Last September, I sat across from a CPA who had 1,100 unread emails.
Not a thousand. 1,100. She showed me her inbox, and I honestly felt a little dizzy looking at it.
She had been meaning to fix the problem for months. Kept putting it off. The backlog grew. She'd come into work, see the number, feel stressed, and do something else instead.
When she finally asked me to help, I did something that surprised her. I didn't open any apps. I didn't start configuring anything. I asked her a question:
“Imagine you hired a real person to manage your inbox. Someone who sits at a desk and reads every email you get. What would you want them to do?”
She laughed a little. But then she thought about it.
Why this question works
Most people trying to build their first AI agent get stuck in the same way. They spend hours reading documentation, watching tutorials, searching for “best AI for email” or “how to automate your inbox.”
The problem is they are asking the wrong question.
“What can AI do?” is a rabbit hole. The answer is complicated, constantly changing, and different for every tool. You end up overwhelmed before you've built anything.
But “what would a real person do?” — that everyone can answer. You don need to know how AI works. You just need to know your own job.
When I started using this framing with clients, everything changed. Sessions that used to take two hours started taking 30 minutes. People went from vague (“I want AI to handle my email”) to specific (“I want it to do these four things”) almost immediately.
This is actually the foundation of what I call Agent Design Backwards. You start with the end state — what does done look like? — and work backwards through decisions, triggers, escalation rules. The human job description gets you to the end state faster than any other method I've found.
What she told me
Back to my CPA client. After I asked her to describe her hypothetical inbox manager, here's what she said:
Anything that looked like fraud or a legal threat should come to her immediately. No delays. Invoices and payment requests should go straight to her bookkeeper. Anything promotional or newsletter-y should be deleted. And anything from an actual client — she wanted to handle personally.
Four categories. Four rules.
That'ss the whole agent.
We built it in about 45 minutes. The agent runs in Lindy. It reads every incoming email, applies those four rules, and either routes it, deletes it, or holds it for her review. She now processes her inbox in about 20 minutes a day instead of avoiding it entirely.
The job description she gave me in plain English became the instruction set. Almost word for word.
The fear that stops people
When I was talking to the Lindy team a while back, they asked me why people hesitate to build email agents. The answer I kept hearing was fear. Specifically, the fear that the agent would do something on their behalf — send an email they didn't approve, delete something important.
That fear is legitimate. But it usually comes from people imagining a fully autonomous robot with no oversight.
The human framing fixes this too. When you think about it as hiring a person, you automatically think about what level of trust you\d give them on day one. You'd probably say: “For the first month, don send anything without checking with me first. Just sort and flag.”
That's a perfectly valid agent configuration. You can start with a “review everything” setup and loosen permissions gradually as trust builds. Treat the agent like a new hire, not a magic button.
How to do it yourself
Here's the exercise. For whatever you
e thinking of automating, write a 5-minute job description as if you were posting it for a real human hire:
- What does this person do each day? What are the inputs they receive?
- What decisions do they make? What's routine vs. what needs your approval?
- What should they escalate immediately? What can they handle on their own?
- How will you know if they're doing a good job?
That's your agent spec. Seriously.
When I did this with another client — a real estate investor — we mapped out his entire research workflow in 20 minutes. He described a hypothetical research assistant who would watch 20 YouTube channels, summarize new videos in Slack, and flag anything relevant to his specific investment thesis. We built that agent the same afternoon.
The human description translated almost directly into the Lindy workflow.
One more thing
Building AI agents isn't really about AI. It's about understanding your own work well enough to describe it clearly.
If you can write a decent job description, you can build a decent agent. You don need to understand machine learning or API calls or any of that stuff.
The technical part is honestly the easy part. The hard part is figuring out exactly what you want. And the “imagine a real person” question gets you there faster than anything else I've tried.
Start with that. The rest follows.
Want to learn how to build your own AI agents? I teach this in my one-day AI workshops in Austin. Check out what we cover at Asian Efficiency.
