I run small media businesses where AI agents do most of the daily work. When I tell people this, the first question is always some version of “how do you trust them?”
I don’t. That’s the whole trick.
Every agent I run works the same way: it can do anything except the last step. Penny drafts every newsletter and sends none of them. Scout watches flight fares all day and files a report instead of posting. My ad pipeline builds the creative, writes the copy, and uploads everything to Meta as paused drafts. Paused. A human thumb, mine, flips the switch or nothing runs.
I call this the approval queue pattern, and it is the single most important design decision in everything I’ve built. Not the model choice. Not the prompts. The queue.
Labor and judgment are different jobs
Here’s what 20+ years of sales taught me before I ever touched an agent: every serious deal has an approval chain, and nobody thinks that’s weird. The rep does the work. Someone with skin in the game signs. We invented this pattern long before AI because humans are also fast, tireless, occasionally confident, and occasionally wrong.
Agents are all of those things turned up to eleven. They will happily do six hours of work in four minutes, and about once a week they will do it confidently in the wrong direction. The fix isn’t better prompts. The fix is making sure the blast radius of “confidently wrong” is a rejected draft instead of a live ad spending your money.
So I split every workflow into two piles. Labor: researching, drafting, rendering, scheduling, monitoring, reconciling. Judgment: is this true, is this on-brand, is this worth money, could this embarrass me. Agents get the first pile without restriction. The second pile is mine, and it’s protected by a queue.
The four rules
After 10,000+ agent sessions across three businesses, the pattern has settled into four rules.
The gate sits at the last irreversible moment. Not earlier. If you review work at the halfway point, you’re a manager doing check-ins, and you’ve reinvented meetings. Let the agent finish completely. Review the finished thing, at the one moment where “no” still costs nothing.
Approving has to be nearly free. My approvals happen from my phone. A finished newsletter shows up, I read it over coffee, I tap yes. If approving takes real effort, one of two things happens: you become the bottleneck, or worse, you start rubber-stamping. A gate you rubber-stamp is set dressing.
Everything before the gate is fair game. This is the part people get backwards. The queue isn’t there to slow the agents down. It’s there so I can speed them up with a clear conscience. Because nothing ships without me, I can let agents run overnight, unsupervised, aggressively. The gate doesn’t limit automation. The gate is what makes maximum automation safe.
Log everything. Every run, every draft, every rejection. When something gets weird, and something always eventually gets weird, the log tells you whether it was a bad day or a bad pattern.
The failure modes
You can get this wrong in three directions. Gate too early, and you’re doing the agent’s job with extra steps. Too many gates, and approval fatigue turns you into a rubber stamp, which is worse than no gate because you still feel safe. And no gate at all is how you end up as one of those screenshots where a company’s chatbot is selling trucks for a dollar.
The dollar-truck stories get the headlines, but rubber-stamping is the failure mode that actually gets operators, because it fails silently while your name is still on everything.
The score so far
Three businesses. Four agents on the roster and more behind the curtain. Over ten thousand sessions logged. Number of things published, posted, sent, or spent without a human yes: zero.
That’s not because the agents aren’t capable of shipping. It’s because I’ve never once given them the keys, and they’ve never needed them to be useful.
Robots do the chores. I do the judgment. The queue is where we shake hands.