The AI Act deadline is a product-design deadline
Transparency rules arrive in August 2026. The durable response is not a compliance banner, but an AI experience people can understand, question and control.
The next important EU AI Act date is close enough to affect work already in delivery. From 2 August 2026, transparency obligations apply in key cases: people must be told when they are interacting with certain AI systems, and some generated or manipulated content must be identifiable or clearly labelled. The European Commission has also published a voluntary code of practice to help providers and deployers prepare.
This is a product-design problem before it is a disclosure-copy problem. A label may satisfy one narrow requirement while leaving the experience confusing, unaccountable or unsafe.
Start with the user's decision
Ask what a person might do differently if they knew AI was involved. Would they check a recommendation more carefully? Avoid sharing sensitive information? Ask for a human review? Challenge an outcome?
The answer determines where transparency belongs. A notice hidden in terms and conditions cannot help someone at the moment they are deciding whether to rely on an output. Good disclosure sits next to the consequential interaction and explains the practical implication in plain language.
For teams building chat, recommendation, classification or generation features, we use four questions:
- What is the system doing, and what is it not doing?
- Which data or instructions materially influence the output?
- How can a person correct context, contest an outcome or reach a human?
- What record will let the organisation reconstruct what happened later?
Make provenance part of the architecture
Transparency becomes expensive when it is added after the data model and event flow are fixed. If generated media must be marked, or an organisation needs to explain which model and policy produced an output, provenance needs to survive the entire path: model call, transformation, storage, export and publication.
That argues for explicit metadata rather than inference. Record the model or service class, policy version, generation time, human review state and any source references that the product is entitled to retain. Keep this operationally useful; collecting everything without a retention purpose creates a different risk.
Design the failure path now
AI experiences fail in more ways than a conventional form. A model can be unavailable, confidently wrong, outside its intended domain or unable to cite adequate evidence. Each case needs a product response.
A mature interface can say that it does not know. It can show sources, narrow the task, preserve the user's work and provide a route to human judgement. Those behaviours are more valuable than a generic warning beneath an otherwise overconfident answer.
A practical readiness review
Before launch, bring product, design, engineering, security and legal ownership into one short review. Map every AI-assisted user journey. Identify disclosure moments. Test the wording with people who were not involved in the build. Verify that logs and provenance actually survive production paths. Exercise correction, escalation and data-deletion requests.
The law sets a minimum. Trust is earned by making the system legible when the answer matters.
This article is practical product guidance, not legal advice. Confirm how the Act applies to your organisation and use case.