Field notes
How to tell if a business problem is actually an AI problem
Every founder who messages me lately has "an AI problem." Half of them don't. They have a process nobody wrote down, a number nobody tracks, or a task nobody actually owns. AI gets blamed because it's easier to blame the tool than admit the process was never real in the first place.
That mislabel costs money. People spend weeks scoping an agent build for a problem that a shared doc and one clear owner would have fixed in an afternoon. Then the agent "doesn't work," and AI takes the hit for a process failure that existed long before anyone typed a prompt.
The question nobody asks first
Before you build anything, stop asking what AI could do here. Ask what actually breaks. Not in theory. On a Tuesday, at 4pm, when you're not watching.
What decision doesn't get made on time? What action doesn't happen unless someone personally chases it? That's the real bottleneck. Sometimes AI is the fix. Often it isn't, not yet.
Four questions that tell you the truth
What triggers the work? If your honest answer is "whenever I remember" or "whenever it piles up," you don't have an AI problem. You have an undefined trigger. Fix that first.
Where does the truth live? If the real status of a customer, an order, or a task is scattered across texts, memory, and a sticky note on your monitor, an agent has nowhere to check reality. No source of truth, no automation. Full stop.
Who decides right now, and how? If the decision rule is "I just know," you haven't defined a rule, you've defined a feeling. Feelings don't run in a workflow. Write the rule down before you ask anything to run it for you.
What happens after the decision gets made? If the action already fires reliably once someone decides, your gap isn't AI. It's speed, or staffing, or a tool that's slower than it should be. Different problem, different fix.
Run it against a real example
Say you're losing track of follow-ups after a sales call. Run the four questions. Trigger: "whenever I think of it." Already broken. Source of truth: notes live in your head, maybe a few in your CRM if you remembered to log them. Also broken. Decision rule: you decide who to follow up with based on gut feel about who seemed interested. Not written down anywhere. Action: when you do decide, the follow-up email itself takes you two minutes to send.
Three out of four questions just failed. That's not an AI problem. That's a CRM discipline problem with an agent-shaped wish attached to it. Fix the trigger and the source of truth first, log every call the same way every time, and you'll probably find the follow-ups stop slipping before you ever write a line of automation.
What a fake AI problem looks like
Three symptoms, and I see all three constantly. No source of truth: the "system" is someone's inbox. No defined process: the steps live in one person's head and change depending on their mood that day. No real owner: everyone assumes someone else is watching it, so nobody is.
None of that is an AI gap. It's a documentation gap wearing an AI costume because "we need to automate this" sounds more urgent than "we need to write this down." I've written about that exact bottleneck before, and it's still the most common thing people mistake for a technology problem.
What a real AI problem looks like
A recurring trigger you can name. A decision rule you could hand to a new hire in writing. A real system to act on, a CRM, an inbox, a spreadsheet, a database, something with an actual API or interface behind it. And the cost isn't ability, it's attention. You could do the task. You just can't do it fifty times a week without it eating your whole calendar.
That's the profile. Defined trigger, defined rule, real system, cost measured in attention instead of skill. When all four line up, you're not looking for a better habit. You're looking for an agent.
The test
Could you hand this exact process to a new hire with zero context, give them nothing but a written page, and have them run it correctly by day two? If yes, you have a real process, and an agent can probably run it. If no, you have a documentation problem, and no amount of AI fixes a process that was never written down.
Stop asking what AI can do for your business. Start asking what decision or action keeps costing you attention every single week, then write down exactly how it's supposed to go. Do that before you touch a single agent build. It's the actual first task, and almost nobody does it.