How to Know What AI Should Automate: The 4-Question Scorecard
A task is ready for AI when it clears four thresholds: it runs the same way at least 4 out of 5 times, one wrong output costs under roughly $500 and is reversible the same day, it happens weekly or more with a payback inside 90 days, and a person can check the result in 60 seconds. Any single no tells you exactly what to fix — and it's almost always fixable in one sitting, not a redesign.
What Business Tasks Are Good Candidates for AI Automation?
It's Wednesday afternoon. Your new hire is in the doorway with the same question she asked Monday. You answer it again — and halfway through you realize you're telling her something slightly different than you did two days ago.
That's the whole diagnosis right there. If the answer changes depending on which day you ask, the process isn't written down. And a tool built on top of it will give three different answers to the same question, in three places, before anybody notices.
So before you build anything, run four questions. Ten minutes. Each one has a number attached, so it's a pass or a fail and not a vibe.
Key Takeaways
Stable: run the task 5 times and compare. More than one run differs and AI will copy every version of it.
Safe: price one wrong output. Under ~$500 and reversible today, let it run. Above that, a person signs off.
Often: hours saved/week × 52 × hourly cost has to pay back the build inside 90 days. 10 hours of setup to save 10 minutes a week is a 4-year payback.
Easy to check: 60 seconds, one source of truth. If checking it takes as long as doing it, you've moved the work, not removed it.
One "no" is a two-hour fix, not a reason to quit. Four "no"s means you're looking at the wrong task.
Map the Real Process First, Not the SOP
Most owners buy the tool before they've watched the work happen. So do this first: pick one task and watch somebody do it, start to finish, once. Write down every step they actually take.
Then compare that list to your SOP. In most $1M–$5M service businesses, the two documents share maybe half their steps. The SOP has things nobody does anymore, and the real process has three workarounds nobody wrote down.
Score the real one. Automating the SOP means automating a document that describes a company you used to be.
The $4,800 a Year They Spent on Tools They Weren't Ready For
A husband-and-wife business I worked with was paying $400 a month across a handful of AI subscriptions. That's $4,800 a year, and they were opening maybe two of them.
They were growing, and onboarding was the thing hurting most. New hires were taking about three weeks to get productive, and every one of them got a different first week depending on who happened to be free.
They didn't have an AI problem. They had a process that was never going to survive contact with a tool. And they'd spent $4,800 finding that out the expensive way.
The 4-Question Scorecard: Is This Process Ready for AI?
1. Stable — does it run the same way 4 out of 5 times?
Don't guess at this one. Pull the last five times the task ran — five onboardings, five quotes, five invoices — and compare the actual sequence.
Four or five match: stable, pass. Three or fewer: you don't have a process, you have a habit that varies by person. AI will learn all five versions and pick one at random.
2. Safe — what does one wrong output cost?
Put a dollar figure and a recovery time on being wrong once. An internal reminder that fires late costs about $0 and 10 minutes. A quote sent 15% under cost on an $8,000 job costs $1,200 and a conversation you can't take back.
Under about $500 and fixable the same day, let it run unattended. Above that, or anything with legal or payroll exposure, AI drafts and a person signs. That's not timidity — it's the same rule you'd use for a new employee in week one.
3. Often — does it pay back in 90 days?
This is where the money leaks. Do the arithmetic before you build: hours saved per week × 52 × the hourly cost of whoever does it, against the hours it takes to set up.
Ten hours of setup to save ten minutes a week returns about 8.7 hours a year. That's a four-year payback, and the tool will change before then. Twenty minutes a day across three people is 260 hours a year — that one pays for itself in a week.
Weekly or more usually passes. Monthly is a maybe. Quarterly is almost always a no.
4. Easy to check — can somebody verify it in 60 seconds?
You need a definition of done that somebody other than you can apply. If it drafts a report, do the totals tie to one system? If it books onboarding, did all five meetings land on the calendar?
The test: one person, 60 seconds, one source of truth. If verifying takes ten minutes and the task took twelve, you didn't automate anything — you moved the work and added a handoff.
What a Scored Process Actually Looks Like
Here's that onboarding process scored. Stable: no — five new hires, five different first weeks. Safe: yes — worst case somebody waits an hour, call it $30. Often: yes — they were hiring roughly every six weeks and burning about six hours of owner time each round, around 50 hours a year. Easy to check: yes — by day three you can tell in a minute whether somebody knows their role.
Three yeses and one no. And the one no was the only thing standing between them and 50 hours a year.
We fixed it in one sitting, about two hours. They talked through everything a new hire needs to know into a voice memo — no polish, just talking. AI turned it into numbered steps and flagged what was missing: the steps they'd stopped noticing after doing the work themselves for nine years.
Then they named the three things a new hire has to be able to do by day three. That's the definition of done, and it's what made the schedule buildable.
Only then did anything get automated: name and start date go in, the shadow schedule builds itself, reminders go to the right people, and a daily check-in confirms nothing changed overnight. Time-to-productive went from about three weeks to nine days, and it stopped depending on who was free that Monday.
AI never decided who trains whom. It ran the schedule, the reminders, and the check-in. The judgment stayed where it belongs.
Four Mistakes That Waste an AI Budget
Automating an unstable process. Five versions in, five versions out. Get to one version first — it's the cheapest step in the whole project.
Skipping the payback math. Building feels like progress, which is why nobody runs the numbers. Run them on a napkin before you open a tool.
No definition of done. If you can't say in one sentence what a correct output looks like, nobody downstream can check it — and they'll stop trying by week three.
Paying monthly for readiness you don't have. $400 a month is $4,800 a year. Cancel what you're not using and spend two hours writing the process down instead.
What to Do This Week
Pick the one process you keep throwing AI at. Score it — four questions, write yes or no with the number next to each.
If Stable fails, stop shopping. Open your voice recorder, talk the process through once like you're training somebody, and paste the transcript into AI with this:
This is me describing how we do [task]. Turn it into numbered steps a new hire could follow. Then list every decision point where I didn't say what the rule is, and every step where I assumed knowledge somebody new wouldn't have.
That list is your process. Answer it, and you've turned a no into a yes in one sitting. And if you're not sure which process deserves it first, the free What to Automate First tool ranks your work by time, cost, repeatability, and risk.
It also helps to give AI your actual numbers up front — that's the difference between a generic answer and a useful one, and it's covered in ChatGPT for Business: How to Stop Getting Generic Answers.
The Goal Isn't More Automation
It's a shorter list of things that stop when you're not there. Count that list today. Then count it in 90 days.
For that couple, it was six hours of owner time per hire and twelve days off their ramp. Not a transformation. Just one "no" turned into a "yes" on a Tuesday afternoon.
You already know this process better than any template does. Go write it down.
Frequently Asked Questions About Knowing What AI Should Automate
How do you know if a process is ready for AI?
Four thresholds. Stable: five runs, no more than one differs. Safe: one wrong output costs under ~$500 and is reversible today. Often: the annual hours saved pay back the build inside 90 days. Easy to check: verifiable in 60 seconds against one source of truth. Any no names the thing you fix first.
What happens if you automate an unstable process?
You automate the inconsistency. Five different sequences in means five different outputs out, faster and in more places. Get the process to one version before a tool touches it.
Should every business process eventually be automated?
No. Run the payback: hours saved per week × 52 × hourly cost, against setup hours. Ten hours of setup to save ten minutes a week is roughly a four-year return, and the tool will change first.
What business tasks are the best candidates for AI automation?
Scheduling, reminders, first-draft summaries, pulling data from one system into a standard format, and routing — anything that runs the same way, weekly or more, costs little when wrong, and checks in under a minute. Quotes, payroll and anything with legal exposure keep a person in the loop.
How long does it take to fix a process that fails the scorecard?
About two hours. Talk it through into a voice memo, have AI turn it into numbered steps and list the gaps, then name the three outcomes that define done. Most single "no" answers close in that one session.
About the Author
Buck Paolino is the founder of The Profit Manual. After building and selling a seven-figure service business, he now helps owners of $1M–$5M service businesses make better decisions, install practical systems, use AI responsibly, and get back to CEO work.
Related Reading
Watch the full breakdown in How to Know What AI Should Automate — the video this article comes from.
Related reading:
What Can AI Do In Your Business?
Most businesses don’t need "more AI tools".
They need to know where AI will actually save them time.
What To Automate First is a free 5–10 minute business checkup. List the work you and your team do each week, and it will identify "3 places to focus first"
—including where your time and payroll may be leaking.
No guessing.
No chasing the newest AI tool.
Just a clearer place to start.
Comments