Post-Enforcement Reality

27 August 2026 — A month into enforcement, the market has begun to separate, and the fault line is not size or sector.

It is whether an organisation can produce a list.

Businesses that can enumerate their AI systems — what they run, what data each touches, who owns it — are moving through the Act's obligations at a pace that surprises them. Having the list makes classification quick, makes vendor review tractable, and makes monitoring a matter of adding rather than inventing. Businesses that cannot produce the list are discovering that every requirement loops back to the same question they cannot answer.

The week's technical guidance pointed in an unusually practical direction: sandboxing strategies for agents. The principle is simple and worth stating plainly, because it is counterintuitive to people who assume agents should be helpful. An agent should be given a box, not a building. The box should contain exactly the tools needed for the task in hand, exactly the data that task requires, and no route out. Every escape documented this year — and there have been several — exploited a permission that was granted for convenience and never revisited.

Some of the framing in vendor material deserves scrutiny. "Sandboxing" as commonly used often means running something in a container with full network egress and broad filesystem access. That is not a sandbox. That is a different office.

ROI benchmarks continued to support the same conclusion: the reliable returns sit in processes with existing baselines. Nothing in this month's data changed that.

What this means if you run an SME. Test the box. Give an agent a task that requires something just outside its intended scope and see whether it stays put. This is a half-hour exercise, it costs nothing, and it will tell you more about your actual exposure than any specification document a vendor provides.

Mid-week review, in parallel with the Monday series.