6 Comments
User's avatar
State of Play's avatar

Thanks for this. The detail that turns the week-one attachment argument from a compliance nicety into a continuous operational problem is the rate: KPMG's Q2 2026 AI Pulse survey (2,145 C-suite respondents, 20 countries) found employee AI-agent adoption jumped from 23% to 56% in a single quarter. A register is a stock control being asked to track a flow that just doubled in three months, so Art 25(1)(c) provider-conversion is happening continuously across the org, not as the rare edge case the register format assumes. The leaver-checklist gap you flag is the same problem on the exit side: the population needing offboarding is growing exactly as fast as the population needing onboarding.

Silvia Stepitova's avatar

Stock control against a flow is a good way to put it. Do you have a link for the 23 to 56 figure though? I went looking in KPMG's Q2 2026 AI Pulse (https://assets.kpmg.com/content/dam/kpmgsites/xx/pdf/2026/06/global-ai-pulse-q2.pdf) and couldn't place it. Page 12 has employee adoption of AI agents going from 25% to 28%, up 3 points. If there's a cut showing 56% I'd genuinely want it, that changes the picture quite a bit.

State of Play's avatar

Sorry about that, I'd conflated two different KPMG reports in my notes— the 23%→56% figure is from the banking industry spotlight of KPMG's Q2 2026 AI Quarterly Pulse Survey (https://kpmg.com/kpmg-us/content/dam/kpmg/corporate-communications/pdf/2026/BANKING_2026_AIPulseSurvey_Q2.pdf): that's 100 US-based C-suite banking leaders, not the 2,145-respondent global sample I attributed it to. Your 25%→28% is the correct number for the broader population.

That changes what the evidence can carry. A single-sector, 100-respondent US spotlight isn't a strong enough anchor for "continuous operational problem" at the scale I framed it — it shows one high-adoption sector moving fast in one quarter, not a global doubling. The stock-vs-flow structure of the argument still holds, it's just a slower flow than I made it sound.

Silvia Stepitova's avatar

Appreciate you going back and checking. A US banking sample moving that fast is interesting on its own, just a different claim. The stock-versus-flow point stands either way.

Mike Schlottman's avatar

This is exactly the piece GRC teams need before their next AI Act gap assessment. The register point is what will actually bite: you cannot classify or disclose agents you don't know exist, and that gap is the same shadow IT problem I chased for years before it had an AI Act price tag. Article 25 turning a deployer into a provider without anyone approving anything is the sharpest trap in the piece.

Silvia Stepitova's avatar

"Before it had an AI Act price tag" is great. I might steal that. It really is the same problem, someone just attached fines to it. One thing to check before the next gap assessment: when Article 25 flips the role, don't count on Microsoft handing over documentation. Their terms may say Copilot was never meant to go high-risk, and then you're on your own.