How I Accidentally Became an AI Builder

I did not enter AI through hype, a bootcamp, or a planned career pivot. I stumbled into it. At the time I was running SaaS operations — managing products, administering tools, training users, and helping the organization wring more value out of platforms it had already paid for. That was my lane, and I was out here managing SaaS products like a boss.

Because I had a reputation for picking up new platforms fast, my boss suggested I take a look at this AI thing and Microsoft Copilot. I mostly brushed it off. It felt like one more SaaS product to be aware of — something to understand at a surface level, maybe administer or support someday. Then I sat in on an internal guild meeting, saw a small glimpse of what was actually possible, and that was enough to get my attention.

Requesting the license I almost didn't request

After the guild meeting I eventually asked for a Copilot Studio license. I wasn't aggressive about it — it took me two or three weeks to even put in the request. But once I got into the product, I started building. And once I lock onto something, I get laser-focused. Sometimes too focused.

I went deep. Long stretches building, testing, publishing, adding tools, writing knowledge files, and rewriting prompts to understand how the thing actually worked. Frontier models were not on my radar yet. Claude existed; I wasn't seriously using it. GPT was the tool I knew, and I used ChatGPT to help me build Copilot agents and write the supporting content. When I heard the phrase "frontier model" back then, it almost sounded like a joke to me. Is that just another LLM?

Learning the hard way

Learning Copilot was clunky. There wasn't much practical information out there, so I dug through YouTube videos, docs, and community posts. I tried the official training and certification paths, but most of it was too broad and too dry for what I needed — 15-hour courses on enterprise administration topics like Intune and security management.

Those things matter. They are not what a business user needs to master to ship their first agent. That should be handled at the enterprise level. A builder trying to make something useful should not have to become an Intune expert to make progress. So I skipped the certification path and learned by building. Piece by piece. Break it, fix it, test it, rebuild it. That became my actual training.

Why it worked: a real use case

The biggest reason I made progress was that I wasn't building to build. I had a real problem to solve. I had a strong partner in HR, and we were working toward a specific Ask HR agent. That gave me a target. I wasn't poking at features at random — I was trying to solve a real business problem for a real department with real knowledge needs.

Every week I had to show up with something better than the week before. It wasn't high-pressure; nobody assigned it. But the weekly rhythm forced progress. A real business problem will teach you faster than any generic training module. It took about six months to get genuine control of Copilot, and I wasn't even doing it full-time. It lived in the evenings and weekends around my actual job.

The pivot, and the part that changed my mind

For a while I believed Copilot was simply where enterprise agents belonged. It was governed, structured, enterprise-connected. That was my whole AI worldview. Then over the last several months I pivoted hard into frontier tools — coding assistants, agent workflows, building actual applications. My Copilot agents went mostly dormant. My perspective didn't.

I now understand agent development differently because I've lived on both sides. I understand the speed, the flexibility, and the power of frontier models. I also understand the tradeoffs — governance, security, the risk of letting something act when it should have asked first. That tension is exactly what I'm building Atlas around now: AI that's powerful enough to do the work and disciplined enough to propose it first.

I didn't come back to any of this as the same builder who started. That turns out to be the whole point.