This wealth firm lets AI do the prep work and humans move the money

Arca, a wealth management firm that calls itself AI-native, has AI agents handle most of the behind-the-scenes work. Only human advisors can make trades or move client money. Arca's lead AI engineer, Aman Singh, described the goal this way: "agents prompt people more than people prompt agents." He spoke on the Latent Space podcast after AI Engineer New York.

What the AI does

Financial work moves slowly and in pieces. Singh said agent tasks at Arca can take up to ten days, with many handoffs and decision points along the way. The agents analyze and propose solutions. Then they nudge a person to act. Arca recently raised $48.5 million in a round led by General Catalyst, with Index Ventures and Venrock also taking part.

Memory for every client

Because tasks stretch over days, Arca built its own memory system. For each client it tracks what has happened so far, what comes next, what is urgent, what is a long-term goal, and which person owns which action. The agents add to this record as they work.

Guardrails that do not depend on the AI

Singh said all of Arca's safety limits are "fully deterministic." That means fixed rules that work the same way every time. They take decisions out of the agent's hands. The agent can run some tools alone, but not ones that put a client at risk. Arca also runs checks for known AI security problems, such as prompt injection, where hidden text tries to trick an AI into following bad instructions.

When agents write computer code to crunch client data, that code runs in a sandbox, which is a walled-off space. The code is converted into a strict format and then thrown away. Guardrails check that format before it reaches Arca's systems. This lets agents sort and combine data to build things like a financial plan.

Own tools, steady improvement

• Arca uses models from Anthropic and OpenAI, with Claude as its main provider.
• Its own harness, the software that directs the AI, picks a model for each task using internal benchmarks.
• Owning the harness lets Arca switch model providers if it needs to.
• Advisor feedback and agent activity logs go to a meta-agent, which suggests changes to the harness.

What to take from it

The source does not say how well the system performs or how many clients it serves. It does show a clear design choice. Let AI do the prep, track the details and flag what needs attention. Keep a person in charge of anything that can cost money.