Why I won't let AI make my investment decisions, and what I built instead

"AI investing agents" are everywhere right now. Tools that pick stocks for you, buy and sell for you. The demos all look the same: the agent reads the news, glances at a chart, and produces a confident paragraph explaining why it is buying.

I started down the road of building one of these and arrived at the opposite conclusion. Handing an AI the whole investment decision is premature. Not because the technology is slightly short, but because the structure of the problem does not fit.

Investment decisions do not come from the information in front of you alone. They come from an insight synthesized across many threads, mixed with a personal belief about the company and an expectation about the future.

Think about someone buying Tesla. Almost nobody buys it on the current financial statements. They are betting on a belief that autonomous driving will arrive and an expectation that this company wins that market. That belief is written nowhere in the present data. Two people read the same balance sheet, one buys and one sells, and both can be rational. They believe different futures.

An LLM, by construction, is a machine that reasons over the information it currently holds. It synthesizes its training data and its context into the most plausible next sentence. That is a genuinely useful capability, and it is a different kind of work from betting on the future. Ask a present-information reasoner to place a future bet, and what you get back is not a bet. It is a summary of present information wrapped in buy/sell phrasing. The prose of conviction without the belief that conviction rests on.

The practical problem I saw firsthand while building. I gave a prototype the same portfolio, the same data, and the same question twice, and got two different answers. Buy today, wait tomorrow, with nothing changed in between. When a human changes their mind there is a reason. This was wobble without a reason, and watching it made one thing obvious: you cannot put money on top of a wobbling judgment.

To be fair, agents that make money do exist. A tiny minority. And the shape of this resembles a trap quantitative investing has known for decades.

Let me scope this precisely, because "quant is a trap" would be wrong. Good quant funds make real money. The trap is this: taking rules optimized to fit past and present data, and automatically applying them to bets on the future. Nothing guarantees the rule survives into the future. Most such rules only work in a particular regime, or are simply overfit. The backtest looks beautiful and melts in production. A rule learned in a low-rate regime quietly breaks the moment rates turn, and you find out it broke long after it did.

LLM agents step into the same trap under worse conditions. A quant's rules are at least explicit, so you can trace after the fact when and how they failed. An LLM's "rules" live somewhere in the weights: it cannot tell you why it decided, and the decision does not reproduce. And on top of that, the arithmetic itself wobbles. Handing your portfolio to a machine that returns a different number each time you ask it to compound is exactly that.

So I came to see the role AI plays best in investing as something other than the decision itself.

Gather the scattered information and shape it into something a human can actually digest. Lay down the raw material of judgment, accurately and consistently: what state the portfolio is really in, where it is concentrated, what distribution falls out of a given set of assumptions. That is the AI's job. Believing in a future and pulling the trigger on top of that material is the human's job.

This division of labor sounds modest, but it is surprisingly powerful, because a large share of investing mistakes happen on the material side, not the belief side. You believed you were diversified and you were not. You confused the average return with the compound return you actually keep. When the material is right, belief can do its work. When the material is wrong, the best belief in the world is walking on a wrong map.

I concluded this model is the best fit for investing at the current state of the technology, and built a tool to match. It is called Opula.

Opula is a hosted MCP connector for Claude or ChatGPT. Two design decisions inside it carry the argument above directly.

One: the numbers are never computed by the LLM. Ad-hoc arithmetic wobbles, so the server computes every figure under the same deterministic rules and the AI only narrates the result. Ask "when do I reach $250k?" and instead of inventing a date, it runs a Monte Carlo simulation and returns probabilities and percentile ranges, because the honest answer is a distribution, not a date. And the return and volatility assumptions behind that simulation are never filled in on your behalf: every projection states the assumptions you chose, verbatim. Assumptions about the future are the domain of belief, and the belief is yours.

Two: no account linking. Instead of an aggregator like Plaid, you record by talking: "bought 10 shares of Tesla." Which means assets a bank feed can never see, say a Korean jeonse deposit, land in one ledger, and ETFs are unpacked via look-through into effective exposure. "I'm well diversified" turns out to be 26% AAPL more often than you would think.

You will notice both decisions serve an AI that lays material accurately, not an AI that decides well. Opula never tells you to buy or sell. It is a diagnostic tool, not an advisor. If the diagnosis is accurate, the decision belongs to the person reading it, made on their own beliefs.

Don't some people make money with AI trading agents? Yes. A tiny minority. And separating whether that minority won by skill, luck, or regime takes more than one regime's worth of time. Quants, whose rules are explicit, need years to make that separation. LLM agents, whose judgment is opaque, need longer.

Is this an attack on quant investing? No. Good quants make money. The trap this essay scopes is "automatically applying rules optimized on past data to future bets," which is the failure mode the quant industry itself guards against hardest. The point is that LLM agents risk repeating that trap without even the transparency of explicit rules.

So AI is useless for investing? The opposite. Gathering and organizing information, consistent computation, concentration diagnostics: AI does the material work better than humans and never gets tired. The one role that does not fit is deciding future bets on your behalf.

Does Opula give buy/sell recommendations? No. It diagnoses the current state of your portfolio, and when you supply assumptions, it computes the distribution of outcomes under those assumptions. The trigger is always yours to pull.