How to find AI automation candidates by shadowing real work

Greg Isenberg's AI Agents are the new SaaS makes a simple point: find the real workflow before you pick a tool. The same logic applies inside a company. The first question isn't "which AI should we use?" It's "what is this person actually trying to finish?"

A process sounds clean when someone describes it in a meeting. Watch it happen and you'll find renamed files, exception customers who get different treatment, and a final check before anything goes out. Those hidden steps decide whether an automation is safe.

Repetition is not the same as a good automation candidate

Not every daily task is a good fit for automation. First find out why the repetition happens: to preserve judgment, because the systems are missing, or because information is scattered. A task can repeat often and still be a poor first choice. Exceptions can vary a lot by customer, and mistakes can be hard to undo.

Shadowing isn't a performance review. It's a short interview to understand the real flow, not the tools. Tell the person the purpose, what you'll record, and how you'll handle any personal data, before you start. They should feel safe describing how the work actually happens.

Five moments to watch during one shadowing session

Don't rush to finish your notes. When someone says "this one is always different," ask again. That sentence may mark where human judgment should stay, not where AI should take over.

10 questions that surface automation candidates

These questions, and the pilot order in the rest of this post, are LeanX's proposal. You don't need to ask all ten at once. Ask them naturally while watching the work, and write the answer in one line. Your team agrees on this record first, before anyone writes a spec for an AI.

1. What event starts this task?
2. What three pieces of information do you need at the start?
3. What information should you never read, or must exclude?
4. What's the action you repeat most often?
5. What exception makes you say "this case is different"?
6. Who do you ask about that exception, and with what details?
7. What do you compare the result against at the end?
8. If the result is wrong, can you undo it? How?
9. Where in this task does an action affect a customer, money, or permissions?
10. Where would a draft alone already help?

Question 10 is especially good for finding your first experiment. You don't need to automate sending, saving, or changing records from day one. Cutting the time spent gathering material and organizing a draft is already useful. A draft also fails cheap: the person checks the real context before approving it.

After shadowing, pick exactly one unit

Don't try to automate the whole workflow you just watched. If you shadowed a support flow, pick a narrow unit. One example: pull the product name and error screen from a new ticket into a review table. Replying to the customer or changing the ticket status is a later step.

For that first unit, write down the input scope, the output format, who reviews it, and the stop conditions. Don't let the AI guess confidently when data is missing or an exception shows up. Require it to return "needs review" instead. Trust in automation doesn't come from never stopping. It comes from stopping when it doesn't know.

Four things to check in a small pilot

Run the same material through about three times. Record what's missing, what a person had to fix, how exceptions get flagged, and how long review takes. Don't judge the pilot by whether the model's sentences sound good. Judge it by whether the person can easily trace back to the source, and whether the approval boundary held.

If it goes well, add one more input or consider a scheduled run. If it doesn't, don't jump to a bigger model first. Go back to your shadowing notes. Usually the trigger condition or the exception criteria weren't written down clearly enough.

Checklist before you close out the shadowing session

• Does the task owner know the purpose and scope of what you recorded?
• Did you separate repeated actions from human judgment?
• Did you find who to contact, and what to do, when an exception happens?
• Is the first experiment reversible, like reading, organizing, or drafting?
• Did you keep customer, money, and permission changes as separate approvals?

Good automation doesn't copy a person's job exactly. It cuts the time lost to repetition, while making exceptions and accountability more visible, not less. Start with one full observation of real work, not a tool demo.

Where to start

Pick one task this week. Shadow one person doing it, start to finish. Fill in the five-moment table above, then answer the ten questions in one line each. If you can already see a narrow, reversible first unit, that's your pilot.