The agent can prepare the decision without owning it (Arca)

Last week I wrote about the evidence that a human approver is the weakest part of agent oversight. The wealth manager Arca runs a different answer in production.

In a Latent Space interview, Arca's Aman Singh describes client work spanning up to ten days, agents carrying the background analysis: mapping where a client's money sits, preparing the plan to fund a tax bill, and tracking every handoff in a persistent per-client memory that records who owns what. The agents cannot move money, liquidate assets or write to the client. Those restrictions are deterministic, Singh says, and take the decision out of the agent's hands. When the plan is ready, the agent pulls in a person to review, adjust and execute.

This is a more serious answer to the oversight problem than another approval queue. A boundary the agent technically cannot cross does not depend on anyone's attention holding up at 5pm on day nine, and a change log of every action, human or machine, gives an audit something real to work with. These are the company's claims, not verified performance, but the design is worth taking seriously.

Yet preparation is its own kind of power. The advisor who receives a finished plan, evidence pre-selected and options pre-ranked, is deciding inside a frame the agent built, and an escalation path chosen by the system can hide the failure that never gets escalated. Final approval is not the same thing as an independent decision. The question for anyone copying this model is how the boundaries are tested: what reaches the person, what never does, and who checks the checker.

A pipeline question sits underneath. If agents carry the background analysis, what trains the next cohort of advisors to recognise a bad plan? Firms adopting this model need a considered way for their juniors to earn that judgment, or the human in the loop becomes the person with the least practice in the room.

None of this argues against agents in advice work: multi-day, stop-start coordination is exactly what software should carry. But "human in the loop" is a slogan until you can say which decisions the loop can still change. The boundary drawn in code is the part worth admiring. The harder question is whether the person holding the pen is deciding, or signing.

#AgenticAI #WealthManagement #AIGovernance

Source: Latent Space, "Agents Prompt People: Using Agents for Long-Horizon Financial Work" https://lnkd.in/giNvRzjM