When NOT to Use AI

The most useful thing I can tell a founder about AI right now is where to keep it out.

That sounds backwards coming from someone who builds AI systems for a living. But I spend most of my time with operators, fractional COOs and founder-CEOs running $1-10M companies, and the ones getting real value aren't the ones who automated the most. They're the ones who got precise about what to hand over and what to hold. The question almost everyone asks is “can AI do this job?” It's the wrong question, and it keeps people stuck. The unit that matters is the part of the job, not the whole job.

The Two Halves of Every Job

Every job splits into two things. There's the labor: the research, the drafting, the structuring, the analysis, the repeat admin that eats a Tuesday. And there's the judgment: the taste, the context, the relationships, the final call. AI is built to carry labor. You keep judgment.

Once you see a job that way, “should I use AI” stops being a yes or no and becomes a dial. How far you turn it depends on two things: what's at stake, and whether a mistake can be caught and undone. I think about it as three zones.

Zone 1: Automate

This is AI in control. It runs the thing unattended and you audit the output.

You turn the dial all the way here when the work is repeatable and rule-followable, and a mistake is cheap and reversible. Lead intake. Client onboarding sequences. The CRM-and-calendar admin that piles up after every meeting. None of it needs your taste. It needs to happen the same way every time, which is exactly what people are bad at and machines are good at.

And some of this shouldn't be AI at all. If the steps never change, a plain rule-based automation (the kind that just moves data from one tool to the next) is more reliable than a model and never surprises you. Save the AI for the parts that need a little judgment, and let a simple script handle the parts that don't.

The test is reversibility. If AI mislabels a lead and you catch it in the audit, you fix it in ten seconds and nothing broke. Low stakes, easy to undo, runs on repeat. That's a zone 1 job, and it's where most operators should start, because the payoff shows up fast and the downside is a rounding error.

Zone 2: Accelerate

This is the one people miss, and it's where most of the value hides.

AI does 80 to 90% of the lifting. You steer and you decide.

You're in zone 2 when the work is big or one-off, it genuinely needs judgment, but underneath that judgment it's mostly labor. Take a go-to-market build. Pulling the market data, sizing the segments, drafting the competitive analysis, structuring the launch plan: that's days of work, and almost all of it is labor. AI can do the lifting in an afternoon. What it can't do is make the bet. You read what it produced, you push back, you redirect, you decide what's actually true about your market.

The mistake I see is treating these jobs as zone 3 because they feel important. Important and labor-heavy are different axes. A go-to-market build matters enormously, and it's still 85% grunt work.

Zone 3: Stay Human

This is where AI stays out of the room.

You hold the line here when trust is made or broken in real time, or when the buck has to stop with a person. Running the sales call. The strategy bet. The hire. The pricing decision. AI can prep you for all of it (brief you on the prospect, model the pricing scenarios, draft the talking points) but it doesn't operate and it doesn't decide. The moment it crosses into the room, you've handed away the one thing that was actually yours.

This zone is small, and protecting it is the whole game. Everything else is negotiable. This isn't.

When Not to Bother at All

There's a fourth answer that doesn't get said enough: don't.

Skip AI entirely when it's a tiny one-off, or when AI can't reach the data it would need to do the job. If setting it up takes longer than just doing the thing, the setup outran the payoff. Half the discipline here is knowing when the dial shouldn't move at all. The operators who trust their own read on this are the ones who don't end up with twelve half-built automations nobody maintains.

Spotting It in a Real Business

When I walk into a company, I'm looking for four patterns. Any one of them is usually a zone 1 or zone 2 opportunity sitting in plain sight.

A founder or an expert is the bottleneck, and their judgment can't reach the team fast enough. The same workflow runs on repeat: intake, onboarding, follow-ups, post-meeting admin. Someone's a “B” at something and needs to be an “A minus”: finance, legal triage, research. Or people are bouncing between five or more tools and just want to talk to their stack instead of clicking through it.

Notice none of those are “replace a person.” They're all “take the labor off a person so their judgment goes further.” That's the framing that lands with owners, because it's the truth.

One Job, Three Zones

Here's what ties it together. One job usually spans all three zones at once.

Go back to that go-to-market build. AI pulls the market data: that's zone 1, it runs the lookup unattended. It drafts the analysis and the plan: that's zone 2, you're steering. Then you make the final market bet and you pitch it to your team yourself: that's zone 3, and it never belonged anywhere else. Same job, three zones, and the skill was never “AI or not.”

That's the whole framework. Split the job into labor and judgment. Hand the labor over as far as stakes and reversibility allow. Guard the judgment. And when you're staring at a job wondering where to start, use the simplest cue there is: when a job is mostly labor, that's your signal to turn the dial.