26 Sept 25 
Everyone's pitching autonomous AI agents for marketing. We run some. Here's an honest sorting of what agents genuinely do well today, what they quietly fail at, and how to tell the difference before you buy.
Somewhere in your feed right now, someone is announcing that AI agents will replace your marketing team by next quarter. We build automation for a living, we use agents daily, and we’d like to offer a calmer scorecard.
What’s Actually Real
Research and enrichment. Agents are genuinely good at the grind: finding prospect websites, pulling public signals, compiling first-draft audits. We run agent-assisted research in our own pipeline, and the output is real. The catch is that everything still passes a human eye before it touches a client or a prospect.
Drafting at volume. Briefs, variant copy, report summaries, first-pass audits. An agent that produces a solid 70 percent draft changes the economics of content and ops work. Our take on which AI tools survived real client work covers the tooling side of this.
Watching things humans forget to watch. Spend anomalies, broken tracking, form failures, a page that suddenly slows down. Machines are tireless monitors, and marketing ops is full of things that break silently. This is the least glamorous use case and possibly the highest ROI.
What’s Still Hype
“Fully autonomous campaigns.” An agent can assemble a campaign. It cannot yet be accountable for one. Brand judgement, budget risk, and the taste calls that separate a good ad from wallpaper still need an owner with a pulse. We wrote about that boundary in our AI creative checklist: the machine proposes, a human disposes.
Agents talking to your customers unsupervised. Support deflection, maybe. Sales conversations and lifecycle email, no. The more automated the send, the more human the words need to be, and today’s agents are confident in exactly the moments they shouldn’t be.
“Set it and forget it” anything. Every automation we’ve ever shipped needed tending in its first month: edge cases, weird inputs, a rule that misfired. Anyone selling zero-maintenance autonomy is selling the demo, not the system.
The Test We Use
Before adopting any agent, we ask one question: if this runs wrong for a week before anyone notices, what breaks? If the answer is “some internal drafts are mediocre,” automate freely. If the answer is “customers get nonsense” or “budget evaporates,” keep a human in the loop, or keep it out of production entirely.
Agents are levers, not employees. Pull them where the downside is boring and the upside is compounding. That’s not as exciting as the LinkedIn version. It’s just what works.
Building out your own marketing ops and wondering what’s safe to automate first? We’ve made most of the mistakes already, and we’re happy to share the map.