EU AI Act logging obligations for high-risk AI shift to December 2027

The date moved, the requirement did not

A dev.to post published in September 2026 reports that the Digital Omnibus on AI, formally Regulation (EU) 2026/1744, appeared in the Official Journal on 24 July 2026 and took effect three days later. Under the amended timetable, obligations for high-risk AI systems — including record-keeping — now apply from 2 December 2027 for stand-alone systems listed in Annex III, and from 2 August 2028 for AI built into regulated products. The post also flags that material circulated over the summer citing an August 2026 deadline should now be treated as outdated.

The practical takeaway, according to the author, is that the deadline slipped but the substantive duty did not. Teams that deferred logging work because the dates were in flux have bought time, not an exemption.

What Article 12 asks for

The post describes Article 12 of the AI Act as short but demanding in effect: a high-risk system must be technically capable of automatically recording events throughout its lifetime, and those records must be detailed enough to reconstruct an incident after something goes wrong, to support monitoring once the system is on the market, and to show how the system has been operating. Crucially, the author argues, this is a requirement about what the system can capture — not about what an analytics dashboard happens to display.

Where current agent logging falls short

Most agent stacks emit logs, the post argues, but very few keep records, and the difference shows up in what a typical log line omits:

• Identity: lines name a service account, not the agent, its permission scopes, or the human behind the request.
• Intent: the prompt is stored, but the actual tool call with its arguments often is not, and nothing links the two.
• Justification: entries show a call executed, but not which policy, at which version, permitted it — or why a blocked call was blocked.
• Outcome: results land in a separate log with a different timestamp.
• Integrity: text logs can be edited, trimmed or lost with no way to detect it afterwards.

On top of that, the author notes, the logs frequently live in a vendor's cloud, so when an auditor asks a question, the answer effectively belongs to someone else.

One open-source response

The post goes on to describe Verax, an Apache-2.0 licensed MCP server positioned between an agent and its tools. In the author's description, every tool call passes a policy gate and produces a signed decision record — including the policy hash and a reason code, with refusals recorded the same way as approvals — before anything runs. Executed calls leave an effect row that is reconciled against the decision that authorised it, records are hash-chained together, calls the policy cannot resolve alone are held for a local human operator, and the ledger stays on the machine rather than in a vendor's cloud.

The authors report engineering work measured on a single laptop: 150 agents generating roughly 200,000 synthetic decisions over 30 simulated days, with the ledger cut into 50,000-row segments so that a restart reads the open segment and a persistent index — about 1.9 seconds and 518 MB at the 200,000-record mark.

The post is candid about limits, and they are worth repeating: the same system both keeps and signs records, with no independent witness yet; there is no tenant isolation; payloads are stored in plaintext; no external audit has been carried out; and whether any given deployment actually satisfies Article 12 is a conformity judgement for an assessor, not a claim the project makes.

Why it matters

The delay to December 2027 reshuffles compliance planning for anyone building or deploying high-risk AI systems in the EU, and the later date of August 2028 for embedded systems matters to product manufacturers in regulated sectors. But the underlying requirement — durable, tamper-evident, traceable records tying identity, request, decision and outcome together — is unchanged, and the dev.to post makes a credible case that mainstream agent infrastructure does not currently provide it. Teams with EU exposure now have a clearer runway to close that gap, and the clock starts running more than a year before the first deadline arrives.