Key Moments
Using Agents for Long-Horizon Financial Work — Aman Singh, Arca
Latent Space PodcastKey Moments
Arca uses AI agents for financial management, handling complex tasks like tax and portfolio planning over days, but requires human oversight for fund transfers and direct client communication.
Key Insights
Arca uses AI agents for a wide range of financial services, including financial advice, wealth planning, tax planning, insurance, and estate planning.
The company's AI agents are designed to augment, not replace, human financial advisors, handling back-office tasks to provide a higher level of client service.
Arca utilizes leading models from Anthropic and OpenAI, primarily Claude, but internal metrics guide model selection, not external benchmarks, to maintain flexibility.
Complex financial processes, such as a tax bill adjustment or portfolio rebalancing, can span up to ten days, with agents handling analysis and humans executing transactions.
Arca implements stringent security measures, including deterministic security constraints, LLM-based checks on data writes, and edge-case validation for prompt injection and data exfiltration.
The system is designed for auditability by default, with every agent and human action logged, which is crucial for regulatory compliance and agent feedback loops.
Arca's AI-native wealth management approach
Arca is building an AI-native wealth management firm, aiming to manage its entire business operations using artificial intelligence. Unlike a typical SaaS company, Arca positions itself as the operational entity, leveraging AI agents to drive services proactively. These agents work around the clock, providing a level of service previously unattainable in the financial industry. The core philosophy is that AI agents direct humans more than humans direct agents, offering a proactive, 24/7 service tailored to client needs. This approach encompasses a broad spectrum of financial services, including financial advice, comprehensive wealth planning, tax planning, insurance, and estate planning, catering to diverse client profiles from business sellers to tech executives planning early retirement and first-time homebuyers.
AI's role in augmenting, not replacing, human advisors
The integration of AI at Arca is strategic: it's designed to replace the surrounding tasks of a financial advisor, not the advisor themselves. Clients interact with a human advisor who provides empathetic and attentive service, while the AI operates in the background. These agents continuously monitor market dynamics, analyze how they impact individual clients' financial situations, and consider personal life events. This allows Arca to be highly proactive and responsive, delivering superior customer service by automating administrative and analytical tasks. Human advisors are thus freed up to focus on high-value client relationships and complex strategic decisions, while AI handles the data processing and initial analysis.
Model selection and platform independence
Arca utilizes state-of-the-art models from leading AI labs like Anthropic and OpenAI, with Claude serving as a foundational model. The company has iterated through various models over time, including Gemini and previous OpenAI offerings. The decision-making process for model selection is driven by internal metrics that assess various platform capabilities against Arca's business needs. This approach allows Arca to remain model-agnostic, avoiding over-reliance on a single provider and maintaining flexibility. For instance, while external benchmarks might praise certain models, Arca's internal testing revealed some were too slow or costly for their specific applications, underscoring the importance of proprietary performance evaluation.
Agentic workflows and long-horizon tasks
Arca's agents are capable of handling complex, long-horizon workflows that can span multiple days. A prime example is managing a client's tax bill, which might involve analyzing asset allocation, determining the optimal way to liquidate a portion to save a significant amount (e.g., $100,000), and executing the necessary steps. While the agents perform the analytical part autonomously, human intervention is required for any actions involving the transfer of funds or liquidation of assets, due to strict internal controls and a requirement for human oversight in such sensitive transactions. The agents present a personalized plan for the human advisor to review and modify, ensuring a blend of AI efficiency and human judgment.
Maintaining context and memory in agentic processes
For long-running tasks, maintaining context is crucial. Arca employs sophisticated memory systems to ensure agents retain the necessary information. If an agent receives a new signal, like an unexpected phone call, a new agent is activated to understand its implications. This is achieved through internal memory systems that track client data, current service goals, responsibilities, past actions, and upcoming tasks. A 'dreaming' process allows agents to reflect on the client's world state and update these memory systems for long-term task management. This ensures that even when new signals arrive, the agent can quickly reference the ongoing processes, such as a large tax bill, and contribute to its state.
Robust security and compliance measures
Security is paramount, especially in financial services. Arca's security framework is deterministic, stripping decision-making power from the agents in critical areas. While agents can use tools for tasks like record-keeping, any data write operation triggers a large language model (LLM)-based validation check. Furthermore, Arca implements edge-case validation to detect prompt injection from incoming data like emails or documents. When data leaves the system, such as for a web search or sending a calendar invite, deterministic checks are applied for Personally Identifiable Information (PII) and to prevent data leakage. This layered security approach, including the use of smaller models like Jev for specific checks, aims to contain risks at the edge of capability.
Auditability and regulatory adherence
Compliance in the financial industry hinges on auditability. Arca's system is designed with auditability as a default feature. Every action taken by an agent or a human generates a change log detailing what was modified, by whom, from which system component, and when. This comprehensive logging extends to human actions via the user interface, ensuring a complete audit trail. This default auditability not only satisfies regulatory requirements but also supports agent improvement by providing feedback on changes and their impact. This transparency is a fundamental aspect of Arca's operational philosophy.
Human-in-the-loop and continuous improvement
The statement 'agents direct people more than people direct agents' highlights a paradigm shift. Agents are tasked with identifying when and whom to involve in a process. Leveraging their long-term memory and daily reviews, agents reflect the current 'state of the world' and engage the right person at the opportune moment. This is facilitated by Arca's nature as an operational firm, not just SaaS, allowing standardization of workflows. Feedback loops are critical for improvement. When human advisors make corrections or cancellations, they must provide a reason. This feedback is logged, linked to the agent that produced the output, and analyzed nightly by a 'meta-agent'. This process identifies errors, misleading tools, or knowledge gaps, leading to system enhancements and prompt refinements, ensuring continuous improvement.
Mentioned in This Episode
●Software & Apps
●Companies
●People Referenced
Common Questions
Arca uses AI agents to automate back-office tasks, enabling financial advisors to be more proactive and responsive. This enhances client service without replacing the human advisor.
Topics
Mentioned in this video
A foundational model used by Arca, developed by Anthropic.
A model tier used by Arca for complex tasks like tax analysis.
A powerful model tier used by Arca for complex tasks.
An AI model previously used by Arca, now superseded by other models.
A smaller model used by Arca for quick checks and analysis, particularly in the context of sandboxed environments and decision-making processes.
An AI model that Arca tested but found to be too slow and expensive, guiding their internal model selection process.
A model tier used by Arca for complex tasks like tax analysis.
A wealth management company that utilizes AI agents for its services. They recently raised $64 million.
A leading AI research lab whose models, such as Claude, are used by Arca. Arca is optimistic about Anthropic's models but less so about their managed agent products.
A leading AI research lab whose models are used by Arca. Arca is optimistic about OpenAI's models.
More from Latent Space
View all 270 summaries
28 minBuilding ChatGPT for Finance — Joseph Kim, OpenAI's Head of Financial Services
107 minHow Anthropic ships the #1 LLM in the World — with Katelyn Lesse and Angela Jiang
27 minThe Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi
86 minWhy AI Compute Comes in Rectangles — Erik Bernhardsson, Modal
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free