EU AI Act deployer obligations now begin on 2 December 2027 — sixteen months after the date the regulation’s own timetable set. The Commission’s implementation page states the new date in a single unremarked line, and the line is worth reading twice for how little it explains: the delay is asserted, not reasoned.
Because the obligations that moved are the ones that would have made somebody keep a record. Article 12 requires a high-risk AI system to log its own events. Article 26(6) requires the deployer — the party using the system, which is what an agency is — to keep those logs for at least six months. The Commission’s own summary of what high-risk obligations contain lists, verbatim, “logging of activity to ensure traceability of results.”
All of it now lands with the Annex III list: biometrics, critical infrastructure, education, employment, access to essential private and public services — credit scoring and insurance pricing sit there — law enforcement, migration and border control, and the administration of justice and democratic processes. Eight areas. Advertising is not one of them. Europe’s AI law has deferred its record-keeping half — and even on arrival that half is addressed to a different industry. The half that went live in August asks for a label.
What the Omnibus moved, and what it left standing
The AI Act entered into force on 1 August 2024 and became applicable on 2 August 2026. Article 113 then carves out the exceptions, and the list is short enough to quote in full:
(a) Chapters I and II shall apply from 2 February 2025; (b) Chapter III Section 4, Chapter V, Chapter VII and Chapter XII and Article 78 shall apply from 2 August 2025, with the exception of Article 101; (c) Article 6(1) and the corresponding obligations in this Regulation shall apply from 2 August 2027.
Read that list for what is missing from it. Chapter IV holds the transparency obligations, Article 50 among them, and Chapter IV is not exempted. The Commission says the same in plainer language: “The transparency rules of the AI Act will come into effect in August 2026.” That half is live — it has been since 2 August 2026.
The half built on evidence is not. The rules for high-risk use cases in certain sensitive areas, the Commission writes, begin “starting on 2 December 2027”, when high-risk systems become “subject to strict obligations before they can be put on the market.” The page then lists what starts on that date, and the third item is “logging of activity to ensure traceability of results.” Sixteen months of extra runway, and not a sentence on the page saying why.
None of that is a scandal — a regulation arriving before its own standards is one nobody can comply with. But an agency owner reading the autumn coverage could conclude the opposite: that Europe had handed the ad account a logging duty. It deferred that duty and left the disclosure duty running.
Read the article that required the record
Europe came close to writing an audit law for the people who operate AI. Read Article 12, which opens with this:
High-risk AI systems shall technically allow for the automatic recording of events (logs) over the lifetime of the system.
Article 12 puts the logging capability in the system. Article 26(6) puts the retention obligation on you:
“Deployers of high-risk AI systems shall keep the logs automatically generated by that high-risk AI system to the extent such logs are under their control, for a period appropriate to the intended purpose of the high-risk AI system, of at least six months…”
Six months is the floor, not the ceiling — “appropriate to the intended purpose,” and longer where other law says so. Then Article 26(6) does something an agency should notice, because it shows how carefully this was drafted for a regulated firm:
Deployers that are financial institutions subject to requirements regarding their internal governance, arrangements or processes under Union financial services law shall maintain the logs as part of the documentation kept pursuant to the relevant Union financial service law.
Brussels knew how to write a record duty for a regulated finance business. It wrote one, named the retention floor, and folded the logs into documentation a firm already keeps. The same instinct runs through Article 26(2), which assigns human oversight to real people — “the necessary competence, training and authority” — and through Article 26(5), where the monitoring duty on financial institutions is deemed satisfied by the internal-governance rules they already follow.
The definitions matter too. Article 3(4) defines a deployer as “a natural or legal person, public authority, agency or other body using an AI system under its authority.” Read the first limb, not the third: in this drafting, “agency” means a public body. What makes a marketing agency a deployer is the legal-person limb — you use an AI system under your authority, on your clients’ accounts. That definition has been live since 2 February 2025.
“Advertising” appears three times in 588,000 characters
That is not a rhetorical flourish; it is a scan you can repeat. Across the 588,000-character text of Regulation (EU) 2024/1689, the string advertis occurs three times. Here they are, in order:
- A recital, reassuring the market that legitimate commercial practice is not manipulation: “common and legitimate commercial practices, for example in the field of advertising, that comply with the applicable law should not, in themselves, be regarded as constituting harmful manipulative AI-enabled practices.”
- A footnote — number 34 — citing Regulation (EU) 2024/900 on the transparency and targeting of political advertising.
- Annex III, area 4(a), where the advertisement in question is a job advertisement: AI systems used “to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates.”
None of the three creates an obligation. Two are exemptions from somebody else’s rule and the third is about hiring. For scale, the same text uses deployer 158 times, logs 19 times, human oversight 15 times, six months 7 times and audit 6 times — and not one of those six “audit” hits is about advertising either.
This is a defensible design choice rather than an oversight. The Act writes its heaviest obligations for systems that decide a person’s access to essential services: can this person borrow, be insured, get the job, the benefit, the visa. Advertising decides none of those things. It decides what someone sees. If you accept the Act’s taxonomy the omission follows — and it still leaves your agency carrying the consequence of what runs in a client’s account.
Article 50 asks for a label — and hands it back when a human signs
The transparency half reached advertising, so read it exactly. Article 50(1) obliges providers to tell people they are interacting with an AI system; Article 50(2) obliges them to mark synthetic output in a machine-readable format. Both sit on the vendor. The deployer duty is Article 50(4), and its first limb is about synthetic media:
Deployers of an AI system that generates or manipulates image, audio or video content constituting a deep fake, shall disclose that the content has been artificially generated or manipulated.
The second is the only limb anywhere in the Act that addresses AI-written text:
Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated. This obligation shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offences or where the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.
Read the two conditions together and you can see why nobody in ad operations has felt a new obligation land. The duty attaches to text published “with the purpose of informing the public on matters of public interest” — a category a performance ad for a funded trading account is not obviously in. And where it does attach, it disappears the moment the content “has undergone a process of human review or editorial control” and a natural or legal person holds editorial responsibility for publishing it.
That is the sentence to sit with. The Act’s one text-disclosure duty is switched off by a human taking responsibility — and it asks for no record of that human. No name, no timestamp, no before-and-after, no export, no scope to a client or a jurisdiction. The transparency requirement is satisfied by the fact that review happened; the review leaves nothing behind. If two people on your team disagree next March about who approved the wording of a promotion, the AI Act will not settle it. It was never asked to.
Meanwhile the systems you connect ship a setting, not a record
The tooling layer has the same shape as the regulation, which is why the two are worth reading side by side. Start with what is genuinely good, because it is genuinely good.
- Supermetrics locks the creation half. Its campaign documentation states plainly: “New campaigns are always created with a paused status so you can review them before they go live.” A campaign cannot be born live — the strictest creation default in this cohort, and not configurable.
- It then publishes the approval menu. Four options, verbatim: “Going live: Approve before a campaign goes live, including unpausing one.” “Budget + going live: Approve before budget changes on live campaigns and before a campaign goes live.” “All live changes: Approve any changes to a live campaign (including enabling the campaign).”
- And the first option is the one to read. “No human approval: Changes apply immediately.”
Across that product’s write documentation the vocabulary scan returns approv 6 times, approver 0, audit 0, compliance 0, jurisdiction 0, retention 0 and export 0. The gate is a setting; the setting has a position that produces no authorisation event, so there is nothing to record. A live budget change on a client’s account then applies immediately, with no artefact behind it. One of the four positions is silence.
Underneath both, the protocol layer moved the same way. The 2026-07-28 revision of the Model Context Protocol deprecated its Logging feature under SEP-2577, with this migration advice: “log to stderr (stdio) or use OpenTelemetry instead of Logging.” Read that as an instruction to the operator: the record, if it exists, is now something you run yourself. Nothing in the revision requires a server to record whose write it was or which human approved it.
The question the Act does not answer
On 14 March, a budget increase goes live on a client’s brand-search campaign. On 2 June, that client’s compliance lead asks three questions: who approved it, what the value was before, and which regime the client sat under. Here is every surface that might hold an answer.
| Surface | What it actually covers | When it applies |
|---|---|---|
| EU AI Act · Article 12 + 26(6) | Automatic logs over the lifetime of a high-risk system, kept by the deployer for at least six months | 2 December 2027 — and only for Annex III systems, which exclude advertising |
| EU AI Act · Article 50(4) | A disclosure that content was AI-generated — switched off by human review plus editorial responsibility | Live since 2 August 2026 |
| Ad platform change history | That a change happened, in the platform’s account — it names no approver and leaves with the account | Always, while the account exists |
| Vendor approval setting | Whether a human gate fires at all — one of four positions is “no human approval” | Per setting, per account |
| Ott · Activity Logging | Every campaign change: the actor the platform reports (including AI-agent writes), before/after values, stored independently of the platform | Always — no retention tier |
The point is narrower than a criticism: a clause telling you to keep six months of logs for a credit-scoring model does not travel to a campaign rewritten on a Tuesday, and a rule that dissolves when a human signs off does not name the human.
What an operations record looks like when it is the product
Ott’s answer is the half of the architecture nobody puts on a roadmap slide, because it does not demo well:
- Activity Logging records every campaign change — pauses, budget edits, creative swaps, bid and review changes — timestamped, attributed to the actor the platform reports, with before-and-after values. It is stored independently of the platform that accepted the change, and carries no retention tier to reconcile.
- Budget Ledger keeps one prepaid ledger per brand: top-ups, manual corrections and balance resets are dated entries with notes, so a question about a client’s balance has a written trail rather than a spreadsheet.
- Agency Hierarchy runs Client → Brand → Account, so a record is filed against the engagement it belongs to rather than against whoever owns the ad account. The AI Act scopes its obligations by list; your record has to scope itself by client.
- Client View lets a brand read its own slice without a shared login and without seeing anything beyond its own accounts, which turns “we can produce a record” into something a compliance lead can verify alone.
To be exact about scope: Ott does not decide whether an AI system is high-risk, and nothing here is legal advice on your obligations under Regulation (EU) 2024/1689 — that turns on what you deploy and for what purpose, which is a question for your counsel. What Ott does is make the record exist before anyone asks for it, on the theory that the statutory question and the operational question eventually point at the same file. Read the earlier argument about why agent governance stopped at the transport layer, and the SEC’s crypto-asset proposal for how ad copy becomes classification evidence in a second regime.
- Which of your deployments is actually Annex III? Creditworthiness, life or health insurance pricing, recruitment screening — if none is on your list, the record duty was never addressed to you.
- If a client’s AI tool does fall in a high-risk area, who is the deployer? The definition turns on who uses the system under their own authority. Settle that in the contract.
- Does any live setting in your stack have a position that produces no evidence? Find the option on every approval menu that reads like silence.
- Can you name the human who reviewed the AI-drafted copy? If the Act’s one text duty turns on human review, the review is the control — and it is the least recorded thing in the workflow.
- Where does the ad platform’s change history live, and what happens to it if the account is the thing that gets restricted? A record that dies with the account is not a record.
- Can you hand one client’s campaign changes to an auditor without a screenshot walk-through? A record you can only read on a screen is not a record.
The honest summary is that Brussels bought everyone sixteen months and left the disclosure rules running, so the compliance question facing a finance agency is not yet a statutory one about AI. It is the older question about advertising: who moved the budget, on whose account, under which rulebook, and what was the value before. That question has an answer in every jurisdiction, and it has never been answered by a retention window or a label.