Everyone can call a frontier model. The API is a credit card away, and the model behind it can read an annual report, work through an earnings call, and reason about a business with real competence.
So why build a product around one?
Because a raw model, dropped into investing, fails in a specific and predictable way. Not because it lacks intelligence, but because it is standing in a void. It does not know what you own, what you believe, or what you concluded last quarter. Every conversation starts from zero, and the burden of supplying all of that falls on you, every single time.
The gap between a model and an analyst is not intelligence. It is everything around the intelligence.
We call that everything the harness. Here is what it is actually made of, and why each piece matters.
Context: the model needs a place to stand
A language model only knows what is in its context window. That is its entire working memory for a conversation. Whatever is not in there might as well not exist. This is why the same model gives a generic answer in a chat window and a sharp one inside a product: the product decides what the model gets to see.
An analyst is only useful relative to your situation. The same piece of news means different things depending on whether you own the company, watch it, or hold its biggest competitor. So the first layer of the harness is deciding, on every question, what belongs in front of the model: your portfolio, your watchlist, your notes on the company being discussed, the thesis you wrote three months ago.
The hard part is selection, not volume. Stuffing everything into the context makes answers worse, not better, because the model has to find the signal in a pile of noise. A good harness curates. It puts the three things that matter in view and leaves the rest out.
PortfolioWatchlistYour notesOld thesisContext windowModelAnswerWhat the model sees, per questionTools: reasoning over real numbers, not recollections
A model's training data is a snapshot of the past. Ask a raw model for a company's operating margin and it will give you a number that sounds right, drawn from whatever it saw during training. It may be a year stale. It may be subtly wrong. It will be stated with total confidence either way.
The fix is not a smarter model. The fix is tools: letting the model fetch the actual numbers instead of recalling approximate ones. Inside Investi, the analyst pulls live fundamentals, financial statements, earnings history, and ownership data as it works. When it cites revenue growth, that figure came from a data call it made seconds ago, not from a training set with a cutoff date. Grounding answers in fetched data is the single most effective way to cut hallucination in a domain where the numbers are the whole point.
Just as important, the work is visible. You see the analyst think, see which data it fetched, and see how the pieces connect to the conclusion. This matters for a practical reason: it lets you catch the model being wrong. An answer you cannot audit is a headline. An answer with its work shown is analysis you can check and challenge.
QuestionRecalled from trainingAnswer, maybe staleQuestionLive data callAnswer, groundedTwo ways to answer the same questionMemory: the thesis is the unit of work
Software development had a natural container for AI to plug into: the codebase. It holds the project's entire history and intent, and every change is made against it. That container is a large part of why AI worked so well for code so quickly.
Investing has an equivalent, but almost nobody treats it as one. The thesis is the codebase of investing. What you believe, why you believe it, and what evidence would change your mind. Most investors keep it in their head, which is exactly where it degrades: memory edits itself, convictions drift to match the price, and the original reason for owning something quietly disappears.
So Investi is built around research that persists. Your notes on a company live next to its financials. The analyst reads them before it answers and refers back to what you concluded before. Ask about a company in October and it knows what you wrote in July. The work compounds instead of evaporating when the chat window closes, and that accumulation is the difference between a chatbot and a colleague.
Jul: noteAug: noteSep: noteOct: questionResearch that accumulates instead of resettingTrust: the model proposes, you decide
Here is a design decision we will not compromise on: the analyst never silently changes your research.
When you ask it to update your notes, say to fold in a new quarter or revise a risk section, it drafts the change as a diff. Here is what your notes say, here is what would change. Nothing is written until you approve it.
This borrows directly from how the best coding tools work: the model proposes, the human reviews, the change is explicit. And the reasoning is not ceremony. Your notes are the record of your judgment, and they only have value if every word in them passed through you. A model that rewrites your reasoning unreviewed would defeat the purpose of keeping reasoning at all. The analyst does the heavy lifting. The judgment stays yours.
Analyst drafts a diffYou reviewNotes updatedThe analyst proposes, you decideWhy the harness is the product
A fair question: will better models make all of this unnecessary?
We think the opposite. A smarter model with no context is still guessing about your situation. A smarter model inside the harness reads your thesis more carefully, uses the tools more skillfully, and catches contradictions you would miss. Every model improvement makes the layer around it more valuable, the way a stronger engine rewards a better car.
And the harness is where the durable work is. Model access is becoming a commodity; everyone will have roughly the same intelligence on tap. What compounds is the layer around it: the context that makes answers relevant, the tools that make them true, the memory that makes work accumulate, and the trust that makes people willing to rely on it.
Software already proved this pattern. The tools that changed who could build software did not invent intelligence. They wrapped it in context, memory, and workflow, and connected it to the actual work.
We are building that wrapper for investing. The engine is ready. This is the car.