This one started in 2023, when the federation was still just me and one machine.
The feature was called Data Analysis. You attached a spreadsheet and it did the analysis. That was the whole thing, and at the time it was remarkable.
Deep research arrived later. Better tools after that. Three years on, the tooling is not what changed most.
What changed is how much I say.
Treat them as fully capable
This is the lesson, and it took me longer to learn than it should have.
Treat an analyst as though it is fully capable. Do not tell it what to do.
A serious analysis in August 2026 looks like this. I attach the dataset. I ask: do we go or no-go on this portfolio?
That is the prompt. That is all of it.
If I want to get fancy I might add that it should include an executive summary — and honestly, that is kind of stupid. It was going to do that.
Attach the data. Ask the question. Stop talking.
You are hiring the firms yourself
When the decision is big — a twenty-seven-million-dollar go or no-go on a portfolio — you do what any company facing that decision has always done. You hire the consulting firms. Plural.
What changed is the roster.
You are no longer hiring McKinsey, Bain, PwC and Deloitte.
You are hiring xAI, Google, Anthropic and OpenAI — yourself.
And you are hiring all of them at once, for a decision that used to justify a single engagement, a six-week timeline and a seven-figure fee.
Each returns its own analysis. They do not see one another’s work. Only after all of them exist does a further pass receive them together — not as opinions to average, but as its starting material.
You do not buy the second opinion for the answer. You buy it for the disagreement.
The mistake that will cost you
Here is where people lose real money, and it is not on model fees.
You have an algorithm in your head. The way you underwrite. The order you look at things. The adjustment you always make on the third pass. And you decide that the job now is to teach that algorithm to all of your AIs.
That is when you are way off.
The moment you set out to teach five machines the algorithm in your head, you have bought five expensive copies of yourself.
It is going to crush you and it is going to cost you a ton — in fees, because every correction drags the entire accumulated context forward again, and in outcome, because you paid for a different geometry and then spent the budget flattening it back into your own.
The whole reason to have five is that they are not you. A room of five analysts who have been trained to reason exactly as you do is not a federation. It is an echo, and you are paying five times to hear it.
If your method genuinely is load-bearing — proprietary, hard-won, the actual reason you win — then encode it as a durable procedure before the work starts. That is a different act from teaching it conversationally through twenty corrections, and it costs a fraction as much.
Interrogate. Do not correct.
When an analysis comes back with a conclusion I did not expect, the instinct is to fix it. That instinct destroys the thing I paid for.
So the questions are always the same. Which fact changed your conclusion? Which source produced it? Which assumption carries the most weight? What would reverse it?
Disagreement usually begins at the index, not the reasoning. Different tools, permissions, retrieval and timestamps mean two analysts given identical instructions were frequently not looking at the same world. One had the signed leases. One had a rent roll eight months stale. One had walked the property.
The most valuable result I have had from this was not a better number. It was discovering that four analyses agreed because all four had read the same stale document — and the lone dissenter was the only one that had gone and looked.
What actually improved
There are roughly 8,400 analyses in this ledger: underwriting, diligence, market reads, contract review, portfolio comparisons.
The tooling improved enormously across those three years. But the change that moved the output most did not come from the model side.
In 2023 I attached a spreadsheet and told it what to calculate.
In 2026 I attach the portfolio and ask whether to buy it.
The machines got better. So did the question.