One anonymous voice is not enough
Ask most AI assistants a wealth question and you get a single, anonymous voice. It sounds confident, but you have no idea which expertise produced it, how sure it is, or what it relied on. For casual questions that is acceptable. For decisions about tax, portfolio construction, or compliance, anonymity is a liability — you cannot evaluate an answer you cannot attribute.
A better model treats AI as a workforce of named specialists rather than one faceless oracle. Each specialist is scoped to a domain — portfolio, tax, research, planning, trading, reporting, compliance — and when it contributes, its contribution is attributed. You see who answered, in what domain, with what confidence, and on what evidence.
A directory you can actually inspect
The starting point is a directory: a place to see which agents exist and what each is responsible for. Functionally, agents fall into a few types:
- Supervisor — coordinates routing, clarification, synthesis, and approvals, and owns the user-facing state.
- Specialist — reasons inside one domain such as portfolio, tax, or research.
- Task agent — performs a bounded, repeatable workflow or background job.
- Utility agent — supports retrieval, formatting, or API work, and usually stays behind the scenes.
Being able to browse this roster — with each agent's domains, status, recent outcomes, and the work attributed to it — turns the system from a black box into something inspectable. You can confirm that a tax question was actually handled by the tax specialist, not a generalist guessing.
Auto-route by default, choose a specialist when you know
Most of the time you should not have to pick an agent. You bring a question and the supervisor decides whether it needs one specialist, several, a workflow, or a clarification first. Auto-routing is the right default for broad or cross-cutting questions where the domain isn't obvious.
But when you already know the domain, choosing a specialist directly is faster and cleaner — a tax specialist for a harvesting or conversion question, a portfolio specialist for drift or rebalancing, a research specialist for a security or fund review. That direct selection should be scoped to the current thread, not a permanent override of the system's routing, unless you explicitly pin it. And critically, even a direct specialist conversation stays under supervisor coordination rather than becoming an ungoverned side channel.
Attribution is the whole point
When specialists contribute to an answer, the result should show its seams rather than melting into one anonymous reply. A well-formed specialist answer includes:
- the specialist's name and domain,
- a concise finding,
- a confidence level or caveat where known,
- citations for material claims,
- a way to inspect the fuller analysis.
And the supervisor's synthesis should be labeled as synthesis — distinct from any single specialist's opinion. When specialists disagree, the interface should preserve that trade-off rather than smoothing it into a false consensus. The disagreement is often the most decision-relevant thing on the screen.