Assistants are judged on their best moment and abandoned on their worst. The best moment is a good answer, and models are increasingly good at those. The worst moment is a confident answer that turns out to be wrong, or a dead end where a person needed help and the machine had nothing.
In institutional finance both failures are expensive, and the second is survivable while the first is not. A confident wrong answer about a price target is not an inconvenience; it is a decision someone might act on.
An answer without a source is a bug, not a style choice.
Draw the boundary first
So the first design decision is not what the assistant says. It is what it is allowed to say. Draw a boundary. For a research product that is published research and nothing else, and make everything outside it a handoff rather than a guess.
That single constraint does more for trust than any amount of tone work. It also makes the interface honest: every answer carries the document it came from, and a client can go and read it.
The handoff is where the design lives
Once you accept the boundary, the interesting surface is the transition. Get it wrong and you have built an expensive phone tree: the client repeats themselves, the expert arrives without context, and both conclude the assistant made things slower.
A few things make it work. Announce it, rather than silently swapping who is speaking. Carry the whole thread across, so the human inherits the question and everything already tried. Keep it in one conversation, so the client is never asked to work out which channel they are in. And set an expectation about time, because silence is the failure mode you are actually designing out.
The status detail people skip is presence. Knowing an analyst is away and typically replies within two hours is a worse answer than an immediate one and an enormously better experience than nothing, because it lets someone decide what to do next.
Protect the humans too
There is a second audience nobody puts in the brief: the experts on the other end. They are usually already at capacity, and an assistant that routes everything to them is a machine for making their week worse.
Which reframes what the assistant is for. Not to seem clever. To absorb the volume that genuinely does not need a person, and to hand over the rest so well that the handover is the best part of the product.