Two products can use a similar model and still create very different risks. A playful writing assistant is not the same kind of decision system as a tool involved in hiring.
The historical milestone
The EU AI Act entered into force on August 1, 2024. The European Commission described a risk-based framework, with requirements depending on the kind of AI system and use. Entering into force did not mean every obligation applied immediately. The Commission's 2024 announcement
This retrospective explains that product-design idea. It is not legal advice or a statement of the current compliance timetable; consult current official guidance and qualified advice for a real deployment.
My product lens: start with the decision
Before discussing a model's benchmark, I would map what the product can do to a person.
Consider two illustrative tools. One suggests a better heading for a blog post. Another produces a score that an employer might use to reject an applicant. Even if their underlying technology overlaps, the consequences and review needs do not.
That is a useful design conversation regardless of the jurisdiction. It replaces a vague "Is our AI safe?" with concrete questions.
Questions I would bring to a planning meeting
- Who is affected by the output, including people who never chose to use the app?
- Is the result a suggestion, a score, or an action?
- Can someone understand and challenge a consequential result?
- Who can override it, and do they have enough information to do so?
- What records would help investigate a complaint?
This is an original planning checklist, not a legal classification test or a complete compliance program.
Transparency that is actually useful
My preference is a disclosure that tells the reader what happened: AI drafted this summary, the editor reviewed it, and the original sources are linked. A generic sparkle icon says much less.
The larger takeaway is that governance belongs in the product brief, not only in a document written after launch. Start with the people, decisions, and possible consequences. Then choose how much automation makes sense.
