Don’t Automate the Wrong Thing

Once Note 02 gets you to a real cause, there are several ways to remove it. Check the cheaper, simpler ones first when they might actually work. Not because they always win, but because skipping past them usually means building technology on top of a step nobody needed.

This isn’t a case against AI, and it isn’t a strict order to follow every time. A rule you can change in an afternoon shouldn’t wait for a software project. But if the cause really is a volume of unstructured information no person could get through (a hundred emails to read every morning, say), that might be the simplest fix available, and it’s fine to start there. The point isn’t to always end at the bottom of this list. It’s to end wherever the cause actually points.

AI is strongest on work that involves reading, judging, and summarizing information, especially at a volume or speed a person can’t match.

• Interpreting unstructured information — emails, notes, documents — and pulling out what matters.
• Classifying incoming requests so they route to the right person or process automatically.
• Drafting a first version of a response, a summary, or a proposal for a person to check and send.
• Flagging exceptions — the unusual case that needs a person’s attention, out of a pile that mostly doesn’t.
• Coordinating information across systems that don’t talk to each other today.

Say employees spend an hour a day writing status reports. AI can cut that to ten minutes. That’s a real improvement — but if those reports then sit unread for two weeks before anyone decides anything, the writing was never the constraint. The decision was. The fastest report in the world doesn’t fix a decision that isn’t getting made.

Before building anything, ask the same question Note 01 used to find the constraint in the first place: does this intervention actually remove it, or does it just make the surrounding inefficiency less annoying? If the answer is the second one, it might still be worth doing. Just don’t expect it to move the business.

Notes 01 and 02 traced a slow proposal process down to a missing pricing rule: every decision defaulted to the owner because nobody had defined which ones didn’t need to. Checked against the list, the fix wasn’t a bigger sales tool. It was a pricing framework the sales team could apply directly for standard cases — a rule change, step four. AI’s role was narrower than it might sound: flag the genuine exceptions that still need the owner’s judgment, and let everything else go straight to the customer. The technology supported the fix. It wasn’t the fix.

Before you build anything, write down what you expect to happen if the diagnosis is right, and how you’ll check it. “Standard proposals should reach customers within one business day, down from an average of 3.2 — measured over the next 30 proposals.” If that number doesn’t move, the diagnosis was probably wrong somewhere, not just the execution. That’s useful to know either way.