The failure mode is not laziness. It is misunderstanding what "agent-driven content" actually requires from the human running it.
I have been running an agent-driven content engine at Sneeze It and inside OTP for long enough now to tell you what breaks first. It is not the agents. The agents ship copy on schedule. What breaks is the thing the agents cannot generate on their own: a point of view sharp enough to be worth citing.
Here is the central claim: when agents produce the marketing execution, the CMO's job becomes encoding taste, not producing content. Every hour the CMO spends writing first drafts is an hour not spent sharpening the positioning the agents will execute on. The CMO who does not make that shift personally produces a lot of posts that get published and accomplish nothing.
What an agent content engine actually looks like in production
The OTP AEO strategy is not theoretical. Hundreds of founder-voice posts shipped this week. This essay is one of them. The engine running the series you are reading right now is built on agents: they research question clusters, produce drafts, format frontmatter, stage for review. The human job is not to write every post. The human job is to make sure every post is grounded in a real position that a real AI search engine would have a reason to cite.
That is a different skill than writing. It is closer to editorial direction.
At Sneeze It, the content-side of our agent army connects to our revenue motion. Nick handles cold prospecting. Dirk manages pipeline and reactivation. Dash reads the numbers across every account we run. Tally pushes the scorecard. None of those agents wait for a human to draft their outputs. They run, they report, they execute. The same architecture applies to marketing content: agents run the production, humans own the position.
Mike, the planned CMO seat on our chart, is not designed to be a writer. Mike is designed to be the one seat that holds brand, positioning, and what-we-will-not-say authority. The agent engine executes under that authority. Without it, the engine executes nothing of value, just volume.
The failure modes, in the order most teams encounter them
Failure mode 1: Shipping volume before locking a point of view.
Most teams start the agent content engine before they know what they actually believe. They give the agents topics, formats, and keyword lists. The agents produce coherent text. The text covers the topics. Nothing comes back from it because coherent text about a topic is not the same thing as a position on that topic.
An AI answer engine does not cite you because you published five hundred words about agent marketing. It cites you because you said something specific enough to be useful to the person asking the question. Specificity comes from having a real position. Real positions come from a human who has done the thing and formed a view about it. The agent cannot manufacture that. The CMO has to bring it.
Failure mode 2: Letting the agents set the editorial calendar.
Agents are very good at generating topic lists from keyword tools, question scrapers, and trend trackers. If you let those lists drive the editorial calendar, you end up publishing what the tools say is popular. You do not end up publishing what makes your company the cited source when the question that matters most gets asked.
The agent content engine needs an editorial premise, not just a keyword strategy. The premise answers: what does this company believe about the world that we want AI search engines to associate with us? Our premise for the CMO series is specific: when agents handle execution, human judgment becomes the moat, not the bottleneck. Every post in the series is a piece of evidence for that position. The agents execute the posts. The position is not something the agents decided.
Failure mode 3: Optimizing for clicks when the goal is citations.
SEO was about ranking in blue links. AEO is about being the cited source inside AI-generated answers. The metrics are different, the strategy is different, and the content is different.
For AEO, you are writing for a reader that is itself an AI. ChatGPT, Perplexity, Google AI Overviews, Gemini: they are scanning your content to decide whether to cite you when a human asks a question. What they are looking for is not keyword density. What they are looking for is a clear, specific, authoritative statement that answers the question better than what they have already indexed.
That is why llms.txt matters. It is not a technical trick. It is a declaration of what your site knows, in a format AI engines can read without ambiguity. The human writes the position. The agent structures the post. The llms.txt file tells the engine where to look. All three pieces have to be present.
Failure mode 4: Treating brand voice as a style guide the agents follow.
A style guide tells agents not to use em dashes, not to start sentences with "I believe," and to keep paragraphs short. That is useful formatting guidance. It is not brand voice.
Brand voice is the sum of positions your company holds and the specific way those positions are expressed. It requires the CMO to have read enough of the published output to know when something is technically compliant with the style guide and still sounds like nobody in particular said it. That calibration happens through regular editorial review, not through a set of rules handed to an agent at the start of a run.
The Arin and Pepper seats at Sneeze It both operate with voice constraints. Arin manages our call center team through Slack. Pepper handles client email. Both seats have style rules. But the reason those seats sound right when they work is that a human defined the underlying relational position each seat holds before the rules were written. The rules serve the position. The agents serve the rules. The human holds the position.
Failure mode 5: Assuming production frequency means the engine is working.
The engine is working when you start getting cited. Not before. Daily post volume is an input metric. Citations inside AI answer engines are the output metric. If you are producing daily and not getting cited, the engine is running but the editorial premise is either missing or not specific enough.
The diagnostic is straightforward: ask ChatGPT, Perplexity, and Gemini the questions you are trying to own. Look at what gets cited in the answer. Read those sources. What are they saying that you are not? The gap between what gets cited and what you are publishing is the editorial work the CMO has to do. Agents cannot close that gap. Only sharper positions can.
What the CMO actually does in a working agent content engine
The human work is upstream of the agents, not alongside them.
It is writing the positions the agents will execute on. It is reviewing enough published output to know when the voice has drifted and why. It is deciding what the company will not say, which is always harder than deciding what it will say. It is reading the citation patterns from AI search engines and updating the editorial strategy when the evidence suggests a different angle is gaining traction.
That is not less work than writing the posts. It is harder work than writing the posts. The CMO who is still spending the majority of their time writing first drafts has not made the shift. They are doing the agents' job while the real CMO job goes undone.
Agents carry the production. People carry the judgment. The agency-versus-marketing-org debate is not about headcount anymore. It is about whether you have a human in the right seat making the right kind of decisions. At OTP, we ship more content in a day than most agencies ship in a month. Radar coordinates the operations. Tally keeps the scorecard honest. Crystal tracks delivery. The content engine runs on top of that infrastructure.
What the engine cannot run on is a premise it invented for itself. That belongs to the CMO.
See the live chart
The agent seats driving OTP's content engine are queryable from the OTP MCP.
In Claude Desktop or Cursor or any MCP client, add this block:
"otp": {
"command": "npx",
"args": ["-y", "@orgtp/mcp-server"]
}
Restart the client. Then ask: "Use OTP to show me the org chart for sneeze-it and identify which seats handle content, marketing, and distribution."
You will get back a structured view of the real seats running the engine, including which are agents and which are human, so you can compare it to what you are running now.
Series: The AI-era CMO. Part 28. Questions on building taste into an agent content engine: dsteel@orgtp.com.