The signals nobody's taught you to read

Ask most business owners why they haven't put AI to work in a real part of their operation, and they'll give you one of two answers. Either "we're not sure it's ready," or "we're not sure where it would even help." Both answers are usually wrong. The question isn't whether AI is ready. It's whether the work is legible enough for someone else to step into it and do it correctly, whether that someone is a new hire, a consultant, or an AI employee.

Most businesses can't answer that question about their own operations, because nobody's ever had to ask it. The people doing the work know how to do it. Nobody's had to write it down precisely enough for someone else to run it cold.

That's the gap. Right now, elsewhere, finding it is a paid, one-on-one service: an outside consultant sits with you for a few sessions, asks the right questions, and hands you a $2,000 to $8,000 readiness assessment at the end. There's nothing wrong with that. But it means the signals that tell you where AI could genuinely help are treated like trade secrets, when they're really just things nobody's bothered to write down for a general audience. Here are four of them.

The work already repeats

If you do something once, it's a project. If you or your team does the same broad thing weekly (the same kind of intake, the same kind of review), it's a workflow, and workflows are where this applies.

This sounds obvious. It's the signal people skip past fastest, because "we do this a lot" doesn't feel like insight. But it's the first filter, and it rules out more than you'd think. Most of what people first suggest automating turns out to be a one-time project wearing a workflow's clothes.

Someone already checks the work before it goes out

Look for the place in the process where a person currently reviews, corrects, or approves something before it moves on: a manager who reads a draft before it's sent, or someone who checks a number before it's booked.

That review point matters more than the task itself. It's where AI can help without anyone having to trust it blindly, because the check that already exists doesn't go away. It just moves earlier, to something faster to produce. If nobody currently checks the work at all, that's not a green light. It usually means the work isn't consequential enough to be worth mapping, or worse, that it should be checked and currently isn't.

People are already doing this with AI, badly

More than one client has come to us with exactly this pattern already running: someone on the team pasting things into a frontier lab's chat interface, quietly, with no structure to it, and getting inconsistent results because of it. If you dig around in your own business, you'll usually find the same thing: a draft email, a summary, a first pass at something, done off to the side because it feels like cheating or because there's no sanctioned way to do it. That's not a discipline problem. That's a business process announcing itself. If people are already reaching for AI on their own, the task has already told you it's a fit. The missing piece is a way to do it that the rest of the business can actually see and trust.

The knowledge to do it well already exists somewhere

Every real workflow runs on some body of stable material: templates, house style, client requirements, past examples of good work, rules that everyone just knows. If that material exists, even scattered across someone's inbox and a shared drive and one person's memory, it can be pulled together and handed to whatever does the work next, human or otherwise. If it doesn't exist at all (if quality depends entirely on one person's judgment with nothing written down anywhere), that's worth knowing too, but it means you're not ready to bring AI into that specific task yet. You're ready to write the material down first.

This matters more than it sounds like it should. Research published by Dwarkesh Patel and Jerry Han in 2026, tracking AI model improvement from 2019 to 2025, found that better training data produced a twelve-times gain in efficiency, against a 3.7-times gain from better model architecture over the same period, more than three times the payoff coming from the unglamorous side of the equation. The same pattern holds here. Getting the right material in front of whatever does the work, human or AI, outperforms a fancier setup built on top of nothing in particular.

What to do with this

None of these four signals is a yes/no test on its own, and this isn't the whole method: reading them well, weighing them against each other, and turning what you find into something a team can actually run is still real work, the kind people currently pay a consultant to do. What changes once you can see the signals is smaller but real. You stop treating "should we use AI here" as a mood or a guess, and start treating it as a question with actual evidence behind the answer.

The next piece in this series covers the question that usually comes right after you've spotted a process like this: whether you can just buy something to handle it instead of doing the work yourself. The short answer is no, and the next piece explains why.