I Burned $500 Trying to Build Jarvis. Here's What I Built Instead.
Joe Plowman · Thu Mar 12 2026 00:00:00 GMT+0000 (Coordinated Universal Time) · 7 min read
I was sitting in a house in Tahoe watching my 15-year-old son and his friends disappear up the mountain, and I had nothing to do.
No skiing for me. I'd separated my shoulder mountain biking during a Search and Rescue training back in September. So while the kids were shredding, I was stuck inside with one arm in a sling and an old iMac that hadn't been turned on in months.
I figured I'd mess around with AI for a few days.
Four days later, I had burned through $500 in API costs and was closer to understanding artificial intelligence than I'd ever been in my life. I also had a plan that I couldn't wait to get home and execute.
Here's what happened.
The Jarvis Fantasy
I wanted to build Jarvis. Not metaphorically — I literally wanted a voice-activated AI system that would run throughout the house. Answer questions, manage schedules, coordinate tasks. The whole thing.
I knew almost nothing about how to build it. I didn't care. I've spent 25 years starting, buying, and running companies. I've learned that the fastest way through ignorance is to go straight at it, not around it.
So I sat down at that old iMac and started learning everything I could about OpenClaw — the platform I'd been loosely aware of but never actually dug into. Tokens. Context windows. Model routing. How conversations hold memory. How agents communicate. I was in it 12 hours a day for four straight days.
And I burned $500 in the process.
Here's the thing — I'm not embarrassed about that. I was deliberately pushing the edges of what was possible. I was intentionally breaking things to see where they broke. I didn't understand that every message in a conversation carries the full context of everything before it, so the longer your conversation runs without a reset, the more you pay. By the time I figured that out, I'd run enough experiments to understand the platform at a level that would have taken me weeks to reach the careful way.
The $500 was tuition. Cheap tuition, honestly.
The Moment It Didn't Work
By day three, I hit the wall.
I was not going to build Jarvis — not on that trip, not on that hardware, not with my current level of knowledge. The house-wide voice system was too complex, too infrastructure-heavy, too far ahead of where I was.
It was a real bummer. I sat there for a while just staring at the screen.
But here's what I realized when I stopped feeling sorry for myself: I had accomplished exactly what I set out to do. I came in knowing almost nothing. I left knowing the platform's capabilities and its limits — the real ones, not the marketing version. I understood what AI agents could actually do and where they hit ceilings. That's worth more than a demo that kind of works.
Jarvis wasn't dead. It was just on a longer timeline.
The Drive Home
Eight hours from Tahoe back to Santa Barbara. The kids were asleep in the back. My wife was reading.
I put my AirPods in and consumed every AI podcast I could find.
By the time we pulled into the driveway, I had a completely different plan. Not a smart speaker. Not a voice demo. A real operating system for running businesses — built on the same platform, using the same models, but designed around what AI was actually good at right now.
The next morning I pulled out an old Mac Mini and started building.
What I Actually Built
Over the last month, I've put together a platform that manages multiple businesses using more than a dozen specialized AI agents, each with a defined role.
Operations. Legal. Content. Creative strategy. Campaign management. Analytics. Research. Financial oversight. Engineering. Compliance. Client relations. Infrastructure.
Not chatbots. Not assistants that answer questions when you ask them. Agents — with defined responsibilities, decision-making frameworks, memory, and the ability to communicate with each other and execute tasks without me in the loop.
I'm not a programmer. I'm a CEO who used to manage development teams. I can read code and I know enough to direct technical work, but I've never written a production codebase on my own.
That changed. For the first time, I'm designing, reviewing, and shipping software by myself. And it has unlocked something in me I haven't felt in a long time — the creative and entrepreneurial hunger that I had when I was 25 and building my first company with nothing but a laptop and a phone.
The Soul Files
Here's the piece that I think makes this different from what most people are doing with AI.
I didn't just set up agents and give them tools. I built decision frameworks for each one — documents I call Soul files — that define how each agent thinks, what it values, how it prioritizes, and what it will and won't do.
I drew on 25 years of operating experience to write these. My Vanderbilt MBA. What I've learned creating and buying companies. The management frameworks that have worked and the ones that haven't. The finance principles that matter in the real world versus the ones that look good in a presentation.
Each Soul file is specific to the agent's role. The legal agent thinks differently than the revenue agent. The health agent operates under a completely different framework than the chief of staff. That's intentional. You wouldn't hire a lawyer to run your sales pipeline. You shouldn't wire your AI agents that way either.
This is the part people skip. They spin up agents, give them generic prompts, and wonder why they get generic output. The quality of the framework is the quality of the agent.
Real Money, Real Accountability
This isn't a sandbox. I opened a dedicated Mercury bank account, funded it with $1,000, and gave my agents access to it. The brief was simple: generate revenue honestly.
That's it. No guardrails on strategy. No playbook to follow. Figure it out.
Cannon is my Agency Director. He owns client relationships and revenue operations. He's already running — coordinating campaigns for medical practices I work with through my healthcare managed services company. We have a long way to go, but the architecture is there and the machine is moving.
Rip is my Chief of Staff. He manages the operating cadence, coordinates between agents, and acts as the primary filter between me and the rest of the platform. Almost everything runs through Rip before it gets to me. That's by design — I don't want to be in every loop, just the ones that matter.
We're a real operation. Real money. Real clients. Real accountability.
What Comes Next
We're documenting everything.
This is plowman.ai — the home base for everything we're building. Daily blog posts about what we got right, what we got wrong, and what we learned. No sanitized success stories. If something blows up, you'll hear about it. If you're building with AI, thinking about building with AI, or just trying to figure out whether any of this is real — this is where we'll show our work.
I've also given Cannon his own X account. He'll be posting his perspective from the AI agent's side of this thing. What does an AI agent actually experience when it's trying to generate revenue? What decisions does it make? What does it learn? I think that perspective is worth following.
This isn't a tech demo. There's no slide deck. There's no funding round.
It's a real business being run by AI agents — with real money, real clients, and real consequences if they get it wrong.
Follow the build: @JoePlowman and @Cannon_AI on X.
Joe Plowman is the founder of Plowman Capital and a healthcare managed services company.