Context

A paraplanner is checking an AI draft of Helen Marsh's annual review letter. The draft describes Helen as a cautious investor and suggests moving part of her portfolio into lower-risk funds to match. The paraplanner knows this is wrong. At last year's review, Helen agreed to a balanced attitude to risk, and her portfolio was changed to suit it.

The whole client folder had been uploaded, and it held two fact-finds: one from 2023 and one updated at last year's review. Both were dated. Nothing told the AI that the newer one replaced the older, and the draft followed the 2023 version. The suggestion to move into lower-risk funds rests on that choice, so correcting one sentence would still leave the letter wrong. The letter needs redrafting from the current fact-find.

Everything the AI was given is its context. Context is the material an AI system is given for one piece of work. It includes the question and the client's documents, and usually instructions added by the software that the user never sees. The AI's general knowledge comes from its training, the process that built it before any firm used it. What it knows about this client comes from the context and nowhere else.

In a firm, context is the file you hand a colleague before they write. The difference shows when the file is wrong or incomplete. A paraplanner handed two fact-finds would ask which one stands, and one given no pension statement would ask for it. Unless the software has been built to stop and ask, the AI carries on with what it has. That is why uploading the whole folder is not the safe choice it feels like. More material gives the AI more to choose from, not a better sense of which document counts.