I had the pleasure of sharing agent building stories with Kyle James on the Agent.AI podcast.
----
"Hey, does anyone have a case study for a manufacturing client... in the $10M-$20M range?"
If you work at an agency or a growing B2B company, you know this message. It’s the "Slack shout."
One person asks, and 100 people get a notification, breaking their focus. People start searching old folders, shouting back, "Ask so-and-so!" It's a massive, hidden time sink.
In a recent conversation on the "Builder Stories" podcast, I broke down how we solved this and other problems by building our own internal AI agents.
It all started with identifying and solving for repetitive, time-wasting tasks.
Solution #1: The Example-Finder Agent
That "Slack shout" was my first target.
I knew we had a decade's worth of client and project examples, but it was scattered. So, we did the hard work: we aggregated all that data into a single, clean database.
The hard part wasn't building the agent; it was aggregating the data.
It took time. We had to go through years of client work, tag it properly, and get it into a structured format. This is the "boring" work that everyone wants to skip, but it's the non-negotiable foundation.
Once we had the data, we fed it to an agent.
The result? The Slack channel went silent. Our sales team could get instant, accurate examples of past clients, projects, and integrations without distracting 100 people. The time savings were immediate.
Solution #2: The $1M Sales Research Agent
After that early win, we aimed higher.
One of the biggest time sinks for a sales rep is pre-call research. A rep could easily spend 2+ hours digging up:
-
Company history and news
-
Technographic data (what tools are they using?)
-
Relevant case studies from our own library
-
Custom-tailored discovery questions
-
Potential follow-up email drafts
We built a "research agent" that connects directly to HubSpot.
Now, when a new company enters our CRM, the agent gets to work. It builds a comprehensive research brief and pins it as a note directly on the company record in HubSpot.
This includes pulling in technographics, identifying 10+ similar projects we've done, and scripting out discovery questions.
The impact is huge:
-
Massive Efficiency: We've created what we estimate to be nearly a million dollars in efficiency by eliminating those manual research hours.
-
Better Conversations: Our sales team goes into calls better prepared. They aren't asking 101-level questions. They're asking 201-level questions that show they understand the prospect's industry and challenges.
My Lessons Learned (So You Can Skip the Failure)
I failed many times in my first attempts at building agents. I was in over my head, the data was a mess, and the AI models weren't quite there yet.
Here’s what I learned:
-
Start Small. Don't try to build a massive, complex solution that solves everything. You will fail. Instead, find one component of a problem (like the "Slack shout") and solve for that first. Gain the confidence and build from there.
-
It's All About the Data. I can't say this enough. AI is useless without a clean, accurate, and foundational data layer. The "boring" work of data aggregation is the most important part.
-
Build for the Future. When I first built some of these, the AI would hallucinate or fail. But I knew the models would get better. And they did, faster than anyone expected. Don't build for the AI of today; build for the AI of six months from now.
