Pack · obligations
EU AI Act Obligations pack
The obligations of the EU AI Act that touch what an agent does, one row per article. Dates follow the Action State EU AI Act obligations mapping, 3 Sep 2026; Reg (EU) 2026/1744 (the Digital Omnibus on AI, in force 27 July 2026). There are 37 rows.
By date in effect
Since 2 Feb 2025 (Chapter I–II)
Art 5(1)(a) · Art 5(1)(b) · Art 5(1)(c) · Art 5(1)(ba)/(bb) — from 2 Dec 2026 · Art 4 (as amended 27 Jul 2026: a duty of effort)
Since 2 Aug 2025 (GPAI Chapter V; penalties)
Art 53(1)(b) · Art 55(1)(c) (systemic-risk models)
Since 2 Aug 2026 (general application)
Art 50(1) · Art 50(2) · Art 50(3) · Art 50(4) · Art 49(1)/(2)/(3) — timing contested · Art 72(1)-(2) — literal 2 Aug 2026, timing contested · Art 73(2)-(4) — literal 2 Aug 2026, timing contested · Art 86(1) — literal 2 Aug 2026, timing contested · Arts 75a–75d (AI Office powers, 2 Aug 2026)
On timing, the pack notes: “Art 50 is fully live (50(2) grace to 2 Dec 2026 for pre-existing systems). Arts 49/72/73/86 apply on their literal text from this date but presuppose the Dec-2027 regime — CONTESTED.”
From 2 Dec 2027 — Annex III high-risk (Chapter III Sections 1–3: Arts 6–27)
Only if the system is classified high-risk: Art 12(1) · Art 12(2)(a)-(c) · Art 12(3) — remote biometric ID only · Art 13(3)(f) · Art 14(3) · Art 14(4)(d) · Art 14(4)(e) · Art 14(4)(b) · Art 15(1),(4) · Art 17(1)(i),(j) · Art 26(1), (2), (4), (5), (6), (7), (11) · Art 27 (narrow trigger) · Art 9(2)(c) · Art 11 / Annex IV
Every article in this group, Art 26 (deployer duties) included, applies only if the system is classified high-risk: from 2 Dec 2027 for Annex III systems, from 2 Aug 2028 for Annex I.
From 2 Aug 2028 — Annex I product-embedded high-risk
Arts 6–27 via Annex I — only if the system is classified high-risk
Recomputed, or judged
Most rows are recomputed. Their evidence is a fact or a stated rule, run over the sealed record, so anyone holding the record gets the same answer. Some rows are document or acknowledgement rows, and those read present or absent. A few rows turn on a word that a person has to judge. For those, the pack records the verdict and who gave it, and does not pretend to compute it. These are the judged words:
- Art 5(1)(a) “materially distorted a person's decision toward significant harm”
- Art 5(1)(b) “exploit age, disability or socio-economic vulnerability”
- Art 5(1)(c) “a social score derived from unrelated context” (graded by a rule, with judgment as the fallback)
- Art 14(4)(b) “approvals were not rubber stamps” (only if the system is classified high-risk)
- Art 26(4) “relevant and sufficiently representative” (only if the system is classified high-risk)
- Art 86(1) “a clear, meaningful explanation” (timing contested; whether it was delivered is recomputed, its quality is judged)
What it is: a versioned file your book is judged against, pinned per action date. Each obligation is stated as something a record can be checked for.
What it is not: not legal advice. It reports each row as present, absent, or judged, and never gives a verdict on the law.
Every row, in plain language →
Pack · outcomes
Outcomes pack
What a buyer of an agent needs to see: did it do the job, how were people treated, and do they come back. Each outcome is declared up front, and each row reads met, not met or not evaluable, with the count it covers. As the pack puts it, “87% confirmed” means nothing without “of 41 assessed, out of 210 published.” The pack is a draft, and it is versioned.
J · the job
Did it work?
- J1End state reached, confirmed by a non-agent system recomputed
- J2Proposal delivered to a named human recomputed
- J3The human responded — accept/decline/defer/no-response all first-class recomputed
- J4The human's later action corresponded to the proposal recomputed
- J5The proposal was later assessed against ground truth recomputed
- J6The assessment agreed with the proposal judged
- J7Finding → confirmed end state within window W recomputed
- J8The agent stayed inside its allowed actions recomputed
- J9Unassessed / unresolved work, aged recomputed
- J10The agent caused the outcome refused
C · the conversation
How was the person treated, and what did they leave with?
- C1A real choice was offered, with a decline path recomputed
- C2The reason came before the ask recomputed
- C3No pressure/urgency without a stated path rule judged over free text
- C4A human was reachable when asked, and the hand-off happened recomputed
- C5Resolved without escalation recomputed
- C6What the human left knowing: the resolution was stated back, and next steps named recomputed judged over free text
- C7What the human left saying: appreciation / frustration in their own words rule
- C8The person, asked directly, reported the experience a person's own answer
- C9Pushback was present and answered rule · a health signal, not a score
- C10The customer was satisfied / trusts us more refused
R · the relationship
Do they return, rely, and grow? Computed from the J and C rows across sessions, with no new instrumentation.
- R1The person returned voluntarily — sessions not preceded by an agent ask recomputed
- R2The same issue did not recur for the same subject within W recomputed
- R3Effort trended down: turns/escalations-to-resolution per subject over time recomputed
- R4Reliance is calibrating, not blind: decline/question rate is non-zero and stable recomputed · a health signal
- R5Experience-tier fairness: outcomes hold across novice and expert cohorts recomputed
- R6The relationship survived a failure — a refuted/failed session followed by voluntary return recomputed
- R7Loyalty / AI-friendliness as a trait of the person refused
The refusals ship as rows. J10, C10 and R7 appear on every report, each with its reason. The pack declines to claim that the agent caused an outcome, that anyone felt satisfied, or that a person has a trait. Those refusals are part of why the other rows can be believed.
What it is: a versioned file your book is judged against, pinned per action date. Every number carries the count it covers, beside the count it confirms.
What it is not: not legal advice, and not a score of the agent or of the people it served.
See it run on real transcripts →