Last August I ran a two-day AI workshop with a client named Evan. Smart guy, entrepreneurial, genuinely motivated to build automation into his business.
When he showed up on day one, he’d already tried building. For weeks. He had 16 apps open at any given time. He was jumping between ChatGPT, Lindy, Zapier, Google Sheets, trying to connect things. Nothing was working. The workflows were a mess. He knew what he wanted… kind of… but the output never matched the picture in his head.
Sound familiar?
The problem isn’t the tool
Here’s what I see with almost everyone who comes to my workshops: they start with the tool.
They hear about Lindy or n8n or whatever was in the newsletter they read last week, and they open it up and start building. Drag this. Connect that. Add a step here. And they’re three hours in before they realize they don’t actually know what they’re trying to build.
It’s like trying to cook dinner by opening the fridge and seeing what’s there, instead of deciding what you want to eat first.
The tool isn’t the problem. The sequence is.
Draw it before you build it
One of the things I pushed hardest in Evan’s workshop: before you open any tool, draw the workflow on paper.
Not a fancy flowchart. Not a Miro board. Literally a piece of paper and a pencil.
Start with: what is the output I want? A summary? A sent email? A row added to a spreadsheet? Get specific. Draw a box. Write it down.
Then work backwards. What has to happen right before that? And before that? Where does the data come from? Where are the decisions — the “if this, do that” moments? What does a human currently do that the agent needs to replicate?
Twenty minutes. That’s usually all it takes.
What the paper reveals
Here’s why this step matters.
When you draw it out, you catch design flaws before you’ve built anything. You realize the single output you imagined is actually three different outputs depending on the situation. You find the step that requires information you don’t have yet. You spot two steps that could be combined into one. You notice one step doesn’t actually need AI at all — a simple filter rule works fine.
Evan messaged me a few days after the workshop. He said: “Going up and just taking a pencil out and mapping it out was really great… I made the process better by actually writing it out.”
That phrase — “made the process better by writing it out” — is the whole point. The drawing isn’t documentation. It’s thinking. It reveals the real workflow, not the workflow you assumed you had.
I use this with every client now
I ran into the same thing with a logistics team earlier this year. They came to me with a long wishlist: automate carrier emails, quoting, market-rate alerts, customer ETAs. Great ideas. But when we tried to map it all out, it became clear they didn’t have a clean picture of any single workflow.
So we slowed down. Drew out the one workflow that would have the most immediate ROI. Carrier emails — where was the data coming from, what happened with it, who needed to see what, and when? Once we had that on paper, the build took about a day.
Before the drawing it would have taken a week and still been broken.
The framework: design backwards, build forward
I think of this as working backwards. Start with the end state you want — the exact output, what it looks like, who it goes to. Then work backwards through every step needed to get there. Only once you can describe the full workflow in plain language do you pick the tools and start building.
This is also how I teach agent design in my workshops. The five questions I always ask before building anything:
- What is the output? (Be specific. Not “a summary” — “a 3-bullet email summary sent to my Slack.”)
- What triggers it? (Calendar event? Email received? Time of day?)
- What are the steps a human would take to do this manually?
- Where are the decisions? (If X, do Y… if not, do Z.)
- What does the agent need to remember across runs?
Answer those five questions on paper first. Then open the tool.
The counterintuitive truth about speed
I know this feels like slowing down to speed up. And yes, that’s exactly what it is.
But here’s the thing. A poorly designed AI agent doesn’t break dramatically. It breaks quietly. It runs, it produces output, but the output is subtly wrong. The summary misses the key points. The email goes to the wrong person 10% of the time. The calendar entries are off by a day.
You don’t catch those until you’ve been trusting the agent for two weeks.
A paper sketch catches most of that before line one of the build.
In my workshops I’ve watched people spend an entire afternoon building something, getting frustrated, then tearing it down. Twenty minutes of drawing at the start would have saved four hours.
Try this before your next build
Next time you want to build an AI agent or automation:
- Put the laptop down
- Get a blank piece of paper
- Write the output you want — specific, concrete
- Work backwards through every step
- Look for anything that doesn’t make sense or needs information you don’t have
- Then open the tool
The paper is not extra work. It’s what makes the build fast.
Want to learn how to design and build AI agents that actually work? I run AI workshops for business owners and teams in Austin and online. Get in touch if you want to learn more.
