Model

For months, a paraplanner corrected the same slip in every suitability report the firm's AI product drafted. It spelled adviser with an o. Each fix lasted one draft. Then, in July, the slip stopped, and the paraplanner decided the product had learned.

It hadn't. In July the vendor had moved the product onto a newer AI model, because the old model's provider was retiring it. The vendor mentioned the change only in its release notes. The new model spelled adviser correctly, but it also rewrote the firm's standard risk warning, which the old one had always copied exactly. In Priya Desai's report, "you may get back less than you invested" became "the value may go up and down". It no longer told her she could lose money. The paraplanner and the adviser checked every figure and skimmed the standard warning. The file check found it in October, after the report had gone.

The model inside this product learned from a vast amount of text before release. The paraplanner's corrections changed each draft, never the model. A model is the part of an AI product that produces its output. A model provider, often not the product's vendor, trains it. It does not change as the firm uses it, and a newer model is a different model, with habits of its own.

Many firms already rely on a model someone else built: the cashflow model. When its supplier updates it for a new tax year, the firm checks which assumptions changed. Nobody can check an AI model like that. It holds what it learned as numbers, not written assumptions, and not even its provider can read them as a list of what it learned. When a new model arrives, only trying it on past cases shows how it handles the firm's work.