Marketing planning picked up a new line this year: agent count. The framing is that headcount and program count are no longer sufficient, so you now budget for the number of agents embedded in daily execution.
The data underneath it is less confident. In the State of B2B GTM survey, n=195, 47% of teams run zero agents in production. Another 32% run between one and three. Two percent run more than twenty. Separately, 53% of GTM leaders report seeing little to no impact from AI so far.
I have twelve scheduled agents configured across my consulting business. Eight are switched on today, and the one doing the heaviest work doesn’t run on my machine at all. It runs as a cron in the cloud. That puts me ahead of roughly eighty percent of that survey and well short of the top two percent. It has been running long enough to say something useful about the gap.
What did twelve agents actually replace?
Not people. The queue.
A site agent audits gtm-consulting.io three mornings a week and deploys the fix to production without me. An inbox agent works expert-network requests on a fifteen-minute cycle, classifies them, checks my calendar, and drafts the reply. A deliverability agent audits SPF, DKIM, and DMARC weekly and stays silent when everything passes. Another one is writing this post’s first draft.
Two of them are one-time watchers: they wait for a specific external event, send me one email, and disable themselves. That is my favorite category of employee.
The infrastructure bill for all of it is in the tens of dollars a month, plus tokens. Nobody got replaced, because there was nobody to replace. What went away is the backlog of small, dated work that used to expire before I got to it.
What is each agent allowed to do without me?
This is the only question that turned out to matter, and it is the entire design.
The site agent deploys to production unsupervised. It is also forbidden to touch positioning, add third-party scripts, or change a form. Those three categories come to me as recommendations no matter how confident it is.
The inbox agent has never sent an email. It drafts, and I send. But a draft that sits unread for seventy-two hours goes out on its own; deleting it is how I say no. Silence is consent; the veto is an action.
The blog agent has no standing permission to publish. It drafts; I decide, post by post. Including this one.
The deliverability agent has a recommendation queued to tighten my DMARC policy to reject. It is explicitly not allowed to apply it.
None of that is in the agent code. It is in a config file per agent, and the config file took longer to get right than the agent did.
Why do half of GTM leaders report no impact?
Because an agent only creates value where a decision rule already exists in writing, and most marketing decisions have never been written down anywhere.
You know which inbound requests are worth your time. You know which site changes you’d ship on a Tuesday and which ones you’d want to sleep on. You have never had to say it out loud, because you were the one doing it. Hand that work to an agent and you discover the rule was never a rule. It was you, deciding, every time.
That’s the actual project. Not prompts, not tooling, not agent count. Writing down judgment you have always executed implicitly, in enough detail that something without your context can run it, and then deciding where it must stop and ask.
What breaks?
Not what you’d expect. The failure mode of an agent fleet isn’t a dramatic wrong action. It’s an absent one.
My inbox agent quietly decided it had already handled requests it hadn’t. No error, no alert, nothing in the drafts folder, and an audit log that cheerfully said “already handled.” It took six separate code paths to close the class, and a test suite whose only job is to keep it closed.
A bad action pages you. A silent no-op doesn’t page anyone. Whatever oversight you build has to look for work that didn’t happen, which is a much harder thing to instrument than work that went wrong.
So is agent count worth planning around?
As a budget line, sure. As a measure of anything, no.
Twelve agents did not compress my judgment. They industrialized the part of the job that was already mechanical and made the remaining part unavoidable, because now it has to be written down before anything can run.
Agent count is easy. The number that matters is how many of your decisions you’ve been willing to put in writing. And right after that, how many of them you were willing to let run without you.