SuperML
← All issues

The EU AI Act High-Risk Deferral Is a Trap, Not a Reprieve

The EU AI Act Digital Omnibus deferred the Annex III high-risk deadline to December 2027 — but that's exactly one conformity assessment cycle, and Article 50 transparency obligations are already live. Banks that filed this as a win are already behind.

Hi there,

Most EU AI Act compliance briefings filed the Digital Omnibus deferral as good news. They are wrong. The new deadline lands at December 2, 2027 — exactly the length of a conformity assessment engagement — and Article 50 transparency obligations are already live and enforceable as of August 2026. Teams that treated the deferral as a reprieve didn't buy more time; they just made December 2027 impossible to hit.

🔥 Featured Post

The EU AI Act High-Risk Deferral Is a Trap, Not a Reprieve

  • The Digital Omnibus moved Annex III standalone high-risk systems from Aug 2, 2026 to Dec 2, 2027 — but that window is exactly one notified-body conformity cycle (12–18 months)
  • Article 50 transparency obligations were NOT deferred — they are live from August 2, 2026 and examiners are already looking
  • Credit scoring, insurance underwriting, and AML transaction monitoring are all Annex III high-risk — there is no "wait and see" for banks
  • Teams starting conformity assessment now are on the critical path to December 2027; teams that haven't started are already late
  • The governance gap: SR 26-2 carved out gen AI, the EU AI Act covers it — financial institutions now run two parallel governance frameworks for the same systems

Read the full post →

📚 In Case You Missed It

Compound AI Failures Live at the Seams, Not in the Model — A 150-incident study of production compound AI systems finds 51% of failures are silent, take 4.2 days to detect, and originate at component boundaries. Circuit breakers, quality gates, and typed interfaces cut recovery time 71%.

Eval-Gated Agent Releases: Passing Evals Is Not a Gate — A VentureBeat Pulse survey finds 61% of eval-running teams shipped a customer-facing failure after passing evals. The fix is risk-tiered release gates, not removing humans or trusting evals more.

Fail-Plausible: When Agents Turn Errors Into Fiction — LLM agents don't just swallow errors, they can rewrite them into confident, plausible prose. Treat error signals as a typed channel that never passes through the model, and validate your guards by sabotaging them.

More posts dropping every week. Stay sharp.

— Bhanu @ superml.dev