The value is obvious
Background on a prospect before a call, a summary of a market, a first draft of a brief. Work that currently gets skipped for lack of time, which makes it new capacity rather than a cost saving.
The risk is equally specific
Research output is prose that reads authoritatively and is consumed by someone who does not know the subject — which is exactly the situation where a confident fabrication survives into a client conversation.
Research is the one job where the reader can’t tell a real fact from an invented one. That’s the whole problem.
Require sources, and check they exist
Every claim should carry a source that can be opened, and the system should be checked for whether the cited page actually contains the claim. Unsourced research is a draft of a guess.
Separate what it found from what it inferred
Retrieved facts and the model’s synthesis should be visibly distinct in the output. Mixing them is how an inference becomes a quoted statistic two meetings later.
Brief it like you would a junior
The question, the audience, the decision it supports, and what to leave out. Vague research requests produce broad summaries nobody can act on — the same specificity that a writing brief needs.
Keep a human between it and the client
Research output should always be read and edited by someone accountable before it leaves the building. This is draft-and-approve territory, permanently, not a stage to graduate out of.