AI agents can already make ad creative, and they can run many of the checks that sit between a draft and a live ad, but they cannot carry the accountability that approval means. UK regulators, industry codes and the agent protocols themselves all leave final responsibility with the advertiser. The practical answer is to let agents generate and pre-check creative, then place a named human at the points where a claim, a likeness or a context could mislead.

TL;DR. Ad platforms and open protocols now let agents generate images, video and copy from a brief. The UK's advertising regulator says the advertiser remains responsible for any AI output used in an ad, even when the ad came from an automated tool. The workable model is tiered: agents handle mechanical adaptations within approved bounds, while a person signs off new claims, new imagery, synthetic people and anything regulated.

Can AI agents create ad creative today?

Yes, and the capability now sits inside mainstream buying tools rather than separate design software. In October 2026, MediaPost reported that Meta is testing agentic capabilities through which advertisers can plan and create a campaign and generate creative images and text by conversational command. The same report says Advantage+ creative can now generate video from existing static assets.

Retail media has moved the same way. AdExchanger reported in August 2026 that Cuisinart beta tested Amazon's video creative agent, with the team fine-tuning one video from a shortlist the agent proposed. Kelsey Smithuysen, Amazon marketing director at Conair, told the publication: "Another year from now, we can definitely see ourselves using the videos just as they are coming out of the AI tool if they continue to progress in the way they've been progressing."

The open standards cover creative too. Alongside IAB Tech Lab's Agentic Advertising Management Protocols (AAMP), which frame how buyer and seller agents trade media, the Ad Context Protocol (AdCP) published by AgenticAdvertising.org defines a build_creative task that "generates, transforms, or retrieves ad creative manifests" from a natural language brief to production-ready assets. Creation, in short, is no longer the bottleneck.

Can an AI agent approve its own creative?

An agent can pass or fail creative against rules, but it cannot take responsibility for the result, and responsibility is what approval means in advertising. The ASA's June 2026 guidance, AI and Deepfakes: Four Things Advertisers Need to Know, states that "the advertiser remains responsible for any output used in their ads". Citing its Stripe & Stare ruling, the ASA adds that even when ads are generated or distributed through automated AI platforms, "responsibility does not shift".

Platform terms point the same way. Digiday reported in June 2026 that OpenAI's updated ad tools terms describe Creative Tools that can "generate, modify, transform" advertising creatives, while making the advertiser "responsible" for reviewing them. The IPA and ISBA principles for generative AI ask for "appropriate human oversight and accountability", including fact and permission checking before AI output is used.

The protocols build the pause in. AdCP's build_creative lets a creative agent return an input-required status with the reason code APPROVAL_REQUIRED, defined as "Creative needs human approval before finalizing". An agent can therefore stop and wait for a person without breaking the workflow.

On the media side, Alkimi is a worked example of the same separation: in its Marketplace humans keep approval rights at the deal level, and in its Gym buyers and sellers set the rules their agents work to and spar against counterparty agents before any real money moves. Creative needs an equivalent line, because an agent that agrees a placement never sees what will render in it, a gap explored in creative delivery as the last human mile.

How do you set up a creative approval workflow for AI agents?

The eight steps below assume an agent that can draft and adapt creative and a team that wants speed without handing over accountability. Each step stands alone, so teams can adopt them in order or pick the ones that close their biggest gap.

Step 1: Write the creative mandate before the agent writes anything

The mandate is the written set of brand inputs and limits the agent works within: approved claims, banned claims, logos, colours, tone, mandatory legal lines and subjects to avoid. AdCP's build_creative accepts a brand reference that resolves brand identity, including colours, logos and tone, from a file published on the brand's own domain. The fewer judgements an agent has to invent, the fewer a reviewer has to catch.

Step 2: Sort creative into risk tiers

Not every asset carries the same risk, so approval effort should follow risk rather than volume. The IAB's AI Transparency and Disclosure Framework Version 2, announced on 18 August 2026, lists images and videos generated from prompts, synthetic avatars and some digital twins among the higher-risk uses that warrant disclosure. A practical four-tier sort looks like this:

Tier 1: mechanical adaptations of an approved master, such as resizes, crops, format conversions and captions taken from an approved script.
Tier 2: new combinations of pre-approved elements, such as approved headlines paired with approved images.
Tier 3: new copy, new claims or new imagery generated from a prompt.
Tier 4: real or synthetic people, synthetic voices, digital twins, regulated categories and anything aimed at children.

Step 3: Map each tier to a stage of earned autonomy

Earned autonomy is a five-stage model in which an agent moves through Observe, Recommend, Draft, Human-approved action and Bounded automatic action as it proves reliable. Tier 1 work can reach Bounded automatic action with sample checks, and Tier 2 can follow once a human has approved the template set. Tier 3 should stay at Human-approved action, and Tier 4 should stay at Draft with legal or compliance review before anything ships; the five stages of earned autonomy explain how promotion between stages works.

Step 4: Automate the mechanical checks

Machines are better than people at the dull checks: file specifications, safe zones, logo placement, colour values and the presence of required legal text. AdCP requires creative agents to confirm that every required disclosure can be placed in the target format before generating, and to fail the whole request if one cannot. Digiday reported that Bayer evaluates and scores every piece of creative against benchmarks before anything goes live, a model that suits agent output well.

Step 5: Put a named person on claims, likeness and context

The human review should concentrate on what rules cannot settle. The ASA's guidance asks advertisers to sense-check AI outputs so they are socially responsible, free from harmful stereotypes and misleadingness, and appropriate for the audience and placement. Placement matters because the same image can be fine on one site and irresponsible on another, which is why context safety and creative safety need separate controls.

Step 6: Decide on disclosure deliberately

Disclosure rules now vary by market, so the decision belongs with a person who knows where the ad will run. The ASA notes there is no blanket UK legal requirement to disclose AI use, and suggests asking whether the audience would be misled without a disclosure. In the EU, Article 50 of the AI Act took effect on 2 August 2026, according to the IAB, and the European Commission's guidance says deployers must clearly label deepfakes.

Labelling everything is not the safe default either. Caroline Giegerich, VP of AI at the IAB, put it this way when the framework was updated: "labeling everything teaches consumers to ignore labels and could negatively impact advertisers."

Step 7: Record who approved what, and when

An approval nobody can retrieve later is an opinion, not a control. IAB Tech Lab's open-source seller agent, one of the AAMP reference agents, captures who decided and why with every human decision on a deal proposal. Creative approvals deserve the same treatment: the asset version, the approver, the time and the tier.

Step 8: Re-review whenever the agent changes the asset

Agents iterate quickly, and an approved asset that is then refined is a new asset. AdCP's creative agent guidance says re-submitting a creative resets the review clock, and the updated creative is treated as a new submission. Teams should apply the same rule internally, whatever tool produced the change.

Where exactly should human sign-off belong?

Human sign-off belongs at the first point where an error would reach the public or commit the brand to something it cannot take back. For most teams, that means approving the mandate and templates once, approving each new claim or generated image individually, and requiring specialist review for synthetic people and regulated categories. Everything upstream of those points, from drafting to resizing, can be delegated to agents with logging.

The common mistake is to place the human at the end of the pipeline as a final glance over hundreds of variants. A reviewer facing that volume approves by default, which is automation with a signature attached. Fewer, better-placed sign-offs give a person time to read the claim.

What does this mean for teams starting now?

Agent-made creative is arriving inside the tools teams already use, so the question is less whether to allow it than how to bound it. A written mandate, a risk tier for every asset, automated mechanical checks and a named approver for anything new or sensitive form a workflow that regulators and clients can follow. Alkimi applies the same principle to media deals: agents can prepare and negotiate, while people keep the approvals that carry accountability.

Frequently asked questions

Can an AI agent legally approve an ad on its own?

An agent can run checks and mark creative as passed, but legal and regulatory responsibility stays with the advertiser. The ASA states that responsibility does not shift even when ads are generated or distributed through automated AI platforms.

Do AI-generated ads have to be labelled in the UK?

The ASA says there is no blanket UK legal requirement to disclose AI use in ads. The test is whether the audience would be misled without a disclosure, and a label cannot rescue a misleading claim.

Does the EU AI Act require labels on AI-generated ads?

Article 50 of the AI Act took effect on 2 August 2026 and requires deployers to clearly label deepfakes. The European Commission's guidance also places marking obligations on providers of generative AI systems.

What should a human reviewer check in AI-generated creative?

The reviewer should check claims for accuracy, people and voices for permission, and the ad for stereotypes or misleading impressions. The reviewer should also confirm the creative suits the audience and placement where it will run.

Can the same agent that made the creative also check it?

The agent can run mechanical checks such as file specifications and required legal text. Judgement about claims, likeness and context should go to a separate check, ideally a named person.

Can AI Agents Make and Approve Ad Creative? Where Human Sign-off Belongs