Chat is the entry point, not the whole story

Chat is the natural front door to AI-assisted wealth work. You ask a question and a supervisor can answer directly, pull in specialists, run research, launch a workflow, or pause for your input. The experience should feel like one coherent assistant — but underneath, what separates a toy from a tool is how it handles research.

A research question deserves more than a confident paragraph. "Why did this portfolio underperform?" or "How does this ETF compare to our current holding?" are not lookups — they require a plan, sources, and a clear line between what is known and what is assumed.

What deep research should actually do

A research-grade answer has structure that a casual answer lacks. It should:

• state the question and the specific entities in scope,
• identify its sources — whether the answer comes from structured portfolio data, internal documents, external knowledge, or a specialist model,
• keep live data out of retrieval — current holdings, balances, prices, and orders come from structured systems, never from retrieved prose,
• cite material claims with source links,
• show assumptions and caveats where the evidence is incomplete,
• leave an artifact — a report, summary, or activity item — when the result needs review or follow-up.

That last point is what makes research durable rather than disposable. A good research answer is not just read and forgotten; it produces something you can return to, review, and act on.

The cardinal rule: prose is not operational truth

Here is the line that protects you from a whole category of AI error: retrieved text is never the source of operational facts. A document might describe a strategy or summarize a holding, but your actual current positions, balances, prices, lots, and orders must come from structured systems of record.

Mixing these is how AI tools produce confident, cited, and wrong answers — citing a document's stale figure as if it were your live balance. Keeping current portfolio data out of retrieval-only sources is not a limitation; it is the discipline that keeps research honest.

Cards that show the work

Because research crosses specialists and sometimes launches workflows, the chat should make the process visible through cards:

• a specialist card showing who is working and why,
• a supervisor synthesis that is clearly separated from individual opinions,
• a workflow progress card when a flow is running, showing the active step and current blocker,
• a human-in-the-loop card when an approval or response is needed,
• an artifact card linking to the report, proposal, or summary produced.

The point is that the reasoning is not hidden behind a single bubble of text. You can see which specialist contributed, where the work stands, and what it produced.

Research that knows when to hand off

Consider "prepare for tomorrow's client review." A shallow assistant writes a summary. A research-grade one assembles client context, surfaces open issues, gathers the relevant report and proposal artifacts, and — because this touches client-facing work — routes through an advisor approval gate. Or "can we harvest losses?": the tax and portfolio specialists run a wash-sale check, propose substitute candidates, and stage an action packet behind an approval rather than just answering yes.

In each case, research is not the end of the line. It flows into the durable work surfaces — activity, workflows, artifacts — that own the actual decision. Chat can summarize and launch work; it should never hide the durable state behind it.

The takeaway

Deep research in a wealth chat should be measured not by how much it writes but by how it reasons: a stated question, sourced evidence separated from assumptions, live data kept out of retrieval, citations for material claims, and an artifact left behind for review. Paired with visible specialist and workflow cards and a clean handoff to durable surfaces, that turns chat from a clever talker into a research partner you can actually rely on.