PMS copilot, chat assistant or agent: classification and choice

Portfolio managers now meet AI in three forms: a copilot inside the portfolio management system, a general assistant such as ChatGPT, Claude or Microsoft 365 Copilot, and an agent that works across systems. The usual question is which tool is best. In our view, that is the wrong question. Which form fits depends on where the data for a task lives and who answers for the result, and the regulatory classification turns on the degree of autonomy, not on the type of tool.

A PMS copilot ends where the data model ends

Many PMS vendors are building assistants into their platforms. They know the vendor's data model, respect its permissions and answer questions on positions, exposures or performance in plain language. For questions that stay within the portfolio, such as which mandates hold a given issuer or how the equity allocation has developed, that is often enough.

The limit is the data model. A PMS copilot does not see the client's emails, the notes from the last meeting, the KYC file in the document store or the house view of the investment committee. Most documentation work in wealth management needs exactly these sources. A meeting brief, a quarterly letter or a periodic KYC review combines figures from the PMS with information that sits elsewhere.

General assistants shift control to each employee

General assistants handle a wide range of requests. Several now offer connectors to mailboxes and file stores, and Microsoft 365 Copilot works within the firm's Microsoft environment and its access rights. The connection itself is therefore no longer the difference. What usually remains open is who decides which client data an assistant may use, how a figure can be traced back to its source and who approves a result before it reaches a client.

Without firm-wide rules, each employee answers these questions alone, in every conversation. Confidentiality then depends on what someone chooses to paste or connect, which is how shadow AI arises. For a portfolio manager, this touches professional secrecy under Art. 69 FinIA as soon as client data reaches a provider that is not bound as an agent. Without structured logs, there is also no record of what was done.

An agent needs rules before it gets access

An agent connects to several systems, plans the steps of a task and executes them. A meeting brief, for example, requires the portfolio from the PMS, recent correspondence from Outlook, open items from the client file and the house view from a shared folder. This reach is the benefit and the risk at the same time.

In our view, an agent in a regulated firm needs four rules before it gets access: which system may see which data, a source for every statement, a log of every step and a named person who approves the result. These rules are not a feature list. They are the conditions under which the firm can explain the agent's work to its audit firm.

Autonomy, not technology, decides whether a task is delegated

Art. 14 FinIA allows portfolio managers to delegate tasks to third parties. Under Art. 15 para. 1 FinIO, a delegation exists where a service provider performs a material task independently and on a lasting basis. Portfolio analysis is a material task under Art. 19 para. 3 FinIA. Whether a delegation exists therefore usually turns on independence.

In our view, a copilot that answers questions and an agent that prepares drafts for approval do not perform a task independently, as long as a qualified person reviews each result before it takes effect. An agent that sends reports without that step does. Where a delegation exists, Art. 17 FinIO requires an agreement in text form with defined control and audit rights. The same platform can run in both modes, and neither FINMA nor the courts have yet addressed AI agents specifically. Each use case therefore belongs in the AI inventory on its own, with its degree of autonomy, as explained in our article on FINMA Guidance 08/2024.

Four questions decide the choice

In our view, four questions about the task determine which form fits:

1. Does the work stay in one system? If so, the PMS copilot may be sufficient.
2. Who checks the figures? If someone compares them against a report, every figure needs a traceable source.
3. What leaves the protected perimeter? If client data is involved, controls must apply before data moves, not only as a policy afterwards.
4. Who owns the result? If it goes to a client, a human approval and a record are required.

Most firms will use more than one form. A PMS copilot for portfolio questions and an agent for the documentation work around the portfolio do not compete. What matters is that the firm knows which tool handles which task and where client data flows.

How it works

Unplex is a cross-system layer for this documentation work. It connects to existing systems such as Assetmax and Infront, to Microsoft 365 and to the document store, and carries out multi-step work such as meeting briefs, quarterly letters and KYC reviews. Every figure keeps its source, client-facing results require approval, every step is logged, and client data is not used for model training.

Conclusion

The right form depends on the task, not on the tool. Work that stays within one system belongs in the PMS. Work across systems requires an agent, and an agent requires clear rules on data access, sources and approval, because its degree of autonomy decides the regulatory classification.

Request a demo

Last reviewed: October 2026. General information, not legal advice.