Letting an AI agent run my marketing, with guardrails
I let an AI agent handle most of my day-to-day marketing. It drafts posts, joins conversations, checks what worked and adjusts. It works because the rules it can't break live outside the AI, and because I can see everything it does.
Why I handed it over
As a solo founder, marketing is the work that always slips. Building feels urgent, posting feels optional, and the result was weeks of silence followed by a guilty burst. An agent doesn't get tired, doesn't get distracted and doesn't need a pep talk. It just shows up every day.
It's my own custom agent, running inside Donna
This isn't an off-the-shelf scheduling tool. It's a custom skill I built for my own copy of Donna, the Mac assistant I make. Donna already knew my calendar, my meetings and how I write. Teaching her marketing was just one more skill: she knows my products, reads what I shipped and posts in my voice. Nobody else's playbook, nobody else's tone.
Show it once, and it learns
I don't program it step by step. I show it what good looks like, and it turns that into a rule it follows from then on:
- An account whose style I like. It picks up the rhythm and the kind of thing worth posting.
- A reason I turned a draft down. The next draft reads that reason before it writes a word.
- A reply I wrote by hand. It counts mine alongside its own, so we never double up on the same conversation.
That's the part that surprised me most. If I can show it what good looks like, it can learn almost any repeatable task. Marketing just happened to be the first job I handed over.
What it does on its own
- Drafts one original post a day about something I actually shipped. Real release notes only, never invented features.
- Joins conversations where it can genuinely help. Most of its replies don't mention my products at all, which is exactly the point.
- Checks how everything performed and rewrites its own short playbook every night.
What still needs me
I decide where it's allowed to promote anything. A community stays off limits until I've read its rules myself. And every post it makes sends me a notification with a one-step way to take it down. Trust, but keep the remote.
The guardrails that matter most
- Rules live in code, not in the prompt. A prompt is a suggestion the model can drift away from. A check that runs before every post is a wall. Daily caps, spacing and where self-promotion is allowed are all enforced that way.
- It fails closed. Not sure it's allowed? Then it isn't.
- One voice, checked every time. Short, plain sentences, plus a list of phrases that instantly make writing sound machine-made.
- It stops itself. If something keeps failing, it tries to fix it a few times, then switches itself off and tells me. A stopped agent is fine. A quietly misbehaving one is not.
Lessons I didn't expect
- "Ran" is not the same as "worked." Early on it reported healthy runs for an entire day while nothing actually got posted. Now I measure the outcome, not whether the job finished.
- Its notes about itself go stale. It writes itself a summary each night and reads it the next morning. One wrong line in that summary became "truth" and held it back for days. Anything it believes about what's allowed is now re-checked against the real rules every run.
- Spam is a pattern, not a post. An account that shows up in one place saying one kind of thing reads like a marketing account, even if every single post is fine. Variety is part of the rules.
Where Donna fits in
Marketing is the newest thing Donna does for me, not the first. She already opens every call the moment it starts, records it on my Mac without a bot joining, writes the summary and drafts the follow-up email when I hang up. Now she also keeps my products in front of people every day. One assistant, more and more of the busywork, and the hours I get back go into building.
Want the meeting half for yourself? Get Donna for your Mac at usedonna.io.
Would I recommend it?
Yes, with one condition: build the walls before you hand over the keys.
Frequently asked questions
Can an AI agent run marketing for a small business?
Yes, for routine and repeatable work like drafting posts, joining relevant conversations and measuring results, as long as the rules are enforced outside the model and a person decides where promotion is allowed.
How do you stop an AI marketing agent from spamming?
Put hard limits in code: a daily cap, spacing between posts and an explicit list of places where self-promotion is allowed, with everything else blocked by default. Vary where and what it posts, too.
Can an AI agent learn a task by example?
Yes. Instead of programming every step, show it what good looks like, such as an example to imitate or a reason a draft was rejected, and have it turn each example into a rule it follows from then on.
Should an AI agent post without approval?
Only content of a kind you have already approved, with a notification for every post and an easy way to take it down.
I'm sharing my entrepreneur journey on this blog as I go. More posts