In December 2022, my leads dried up completely.
Not slowly. Not gradually. One month I had a pipeline of crypto development work; the next month there was nothing. The successive implosions of that summer — Ohm, Terra/Luna, Sam Bankman-Fried — had already tightened things considerably. Then ChatGPT launched, and the remaining leads evaporated overnight.
I figured out why pretty quickly. The kind of clients I'd been building for — smart people with budget and product requirements but no coding experience — could suddenly skip the developer entirely. ChatGPT could interpret the dense, hostile documentation of the Ethereum ecosystem well enough that a non-technical founder could just build the thing themselves. Save $50k, ship in a weekend. Why hire me?
I was watching my own market collapse in real time, via screenshot threads on X.
Most people would have panicked. I opened the OpenAI docs instead.
Why do most AI products deliver generic results?
Something bothered me about the screenshots people were posting. The outputs were impressive, sure — but they were generic. Same voice, same structure, same limitations, regardless of what you were trying to do. People were excited about the technology, but nobody seemed to be getting genuinely good work out of it for specific professional contexts.
Buried in the docs, I found a note about manipulating what we now call the system prompt. The model's behaviour, tone, expertise, and persona could be shaped before a user ever typed a word. The quality delta between a raw ChatGPT conversation and a well-prompted one was enormous.
I saw the gap immediately.
The question was what to build. My first instinct was narrow — a specialized interface for a "crypto expert" persona, something useful for my existing audience. But it felt small. I kept pulling at the thread.
What if you could have an entire company in your pocket?
Not one persona — a whole virtual office. Different AI employees for different functions, each one tuned for their domain, each one accessible through a single interface. A growth team: marketing coach, sales coach, copywriter. A product team: designer, developer. Operations: accountant, lawyer. Seven employees to start, covering the core functions of any small business.
I called it GPTBoss.
What happens when a product launch goes viral before it's ready?
I built it on Next.js and Supabase's free tier, wired up Stripe, and filmed a TikTok in my living room.
Single shot. No editing. Just me, animatedly walking around my apartment explaining what I'd built and why it worked better than ChatGPT for real work.
"Hi, my name is Mackenzie, and I made an app that works better than ChatGPT. Here's how."
By the end of the third day, the video had a million views. I had 60,000 signups.
I was expecting maybe five users over a few weeks.
What followed was two months of the most stressful firefighting I've ever done. The free trial was broken — people had unlimited access while I was paying the OpenAI bills. Users were getting double-created in Supabase. Chats were disappearing. The whole thing was held together with the digital equivalent of duct tape and optimism, and tens of thousands of people were actively using it.
I hadn't been ready for showtime. Showtime showed up anyway.
How do you turn viral chaos into a real business?
The first thing I did was turn off the free trial.
A few hundred of those 60,000 signups converted to paid — which sounds like a bloodbath, and by conversion rate standards, it was. 99.5% failure. But I wasn't looking at the denominator. I was looking at the fact that a few hundred strangers had used a half-broken product for three days and still handed over their credit card. That's not a lukewarm signal. That's someone saying the problem is real enough to pay to solve even when the solution barely works.
I'd applied the same value ladder logic I'd used at Marley Baird: tiered monthly pricing at roughly $17, $47, and $97 CAD. Low enough to convert curious users; high enough to filter for people who were serious. That first weekend, a few hundred people upgraded.
The business felt real when I could draw a direct line between content and conversions — post a specific TikTok tutorial, watch signups move, watch the paid tier uptick a few days later. It wasn't luck. It was a measurable system.
By December 31, 2023, GPTBoss had done $240,000 CAD in revenue at roughly 50% margins. Solo. Zero paid acquisition.
What do users actually want from AI tools?
Nobody had figured this out yet in 2022, including me: the personas worked too well.
The "AI Developer Employee" felt real enough that users expected it to behave like a real developer. And real developers don't paste their work into a chat window — they promise to email it over. So that's what the persona did. It promised to send an email.
The email never arrived. Tool calling wasn't a thing yet. The model couldn't actually do the work; it could only think and write about it.
I was building research and thinking tools. Users wanted a "do the work for me" button. The gap between those two things is enormous, and the more convincing the persona, the more frustrating the gap.
I published tutorials. I added warnings. I made posts explaining the limitations. None of it worked — because when the interface is effective enough to make someone forget they're talking to a language model, no amount of educational content will override that feeling when it fails to deliver.
I felt stupid about it for a long time. The product worked exactly as designed, and that was precisely the problem.
How do you know when to shut something down?
In November 2023, OpenAI launched GPTs.
Custom personas, collocated with tool calls, inside their own free app. Overnight, the core value proposition of GPTBoss was available for free from the company whose API I was reselling.
I didn't fight it. I filmed a tutorial showing my users exactly how to replicate GPTBoss-style functionality in the free version. I published all seven of my original AI employees as public GPTs. Then I watched the churn.
It was significant.
What I didn't know how to say at the time — what I can say clearly now — is that GPTBoss was never really the product I wanted to build. It was the product I could build with the technology that existed in 2022. The product I actually wanted was an AI employee that could do real work: take a task, execute it autonomously, and deliver a result without a human in the loop at every step.
The technology wasn't there. So I built what I could, proved that people wanted it, and paid attention to what they were actually asking for.
When Huzi came along — a company trying to build real AI-assisted workflows for a specific customer profile — I took the job. I needed to understand how to close the gap between "thinking tool" and "doing tool." I couldn't get there alone at GPTBoss, and I couldn't keep running GPTBoss while trying to figure it out somewhere else.
So I shut it down. Intentionally. Cleanly.
Was GPTBoss a failure?
No. It was correct.
Everything I built GPTBoss to be — a virtual office of specialized AI employees that handles real business functions autonomously — is now table stakes in the AI industry. Agentic systems, tool-calling frameworks, multi-agent orchestration: the infrastructure caught up to the vision. It just took two years.
I'm writing planning docs for actually doing what those customers wanted all along. Not a persona. Not a thinking tool. An employee that takes a task and closes it.
GPTBoss taught me what people actually want. They want results, not conversations. They want an employee, not a consultant.
Everything I build from here is pointed at that.
What did building an AI product in the early days actually teach you?
Being early is genuinely hard to distinguish from being wrong.
The signals look identical from the inside: the product doesn't do what users want, churn is high, the gap between promise and reality is embarrassing, and the technology keeps falling short. Early-and-right and wrong-and-persistent feel the same until they don't.
The difference, I think, is whether the problem is real — not whether your solution is ready.
The problem GPTBoss was solving was real. Small business owners don't have access to the expertise they need. They can't afford a lawyer, a growth strategist, a developer, a designer on retainer. They make expensive decisions without good counsel and expensive mistakes without anyone to catch them.
That problem exists today. It will exist in ten years. The solution in 2022 was limited. The solution in 2025 is considerably less so.
I was right about the problem. I was early on the execution. Those aren't the same thing as failure.