How ProAI was built by fine-tuning a firm’s data and sold after seven figures
Chase W. Hughes is an AI product leader and three-time founder who runs an independent consultancy and previously built and sold ProAI. He turned a New York college consulting shop into a global, web-first business, then fine-tuned its data into ProAI, which reached 300,000+ users and generated seven-figure revenue before he sold it. The story is worth reading for his skeptical-but-practical approach: make AI prove itself and keep anything expensive-to-get-wrong outside the model.
Decide where a probabilistic system is allowed to be wrong before you automate: give models the narrow, checkable work and put anything expensive to get wrong on deterministic rules or human oversight. That discipline, not the model choice, is what prevents costly failures in real deployments.
What’s the quick origin story of your business, and what makes what you do genuinely different from the other options your customers are looking at?
I started a consulting firm in college — business plans and financial models for people raising money. It worked, and it was completely dependent on being in a room in New York. When COVID hit I rebuilt it as a global, web-first business instead of waiting it out, and it grew about 5x.
Then in 2021 AI came for that business and I didn’t defend it. I fine-tuned the firm’s own data on GPT and built ProAI, one of the first commercialized GPT products: 300,000+ users, from the Abu Dhabi Investment Authority and Keiretsu Forum to national Small Business Development Centers and thousands of CPAs and startups. Seven figures in about 18 months, entirely bootstrapped, then sold.
What makes it different is probably unfashionable. I’m a hard skeptic about AI and all-in on it at the same time. I make it prove itself and reach for it last — and last resort turns out to be exactly where it belongs.
Where does most of your new business come from today, and where do you wish it came from?
Referral, almost entirely — people I’ve built something for before, or people they send. That was true of ProAI too: bootstrapped means there is no customer-acquisition budget, so the product either travels by word of mouth or it doesn’t travel at all.
Where I wish it came from: people who’ve read the argument first. The conversations that open with "I read the thing about why pilots fail and I think you’re wrong about the second part" are worth ten inbound leads, because the work has already started. That’s most of why I write anything.
What’s one thing you’ve actually handed over to AI in the last year, and one thing you tried it on and went back to doing the human way?
Handed over: first-pass research. Reading everything new in a field, pulling out what actually changed, and telling me what I need to go look at myself. That’s the thing I filed on — a patent-pending multi-agent research system, filed early 2023, where a supervisor hands work down to cheaper, faster agents. In practice it feels less like using a tool and more like managing a team.
Went back to the human way: anything where being wrong isn’t survivable. ProAI’s financial projections looked like the perfect AI job and they weren’t — a hallucinated number in somebody’s fundraising model is not a cute mistake. So the model did the narrow, checkable parts and everything expensive to get wrong went to a deterministic rules engine underneath it. That’s still my rule: decide where the system is allowed to be wrong before you decide what to automate.
Have you noticed a change in how customers find you or what they already know before they reach out? Is anyone arriving through AI assistants like ChatGPT yet?
Honestly, I don’t have clean attribution for it, and I’d be a little suspicious of anyone who says they do. Assistants don’t hand over a referrer the way a search engine does, so most of what gets called "ChatGPT traffic" is a guess wearing a number.
What I can see is the second half of your question. People arrive already briefed. They’ve read something, or something read it for them, and the first conversation now starts three steps in — nobody asks what a multi-agent system is anymore, they ask why theirs stalled after the demo. That’s a real change and it raises the bar, because the generic version of your answer is already free.
If you were advising someone in your industry on all of this for the next 12 months, what would you tell them to actually do, and what would you tell them to ignore?
Do: take one workflow you genuinely understand and find the exact point where a probabilistic system is allowed to be wrong. Put the checkable work on the model and put anything expensive to get wrong on something deterministic. It’s unglamorous and it is most of the job.
Then look at who is running your pilot. Roughly 95% of enterprise AI pilots never show a measurable return (MIT Project NANDA, "The GenAI Divide," 2025), and the cause usually isn’t the model — it’s that the person accountable for it has a job at risk whether it succeeds or fails. No tooling decision survives that.
Ignore: the model of the week. And ignore anyone, me included, who sounds certain about 2028. We are still very early, today’s assumptions have a one-to-two-year shelf life, and building as though you can see further than that is the actual mistake.
Thank you to Chase W. Hughes and the team at Independent for sharing what actually worked with Leaders Perception readers.
Want to share your playbook?
Leaders Perception publishes founder and operator interviews every week. A few questions, about five minutes, no calls. It is free and everyone we accept gets published.