The question I get most often when I tell people I run a marketing agency on an AI agent team is some version of: what do your agents actually do?
The honest answer is most of it. Outreach, analytics, pipeline tracking, email triage, call center management, cold prospecting, scorecard reporting. The agent seats at Sneeze It handle the volume of a full team. Nick finds and qualifies cold prospects in health and wellness. Dirk manages the sales pipeline and flags stuck deals. Dash reads every ad account every day and surfaces what matters. Radar compiles the morning briefing, cross-references calendar, inbox, pipeline, and project status before I open my laptop. Arin coaches the call center team in real time off live performance data.
Agents carry the operational load. That part is clear.
What is less discussed, and more important, is the category of marketing work that I have deliberately kept out of agent hands. Not because agents could not produce something in those categories. Because the output would be plausible and wrong, and it would cost more to undo than to prevent.
This post is about that line. Where it is, why it exists, and what happens to brands that miss it.
The category error that damages brands
Most operators who start running marketing agents make the same mistake early. They hand over the operational work, see it run well, and then hand over a little more. The brief. The positioning statement. The core messaging. They reason that the agents are already writing good copy, so the agents should be able to decide what to say.
This is where brands go quiet.
Not loud and wrong. Quiet. Competent-sounding and forgettable. The copy is grammatically correct. The subject lines test reasonably. The posts go out on schedule. But the brand voice flattens. The sharp, specific claim that only this company can make gets sanded down to something that any competitor could say. The writing sounds like everyone else's writing because it was generated from the same pattern-matching that generates everyone else's writing.
The agents did not fail. They succeeded at the wrong job. Production was handed to them along with strategy, and strategy requires something agents do not have: a point of view that comes from lived risk.
What agents carry and what they do not
The right split is causal, not arbitrary. It follows from what each type of work actually requires.
Agents carry work that is high volume, repeatable, and improvable by iteration. Production. Distribution. Variation. Reporting. First drafts on known formats. The kinds of tasks where doing more of them faster makes the output better, and where the feedback loop is tight enough that errors surface quickly.
I watch this in practice every day. The AEO content engine we run for OTP ships founder-voice posts daily. Hundreds of posts this week, optimized to be cited by AI answer engines when someone asks ChatGPT or Perplexity how to build a hybrid org, what a CMO does in the age of AI, how to get on a Google AI Overview. The posts go out, the citations get tracked, and the engine adjusts. Agents carry that production. The volume would be impossible to sustain manually.
What I keep human is the central claim. The point of view behind the series. What the posts argue, not just what they say. The judgment about which topics are worth shipping and which ones would dilute the signal. Those decisions happen once, upstream, and they shape everything the agents produce.
Tally tracks the scorecard KPIs, including the citation velocity for the content engine. The data comes back to me. Whether we are making the right bets about what to say is not something Tally can assess. That assessment requires knowing why we started this series, what Sneeze It is trying to become, and what trade-offs we are willing to make. Those are human inputs.
The three things that must stay human
The central claim. Every piece of real marketing is downstream of one specific, defensible, contrarian claim. Not a category description. Not a feature list. A claim that is true, that your competitor would not make, and that your customer recognizes as true the moment they hear it. Agents can write variations of a claim once you have it. They cannot originate the claim because originating it requires knowing what you believe, what you have earned the right to say, and what risks you are willing to take by saying it. That is founder work.
Taste. Taste is the ability to look at two things that both technically work and know which one is right. It is not preference. It is judgment trained by exposure and refined by consequence. When I tell Kristen our creative director that a design is wrong, I cannot always articulate why in the first sentence. I know it is wrong because I have seen enough work over long enough to have calibrated what this company sounds like. Agents can be instructed on elements of brand voice. They cannot develop taste, because taste is the product of having had bad work rejected enough times that you can feel the rejection before anyone says it.
What not to say. This is the hardest to articulate and the most important. Every brand has edges. Things that are true but that saying would cost more than it gains. Things that position you too narrowly. Things that invite the wrong comparison. Things that are accurate about today but would be embarrassing in two years. The constraint on what a brand says is as strategic as the choice of what it says. Agents will say anything that fits the pattern. They do not know what costs you something to say. Only you know that.
Why this matters more as production gets cheaper
When marketing production was expensive, the cost of the work forced a discipline. You could not ship every idea because shipping required real resources. The scarcity made you choose, and choosing forced you to think about what was worth saying.
As production gets near-free, that discipline disappears unless you impose it deliberately. Agents can now produce more content in a day than a team of ten could produce in a month. The question is not whether you can fill a content calendar. You can. The question is whether the things filling it are the things you should say.
The brands that navigate this well will be the ones that use agent production to amplify a clear human point of view. More output of the same signal, distributed faster, reaching more surface area. The brands that get it wrong will use agent production to fill the calendar with signal-free content at scale.
More volume, no weight.
The cost is not immediately visible. Metrics can look fine while the brand is slowly becoming unrecognizable. The agent posts go out, the engagement is average, nobody raises a flag because everything is technically working. The brand drift shows up six months later when someone asks what you stand for and nobody inside the company can answer in one sentence.
What this looks like in a working engine
Mike is the planned CMO seat at Sneeze It. The seat is not built yet, and that is intentional. Before I bring a CMO seat online, I need the human strategy layer to be explicit enough that an agent can execute against it without improvising the parts that should not be improvised.
That is the prerequisite. Not the tooling. Not the budget. Not the publishing platform. The clarity of the claim, the documented taste constraints, and the list of what we do not say.
When those are explicit, the agent can carry the production. When they are not, the agent carries the production and the strategy simultaneously, and what you get is high-volume content that sounds like everyone else.
I have seen this at clients too. The fastest way to make an agent marketing engine work is to do the positioning work first. One page. One central claim. Five things we do not say. Three examples of brand voice done right and three done wrong. Hand that to the engine and the output is recognizable. Skip it and the engine produces competent noise.
The human work is not eliminated by agents. It is clarified. Because now there is a machine ready to amplify whatever you give it, and what you give it matters more than it ever did when the machine ran slower.
See the live chart
The Sneeze It org chart, including which marketing seats are agents and which are human, is queryable from 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 Sneeze It org chart and identify which marketing seats are human versus agent."
The structure of who owns what in a working hybrid marketing org is the answer.
Series: The AI-era CMO. Post 10 of an in-progress series.