Arca's Aman Singh: How Agents Run a Wealth Management Firm
How Arca Uses AI Agents to Run Wealth Management
Aman Singh leads the AI platform team at Arca, an AI-native wealth management firm that has raised $64 million. His team builds the in-house harness that Arca uses to run the business. Speaking at the AI Engineer conference in New York, Singh gives one of the most concrete accounts so far of agents doing long-horizon work in a regulated industry: tax plans, portfolio liquidations and client onboarding that span days and involve several people.
On being the firm, not the vendor: "We're actually not a SaaS company. We are the firm ourselves." Arca gives financial advice, financial planning, tax, insurance and estate planning directly to clients. Because it owns the business, it can standardize how the work is done so agents can take part in it, and it can require its own staff to give feedback on agent output. A software vendor can do neither.
On replacing the back office, not the advisor: "The role of AI at Arca is not to replace the advisor. It is to replace all of the work around the advisor." Clients rarely deal with AI directly. They get faster, more proactive service from a human advisor while agents watch the market, the client's life events and upcoming deadlines in the background.
On a 10-day tax bill: When a client emails about a $100,000 tax bill, agents work out where the client's money sits and the best way to liquidate assets to cover it. They can't move money themselves. Arca has an API integration with its custodian, but anything that moves money, liquidates assets or contacts a client goes to a human, who gets a plan to review and adjust. New signals, like a phone call halfway through the process, wake a fresh agent. That agent catches up from a per-client memory of what is in progress, who owns each step and what is due next. A nightly "dreaming" process reflects on each client's situation and updates that memory.
On safety at the edges: "All of our safety restrictions are fully deterministic." Simple record-keeping tools run on their own. Fast LLM checks verify every write to the database, and incoming emails, documents and calls are scanned for prompt injection. Anything leaving the system, from a web search to a calendar invite, passes deterministic PII checks plus a consensus of small models looking for data exfiltration. Client artifacts such as financial plan PDFs are checked for raw account numbers and for non-compliant wording like promised returns. Singh argues this makes compliance stronger than human review because the AI never tires of looking.
On model choice: Claude is Arca's main model today, after earlier phases on Gemini and OpenAI. Internal benchmarks decide, not public leaderboards: one recent model that looked strong on external benchmarks turned out too slow and too expensive for Arca's work. Tasks are tiered by complexity, so tax analysis gets the heaviest models and background work gets cheaper ones. Advisors never see these routing decisions. Arca also uses its own harness rather than a lab's managed agents product, because it wants to own the memory and operating knowledge its agents build up.
5 Lessons from Arca on Agents in Regulated Work
- Agents prompt people - "Agents prompt people more than people prompt agents." Agents track a multi-step process and bring each person in only when their step arrives.
- Every staff edit becomes an eval - Staff feedback is stored next to the agent trace that produced it. Each night a meta agent reads real traces and suggests fixes to tools and the knowledge base. A human engineer still approves prompt changes, because agents left alone turn every small fix into a rewrite of the whole prompt.
- Auditability by default - Every change, by an agent or a person in the UI, is logged with what changed, who changed it, from where and when. The same log satisfies compliance and tells agents what happened.
- Knowledge is a product tab - The firm's operating knowledge is a set of files that advisors edit directly in the product. Agents cite it, and trace analysis flags entries that are wrong.
- Models are overconfident about finance - Models will happily use the previous year's tax brackets. Arca detects when a model is about to make a financial recommendation and pushes it to check a knowledge base of verified current facts or search the web.
What Arca Shows About AI-Native Services Firms
Arca's model is a services firm rebuilt around agents. Humans keep authority over money and client relationships, while agents do the analysis, track every open process and surface what people would miss. In one example, agents reviewing a prospect's statements found fees the prospect didn't know about and showed that the current advisor wasn't a fiduciary. The approach rests on three engineering choices: own the agent harness, keep human-in-the-loop gates at the points where money moves, and turn everyday staff feedback into evals. For related approaches in finance, see how Optiver ships new models the same day and how Two Sigma controls AI agents.