TL;DR: Don't transform your whole business with AI. Just fix that one annoying process everyone hates. Success is contagious when you start small and smart.
Start with your most annoying process. You know, the one that makes everyone groan in your Monday stand-up. The one where someone always says "there's got to be a better way."
That's where you start with AI agents.
Not the mission-critical stuff. Not the complex workflows that touch seventeen systems. Not the thing that'll transform your entire business model. Slow your roll a bit.
It's the annoying thing. The stupid thing. The thing that wastes two hours of someone's Tuesday.
I was just talking to a CFO who deployed their first AI agent on expense report approvals. Expense reports. The most boring, annoying, time-sucking process in corporate America.
His AI agent now handles the under-$500 ones. Checks receipts, validates against policy, approves or flags for review. Takes 8 seconds instead of 8 minutes per report. Nobody's job depends on it. Nobody's scared of it. But everyone loves it.
Here's why this works:
When you start with the annoying stuff, failure doesn't matter (or at least matters less). If your expense report agent screws up, you go back to the old way for a week. Nobody dies. The company survives. But you learned something.
When it works (and it usually does because annoying processes are often simple processes), something magical happens. That guy who was skeptical? He's asking if the agent can handle travel bookings too. The security team that was worried? They're proud of the controls they built.
Success is contagious. AI is contagious. But you have to start somewhere.
Michelle (my wife and co-author) always tells me I overcomplicate things. "Just pick something everyone hates," she said when I was mapping out complex AI transformation strategies. "Make that better. Then pick the next thing."
She's right. (Don't tell her I said that.)
Your most annoying process is probably:
Manual data entry between systems
Scheduling that requires 47 emails
Approvals that sit in inboxes
Reports that take hours to compile
Answering the same customer question 1,000 times
Pick one. Deploy an agent. Make life 10% better. Then 20%. Then watch your biggest AI skeptics become your biggest champions.
The companies with 200 agents didn't start with 200. They started by fixing that one stupid thing that drove everyone crazy.
What's your stupid thing?
Frequently asked questions
Which process should my first AI agent handle?
The one that makes everyone groan in your Monday stand-up. Not the mission-critical work, and not the workflow touching seventeen systems. Pick the annoying thing that wastes two hours of someone's Tuesday. Annoying processes are usually simple processes, which is a large part of why they tend to work on the first try.
Why start with a low-stakes process instead of a high-value one?
Because failure costs you almost nothing. If your expense report agent gets something wrong, you go back to the old way for a week. Nobody's job depends on it and the company survives, but you learned something. Start on the mission-critical workflow instead and one mistake ends the experiment.
What does a first AI agent on expense reports actually do?
One CFO's agent handles approvals under $500. It checks receipts and validates them against policy. Then it either approves the report or sends it on for review. That takes 8 seconds instead of the 8 minutes a person spent. Nobody's job depends on it, and everyone likes it.
What kinds of processes are usually the right starting point?
Look for manual data entry between systems, scheduling that takes 47 emails, approvals sitting in inboxes, reports that take hours to compile, and the same customer question answered a thousand times. Pick one. Make life 10% better, then 20%. The companies running 200 agents didn't start with 200. They started by fixing one thing that drove everyone crazy.
How does one small agent lead to wider adoption?
Success spreads. The person who was skeptical starts asking whether the agent can handle travel bookings too. The security team that worried about it ends up proud of the controls they built. That's why the first agent should be easy to like. At that stage you're not buying efficiency, you're buying permission to do the next one.