How Do You Know If AI is Working In Your Service Business?
Test an AI workflow the way you'd test a new service: one job, one number, two weeks. Pick a single process, write down today's number before you change anything, set the pass mark in dollars in advance, then run it on real work for two weeks. If the number clears the mark, scale it. If it doesn't, you just saved yourself a subscription and found out for the price of an afternoon.
How Should You Test an AI Workflow Before Building It?
A plumber asks his software rep about AI dispatching. An accountant hears a competitor is "using AI now." A contractor watches a video promising 20 hours a week back.
All three buy something. Six weeks later, ask any of them what changed and you get the same answer: "we're still figuring it out."
They can't answer because nobody wrote down the starting number. Without a baseline, there's no such thing as an improvement — only a feeling, and a charge that renews on the 1st.
Adding AI is a new offer. You'd never launch a new service without a price, a target, and a way to tell if it sold. Same rules here.
Key Takeaways
Write the baseline down before you touch anything. "Last week we missed 14 calls and returned 6" is a baseline. "We miss a lot of calls" isn't.
Set the pass mark in dollars, in advance, so you can't move it after you've grown attached to the tool.
Two weeks on real work — real calls, real clients — not a vendor demo with clean sample data.
If five people do the job five ways, standardize first. I saved 15 hours a week doing that before I bought anything.
Miss the mark, kill it that day. A pilot with no end date is just a subscription with extra steps.
Why "Should We Use AI?" Is the Wrong Question
You already know AI matters. You've used it to clean up an email or summarize a call. That's one person, one task, one time — useful, but it doesn't show up anywhere you can count.
The better question is smaller and harder: which single workflow is losing you the most money this month, and can you prove a fix moves it?
Most owners skip that and go straight to the tool. I did too — and the tool I went straight to cost $35,000 a year.
What a $35,000 Software Quote Taught Me About Structure
Before AI was part of the conversation, I was looking at project management software for my bookkeeping company. $35,000 a year. I was treating it like a hire — the thing that would finally fix the chaos.
Before signing, I sat with my team and watched a month-end close. Some people worked in Word. Some in Excel. Files lived across ten folders. Every person had their own way of closing a month — and every one of those ways worked, which is exactly why nobody had ever fixed it.
The software wasn't the problem. Five versions of one process was.
So we rebuilt month-end close as one sequence, and agreed on details as small as what a highlighted cell means — yellow is waiting on the client, green is done. Tiny stuff. It saved the team 15 hours a week, which at roughly $30 an hour is about $23,000 a year. Cost: two meetings and a written standard.
We did buy real software — months later, at 20 people, once everybody was running the same process. If I'd bought it first, I'd have spent a year forcing a $35,000 system onto five different workflows and blamed the software.
You've hired people, priced jobs, and made calls on incomplete information for years. Running a two-week test isn't a new skill. It's the same instinct, pointed at a smaller question.
How to Test an AI Workflow Before You Build It
1. One job — pick the workflow that's bleeding, not a department
A plumber isn't "adopting AI." He's testing whether missed-call text-back books more jobs this month. An accountant isn't "implementing AI." She's testing whether AI intake cuts new-client setup from 90 minutes to 30.
One workflow. If you're not sure which one is costing most, that's a ranking problem — Which Workflow Should You Automate First? solves it before you get here.
2. One number — baseline today, pass mark before you start
Write down where you are right now. Not a range, one number: last week we missed 14 calls and returned 6 of them. Intake took 90 minutes across 4 new clients. We sent 22 quotes and followed up on 9.
Then do the arithmetic that decides the test in advance. Text-back tool: $89 a month. Your average job: $420. Two recovered jobs a month is $840 against $89 — so the pass mark is two. Write it down where you'll see it on day 14.
Setting the mark in advance is the entire discipline. Set it afterward and you'll find a reason two was really one, and one was really "promising."
3. Two weeks — on real work, with an end date
Real calls from real customers, not the vendor's demo data. Two weeks is right for anything daily — enough volume to see a pattern, short enough that a failure costs you almost nothing. For a weekly task, run six to eight cycles instead.
Put the end date on the calendar now. And if a vendor offers to run it for you, fine — but they name one workflow and one measurable result, or they're selling you a transformation. Start with something on the shortlist of easiest AI wins if you want the shortest path to a yes or no.
The Test on Day 14
Put the two numbers side by side and say the sentence out loud: "We went from 6 of 14 missed calls returned to 13 of 15, and that booked 3 jobs at $420. The mark was 2."
Cleared it — scale it, and pick the next workflow. Missed it — cancel today, not at renewal. And if you catch yourself explaining why the number doesn't reflect how well it's really going, that's the answer. You've written a case for something that didn't work.
There's a third outcome worth naming: it worked, but only because you fixed the process to run the test. That happens constantly. Keep the process, drop the tool, bank the savings.
Four Ways These Tests Go Wrong
No baseline. Without a starting number, every result is "feels better," and nothing ever gets cancelled.
Testing three things at once. New tool, new process, new person in the same two weeks means you learn nothing about any of them.
Testing on a tool the team hasn't been shown. Fifteen minutes of setup with the person who'll use it is the difference between testing the tool and testing whether they guessed right.
A pilot with no end date. Month seven of a two-week test is just a subscription you stopped thinking about.
What to Do in the Next Two Weeks
Pick the task that drops the most leads or eats the most hours — missed calls, quote follow-up, intake, overdue invoices. Today, write down its number. This week's, not a guess.
Then set the mark:
I'm testing [fix] on [workflow] for two weeks. Today's number is [baseline]. The tool costs [$/month] and my average job is worth [$]. Calculate what result would make this worth keeping, and tell me the smallest number of [jobs/hours] that clears it. Then list what else could move that number in two weeks so I don't credit the tool for something else.
Put the end date on the calendar. That's an afternoon of setup and two weeks of watching one number — not a six-month AI strategy.
What This Means for Your Next AI Decision
You don't need an AI strategy. You need one workflow, one number, and two weeks of proof — the same test you'd run on any new service before you built it out. AI isn't a special case.
Run it four times this year and you'll have four decisions made on evidence, and you'll know the exact dollar figure behind each one. Two of them will be no. That's the point — a no in week two costs $89, and a no in month nine costs $801 plus everything you didn't try instead.
Prove it small before you build it big. Then take a Tuesday off, and the missed-call follow-up still runs without you.
Frequently Asked Questions
How do you know if AI is right for your service business?
One job, one number, two weeks. Pick a single repeating task, write down today's number before you change anything, set the pass mark in dollars in advance, and run it on real work. If it clears the mark, scale it. If not, you saved yourself the purchase.
How do you set a pass mark for an AI test?
Convert the result to dollars per month and compare it to the tool's monthly cost plus the time to run it. An $89/month text-back with a $420 average job needs two recovered jobs to be worth keeping. Write that number down before the test so you can't move it later.
What should you automate first in a service business?
Whatever repeats most, eats the most hours, and costs the most when it slips — usually missed calls, slow quote follow-up, or unpaid invoices. Score your repeating tasks on those three and start with the highest number.
Is AI worth it for a business under $3 million in revenue?
Yes, if the process is standard first. Five people each doing the job their own way means AI just makes five versions of the mess faster. Getting everyone onto one process is often the bigger win — it cost me two meetings and returned 15 hours a week.
How long should an AI pilot run?
Two weeks for anything daily — enough volume to see a pattern without much exposure. Six to eight cycles for a weekly task. Set the end date before you start so the test ends in a decision instead of drifting into month seven.
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
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