AI in Trading: How It Actually Works
"AI in trading" gets used for four distinct jobs that have almost nothing in common except the label. Here is what each one actually does, where each one actually fails, and what has to sit outside the model before any of it is safe to run unattended.
Ask ten people what "AI in trading" means and the answers overlap without agreeing: a model summarizing earnings calls, a model predicting tomorrow's move, an optimizer rebalancing a portfolio, a bot placing live orders. All four get called the same thing. Only one of them is actually deciding what happens to real money without asking first.
Four jobs, one label
Each does something specific, with a specific way it breaks.
1. Sentiment and news analysis
A language model reads headlines, filings, earnings-call transcripts, or social posts and scores them — bullish, bearish, or somewhere between. It is genuinely useful for surfacing that something changed faster than a person scanning the same sources by hand.
Where it fails: sarcasm and context are still hard for these models to get consistently right; a lot of what gets scored has already moved the price by the time the score exists, making it a coincident read rather than a leading one; and a model is only as good as the sources it is fed. Low-quality or manipulated input produces a confident, wrong score.
2. Signal generation
A model trained on historical price, volume, and derived features outputs a prediction or a trigger — buy, sell, a probability, an entry level. This is the category most people picture when they hear "AI trading."
Where it fails: overfitting to noise in the training window is the dominant failure mode. A model can find a pattern that explained the past perfectly and predicts nothing about the future. Even a genuinely real pattern can suffer alpha decay once enough capital trades against it. And a signal is not a plan: it says nothing about position size, stop placement, or what to do when it is wrong.
3. Portfolio and risk optimization
A model adjusts position sizing or allocation across a portfolio based on estimated correlation and volatility, aiming for a defined risk profile rather than picking individual trades.
Where it fails: correlation estimates are built from a historical window, and the moment they matter most — a market-wide shock — is exactly when correlations spike and the diversification the model assumed is not there. An optimizer tuned on a short or calm window looks precise and is often just precisely wrong.
4. Full execution agents
The model observes market state and independently opens, manages, or closes positions — no person confirming that specific decision before it reaches a broker.
This is the highest-stakes category, and not because the model is smarter or dumber than the other three. It inherits every failure mode above — sentiment misreads, overfit signals, bad correlation assumptions — and adds one more: there is no human between "the model concluded X" and real capital moving. Whatever the model gets wrong, it gets wrong at full speed.
"The AI decides everything, no oversight" is the wrong mental model
For the first three categories, a person is still the last check before anything happens to real money: the model informs a decision a human still makes. Full execution agents are where that check has to be rebuilt somewhere else, because it no longer happens by default. Treating any of the first three like the fourth understates the risk. Treating the fourth like it does not need the same safeguards as the first three is how accounts get damaged in ways that are entirely avoidable.
What actually has to sit outside the model
For a full execution agent to be survivable, the safeguards cannot depend on the model policing itself. They have to be enforced by something the model does not control:
- Hard risk limits set in advance. Maximum position size, a daily loss cap, a maximum number of concurrent positions — decided by a person ahead of time and enforced by code the model cannot negotiate with in the moment.
- Approval gates above a defined threshold. The model proposes; a human, or an explicit rule the human wrote, confirms before the order reaches a broker.
- A complete, plain-text audit trail. Every decision logged with its reasoning, before and after, so a bad trade can be diagnosed and fixed rather than shrugged off as unexplainable.
- Staleness checks and a kill switch. If the data feed goes quiet or a connection drops, the safe default is to stop and flag it — not to keep trading on a guess about current price.
If the question underneath this is about frequency
Some of what gets asked as "how does AI trading work" is really a question about pace — is this a system trading every few seconds, scanning dozens of symbols at once, firing constantly through the day? That is a separate, narrower question from what is covered here, and it deserves its own answer: see [AI Day Trading: How Cortiq Trades Intraday](/ai-day-trading) for how a wake-based, single-instrument model differs from a system that never stops moving.
Where Cortiq fits
Cortiq is a full execution agent (category four above), built around the guardrails in the previous section rather than around a claim that the model does not need them. Risk limits, trade approval, and the audit trail sit outside the AI and cannot be reasoned around by it. Every session runs on a single instrument, a deliberate constraint that keeps the evidence behind each decision deep enough to actually inspect, rather than spread across a basket of symbols. And the AI itself is bring-your-own: you connect the provider you choose. Cortiq does not run a managed model on your behalf.
The full trade cycle — how a decision gets made, checked, and logged from wake to close — is in [How Cortiq Actually Works](/how-it-works).
Key takeaways
- "AI in trading" covers four different jobs — sentiment analysis, signal generation, portfolio optimization, and full execution — with almost nothing in common except the label.
- Each category has its own failure mode: sentiment lags or misreads context, signals overfit or decay, optimizers assume correlations that break exactly when they matter most, and execution agents inherit all three at full speed.
- "The AI decides, no oversight" is the wrong model for anything serious. A person is the last check in three of the four categories by default — the fourth has to rebuild that check deliberately.
- Survivable automation puts risk limits, approval gates, and an audit trail outside the model, enforced by rules it cannot override — not inside the model's own judgment.
- None of this is about finding a smarter model. It is about what sits around it.