Stop Asking "Can AI Do This?" Ask "What Happens When It's Wrong 5% of the Time?"

"Can AI do this?" is almost always yes, and the yes is almost useless.

Yes, it can draft the email. Yes, it can categorize the ticket. Yes, it can read the invoice and pull the line items. The demo will work. The demo always works, which is why demos are such bad evidence.

Here's the question that predicts whether the project survives: when this is wrong 1 time in 20, what happens?

Every automation call is an error budget call

Nothing you automate will be right every time. Not the AI, and not the humans doing the work today, whose error rate nobody measures because it never came up.

So accuracy in the abstract tells you very little. What matters is whether the workflow absorbs the misses.

Some absorb plenty. If the agent drafts a first-pass reply that a human reads before sending, a 5 percent miss rate costs a few seconds of editing. If it tags support tickets and a wrong tag parks something in the wrong queue for an hour, you'll live, and you'll find out cheaply.

Some absorb nothing. If it approves refunds, 5 percent wrong is money walking out the door. If it writes to customers unreviewed, 5 percent wrong is your brand, at scale, in public. If a regulator reads the output, one bad answer is the whole story regardless of the other 99.

Same model. Same accuracy. Completely different decisions.

3 questions that replace "can it?"

How do I find out it was wrong? This separates workable from dangerous, and it has nothing to do with accuracy. An error you catch in an hour is an inconvenience. The same error running undetected for 3 weeks is an incident. If your answer is "a customer will tell us," then your error detection is your reputation.

What does one miss cost? Put a number on it. Time to fix, money at risk, relationship damage. Multiply by your honest expected error rate, not the demo's. If that's smaller than what the automation saves, you have a project. If it's bigger, you have an expensive way to feel modern.

Can I make the failure smaller instead of rarer? Most teams give up here too early, because chasing the last few points of accuracy is slow, expensive, and never finishes. Changing what happens when it's wrong is usually faster. Have it flag low-confidence cases instead of guessing. Have it draft instead of send. Cap what a single action can affect. Batch the outputs so a human scans 40 of them in 90 seconds.

Why the hybrids are the ones in production

There's a pattern in what's surviving contact with real users this year. The deployments getting from pilot to production aren't fully autonomous. They're hybrids: the agent handles routine cases and escalates the edges to a person.

That's the design. It works because it separates two things founders keep gluing together, how often the system is right and how bad it is when it isn't.

[ROGER: a case where you deliberately narrowed what an agent could do rather than trying to make it more accurate, and what that bought you.]

Where I land on this

I've talked founders out of AI projects that would have worked technically, because the error economics were wrong and no model improvement was going to fix them. I've also greenlit projects running on a frankly mediocre model, because a miss cost 11 seconds and got caught immediately.

That's the 70/30 split applied before you build anything. The mechanical work goes to the agent. The judgment about where its mistakes land stays with somebody who understands the business.

Capability is table stakes now. Everything can do everything, badly, some of the time. The advantage goes to whoever's honest about the "some of the time."