Why I keep human in the loop

I don't trust automation to talk to my customers.

That's not a philosophical stance. It's a business decision.

I run a small operation. When someone emails, they're not reaching "support" — they're reaching me. If I let a model answer on its own, I save ten minutes. If the model says something wrong, or weird, or off-tone, I lose the customer and don't know why.

The math is simple: one lost customer costs more than ten saved minutes.

Human-in-the-loop isn't about whether AI can do the thing. Most of the time, it can. The question is what happens when it can't, and whether you'll know before the damage is done.

For a Fortune 500 company, a model that's right 95% of the time is a miracle. Five percent edge cases get routed to a human somewhere, a ticket gets filed, someone fixes it later. The system absorbs the error.

For me, five percent is one in twenty. If I send twenty emails this week and one of them is confidently wrong, I've probably just alienated someone I can't afford to alienate. There is no "system" to absorb it. There's just me, finding out three days later when they don't reply.

So I keep the human in the loop. Every outbound email, every published post, every calendar invite. The agent drafts it. I read it, edit it, and hit send.

This is not perfectionism. It's risk management.

The agent doesn't know that this particular client hates exclamation points. It doesn't know I promised someone a reply by Thursday, not Friday. It doesn't know the subtext of the last conversation, or that the relationship is newer than it looks, or that this person once told me they nearly went with a competitor because our emails felt robotic.

I know those things. The model doesn't. And if I let it run unsupervised, I'm betting my business that it will never need to know them.

That's a bad bet.

Here's what actually happens when I use the agent: I get the first draft in thirty seconds instead of thirty minutes. Then I spend five minutes fixing the two sentences that would've embarrassed me. Then I send it.

Net time saved: twenty-five minutes. Net risk introduced: zero.

That twenty-five minutes goes back into the work only I can do. Strategy. Client calls. The decision about what to build next. The model bought me that time, but it didn't make the decision for me.

This is the thing most takes on AI miss. The question isn't "can it do the task?" The question is "what's your tolerance for errors, and who pays when one slips through?"

If you're big enough, you can tolerate errors. You have brand equity, legal cover, customer support infrastructure. You can say "sorry, system glitch" and most people shrug.

I don't have that cushion. My brand is me. My customer support is me. If the model screws up in my name, "system glitch" doesn't land — it just sounds like I wasn't paying attention.

So I pay attention.

The irony is that keeping human-in-the-loop makes me more willing to use AI, not less. I'm not afraid of the agent drafting something weird, because I'll catch it before it ships. If I had to choose between "AI does it all" and "I do it all," I'd pick me every time. But I don't have to choose. I get the speed of the model and the judgment of a human who actually cares whether this specific email works.

That's the whole trade. Speed for volume, judgment for quality. The model handles the volume. I handle the judgment.

And in a small business, judgment is the only moat you have.

I'll automate the part that doesn't matter. I will never automate the part that does. The line between those two is the entire strategy.

Right now, that line is: the agent drafts, I decide.

It might move someday. Models will get better. My tolerance for risk might shift as the business grows. But today, for this business, at this scale, I'm not running anything unsupervised that talks to a customer or ships under my name.

Not because I don't trust AI. Because I can't afford to.

— Mir