Finance is where agents have to prove they can last, not just answer
Arca runs proactive agents under strict audit and approval rules; OpenAI's finance product bets on the same human-in-the-loop shape, while AI spending keeps rising behind both.
- Arca raised $64 million to build an AI-native wealth management firm, according to the firm's account on the Latent Space podcast. Self-reported m:…4291
- Arca keeps humans in the loop for all financial transactions, liquidations and client communications, logs every agent and human action, verifies every database write, and does not currently perform autonomous trading. m:…4291
- OpenAI's Joseph Kim said some generic financial-data connectors carried error rates above 40%, and that Morgan Stanley and Evercore are design partners for ChatGPT for Financial Services. Self-reported m:…1320
- Kim said 95% of what OpenAI built for the finance product could have been replicated by a startup using existing plugin primitives, and that Codex adoption in finance, legal and sales jumped to 90% after the desktop app arrived in February 2026. Self-reported m:…1320
- BofA's Vivek Arya said hyperscaler spending forecasts have been revised upward by about 20% within the quarter. m:…4401
Proactive, not prompted
Aman Singh, who leads the AI platform team at Arca, told the Latent Space podcast that his firm's agents are built to prompt people rather than wait to be prompted, running a proactive service around the clock. The agents track multi-day financial processes using long-term memory and daily reflection, and a model-agnostic internal harness routes each task to different tiers of language models based on complexity. m:…4291
A meta-agent suggests improvements to that harness from feedback traces collected in real operations, Singh said. Arca, a wealth management firm that has raised $64 million to build an AI-native service, is deliberately avoiding dependence on managed agent products from model labs, Singh said, to keep control of its business intelligence. m:…4291
Guardrails before autonomy
The guardrails are strict and stated plainly. Arca keeps humans in the loop for all financial transactions, liquidations and client communications, produces audit logs for every action taken by agents and human actors, and uses language-model verification to analyze every write action to its database. Decision models nudge agents toward verified, current knowledge when making financial recommendations. m:…4291
And there is a line the firm has not crossed: its agents are not currently performing autonomous trading. m:…4291
The enterprise version of the same test
OpenAI's Joseph Kim, head of financial services, described the same test from the vendor side with the launch of ChatGPT for Financial Services, an enterprise product built with design partners Morgan Stanley and Evercore. The product assumes a human in the loop for auditing and review, mirroring the hierarchical review processes already used inside financial institutions, and it integrates firm-specific templates and data sources. m:…1320
Kim's candor is the useful part. Generic connectors for financial data can carry error rates above 40%, he said, and 95% of what OpenAI built could have been replicated by a startup using existing plugin primitives. Codex adoption in finance, legal and sales jumped to 90% after the desktop app arrived in February 2026, he said, a reminder that distribution, not just capability, decides whether a tool gets used. m:…1320
Kim predicted that by 2027, models will handle tasks now done by senior financial professionals, including generating investment theses, and said OpenAI is training models on human taste through pairwise feedback on which of two slides looks better. That is a prediction, not a result. m:…1320
OpenAI structures its operations as a research-first, deployment-second company, Kim said, and its financial models are trained with financial reasoning data upstream into the main model. m:…1320
Why the test matters now
The money behind the test keeps growing. Vivek Arya of BofA Global Research told a BofA podcast that hyperscaler spending forecasts have been revised upward by about 20% within the quarter, and that AI contributed about 70 basis points to 2026 GDP growth. Memory demand outstrips supply by three to four times, he said. If that spending is real, the question for financial firms is not whether agents can demo, but whether they can persist, verify and stay auditable over long horizons, the test Arca and OpenAI are both, differently, trying to pass. m:…4401
The shared shape is the point: both firms keep humans approving, logging and reviewing, and both treat persistence over days as the capability that still has to be proved. m:…4291 m:…1320
What is still unresolved
- How Arca will handle agentic failures in complex, multi-step financial trades as it scales.
- The timeline for rolling out financial services features to the Pro plan.
- Whether models will in fact handle senior-level financial tasks, including investment thesis generation, by 2027, as Kim predicted.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.