I Built an AI Agent in 20 Minutes Before a Client Meeting. Here’s What Happened.

Before a client meeting last July, I had about 20 minutes to spare.

I used them to build a prototype AI agent specifically for that meeting.

This isn’t a story about some polished product I’d been working on for months. This is a story about what you can actually put together in the time between finishing lunch and getting on a call.

The Setup

Seth Gilford and Steve Leathers are commercial real estate brokers at Transwestern. They specialize in medical office properties — physician-owned buildings, healthcare real estate across Florida and Texas markets.

Their prospecting workflow, as Seth described it to me a couple of weeks before, was entirely manual. A virtual assistant in the Philippines compiles leads into a spreadsheet — property address, owner name, company, contact information. Then Seth and Steve each open CoStar on one screen, Google Maps on another, spend 30-45 seconds looking at the aerial view of the property, then pick up the phone.

For every call they make, they’ve done manual research. For every email they send, they’ve written something personal — or tried to.

The volume problem: they could only realistically contact a small fraction of the people in their database. There weren’t enough hours.

The Prototype

I had a spreadsheet of their Naples prospect list from a previous conversation. With 20 minutes before the call, I opened Lindy, their prospect sheet, and started building.

The agent I put together did three things:

1. Read each row from the spreadsheet
2. Used Perplexity to research the person and their company — looking up their background, their company’s profile, any transactions or relevant news
3. Generated a personalized LinkedIn message that referenced what it found

That’s the basic structure. Nothing architectural. More like a working sketch.

When the call started, I shared my screen and ran it live.

What Happened in the Meeting

The agent started processing. For the first prospect, it came back with a message that was noticeably more specific than anything you’d get from a generic template. It mentioned the person by name, referenced their company’s work, and framed the outreach in a way that connected to their specific situation.

Seth’s reaction was positive. But then we got to the third or fourth prospect — a physician who owned a medical office building.

The agent had found something specific. It mentioned a real estate transaction involving a German investment group. It cited an exact dollar figure. It identified that the structure was a sale-leaseback.

Nobody told it the person was a physician-owner. The spreadsheet just had a name, a company, and an address. Nobody programmed it to look for sale-leasebacks or to connect physician ownership to that kind of transaction.

Seth stopped the demo.

“We didn’t even specify that the list had physician owners and real estate owners. We just gave you a list.”

Then: “That’s pretty incredible.”

Steve: “It kind of found proprietary stuff.”

I was surprised too. Not because I thought it couldn’t do it, but because I’d built this in 20 minutes and it was already connecting dots I hadn’t drawn for it.

What This Actually Means

I tell this story not to suggest that 20-minute prototypes are always this good — they’re usually rougher than this. But to push back on a belief I hear constantly: that custom AI tools take a long time to build, require a developer, or need weeks of planning before they can do anything useful.

The reality is that the tools available now — Lindy, ChatGPT with custom GPTs, Claude, Perplexity — are powerful enough that a working prototype is genuinely a matter of hours or even minutes if you know what you’re building.

The harder part is understanding the workflow. That took two weeks and a prior conversation with Seth to understand. The actual build — once I knew what I was building — was 20 minutes.

This is true in most of the consulting work I do. The AI part is usually fast once you know what to automate. The slow part is the conversation that gets you to clarity about what actually needs to happen.

The Question to Ask

If you’re sitting across the table from someone who could use an AI workflow — a client, a colleague, your own team — the question isn’t “how long will this take to build?”

It’s “what exactly needs to happen at each step?”

Once you can answer that clearly, the build tends to be much faster than people expect.

And sometimes it surprises you too.

Want to learn how to build agents like this? The 4-Day AI Sprint covers the fundamentals of workflow design and agent building — starting from scratch, no technical background needed.