At Arca, AI does the back-office work and people judge what it proposes

Aman Singh, head of AI at the wealth firm Arca, argued that an AI-native firm should be the service itself, not software sold to others. He said so on the Latent Space podcast. In his telling, staff work shifts from doing back-office tasks to judging what agents propose.

What was said

Singh said Arca is "not a SaaS company." In his words, "We are, um, we are the firm ourselves." A firm that merely has AI tools, he said, needs its people to prompt the agents. At Arca, he said, agents prompt people more than people prompt agents. Today agents do the back-office work while humans advise clients, approve trades and review prompt changes.

Singh said AI is meant to take over the work around the advisor, not the advisor. Clients do not deal with AI directly. Behind the advisor, agents track each client's situation and prompt staff to act. Arca does no autonomous trading, he said. Agents propose trades, and the firm's experts decide whether each is right and then execute it.

Singh said the safety limits are fixed rules, so the agent cannot decide them. A human must be involved before money moves, assets are sold or clients are contacted directly. Automatic checks also run on a drafted financial plan. Some are fixed rules, such as catching a raw account number. Others use smaller models.

Judging also trains the system. Singh said each staff edit or cancellation comes with a reason, stored beside the agent's trace, the record of what it did. Each night, another agent reviews real traces and suggests changes. A human engineer still checks any prompt change before it ships.

Why it matters

Our reading: the skill here is review, not production. A staff member who approves a trade plan must know what a good plan looks like and explain why they rejected a bad one. If agents do the groundwork, juniors may get fewer chances to learn it by doing. Singh did not address this, so managers adopting the model must decide how people build that judgment.

The other side

Singh described a partly automated firm. Staff approval and engineer checks exist today, which leaves open what happens as agents improve. He did not say how many people review proposals or how well they catch errors.

An open question is what the fixed rules and automatic checks do not cover. A reviewer shown a polished plan might simply approve it, and the conversation did not test whether scrutiny holds up.

Singh also said agents left alone make bad prompt changes: "every little fix becomes the whole prompt." That is one engineer's observation, not a measured result.