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.

The same specialist, adapted to who is asking

A specialist's identity is stable, but its delivery should adapt to the user. A tax specialist explains differently to a self-directed investor than to a fund manager — plainer language and clearer education-versus-action separation for the former, mandate and constraint context for the latter. The expertise is the same; the framing meets the user where they are. That is persona adaptation done in service of clarity, not a different agent pretending to be the same one.

Transparency is not a substitute for gates

A directory of attributed specialists is a transparency and control surface — it is not a replacement for workflow governance. If a specialist proposes a financial action, the result still flows into an action packet, approval, or workflow when it affects trading, taxes, compliance, billing, or external delivery. Seeing who recommended something and being able to inspect their reasoning makes the approval more informed; it does not remove the approval.

The takeaway

Treating AI as a directory of named, attributed specialists — coordinated by a supervisor, routed automatically by default, and selectable when you know the domain — turns an anonymous assistant into a workforce you can inspect and hold accountable. Attribution, preserved disagreement, and governance gates are what make the difference between an answer you have to trust blindly and one you can actually evaluate.