Share with your CISO
Hong Kong’s Privacy Commissioner for Personal Data just published its 2026 AI compliance check findings, covering 60 organisations across 16 sectors. AI adoption jumped from 80% to 95% year-over-year, and data minimisation practices improved sharply, with organisations retaining personal data collected through AI dropping from 79% to 29%. The concerning reversal is at the governance layer: the share of organisations with formulated AI policies fell from 63% to 50%, and board-level AI discussions dropped by 25 percentage points, even as deployment accelerated. Agentic AI received its own risk category for the first time.
What this means for your business
The organisations most exposed here aren’t the ones ignoring AI governance entirely; they’re the ones that stood up AI programs fast, assumed operational controls were sufficient, and let formal policy documentation slide. If your organisation runs multiple AI systems, has employees using generative AI tools, and still lacks a written enterprise AI policy signed off at board level, you’re in the majority of the 2026 cohort, and the PCPD has now documented that gap explicitly as a governance regression, not an acceptable tradeoff.
The agentic AI flag deserves more attention than it typically gets in compliance roundups. A conventional chatbot processes a query and returns text. An agentic AI system, by contrast, is granted standing access to files, email, system resources, and third-party plugins, then executes multi-step tasks autonomously between human check-ins. That architecture means a misconfigured permission set or an unvetted plugin isn’t a data exposure risk in theory; it’s a live attack surface. The PCPD’s guidance on least-privilege access, environment segregation, and plugin vetting isn’t aspirational. It’s the minimum viable control set for any agentic deployment touching personal data.
The governance gap the 2026 data reveals has a predictable failure mode: organisations treat technical controls as a substitute for documented accountability. Security measures, privacy impact assessments, and human-in-the-loop oversight all held steady or improved, but the policy and board-level structures that would allow a regulator or an auditor to reconstruct who decided what, and why, weakened. That asymmetry matters most when something goes wrong. A breach response plan that covers AI incidents specifically (only 41% of organisations with plans had this) is worth auditing now against your actual deployed systems, not the systems you had when the plan was written.
Concept deep-dive: Human-in-the-loop
Human-in-the-loop means a human must review and approve an AI system’s output before it triggers a real-world action, analogous to a surgeon reviewing imaging analysis before making an incision rather than after. The alternative, human-in-command, allows AI to act autonomously but keeps a human empowered to intervene. The distinction matters operationally because agentic AI systems, which execute sequences of decisions across files and systems without pausing, are structurally incompatible with genuine human-in-the-loop control unless that checkpoint is explicitly engineered into the workflow.
Based on reporting from Hong Kong Privacy Commissioner for Personal Data Completes its 2026 AI Compliance Checks: Findings, Trends and the Rise of Agentic AI | Insights, originally published 2026-07-03 07:40:00.
