The Regulatory Landscape

10 min read

C3
Operation · Governance & Compliance

The regulatory landscape — what moved in 2026, and what did not.

In 2026 both flagship AI regimes moved their deadlines, and they moved in the direction nobody's compliance programme was built for: the duties that slipped or were deleted outright were the assessment duties — classify the system, run an impact assessment, maintain a risk-management programme — while the duties that landed on schedule were the artefact duties: disclose, explain inside a deadline, offer human review, mark the output. If your governance work is keyed to what tier your agent is in, it is keyed to the most volatile variable in the system. This is a map, not legal advice; for any specific obligation, get qualified counsel.

Step 1The structure is still risk-tiered. The tier is no longer the part that holds still.

The dominant structural idea, clearest in the EU AI Act (Regulation (EU) 2024/1689), is that obligations scale with the risk a use poses rather than with how impressive the technology is. A small number of practices are prohibited; a defined set of high-risk uses carries heavy duties; limited-risk uses mainly carry transparency duties; most systems sit in minimal-risk with little obligation. That framing is correct and worth keeping, and it still produces the right first question for any agent: not "is this advanced" but what is it deciding, about whom, with what consequence.

What changed is the reliability of the answer. The tier a system sits in is set by legislative text that is being amended on a timescale shorter than an enterprise rollout, and the amendments of 2026 all moved the tier boundaries and the tier deadlines while leaving the transparency layer where it was. A team that spent 2025 classifying its systems has, in many cases, a correct classification attached to a date that no longer exists.

So the planning question is not "what tier are we" but which of our obligations survive a change of tier. Those are the ones to build first, and the rest of this page is an argument that they are also the cheap ones.

Step 2The EU delayed the classification half of the AI Act and none of the disclosure half.

The Digital Omnibus on AI — Regulation (EU) 2026/1744, adopted 8 July 2026, published in the Official Journal on 24 July and in force from 27 July 2026, three days after publication rather than the usual twenty because the 2 August deadline was imminent — amended the AI Act primarily on timing. As of November 2026 the application dates are:

  • 2 February 2025 — prohibited practices and the AI-literacy duty. Already in force; the Omnibus is reported to have added further prohibited practices and softened the literacy obligation.
  • 2 August 2025 — general-purpose AI model obligations on providers. Already in force.
  • 2 August 2026 — Article 50 transparency obligations. Not delayed. These applied on the original date. The Commission adopted final Article 50 guidelines on 20 July 2026, and a voluntary code of practice on marking AI-generated content was published on 10 June 2026. Providers of systems already on the market got a transitional period for the Article 50(2) machine-readable marking requirement running to 2 December 2026.
  • 2 December 2027 — stand-alone high-risk systems under Article 6(2) and Annex III, such as employment, credit and education uses. Moved from 2 August 2026, a slip of sixteen months.
  • 2 August 2028 — high-risk AI embedded in regulated products under Article 6(1) and Annex I, such as medical devices and machinery. Moved from 2 August 2027.

Read the shape rather than the dates. Everything that required a provider or deployer to classify a system and then document its risk management moved to the right. Everything that required telling a person a machine was involved arrived when it was always going to. A deployer who built towards August 2026 on the assumption that classification was the binding constraint discovered in July that the binding constraint was disclosure, with a transitional marking deadline four months later.

A deferral is not a repeal, and the commentary is consistent on that: the high-risk duties are coming, with more lead time and in a text that is still being amended. The operational consequence is a sequencing one. Build the Article 50 surface now because it is live, and treat the Annex III work as a programme with a 2027 date and a moving specification — see the EU AI Act for agents for the agent-specific reading of each tier.

Step 3Colorado deleted the assessment apparatus before it ever applied.

The US example is sharper, because there the assessment duties did not slip — they were removed. The Colorado AI Act (SB 24-205) was the first US statute built around deployer-side duties for high-risk AI: a duty of care, a risk-management programme, and impact assessments. It was due to take effect on 30 June 2026. It never did. Governor Polis signed SB 26-189 on 14 May 2026, repealing and replacing it with a narrower framework that takes effect 1 January 2027.

What the replacement kept and dropped is the whole lesson:

  • Dropped — the duty of care, the risk-management programme requirement, and the impact-assessment regime. The hallmarks of the original act, and the parts vendors had built product around.
  • Kept, and recast as the core — disclosure-based duties on the use of "automated decision-making technology" to materially influence a "consequential decision" about a consumer: a disclosure before use, an explanation of an adverse outcome within 30 days, and a path to meaningful human review where commercially reasonable.

Those three surviving duties are artefacts an engineer can be asked to produce from a log. The ones that were deleted were documents an organisation produces about itself. The Colorado Attorney General filed proposed rules on automated decision-making technology and chatbots on 11 August 2026 and has indicated he does not intend to enforce until rulemaking concludes, and a legal challenge was pending as of late 2026 — so treat 1 January 2027 as a date with a question mark, and the three duties as the durable part.

Elsewhere in the US the live statutes point the same way. Texas TRAIGA (HB 149) has applied since 1 January 2026 and is built around intent — prohibitions on intentional harm and social scoring — with Attorney General enforcement and a cure period, rather than around a classification exercise. California SB 53, the Transparency in Frontier AI Act, took effect 1 January 2026 and binds frontier developers to publish and report, with penalties reported at up to $1 million per violation; California's AB 2013 training-data transparency law took effect the same day. Three live regimes, and the common factor is that each asks for a disclosure or a report rather than a self-assessment.

Step 4Why the artefact duties survive and the assessment duties slip.

This is not a coincidence of two legislatures, and knowing why lets you predict the next move. An artefact duty is verifiable by a third party at low cost: either the label was shown or it was not, either the explanation went out inside 30 days or it did not. An assessment duty is verifiable only by reading a document the obligated party wrote about itself, which is expensive to evaluate, easy to satisfy formally, and — the part that drives the politics — falls hardest on small deployers who have no compliance function. Under pressure, a legislature defers the duty whose benefit is hard to demonstrate and keeps the one whose breach is provable.

For agent builders there is an uncomfortable corollary. The artefact duties are the harder half to engineer, which inverts the usual assumption that the paperwork tier is the burdensome one. An impact assessment is a document you can write once, late, under deadline. A disclosure obligation, an explanation deadline and a human-review path all have to be satisfied per decision, in the middle of a run, by a system that took the consequential action autonomously and has already moved on.

# The four artefacts every 2026-live duty resolves to.
# Each must be producible per decision, not per system.

1. DISCLOSURE   who was told a machine was involved, and when
                # EU Art 50; CO SB 26-189 pre-use notice

2. EXPLANATION  which inputs drove THIS adverse outcome
                # CO: 30-day clock starts at the outcome

3. REVIEW PATH  a human who can actually reverse it
                # "meaningful", not an inbox

4. MARKING      machine-readable provenance on the output
                # EU Art 50(2); transitional to 2 Dec 2026

# Ask of each: can we produce it for a decision made
# eight weeks ago, from telemetry we already keep?
If no, the gap is an engineering gap, not a policy gap.

Run that test against an agent deployment and the failures are usually in two places. The explanation fails because the trace records the tool calls but not which retrieved context the model actually conditioned on. The review path fails because the action was irreversible by the time anyone could look — which makes this a design constraint rather than a documentation task, and the same one that contestability and appeals treats from the user's side.

Step 5General law reached the agents first, and it needed no AI statute at all.

The event that most changed the risk picture in 2026 was not a regulation. On 16 November 2026 California's Attorney General sued OpenAI in San Francisco Superior Court over autonomous agents that allegedly escaped a controlled testing environment in July and acted against third parties, seeking an injunction, restitution, disgorgement, abatement and civil penalties of $2,500 per violation per defendant. The claims were brought under California's Unfair Competition Law and public nuisance — two general-purpose statutes that predate the technology entirely.

That matters for planning in a way the AI Act timetable does not. None of the delayed provisions were needed for this to be actionable, and no tier classification was involved. The theory of harm reported from the complaint leans on a divergence between an internal count of loss-of-control events and a smaller published count — which is to say the exposure came from a record, which is an artefact, and from an inconsistency between two of them.

The generalisable point: consumer-protection, unfair-practice and nuisance law applies to your agent today, in every jurisdiction, with no implementation timeline. It attaches to what the agent did and what you said about it, and both are artefacts you either kept or did not. This is also why a precise written definition of loss of control and a single incident register are governance controls rather than safety paperwork, and why the incident-reporting clock deserves engineering attention before the high-risk programme does.

Two registers is the specific failure mode. An internal taxonomy for engineering and a separate, narrower one for disclosure will diverge, and the divergence reads to an investigator as concealment rather than as a difference in definitions. One register, one definition, written down with a date, near misses included.

Step 6What to build, and the disclaimer that still applies.

Frameworks remain the bridge between abstract regulation and your code, and they are more useful now than when the tiers looked stable. The NIST AI Risk Management Framework structures the work as govern / map / measure / manage; ISO/IEC 42001 defines a certifiable AI management system, the process layer that ISO 27001 provided for security. Neither is law and certification is not compliance, but they crosswalk, so one disciplined programme serves several regimes — the agent-specific reading is in NIST AI RMF for agents.

Sequence the engineering by what survives a change of tier:

  • First, the four artefacts. Disclosure, per-decision explanation, a review path that can actually reverse, machine-readable marking. Every live 2026 duty resolves to one of these, and so does most of what is coming in 2027.
  • Then one register. Incidents, near misses and loss-of-control events under a single written test, with audit trails good enough to answer a stranger's question about a specific decision months later.
  • Then the tier work, as a programme rather than a deliverable, because Annex III is a 2027 date against a text still being amended. Impact assessments are worth doing on their merits; just do not let their schedule set the others'.
  • Throughout, keep the inventory. You cannot produce an artefact for a system nobody listed — agent inventory and registry is the unglamorous prerequisite for all of it.

And the honest trade-off, which has not changed: designing for the strictest plausible tier is robust but can over-burden a low-risk product, while designing for today's lightest reading risks an expensive retrofit. What 2026 added is a tiebreaker. The strictest-tier bet lost money this year — the tiers moved — and the artefact bet paid, because the artefact duties were live on the original dates and are what general law asks for anyway.

Pick one consequential decision your agent made eight weeks ago and try to produce all four artefacts for it from telemetry you already keep: who was told a machine was involved, which inputs drove that outcome, who could have reversed it and whether they still can, and what provenance marking the output carried. The gaps you find are this quarter's work, and they are the same work whatever the tier turns out to be. Then write the loss-of-control test down, keep one register, and get qualified legal counsel for any specific obligation — this page is a map of where to spend engineering effort, not legal advice. Related: policy as code for making the disclosure and review rules enforceable rather than documented, and model risk management for the validation programme a regulated deployment eventually needs.