The regulator has said what "human oversight" does not mean. Reviewing thirty UK employers' use of automated decision-making, the ICO found that many employers believed they were using their AI tools only for decision support, when in practice there was no meaningful human involvement in the outcome at all [ICO, Recruitment Rewired, March 2026]. This applies a standard the ICO has held since before the 2025 reforms to UK data protection law: human review which amounts to a "token gesture" does not satisfy the safeguards required under UK GDPR. Its 2026 draft guidance tightens the test further, requiring the reviewer to be suitably trained and to hold genuine authority to depart from the system's recommendation. That draft is still under consultation; the finalised wording may differ, and the statutory code itself is not expected until 2027.
The exposure is no longer prospective for solicitors. On 17 August 2026 the SRA issued a formal warning notice on AI misuse, disclosing alongside it that it had received 42 related reports in the preceding year, with live investigations into inaccurate citations, supervision failures and confidentiality breaches [SRA, Warning Notice: Misuse of AI, and accompanying press release, 17 August 2026]. The notice cites four judgments from 2025-26 in which courts held authorised persons personally responsible for unchecked AI hallucinations, regardless of the tools used to prepare submissions.
The FCA has confirmed its direction without confirming new rules. The Mills Review, published in full on 6 July 2026, concludes that existing frameworks including SM&CR and Consumer Duty remain the right foundation but signals they will come under increasing pressure as AI moves along a five-stage autonomy spectrum, from Operator through to Observer [FCA, The Mills Review, July 2026]. Among its seven recommendations is a proposed Agentic Supervisory Model for continuous, AI-enabled supervision, in place of today's episodic, document-based approach. These are recommendations to the FCA Board, not yet adopted policy. The FCA's formal response, and a promised publication on "good and poor practice," are both still pending.
The frameworks firms are pointed toward were not built for what AI is becoming. ISO 42001 and the NIST AI Risk Management Framework remain the reference standards firms are told to align with, but both were finalised before agentic AI deployment scaled, and neither specifically addresses risks that come with delegated decision-making, tool authorisation, or action taken without a human in the loop at each step [Cloud Security Alliance, Agentic AI Governance research, 2026]. NIST's own AI Agent Standards Initiative, launched February 2026, exists because the current standards landscape has not caught up.
The obligations firms assumed were imminent have moved, not disappeared. The EU AI Act's high-risk (Annex III) obligations, originally due 2 August 2026, are currently set to apply from 2 December 2027, following the Digital Omnibus's entry into force on 27 July 2026 [Regulation (EU) 2026/1744]. Article 50 transparency obligations have not been deferred in the same way as the high-risk provisions, though the amended Regulation includes transitional arrangements for certain systems already on the market. This timeline has already shifted once and remains politically set; treat it as the current position, not a fixed one.
International direction is converging on proportionality, not uniformity. Singapore's updated Model AI Governance Framework for Agentic AI (May 2026) applies risk-proportional oversight scaled to the deployer's context rather than a single bar for every organisation [IMDA, May 2026]. No AI governance standard has yet done the same for the UK's regulated SME deployer, the firm too small for enterprise consultancy and too exposed to do nothing.