Can an AI Agent Do My Marketing?
"Just have your agent run our newsletter." Some version of this arrives more often than you'd think — and it's one of the few requests we decline on principle, not because we can't build it. The reason explains almost everything about how autonomous business agents should be designed.
The line is drawn at who spoke first
Our agents are inbound workers. They answer calls people place, triage mail people send, book slots people request, chase invoices your business is owed, and follow up on documents clients owe you. Every one of those conversations starts with the other party.
Marketing — newsletters, social posting, cold outreach, ad campaigns — means starting conversations with people who never asked for them. That's not a harder problem; it's a different-shaped one. So it isn't a feature gap we'll close later. It's a structural boundary, enforced in code: the lead qualification agent scores and routes, and its action space contains no way to contact a lead unprompted. Not "discouraged." Not "requires approval." Absent.
Why inbound-only is the safety property, not a limitation to apologize for
Handing an autonomous system the power to speak to strangers unsupervised is where agent deployments go wrong: wrong recipient lists, tone drift at scale, replies nobody can unwind. An agent whose entire world is "things that arrived today" has a much smaller blast radius. When something looks off — a dispute on an invoice, an angry caller, a request outside its scope — it stops and routes to a human instead of improvising.
This is also why our dunning agent can touch each overdue invoice at most a fixed number of times, why email drafts queue for approval before sending, and why disputes freeze collections entirely. The same architecture that says "no marketing" says "no runaway reminders."
What the operations side of "growth" actually looks like here
If what you really want is growth, the honest decomposition is: marketing fills the top of the funnel (that's your job, or another tool's), and operations makes sure nothing that enters leaks out the bottom. In a simulated 14-day dry run across our pilot fixture — labeled simulated estimate, not results — the AR agent reviewed 70 invoices, drafted 34 polite payment reminders through its tone gate, paused on 8 promised-payment cases, and escalated 28 aged accounts to humans rather than pressing further. Follow-up discipline is an operations task. Asking for more of it is reasonable. Asking it to invent demand is a category error.
A fit checklist in one paragraph
An autonomous business agent probably fits if your pain sounds like: calls going unanswered, an inbox that eats mornings, scheduling ping-pong, slow lead follow-up, awkward invoice chasing, documents stuck with clients, or Sunday-night report building — repetitive, rules-based, inbound work. It doesn't fit if you need outbound campaigns, creative strategy, licensed advice, or contract judgment. Our service fit finder maps plain-language complaints to specific agents row by row.
Found your row? Watch it in the live demo or start here →.