How Autonomous Are Marketing Agents Really? Gartner Says Under 5 Percent – and Who Is Liable for Mistakes?

A vendor promises you a marketing agent that “fully autonomously” distributes your ad budget, runs campaigns and optimizes results. All you would have to do is watch. It sounds tempting, especially for a business without its own marketing department. But how many companies actually use agents in marketing like this? And if the agent shifts budget and performance drops afterwards: who is accountable?

A trade article from Adweek puts both questions in context, with figures from the analyst firm Gartner. The article is neither a product launch nor news of the day, but a sober assessment of the market. That is exactly why we consider it useful for mid-sized companies: it helps you calibrate your expectations before you spend money. This article sets out what the source supports, what it does not, and how you can decide in advance where a human approves.

What this is about: an assessment, not a novelty

The primary source is an article in the series “Adweek Tech Advantage” titled “AI Agents Are Increasingly Getting Their Hands on Ad Dollars”. We read it in full. The page gives no publication date. The text refers to events in September 2026 and to a campaign from the second quarter of 2026, so it was written roughly within the past few weeks. We cannot narrow it down further, and we therefore do not present it as a news item of the day.

The topic is an “evergreen”, meaning one that will still hold up in a few months. The article describes how advertisers and agencies in the US are beginning to let AI agents buy media and distribute budgets, and it contrasts this with the state of actual practice. The tension between promise and reality is the core.

The number: under five percent truly autonomous

The most important sentence in the article reads, in the original: “Highly autonomous deployments represent less than 5% of all marketing use cases in 2025, according to Gartner.” In other words: according to Gartner, highly autonomous deployments accounted for less than five percent of all marketing use cases in 2025.

The article also cites a forecast. Gartner predicts that by 2028 only 15 percent of agentic AI deployments will be highly autonomous. Note the wording: the source says “only 15%”, so the word “only” appears there. And the forecast, going by the wording, refers to deployments of agentic AI overall, while the 2025 figure refers to marketing use cases. The two figures therefore do not measure exactly the same thing, and we do not mix them up. Anyone who quotes them in one sentence should know that.

• What the figures say: Most deployments do not run fully autonomously, but with humans in the loop.
• What they do not say: They say nothing about how well or badly the deployments work. They also say nothing about the German-speaking region.
• What is behind them: An assessment by an analyst firm. How Gartner defines “highly autonomous” in detail is not stated in the article.

For your expectations, this is still a helpful calibration. If a vendor claims that “everyone” already uses fully autonomous marketing agents, the source clearly contradicts that, at least for 2025.

“Autonomous” is a label, not proof. Ask every vendor: what does the agent decide on its own, and what does it put in front of a human?

The accountability gap: who is accountable when the agent shifts budget?

The second part of the article matters more for managing directors. It quotes Andrew Frank, an analyst at Gartner, with a question that every company handing decisions over to software should ask itself:

• “The problem is agents are supposed to make strategic decisions about budget allocations.”
• “If agents reallocate, say $10 million from TV to creator inventory, and performance drops, who’s accountable?”
• “Agentic buying introduces this whole new decision layer whose recommendations may not be fully transparent.”

In essence: the problem is that agents are supposed to make strategic decisions about budget allocation. If they move ten million dollars from television to creator inventory and performance drops, who is accountable? And agentic buying brings a new decision layer whose recommendations may not be fully transparent.

The amounts in the example are American and large, but the structure is not. A business that spends 3,000 euros a month on advertising also knows the situation: someone decides that budget moves from one channel to another, and later management asks why. With a person, you can ask. With an agent, the answer depends on whether you defined beforehand what it may do, and whether its actions are recorded.

We cannot answer a legal liability question here, and the source does not either. It identifies a gap in the understanding of responsibility. Whether and how this gives rise to obligations toward customers, tax advisors or authorities is something you clarify with your legal counsel. What you can do without lawyers: write down who is internally responsible before an agent acts on its own.

Two practical examples from the US and how to read them

The article names two examples that show how far practice has come so far. Both come from the US, and both are claims reported in the article.

Apostra: According to the article, the company handled automated media purchases over the past three months for several “very large” brands, amounting to around 1.2 million dollars in total. The figure seems large, but it is modest relative to the budgets of very large brands. What the article intends to show with it can only be said cautiously: there are deployments, but not on a scale that shapes the market.

Rouge Care: According to the article, the brand attributes a fivefold return on ad spend to AI agents. It says it spent 25,000 dollars on connected TV ads and thereby generated over 125,000 dollars in attributable revenue. The campaign was run in the second quarter of 2026 via PubMatic’s “AgenticOS”.

How should this be read? It is a single case, reported by the company itself, with a budget of 25,000 dollars. How the revenue attribution came about, whether there was a comparison campaign without agents, and how long the effect lasted is not stated in the article. As proof that agents “work” in ad buying, that is not enough. As an indication that attempts exist and that some of those involved are satisfied, it suffices.

Availability: not a feature to try out

Unlike a product announcement, there is nothing here that you could activate today.

• Status: Market assessment with analyst figures and individual cases, not a product.
• Region: The examples come from the US.
• DACH: unknown. The source says nothing about Germany, Austria or Switzerland.

We therefore frame the topic not as “You can use this now” but as “This is coming for companies”. Platforms such as PubMatic are building agentic functions into their advertising infrastructure, and sooner or later someone will offer you an agent meant to manage your budget. You should be prepared when that happens.

Autonomy or automation? The difference in two sentences

Much of the confusion arises because very different things are sold under the name “agent”. Classic marketing automation executes rules that a human has set in advance: if someone fills in the form, email A goes out, and after a click email B follows. An agent, by contrast, is given a goal, for example “more qualified inquiries from target group X”, and chooses the steps toward it on its own, within limits set by a human.

The difference therefore lies not only in the technology, but in who defines each individual step. With automation, you do. With an agent, you define the goal and the limits. And these limits are exactly where the accountability question is decided. The less you specify, the more important it becomes that the limits hold and that someone keeps an overview.

We cannot show how many of the tools sold as “agentic” are in truth automation under a new name; the source gives no figure on this. As a test question it is nonetheless useful: can the vendor explain which decisions the tool makes itself and which it does not?

What this means for a mid-sized company

First: do not let the label guide you. If only a small share of deployments truly run highly autonomously, it is highly likely that a product advertised as “autonomous” also involves people in practice. That is not bad news but an opportunity: you can design your approvals deliberately instead of leaving them to the vendor.

Second: decide accountability in advance. Who approves a budget shift from what amount? Who reads the report when performance drops? Who may pause the agent? These questions cost nothing if you answer them before introduction, and they are expensive if you ask them after the first mistake.

Third: demand traceability. Andrew Frank speaks of recommendations that “may not be fully transparent”. Require every vendor to log the agent’s decisions, with the time, the reasoning and the state of the metrics before and after. Without this trail, you cannot later explain what happened.

Fourth: start small and measurable. A limited test budget, a clearly defined metric and a fixed period. Without a baseline, you will not know after four weeks whether the agent was better or merely different.

Which team at Amazing Agents would be responsible: Daniel and Lotte

On amazing-agents.com, the topic belongs in the Research & Analytics department. Agent Daniel is responsible for analytics there. The website describes him as compiling performance data across the social media channels, building reports and flagging anomalies before they become a problem. He delivers a weekly report with a recommendation for the following week. The question of which metrics need an approval limit and when a deviation is flagged fits this task.

Agent Lotte is the trend scout on the research team. According to the website, her trend reports assess relevance and recommend whether to adopt a trend, watch it or ignore it. For the question of how seriously to take a promise like “fully autonomous”, this assessment is the right frame: watch, do not adopt blindly.

For the review before publication, the website names agent Magdalena in the Tech & Ops department: no content goes live without her brand check. We would carry that idea over to advertising: an ad or campaign change that appears in your name goes through a review before it goes live.

We are describing an approach, not a result from a client project. In an audit, we would work with you to define which marketing decisions an agent may prepare, which metrics Daniel evaluates and at what threshold a human approves.

Agent with limits and an approval step: the difference

Without an approval step: The agent shifts budget on its own. Mistakes only come to light when the invoice or the monthly report arrives, and by then the money has been spent. Who was responsible is unclear, because nobody saw the decision.

With an approval step: The agent prepares the shift and gives its reasons. Up to a defined threshold it may act; above that, a human confirms. The decision is documented, and the responsibility has a name. This costs time on the large amounts and saves it on the small ones. An agent with limits and an approval step works fast on preparation and leaves the consequential decisions to you.

On amazing-agents.com, this framework applies to all agents: approval steps for critical actions, and emails are created as drafts, not sent directly. The principle can be applied to budgets: a shift is a draft until someone approves it.

What we do not know

• The publication date: The page gives none. We can only say that the text refers to September and the second quarter of 2026.
• The definition of “highly autonomous”: How Gartner delimits the term is not stated in the article.
• The results of the examples: Rouge Care’s return is a self-report of a single case.
• Germany, Austria, Switzerland: The source says nothing about these markets.
• Legal consequences: We have not checked whether liability or documentation obligations arise.

What you should prepare now: three steps

Because this is a development coming toward companies, and not a product to switch on, the steps consist of preparation.

1. List the decisions: Write down which marketing decisions at your company move money today, such as budget between channels, bids, target groups or ad copy. For each, note who makes it and how often.
2. Set thresholds: For each decision, determine an amount or a change above which a human must confirm, and name the responsible person with a deputy.
3. Create a question list for vendors: Keep four questions ready to ask every vendor of an “autonomous” tool: What does the agent decide on its own? What does it put in front of a human? Is every decision logged? How do I pause it immediately?

Conclusion

The source shows: according to Gartner, highly autonomous deployments accounted for less than five percent of marketing use cases in 2025, and by 2028 only 15 percent of agentic AI deployments are forecast to be highly autonomous. A Gartner analyst raises the question of accountability when agents reallocate budgets, and warns of a decision layer that may not be fully transparent. The source does not show how this plays out in the German-speaking region, and it provides no proof for or against the benefit of agentic buying.

For you, this means: do not rely on the label “autonomous”, but define what an agent may do on its own, at what point a human approves, and who puts their name to it. An agent with limits and an approval step is a better foundation than quietly handing responsibility over to software.

Sources: Adweek, “ADWEEK Tech Advantage: AI Agents Are Increasingly Getting Their Hands on Ad Dollars”, adweek.com (publication date not stated on the page; retrieved on October 24, 2026; Gartner figures, quotes from Andrew Frank, and Apostra and Rouge Care claims as reported there). Assignment to the agents: amazing-agents.com, team pages (Daniel, Lotte, Magdalena; approval steps for critical actions, emails as drafts).