A few weeks ago, I was explaining to a friend what a family office is, and halfway through, I realised most people have never heard of it. This is strange, because for the last decade, the entire fintech industry has been claiming to democratize wealth management, and none of it has reached the point where a normal person even knows what’s being democratized.
Here’s why. What the industry has actually been selling isn’t wealth management. It’s SaaS tools dressed up as wealth management. And the thing that separates a real family office from a Robo-Advisor isn’t sophistication. It’s labor.
A single family office costs between $500K and $10 million a year to run.
That’s the number a single wealthy family pays to employ a small team of humans, usually four to twenty people, whose entire job is to pay continuous attention to that family’s money. A Chief Investment Officer deciding allocations. A tax strategist watching for harvesting opportunities in real time. A controller reconciling twelve accounts across four jurisdictions. A bill-pay operator. A trust and estates lawyer on retainer. Someone to answer the phone when a bank calls about a wire at 4 p.m. on a Friday.
The minimum to justify hiring such a team is usually quoted at around $100 million in investable assets. Below that, the math stops working — you’re paying more for the office than you’d save from its decisions.
This is the number I want you to hold in your head: between $500k and $10 million a year for continuous, human-grade financial attention.
Now here’s the claim: that number is about to become $500 a year.
Not because wealth management is getting cheaper. Because the thing a family office actually sells was never sophistication, it was labor. And labor is what agents do.
What a family office actually does
The misconception most people carry about family offices is that they’re buying better ideas. Secret hedge fund access. Private deal flow. A quant edge. Tax loopholes retail investors can’t reach. Some of that exists at the very top end, but it’s not the core product. The core product is much more boring and much more expensive: somebody paying continuous attention to everything at once.
Break down what that actually means on a line-item basis, because this is where the $500 number comes from.
Tax optimization: A family office doesn’t do your taxes in April. It watches your tax position every week, harvests losses the day the opportunity appears, structures charitable gifts at the exact moment the appreciated asset peaks, prepays estimated taxes in the state that minimizes your total bill, and coordinates all of this across entities. For a family with $50M in mixed assets, this alone can be worth six figures a year.
Cash management: Moving idle cash between yield venues as rates change. Negotiating deposit rates with multiple banks. Making sure nothing sits in a 0.01% checking account when a 5% treasury ladder is one wire away.
Account and entity coordination: A wealthy family has LLCs, trusts, brokerages, foreign accounts, and operating companies. Reconciling all of it and producing a single picture of net worth is a full-time job.
Opportunistic action: Refinancing when rates cross a threshold. Rebalancing when allocations drift. Exercising options at the optimal moment. Filing the form before the deadline. These aren’t decisions that require genius; they require someone whose job is to notice.
Exception handling: The wire that didn’t arrive. The tax notice from a state that the family forgot they had exposure to. The margin call at an awkward moment. Someone has to pick up the phone, understand the context, and act.
Add it up, and a modest family office runs in the low millions a year, minimum. That cost is why the service has a $100M floor. Not because everyday people don’t deserve continuous attention, but because continuous human attention doesn’t scale below a certain asset base without destroying the economics.
Notice what every one of those line items has in common. None of them is hard thinking. They’re all sustained watching. The family office’s moat was never intelligence. It was payroll.
Why didn’t robo-advisors touch any of this
The 2010s robo-advisor revolution was sold as the democratization of wealth management, and it was not. Betterment and Wealthfront democratized exactly one line item on the list above: portfolio construction, and left the other four untouched.
Robo-advisors gave you a model portfolio and a quarterly rebalance. They did not watch your tax position in real time. They did not coordinate across institutions that you didn’t host with them. They did not negotiate with your bank. They did not refinance your mortgage when rates moved. They did not notice when a state filing deadline was two weeks out. They did not pick up the phone.
They automated the cheapest, most commoditized piece of the family office stack, the piece that was already mostly a spreadsheet, and left every labor-intensive, attention-intensive piece exactly where it was. Then they called it a revolution and charged 25 basis points.
This is why a Betterment customer with $200k and a family office client with $200M are not, despite a decade of marketing, getting the same service. They’re getting two services that share one overlapping component. Everything that actually matters about wealth management: the continuous watching, the exception handling, the cross-account coordination, the tax gymnastics, remained strictly a function of how many humans you could afford to employ.
Until the humans stopped being the only option.
The $500 number
Here’s where the math gets uncomfortable for an entire industry.
An autonomous agent doesn’t sleep, doesn’t take PTO, doesn’t need a corner office, and doesn’t bill hourly. It watches twelve accounts with the same effort as one. It reads a tax code change and applies it across every entity the user owns before the user has finished their coffee. It picks up the metaphorical phone at 4 p.m. on a Friday because it doesn’t know what Friday is.
The cost structure of that agent is not a salary. It’s a model invocation bill plus some API fees plus a thin margin for whoever operates it. Today, running a reasoning-capable model continuously against a single user’s financial life, with the context loaded, the tools connected, and the integrations live, lands somewhere in the low tens of dollars a month in raw inference cost, and that number is falling fast. Add tooling, data feeds, compliance overhead, and a reasonable operator margin, and you get to a product that looks like $500 a year for the core service, maybe $2,000 for a premium tier with insurance and a human backstop.
For that price, you get the line items a family office used to charge seven figures for. Not some of them. All of them, because the constraint was never that the strategies were secret; it was that watching required payroll. Remove the payroll, and the exclusivity collapses on its own.
The $100M floor becomes a $50k floor. Maybe a $10k floor. Maybe no floor at all, because even someone with $10k in assets can benefit from continuous tax-efficient cash management when the marginal cost of running the agent rounds to zero.
What does autonomous actually mean here
The word “agent” has been so thoroughly laundered by every SaaS vendor with a chatbot that I need to put a fence around it. An autonomous financial agent, in the sense I mean it, does three things a robo-advisor structurally cannot:
It holds goals, not instructions.
You don’t tell it to buy AAPL. You tell it you want to retire at 55 and it figures out the rest, including the parts you didn’t think to specify.
It initiates across boundaries.
It doesn’t ask permission for every step. It opens the account, moves the money, and files the form. Per-step approval is the opposite of autonomy; it’s just a slower GUI.
It handles exceptions by reasoning, not escalation.
When something weird happens: a rate changes, a transfer fails, a tax rule shifts mid-year, the agent figures out what to do. It doesn’t fire off a “we need your attention” email unless you explicitly ask it to.
Once you have those three properties, the unlocks are the exact line items a family office was selling: continuous tax-loss harvesting across every account you own. Cash that moves hourly between yield venues because the spread is finally worth chasing when the cost of moving is zero. Refinancing that triggers the second rate to cross a threshold, not the week after your neighbor texts you about it. State filings that happen on time because the agent reads the calendar.
These are not features. They are outcomes that were structurally impossible when a human had to be in the loop, not because humans weren’t smart enough, but because humans weren’t cheap enough.
Who captures the $500
Once the labor constraint is gone, the interesting question stops being whether this happens and becomes who collects the margin when it does. And here’s where the entire consumer fintech industry is about to misread the moment, the same way SaaS companies are currently misreading agents.
I wrote in SaaS is Dead. Long Live APIs as a Service, that when agents become the users of software, the stack inverts. Finance is going to feel this harder than any other vertical, because consumer fintech has spent fifteen years building moats out of exactly the things agents don’t care about.
Think about what modern fintech competes on: onboarding friction, UI polish, brand trust, referral loops, push notification design, and app store ratings. Now imagine your customer is a piece of software that cannot see your app, does not care about your brand, will never open your push notification, and makes decisions based on which endpoint returned cleaner JSON.
Every moat evaporates. What’s left?
Which APIs the agent trusts: Not “has heard of”, trusts, as in: track record, clean auth, deterministic responses, no weird rate limits that break a workflow at 3 a.m.
Which rails it can move value across: Stablecoins are winning this quietly and then all at once, because an agent holding USDC can transact on Sunday at 2 a.m. and an agent holding dollars in a checking account cannot.
Who underwrites the decision: Because when something goes wrong, and it will, someone has to be on the hook, and that someone is going to charge a premium.
Notice what’s not on that list: anything a marketing team can influence.
The incumbents are already shipping the wrong product
The banks and neobanks have heard “agentic AI” and translated it the way incumbents always translate disruptive shifts: as a feature to bolt onto the existing product. Revolut just shipped an in-app assistant. Starling shipped one. Every neobank’s roadmap has “AI” on it, and every one of them means the same thing: a chatbot glued to the existing app so users can say “freeze my card” instead of tapping a button.
This is Clippy. It is being celebrated as the agentic revolution, and it is Clippy.
It is also priced like sophistication. Revolut’s Ultra tier is €55 a month for a bundle of consumer perks and a chatbot that summarizes your spending. That’s €660 a year for a product that does approximately none of the line items a family office actually sells. The pricing implies they still think wealth management costs what it used to.
The actual shift happens one layer down, and it happens around these apps, not through them. Autonomous agents don’t want to live inside Revolut. They want to talk directly to the underlying rails: clearing houses, stablecoin networks, custody providers, and tax APIs. The consumer app is a detour. Every additional layer between an agent and a settlement rail is a tax the agent will eventually route around.
The banks that survive this will stop thinking of themselves as consumer brands and start thinking of themselves as machine-readable infrastructure. Clean APIs. Deterministic responses. Auth an agent can hold. Trust boundaries that a non-human customer can reason about. Nobody’s marketing department wants to hear this, which is exactly why the opportunity is real.
The hard problems are legal, not technical
I don’t want to sell a frictionless future, because the frictionless future is not the interesting part. The interesting part is that the blockers here are almost entirely legal and institutional, not technical. The tech mostly works. The law does not.
When an agent executes a bad trade, who pays? The developer who built the reasoning loop? The user who deployed it with too much latitude? The API that returned stale data? Tort law assumes a human decision somewhere in the chain. Agents break that assumption by design.
KYC is unsolved
Anti-money-laundering regimes were built for humans with faces and addresses. Agents have neither. The workaround of treating the agent as a proxy for its human owner works until the agent is acting across twelve institutions in four jurisdictions at machine speed, which is roughly day one.
Reversibility is unsolved
Most AI lives in forgiving environments. A bad marketing email gets edited. A bad line of code gets reverted. Finance is not forgiving. When capital moves, it’s moved. The feedback loop is immediate, and the undo button does not exist.
Fiduciary duty is unsolved
Can software hold a fiduciary obligation? The question sounds academic until an agent has to choose between two legally defensible actions and picks the one that happens to benefit its operator. Nobody has a good answer. Nobody is close to one.
These aren’t reasons the shift won’t happen. They’re reasons the shift will be ugly, a decade of regulatory scrambling, test cases, and at least one public disaster that becomes the reference point for all subsequent rules. Build accordingly.
What to build
If you’re a builder reading this, looking for a signal, here’s where I’d put attention:
The agent-native wealth product: Nobody has shipped this yet as a product. Everyone is shipping it as a feature bolted onto something else, a chatbot inside a neobank, a copilot inside a brokerage. The first team that ships The $500 Family Office as a standalone product, an agent with custody-adjacent permissions, tax integration, multi-institution reach, and a price tag that matches its cost structure, captures a category that doesn’t exist yet.
Agent-native brokerages: No GUI. No app. Just an API surface designed from day one to be called by software, with auth, audit, and reasoning-friendly error messages. The first one that gets regulatory clarity wins an absurd amount.
Programmable identity: Not “sign in with Google”, cryptographic credentials an agent can carry across institutions, with revocation and scope built in. Whoever solves this becomes the Stripe of the agentic economy.
Stablecoin-native settlement: Because agents don’t respect banking hours, and ACH is a relic. Stripe’s Bridge acquisition is the canary here, not the ceiling.
Liability underwriting: Somebody has to insure this. The first serious product that prices the risk of agent actions, algorithmic hallucination, bad reasoning, A2A negotiation failure becomes critical infrastructure within five years.
None of these are consumer plays. All of them are boring-sounding infrastructure.
The real shift
The family office was never a product. It was a labor arrangement that got called a product because only rich people could afford it. For forty years, the industry sold exclusivity as sophistication, and the marketing was so good that even the people working inside family offices started to believe it. They were not selling secret strategies, but twelve people paying attention.
Take away the twelve people, and the whole edifice collapses into something that costs $500 a year.
Robo-advisors were a rehearsal. They automated the one-line item that was already almost a spreadsheet, called it a revolution, and left the actual expensive parts of wealth management untouched because those parts needed humans. Autonomous financial agents don’t need humans. That’s the whole difference, and it’s enough to collapse an industry.
Money is learning to manage itself. The fintechs building better dashboards are going to be very surprised when they look up and notice their customers are hiring agents instead.