The AI Act stopped being a slide deck in 2025. Now that general-purpose-model obligations, high-risk system conformity assessments, and the national market-surveillance authorities are live, every model that leaves a LIACC repository passes a dozen-item gate before it sees a user.

Classifying the risk tier

Each project gets a risk classification memo before any code is written. The memo answers three questions, mapped to the Act's four tiers:

  • Prohibited: would this system score or categorise people in a way Art. 5 forbids? If yes, we stop.
  • High-risk: does it land in Annex III? (health, education, employment, law enforcement, critical infrastructure, administration of justice, democratic processes, migration, access to services.)
  • Limited-risk or minimal-risk: transparency obligations only.

Most applied AI in health and public administration falls in the high-risk tier, so the rest of the checklist is written against that bar.

The 12-item gate

  1. Documented intended purpose — and documented out-of-scope uses.
  2. Data governance record: provenance, relevant populations, bias analysis.
  3. Technical documentation in the Annex IV format.
  4. Logging and traceability (Art. 12) — not an afterthought.
  5. Transparency to deployers: instructions of use, known limitations.
  6. Human oversight: what a reviewer can do, and when.
  7. Accuracy, robustness, cybersecurity (Art. 15) with documented metrics.
  8. Post-market monitoring plan.
  9. Conformity assessment path identified (internal vs. notified body).
  10. Registration in the EU database once CE-marked.
  11. Incident reporting contact and procedure.
  12. Ongoing compliance review, not a one-off.

What's changed in practice

Three shifts we've measured inside our own delivery process:

  • Documentation is now code-adjacent. Model cards, data sheets, and risk memos live in the repository next to training scripts. If the docs don't exist, CI refuses to ship.
  • Evaluation budgets tripled. Accuracy on the headline metric isn't enough — we measure calibration, subgroup performance, adversarial robustness, and deployment drift.
  • Procurement conversations are longer. Clients want evidence, not assurances.

The one rule we repeat

Compliance isn't a checkbox exercise, and it isn't a marketing exercise either. It is an operational requirement. If your team can't articulate the failure modes of the system in a five-minute standup, you haven't yet designed it to be deployable.

—

If you're preparing a high-risk deployment in Portugal, the LIACC compliance advisory desk has worked through several Annex IV packs this year and is happy to share templates.