When you ask an AI assistant what the dealer map says about a ticker, something specific happens. The model calls a tool, receives a block of structured data, and then writes prose about it. Everything useful or wrong in the answer traces back to that block.
It reads fields, not pictures
A trader looking at a gamma profile sees a shape: a fat wall here, a gap there, the flip sitting below price. A model receives a list of strikes and values and has to reconstruct the shape from arithmetic. It is surprisingly good at this, and it is completely dependent on the data being labelled clearly enough to know what each number is.
Which means field names are part of the analysis. A value called gex is ambiguous; one called gamma_notional_per_one_percent is not. Vague naming produces confident misinterpretation.
It cannot tell that a number is wrong
A human who has watched a ticker for a year notices when the flip is in an impossible place. A model has no such prior. If the feed hands it a stale value or a zero standing in for a failure, it will reason over that value as though it were true and produce a perfectly fluent wrong answer.
This is the argument for feeds that refuse rather than substitute. A model handles a named error well — it says it could not get the data. It handles a plausible lie not at all.
What it is genuinely good at
Combining sources. A model can hold the dealer map, the volatility percentile, the funding picture and what the wire is reporting in one context and notice that three of them point the same way. That is real work, and it is the kind a person does slowly across four browser tabs.
It is also good at saying what it read. An answer that names its sources can be checked, which is the only property that makes an automated read safe to act on.