The best business advice already exists. The people who built real companies have been giving it away for decades. It sits in thousands of hours of talks and interviews, more than any one person could hold in their head. I wanted to know if a machine could. So I built Thoth. I gave it that whole library, and a perfect memory for where every word came from. Ask it a hard question and it does not guess. It finds what the experts actually said, and it shows me where they said it.
It runs all day on its own server. My other AI agents ask it questions, the same way I do. Almost every problem I have hit in business, someone in that library has already faced it, solved it, and explained it. Every question I ask comes back with knowledge I did not have, from people with real results behind them. I use it every day. I struck gold with this one.
What it is
I had thousands of hours of business teaching sitting in transcripts, and no good way to ask it questions. A normal AI answers from memory and guesses when it should not. So I built Thoth on a method called GraphRAG, where the AI reads every transcript once and builds a map of everything inside. Every person, method, and number, and how they all connect. The map holds 101,556 entities and 73,187 relationships. In plain terms, over a hundred thousand distinct things it found, wired together by how they relate. When I ask a question, it walks that map and hands back the exact words the expert said, with the source attached.
The library underneath is 4,639 transcripts, about 30 million tokens. Tokens are the units AI reads, roughly three quarters of a word each, so that is hundreds of novels' worth of teaching. The software is my own maintained fork of Microsoft's GraphRAG, taught to look for business concepts instead of the news-style names and places it ships with.
What you get back
The easiest way to understand Thoth is to watch it answer. Below are five real questions, the kind a business owner actually asks. Each question went to Thoth and to GPT-5.6 Sol, one of the leading AI models. Both got the same instruction, answer in one short paragraph. The answers are shown exactly as they came back.
Watch what the answers are made of. A normal model gives clean advice that could fit anyone. Thoth answers with specific plays from specific people. Each claim carries a small tag that points at the exact spot in the map it came from. When I want to check a claim, I pull up the raw source behind the tag and read what was really said.
GPT-5.6 Sol · from memory
Grow membership by defining your ideal customer, offering a compelling low-friction trial, and promoting it through member referrals, local partnerships, social proof, and targeted local ads. Convert leads quickly with friendly follow-up and a clear onboarding experience, then retain members through measurable progress, strong community, clean facilities, and consistent coaching—because better retention makes every marketing dollar more effective.
Thoth · from the library
Short answer: advertise far more than you currently do, then fix conversion. Spend the first 4 hours of your day, 6 days a week, on one of the core acquisition activities — paid ads, outreach (the "rule of 100," which is "as consistent as gravity"), content, or referral partners — because for the vast majority of gym owners "you are not advertising enough" Data: Sources (15373, 15374). Do the math backwards from your goal: if churn is ~10% and you want 300 members off 10 trials/week, you'd need to convert 75% week in week out, which means ~4 calls a day closing at 50% and 8 scheduled appointments daily — most owners' bottleneck is inflow they never operationally prepared for, and "if you want to scale, your inflow never stops" Data: Sources (20014). Improve efficiency at the three conversion steps (schedule → show → close) by role-playing scripts daily, and remember there are only two levers: more customers (traffic + conversion) or making them worth more (raise price, decrease churn) Data: Sources (15374). For fast cash from existing members, run a "reason why" campaign to 3 people at $3,000, an internal challenge ($100–$200 at a 20–25% take rate), or give the challenge away free and upsell food/supplements (~85% take rate, 80–90% close) — and add one-on-one nutrition/accountability, which drops churn hard Data: Sources (24231). Structurally, use proven mechanisms like pre-selling so a facility opens at full capacity, supplement cross-sells to cover acquisition cost, and dedicated 1-on-1 onboarding (six sessions) before group graduation Data: Entities (61153, 33206, 42579); Relationships (42010, 21914). Finally, raise your standards and fire your "Susan" — mediocre staff run your delivery and repel good talent Data: Sources (15380).
Both answers are shown exactly as they came back, scroll to read them in full. The tags are Thoth's own citations. Every number points to a real spot in the map, backed by a real transcript.
What a question costs
Thoth runs around the clock on its own server. A fully answered question costs about ten cents and comes back in about two and a half seconds. Adding new transcripts costs about nine cents an update.
The cheap answers sit on a heavy foundation. Building Thoth consumed over 680 million tokens across more than 365,000 model requests, against a library of about 30 million. Between extraction, rebuilds, and testing, the library was effectively read more than twenty times over, so that every question after it costs a dime.
How I prove it
Anyone can claim their system works. I make mine win blind tests. A blind test is simple. Two versions of Thoth answer the same 40 questions, every answer is written down with no name attached, and judges pick winners without knowing which version wrote what. The current Thoth beat the first version 63 to 11 under those rules, winning every category of question. The judges are AI models too, from two different families, because one judge is an opinion and two is a measurement.
The numbers get a harder check. All 27,469 metrics the graph extracted are verified by a script against the transcripts they came from, and anything that does not match gets flagged for human review. That script needed checking too. Its first version flagged three times too many, because spoken numbers like a million five confused it. Validators need validators.
The rebuild that lost
The best story in this project is a failure. After the first version shipped, I planned a rebuild with clear goals. I wanted a smaller, more focused graph, tuned to my needs, cheap and fast enough to stay live around the clock. So version two followed stricter rules. The extraction prompt told the AI to stay sparse, a few concepts per chunk and no more, and I fed it a leaner set of sources on purpose. In small tests, the theory held up. At full scale, it lost its blind test to the messy first version, 8 to 2.
Instead of guessing why, I did forensics. Run on identical text, the messy first version had found six times the entities and twenty times the connections. And when I measured its supposed noise, it was under three percent junk. The big messy graph was almost all real. My theory was wrong, and the test is what caught it.
The fixes went cheapest first. An AI model can only read so much at once. I changed the split so the words pulled from the library got a bigger share of that room. I also copied over the first version's better answer prompt. Those two small changes flipped the blind result from losing to winning. Only then came the expensive fix, a full rebuild. Even that was proven on a small slice of documents before the real money was spent. When a measurement surprises you, it is trying to tell you something.
The rebuild delivered. It took days of iteration, careful small tests, and a slow scale-up, and two build attempts burned down along the way. But out the other side came a graph bigger than the messy first version, with all the discipline of the second. That is the Thoth running today. It won the final blind test 63 to 11, and it stayed cheap enough to keep running around the clock.
The model under the hood
Under the hood, Thoth is driven by one of the strongest AI models in the world. It did not get the job by name. Every model had to try out for it, blind. I sent the same questions to eight of the top models, from every major AI lab, and hid the names before any answer was judged. Across 63 blind answers, one model kept winning. It took 10 of the 12 questions, so it got the seat.
The seat is not safe either. The day a newer model came out, I ran the same blind test again. The newcomer won all 12 questions, so the job changed hands. Whoever wins the test writes the answers. That stays true no matter whose name is on the model.
What would your team do with your whole industry's knowledge on call?
Discuss your project