13 min read
Arca, a wealth management firm based in Portsmouth, New Hampshire, has 28 employees. Eleven of them perform investment advisory functions, and together they look after $1.27 billion belonging to 1,108 individual clients and six retirement plans. Those figures come from the Form ADV the firm filed with the Securities and Exchange Commission on September 28, the public disclosure every US investment adviser has to keep current.[2]
People handle the client relationship and the final calls. The work around those calls is handled by AI agents, software that carries out multi-step tasks without being prompted at each step. Aman Singh, who leads Arca’s AI platform team, described the firm at an AI engineering conference in New York in late October as a place where “agents prompt people more than people prompt agents.”[1]
Range, an advisory firm in Virginia registered with the same regulator, went the other way. Chief executive Fahad Hassan told InvestmentNews in March that “over the next one to three years, our plan is to eliminate our own advisor base.”[4] Both firms buy their models from outside labs, and both keep client money at Altruist. They diverge on what they leave to humans.
What Arca leaves to people
The example Singh walked through involves a client who emails about a $100,000 tax bill. The agents work out where the client’s money sits and which holdings to sell to raise that amount with the smallest tax hit. One case of this kind runs about ten days and pulls in several people along the way. Selling the assets, moving the money and contacting the client are blocked off for humans.[1][3]
“All of our safety restrictions are fully deterministic, and they take the decision making off the agent’s hands,” Singh said.[1] Deterministic here means the block is a rule in code rather than a judgment the model makes. The model can produce a persuasive argument for the trade and the order still will not go out.
Each time an agent saves something to the firm’s records, a small, fast model reads what is being saved. Material arriving from outside, including email, documents and call recordings, is screened for prompt injection, an attack that hides instructions inside a document so the AI follows them. When data leaves the system, whether through a web search or a calendar invite to a client, rules in code look for personal details such as account numbers and several small models then vote on whether anything is being leaked.[1]
Singh said models will pull up 2025 tax brackets and present the result as settled. Arca runs a separate model that notices when a financial recommendation is about to be made and pushes the agent toward the firm’s verified reference material or a web search. The model in that slot answers with a choice and a confidence score rather than a paragraph of text, the kind of system we covered when Jev launched.[1]
Agents produce financial plans inside a sandbox, an isolated space for running code, and the checks start as soon as the document is written. Rules in code catch things like a raw account number in the text, and small models look for language that would be non-compliant, such as promising returns. Singh argued AI is better than people at this particular job because it never gets tired of reading every line.[1]
Holding a ten-day task together
An email goes out, a reply comes back days later, a document arrives, an approval is needed. Ten days of elapsed time is not ten days of continuous work. Arca keeps a running record for each client covering what is happening now, what has to happen in the next few days, what is being tracked long term, and who owns which action. Agents write into that record, and once a day a separate pass rereads the client’s whole situation and updates it.[1]
A reply to an open email thread goes back to the agent already reading it. A phone call that belongs to no existing thread starts a fresh agent, which reads the record, works out that a large tax liquidation is underway, and files what it learned into that case.[1] How agents hold onto context over long horizons is covered in more detail in our piece on agent memory.
“The industry cares about outcomes, but compliance cares about auditability,” Singh said. Every action produces a change log entry recording what changed, who changed it, from which part of the system and when. That applies to staff edits in the user interface too, and those same entries double as signals that tell an agent something just happened.[1]
Arca built its own harness, the software that wraps a model and gives it tools, memory and a place to run. The harness picks which model handles a given task, and it makes that choice using tests the firm built from its own work rather than public leaderboards. Singh said one recent model with strong public benchmark scores failed those internal tests because it was too slow and too expensive. Tax analysis goes to the most capable models, and routine background jobs go to cheaper ones.[1] We compared the model and the harness directly in an earlier piece.
Singh also said the firm has no plans to adopt a managed agent product from a model provider. Memory and operational knowledge sit at the center of those products, and handing those over to a vendor means giving up control of the thing that makes the firm’s service different.[1]
The firms cutting advisors instead
Range manages roughly $700 million and employed about 25 advisors as of March. Instead of charging around 1% of assets a year, it sells flat annual tiers of $3,000, $6,000 and $10,000. It reported more than 6,000 clients across about 3,100 households, and says Rai, the in-house AI tool it sells access to, scored 95% on practice questions from the Certified Financial Planner exam.[4]
Seven advisor roles were posted on Range’s site in March, months after Hassan said the firm would do away with the job. He said the firm hired planners early on to learn “what they do, how they think,” and that about six months before the interview it switched to having AI operate the tools the advisors had been using.[4]
In September, InvestmentNews reported layoffs at Domain Money, which sells similar flat-fee tiers. Former employees said roughly a third of the company and about half of the financial planning team were cut in August. Chief executive Adam Dell said improved automation and efficiency now let one advisor serve up to 150 clients, and that the firm sized its advisor team accordingly. Domain Money has about 1,500 clients and roughly $43 million in regulatory assets under management.[5]
One advisor laid off there described using an AI tool that recorded client conversations and drafted plans. “You couldn’t make it write it well enough that it didn’t still take hours of review to get it finalized effectively,” the advisor said, adding, “I think we were asking it to do more than it’s capable of at this point.” Some of those laid off had started between April and June, in what the company had called its largest hiring class.[5]

The same promise, three times before
Michael Kitces and Ben Henry-Moreland pointed out in an April column that this promise has a history. Discount online brokerages made it in the 1990s, search engines in the 2000s, and robo advisors in the 2010s, those being services that pick and rebalance a portfolio automatically for a fraction of a human fee. None of them removed human advisors, and robo advisors ended up taking manual portfolio work off planning-focused advisors instead.[6]
The obstacle they identify is in the arithmetic of the fees. Kitces Research on advisor productivity puts the floor for a profitable human advice relationship at roughly $3,000 per client per year, while a consumer software business works by cutting price to add users. Range currently charges human-advice fees, and whether clients keep paying them once no human is attached is untested. Retention for firms with human advisors typically runs 95% to 98%. At 80%, the total a firm collects from a client over the life of the relationship drops to a quarter of what it was.[6]
Domain Money hired heavily in the spring and cut in the summer, and the advisor who was cut said the AI tooling did not save time. Nothing in the public record separates a correction to over-hiring from automation actually reducing headcount, which is what makes a single round of layoffs hard to attribute to AI. On Range’s direction, Kitces looks at who owns the company: investors who backed a scalable technology business are unlikely to accept the thinner profit margins of a service firm.[6] The ways AI and employment numbers get conflated are laid out in our piece on AI job statistics.
Anthropic describes its own advisor product in similar terms. It connects Claude to the systems a firm already uses for custody, portfolios, planning and client records, and adds a package covering onboarding, meeting prep, portfolio reviews, proposals and compliance reviews, with human review steps and an audit trail built in. The company’s summary of its financial services work is that “Claude does the assembly, and the professional keeps control of the judgment calls.”[7]

South Korea certifies the algorithm first
In Korea the same question was settled by the regulator rather than by each firm. In December 2024 the Financial Services Commission approved applications from 17 investment managers to let robo advisors manage retirement accounts on a saver’s behalf, granting them innovative financial service status, which suspends existing rules for a limited period. The first service launched on March 28, 2025, run by Fount Investment Advisory with Hana Bank as the plan provider. Savers in individual retirement pension accounts, who previously had to issue investment instructions themselves, can now hand that over.[8]
The conditions attached are narrow. Only algorithms that have passed both the operational and systems review at the testbed run by Koscom, the securities industry’s IT operator, qualify. Contributions are capped at 9 million won a year per account, about $6,700, rising by the same amount annually, and the designation lasts at most four years, after which the regulator plans to amend the law if returns hold up.[8][9]
A certified robo advisor and the agents these US firms run do different jobs. The robo advisor submits its allocation rules for review and then moves money inside the limits it was approved for. Arca’s agents read, calculate and draft, and have no authority to place a trade at all.
Questions worth asking
If a firm handling your money, taxes or insurance says it uses AI, a few questions separate the designs.
- Who makes the final call? A firm that stops at analysis and drafts carries different risk from one that automates orders and transfers.
- Are actions logged? If every change is attributed and timestamped, a problem can be reconstructed afterward.
- Where do figures that change every year, such as tax brackets and contribution limits, come from? A model will state last year’s numbers without hedging.
- What happens to documents you upload when they leave the firm’s systems, and what gets filtered on the way out?
The same test applies to anyone handed one of these tools at work. If reviewing an AI draft takes longer than writing it did, the tool is not yet saving work. That was the experience the laid-off Domain Money advisor described.
Arca removed the work around the advisor’s job, and Range says it will remove the job. Which reading holds up will show in client counts, retention and fee levels one to three years from now.
This article is for information only and is not investment advice. Decisions and their consequences are your own.
Sources
- Agents Prompt People: Using Agents for Long-Horizon Financial Work (Latent Space, 2026-10-31)
- ARCA WEALTH, LLC Form ADV (SEC IAPD, filed 2026-09-28)
- AI-enabled RIA Arca raises $64M, works with an Anthropic client (InvestmentNews, 2026-06-26)
- RIA startup Range plans to eliminate its advisor workforce as AI takes over (InvestmentNews, 2026-03-06)
- Adam Dell’s Domain Money lays off advisors, citing AI and efficiency (InvestmentNews, 2026-09-15)
- Range reiterates plan to eliminate human advisors in favor of AI (Kitces, via InvestmentNews, 2026-04-24)
- Financial Services Commission press release on retirement robo advisor services (2025-03-27, Korean)
Notes
- Latent Space podcast, “Agents Prompt People: Using Agents for Long-Horizon Financial Work,” hosted by Richard MacManus, published 2026-10-31. Checked against the full episode transcript: the ten-day $100,000 tax case, “All of our safety restrictions are fully deterministic,” LLM verification on database writes plus prompt injection and PII screening, sandboxed plan generation and verifiers, stale tax brackets and the decision model, per-client memory with a daily reflection pass, “the industry cares about outcomes, but compliance cares about auditability,” the in-house harness and internal benchmarks, the position on managed agents, and “agents prompt people more than people prompt agents.” The transcript names the model that failed Arca’s internal tests, but the name could not be matched to a specific release, so it is left out here. Source ↩
- ARCA WEALTH, LLC Form ADV Part 1A, filed 2026-09-28 (CRD 174765, SEC file 801-92174), read directly from the filing. Regulatory assets under management of $1,269,642,461 across 4,909 accounts ($1,196,353,484 discretionary over 4,713 accounts, $73,288,977 non-discretionary over 196 accounts); 28 employees of whom 11 perform investment advisory functions; 671 individual clients, 437 high net worth individuals, and six pension or profit sharing plans. Filing ↩
- InvestmentNews, 2026-06-26. Arca came out of stealth with $64 million from General Catalyst, Index Ventures and Venrock. Founder and chief executive Rron Rexha: “Every Arca client gets a human advisor, they do not interact with AI directly.” Fees of 0.50% to 1.50% a year per the firm’s Form ADV, custodians Schwab, Fidelity and Altruist, and a $500,000 investable asset minimum that may be waived. Source ↩
- InvestmentNews, 2026-03-06. Hassan’s quote in full: “Over time, let’s say over the next one to three years, our plan is to eliminate our own advisor base, or maybe have them work with ultra-high-net-worth individuals who no matter how good the AI is, they just want to talk to a person.” About $700 million in client assets, roughly 25 advisors (17 investment adviser representatives per the August 2025 Form ADV), three flat tiers, more than 6,000 clients across about 3,100 households, seven open advisor roles, and the 95% CFP practice result as claimed by the company. Source ↩
- InvestmentNews, 2026-09-15. The scale of the August cuts, about a third of the company and half the planning team, comes from former employees granted anonymity. Dell responded in writing that automation lets one advisor serve up to 150 clients. About 1,500 clients, roughly $43 million in regulatory assets, 39 employees including 14 in advisory functions per the 2026-05-15 Form ADV, tiers of $3,900, $5,200 and $9,000, and eight planning hires who started between April and June. Source ↩
- Michael Kitces and Ben Henry-Moreland, published on Nerd’s Eye View at Kitces.com and reprinted by InvestmentNews, 2026-04-24. Covers the brokerage, search and robo advisor precedents, the roughly $3,000 per client per year floor from Kitces Research on Advisor Productivity, Range’s $2,950 to $9,950 fee range, 95% to 98% retention for human advisors and the quarter-of-lifetime-value effect at 80%, and the investor-expectations argument. Source ↩
- Anthropic financial services page, read 2026-11-03. “Claude for Financial Advisors connects to custodian, portfolio, planning, and CRM systems. A plugin covers onboarding, meeting prep, portfolio reviews, proposals, and compliance reviews with review steps and an audit trail built in,” and “In each case Claude does the assembly, and the professional keeps control of the judgment calls.” Source ↩
- Financial Services Commission press release, 2025-03-27 (Korean). December 2024 innovative financial service designation covering 17 investment managers, first launch on 2025-03-28 by Fount Investment Advisory with Hana Bank, algorithms limited to those passing Koscom testbed review, a 9 million won annual cap per IRP account rising by the same amount each year, and a designation period of up to four years with legislative changes to follow if results hold. Source ↩
- Won to dollar conversion at 1,344.49 won per dollar, the November 2, 2026 close (Yahoo Finance). ↩
Source: https://www.latent.space/p/arca