, once a week.
The guides worth reading and the tests worth trusting. One email, every Friday.
No spam. Unsubscribe in one click.