I Spent More Than $50,000 of My Own Money on AI. Here Is What Actually Works.

I did not raise venture capital to spend that money. I paid for it out of my own paycheque while working full time in Toronto. Over that time I worked through thousands of conversations with AI models³ and took 52 of my 99 ideas to something built or shipped.⁴ None of them earns money yet.

When you spend a venture-backed seed round, you can afford to buy the industry narrative. When you spend your own after-tax savings, every failed experiment hurts. You stop caring about what looks impressive in a demo video, and you start caring only about what survives contact with real work.

I came out of it with four findings that run directly against what most of the technology industry is preaching today.

The loudest story in technology right now is the replacement story. Large software vendors are telling big companies that artificial intelligence will replace their analysts, designers, support staff and junior engineers.

That story is wrong about human nature and wrong about how the technology actually behaves.

A language model has no skin in the game. It does not care if an invoice is wrong, if a tenant gets bad legal information, if a student learns a false formula, or if a kitchen ruins a Saturday catering order. Left alone, a model will cheerfully invent a number just to finish a sentence.

What AI actually does, when you wire it properly, is remove the mechanical friction between one capable human mind and the work that mind knows how to judge.

Before I started building software, I worked from the floor up in finance, healthcare, market research, retail operations and employee benefits.⁵ At 23, I became an operations supervisor at The Beer Store in Ontario within six months of joining, running a team of 18 people who were all senior to me.⁶ Earlier, in India, I taught myself accounting from a science background to clear the Chartered Accountancy CPT exam on my first attempt,⁷ and later taught myself R and Python to get into a postgraduate data science programme that asked for an engineering degree I did not have.⁸

In every one of those rooms, I saw the same waste. Smart, capable people were spending most of their week acting as human routers: copying numbers between spreadsheets, formatting documents, chasing status updates, and repeating rules that software can execute by rule.

AI does not replace the operator who knows the floor. It takes away the clerical tax so one person can do the work that used to require an entire bureaucracy.

Most AI companies chase contracts with large corporations, because big companies have big purchasing budgets and forgive mediocre software.

I have no interest in building for them. Large corporations already have compliance departments, engineering teams of their own and consultants on retainer. When a big company buys AI, it usually uses it to protect a slow process or to cut jobs.

The real revolution is on the other side of the street.

A big company that buys an AI system gets slightly cheaper. An individual or a small business that gets the same system can do work that used to need a much larger team. The bet behind everything I build is that one person with the right system can do what took 300 people in 2024.⁹

That is the mission of The AGI Studio and Million Dollar Studio. We build software and operating systems that help individuals, solo operators and small businesses reach their full potential. The plan is for the software we sell to small businesses to pay for free access for people who cannot pay at all: free exam preparation for government school students, and free legal information for tenants.¹⁰ The tenant tool, NestIQ, is free to use today.¹¹ The rest is still being built.

Here is the technical lesson the money bought me: raw language model intelligence is hitting a wall.

Each new model generation is slightly better at benchmarks and slightly faster at writing prose. Yet if you hand a raw frontier model a complex, multi-day business problem, it still fails in the same ways it failed two years ago.

Waiting for a bigger language model to fix these failures is like waiting for a bigger engine to steer a car that has no steering wheel, no brakes and no road map.

Superintelligence will not arrive as a single giant model. Superintelligence is a cognitive harness wrapped around any capable model.

When you wrap a model inside an operating system that holds persistent memory, runs calculations in tested code, separates planning from execution, checks every claim against its source and sends irreversible decisions to a person, the same model starts doing work it could not do alone. You can swap the model underneath and the company keeps getting smarter, because the intelligence lives in the harness.

The most expensive mistake I made early on was putting generative AI inside the hot path of execution.¹²

If a user asks for a calculation, a tax rule, a planetary longitude, a chemical equation balance or an invoice total, calling an AI model to compute that answer is a defect. It costs money on every click, it takes seconds instead of milliseconds, and sooner or later it lies.

Our rule at The AGI Studio is simple: the crane is not the building. AI builds the system; the system runs without AI.

Every product we build is designed around three tiers:

If you are an individual creator, a solo operator or a small business owner, you do not need a fifty-person team, and you do not need to wait for a mythical future model.

You need systems where deterministic software does the heavy lifting, AI amplifies your judgment, and your own taste and moral compass stay in command.

None of what I have built earns money yet. Building turned out to be the easier half. The work now is putting it in front of the people it was built for.

If you run a small business or work on your own and something here sounds like your week, write to me.