After Assembly Line Finance

In a 1909 sales meeting, Henry Ford told an executive that a customer could have a Model T in any color, “so long as it is black.” Black paint dried fastest and cost the least. The assembly line made cars affordable by making them identical, and color was one more thing to standardize.

Financial services run on the same principle. The checking account, the index fund, the 30-year mortgage, the renters policy, the certificate of deposit, the PFM dashboard, the cashback card, the co-branded card, the debit card are all Model T financial products, the output of the industry's own assembly lines. As community banks, local carriers, and regional lenders have consolidated into national institutions, personalization has been the price of liquidity, portability, and a far lower cost to serve.

But the cost curve is changing. Designing, pricing, and servicing a financial product used to be expensive. AI does that work at a fraction of the cost, so a product no longer has to be designed once and sold a million times to pay back the spend. A product can be generated for one person, for one purpose, at the moment of need. N-of-1. The assembly line runs in reverse.

If products can be generated on demand, value shifts to whoever enables, governs, and capitalizes that generation. This issue is a tour of where I think those opportunities sit. First, the building blocks underneath the shift, then an attempt to disentangle the stack into three layers, and last a simple case study to make it concrete.

Building blocks

Each of these building blocks removes a fixed cost that once made n-of-1 product creation uneconomic. I see them both as the foundation that the later opportunities build on, but also an opportunity in their own right.

Tokenization: A tokenized asset or liability is divisible, transferable, and programmable. The most important constraint it removes for our purpose is minimum viable size. Every fund, pool, or position carries fixed administrative costs like recordkeeping, transfer, reconciliation, compliance, which set a floor on how small each unit can economically be. An asset manager cannot run a million bespoke accounts, because each custom mandate carries its own fixed overhead. Even if a company lowered access to standard products (e.g., free trading, fractional shares), tokenization is different because it lowers the cost of construction, i.e., the ability to assemble product to the customer’s spec that can be split, combined, or unwound as a software operation at near-zero marginal cost.

Synthetic instruments: A synthetic instrument delivers exposure to an asset through a contract rather than ownership of the asset itself. The perpetual future ("perp") is a good example because it requires holding none of the underlying asset, and a funding rate that balances longs and shorts keeps the contract tracking the asset's price. The product is fundamentally a formula that removes the fixed cost of holding assets to sell exposure to them, so an instrument can be generated on demand, in any size, with nothing warehoused in advance.

Agentic payment protocols: For agents to transact, someone has to define what an agent may do with money and prove it did only that. Today's infrastructure is built around identity, permissioning, and audit tied to human actors, whether a consumer or enterprise, that do not translate well to the future. Rather, the agent economy will perform better on rails with machine speed settlement, the ability to settle sub-cent transactions, and accountability by cryptographic provenance since no human is in the loop at the moment of payment. The emerging protocols (e.g., Coinbase-led x402, Google-led AP2) differ in architecture but converge on a verified mandate for the agent and its principal, spending bounded by amount, category, and purpose, a record of every transaction.

Stablecoins: A stablecoin is a dollar that settles in seconds, at any hour, callable by code. It eliminates slow correspondent banking and card rails that charge per transaction, have working hours, and generally are more difficult to integrate programmatically. Stablecoins underlie the other three building blocks, which all need money to move at the speed of the code composing them.

A three-part framework for n-of-1 financial products

The simplest way I can try and disentangle the emerging category of n-of-1 products is in three layers. The capital layer is who can hold and pool risk like the balance sheet, the license, the custody, the regulated capacity to take a position. Technology can make this layer more efficient, e.g., tokenized pools, programmable settlement, faster capital velocity, but it cannot manufacture the scarce input, which is the licensed, regulated capacity to bear risk.

The production layer is the cognitive work of actually making the product, such as underwriting, pricing, drafting the policy or documents, adjudicating claims, running compliance, servicing the account, and so on. This is the human labor and expertise that has (until now) gatekept bespoke products and intelligence from consumers and enterprises outside of the wealthiest or largest.

The human perimeter is who owns the customer. The perimeter is what cannot be generated just by cognition like trust, accumulated context, and permission to take delegated agency. When manufacturing an instrument is cheap, the scarce inputs are knowing which instrument to manufacture, for whom, and being trusted to do it.

Where opportunities emerge in the new financial stack

Using that framework, let’s look at opportunities in each layer.

1. Collapsing the production layer

A handful of well-capitalized labs that supply the base capability and improve it on a cadence few application companies can match have so far captured most of the value in collapsing the production layer. So is there still space in this opportunity set for early-stage teams? Two opportunities come to mind.

The first is the domain-specific model. Trained on the public commons, frontier models lack the expertise of a model trained on proprietary financial outcome data. Domain-specific models can be more effective in that they learn from the ground truth and spend no compute on subjects that do not pertain to finance. In 2026, this isn’t novel. My mind goes to Revolut's PRAGMA, which is a family of foundation models built on some 24 billion banking events. It tokenizes tabular financial records and applies LLM-style training to produce a general-purpose representation of a customer that gets reused across tasks. Although less public, I also think of Nubank, which is training billion-parameter transformer models on transactions to learn general-purpose customer representations that replace hand-engineered features like risk and collections.

These examples are encouraging and discouraging for our purposes. Encouraging because the value is clearly distinct from generalized LLMs in that a learned representation of financial behavior makes every downstream underwriting, pricing, and servicing decision better and cheaper. Discouraging because both examples are incumbents that have a decade of first-party data that a startup cannot replicate (and capex budget). The open question for startups is whether that prerequisite can be assembled another way, like acquired niche datasets, synthetic data, data partnerships, or (most likely) a route I have not thought of.

The second opportunity in this layer is the scaffolding or harness between the model and a live financial action that makes the action permissible and correct. This layer matters more in finance than elsewhere because regulation and trust require audit trails, permissioning, verification, and connectivity… and because the cost of error when handling money is disproportionate. In practice, the harness sits between the model's output and the money. In wealth management, for example, the model might propose a new investment in a client's portfolio; however, it’s the scaffolding that checks the trades against the client's specific mandate and files a record of why each trade was allowed. Or in insurance, the model might quote a policy, and the harness confirms that the output used no prohibited variables before going out to the customer.

2. A new unit of underwriting at the capital layer

The obvious opportunity at the capital layer is balance-sheet-as-a-service for the agentic market, i.e., to rent scarce, licensed capacity to perimeter players like a sponsor bank or MGA model. That captures real, utility-like value, but I believe it is a fraction of the larger opportunity. The deeper shift is that the unit of risk is changing. Incumbent capital machinery underwrites a stable counterparty, i.e., a person or a business, and holds capital against a book of standard, comparable contracts. The new unit of risk is agentic (so removed from a stable counterparty) and comprises a book of one-off instruments rather than a portfolio of like contracts. That is what incumbent machinery cannot price, but a first mover could.

On the asset side, the unit of underwriting shifts from the person to the agent. Underwriting today attaches to a legal person or business, which is reasonable since people and businesses were the only actors. As agents transact, the more natural atomic unit is the agent, or more precisely its structure of:

• Mandate: the authority to spend up to a limit, on a defined category, for a defined purpose

• Provenance: who deployed the agent, on which model, under what controls

• Record: the audit trail of what it actually did

The opportunity is to be the first entity that can hold capital against that new structure.

This also has implications on the liability side. When products are generated on demand, the book becomes a continuously recomposing one of distinct, n-of-1 instruments rather than a portfolio of standard contracts. Incumbent capital frameworks cannot describe that book, because statutory reserves, risk-based capital, and RBC/Solvency II all group similar contracts into cohorts, observe how losses developed on past cohorts, and hold capital against the pattern. But a per-instrument n-of-1 book can’t really be bucketed, because no two instruments are like-for-like, almost by definition. There also isn’t really a loss history, because each instrument is new. The opportunity then is to deploy and translate that new unit of risk into the language and systems that regulators and the rest of the capital stack can read and engage.

3. Owning the human perimeter and defensible context

The human perimeter is where automated finance meets an actual person. Relative to the underlying technology, it represents the standing relationship that accumulates context of a person's financial life and the permission to act on it, neither of which the model alone can generate.

This relationship of course compounds. The more a customer does within that human perimeter, the larger that perimeter grows, and the better it serves the customer.

I see winning the perimeter through three questions about context.

Is the context defensible? Anything a competitor can buy or pull (e.g., bureau files, credit scores, open-banking feeds, public records) confers no advantage. The defensible context exists nowhere else, much of which is messy and qualitative (e.g., the customer mentions an inheritance coming, a job offer in hand, private investments tracked on a desktop Excel file). Those signals often only appear through conversation and use on a surface the customer already trusts.

Is the context legitimate? Not everything collectible is priceable. Protected characteristics like race, religion, national origin, and sex, along with their proxies, cannot be priced on, directly or indirectly, under fair-lending law and state insurance codes. Regulation is catching up to the technology, slowly and unevenly across states. Colorado, for instance, now requires insurers to test their pricing models for discriminatory outcomes. The first challenge is operationalizing this shifting, state-by-state set of rules. The harder one is the gray area. Technology outpaces regulation, so a company will have to decide what to price on before a regulator has ruled.

Is the collection acceptable? Not all data a company could price on makes for a good customer experience, even when it lowers the price. I think about telematics in auto insurance, which is both defensible and legitimate. But the product is still poor because, frankly, people do not like being watched while they drive, even if it saves them money.

AI will magnify this challenge. A model that manages someone's financial life can absorb far more than driving data. Some of that intimacy the customer will welcome, because it produces a visibly better product at a visibly better price. But some of that proximity, customers will refuse. No matter the discount, there are simply parts of life people customers do not want a company to see. Ironically, this frames the question of AI in unmistakable human terms, because companies that win in this domain will have management that can distinguish between context that earns a better product and context that feels like surveillance.

Case study: n-of-1 renters insurance

What would this look like in practice? Naturally, the biggest delta are in the most gatekept, expensive categories like wealth management or complex commercial underwriting, because collapsing the most production costs saves the most money. But I think the opportunity is much more pervasive than that, so let's test the framework on one of the simplest products in finance, my own renters insurance policy.

Start with why the product exists. In my case and the majority of tenant contracts, landlords require renters insurance, mostly for their own benefit, specifically:

• Liability transfer: If a tenant starts a kitchen fire or a guest is injured in the unit, the tenant's liability coverage pays first i.e., before the landlord’s policy

• Protecting the landlord's own policy: Generally every claim kept off the landlord's policy protects its own loss history and therefore its premium

• Loss of use: If the unit becomes uninhabitable, the tenant's policy usually pays for temporary housing. Without it, a displaced tenant could withhold rent, demand housing, or litigate

Without the clause in my tenant contract, I would not buy a policy. My building has a doorman and no history of incident, so theft risk is low. It is a new development, so water-damage risk is low. And it is hard to construct the scenario in which a guest brings an injury claim against me in a one-bedroom New York apartment (no chance for slips and trips, no falling objects, no decks or trampolines or pools). For me, the product is a tax paid to the landlord, albeit a small one.

The policies themselves are near-identical. Each more or less covers three things: personal property (repair or replacement of belongings that are stolen or destroyed), personal liability (injury to others or damage to their property), and loss of use (temporary housing if the unit is unlivable). The only real levers are price and deductible. I am a price-taker. I shop briefly and choose Lemonade.

I find a few structural problems with the product:

1. The covered risks rarely present: Renters’ losses do not occur the way collisions do in auto or fire does in home. The events the policy is built around are improbable for most tenants most of the time. The actual unbudgeted expenses among younger New York renters are incidents like an emergency locksmith after lost keys, pest control, a security deposit dispute (which have cost friends more than water-damage ever could).

2. There is no underwriting: The economics do not justify underwriting a ~$170 policy, so pricing runs on zip code and coverage tier. Whether you live behind a doorman or keep your key under the mat, you pay virtually the same.

3. No value flows back: Cards return rewards; deposits return yield. A renters policy in a claim-free year returns nothing; my customer's best case is paying for a product they never use.

All of this is somewhat rational because precise underwriting and servicing cost more than a $170 premium can support. But what if that constraint collapsed? Let’s run the example through the three layers.

The production layer: Calling a general model to draft a policy is the low-hanging fruit. What about a renters’ version of PRAGMA, a model trained on claims histories, building incident records, water-damage data, record of what pricing regulators have accepted and so on, until it learns what a risk looks like and can incentivize siloed parties (e.g., the carrier, the landlord, the city) to share necessary data?

The capital layer: The capital layer is complex because insurance works by pooling risk. In other words, the many who do not claim pay for the few who do, which works only because no one knows in advance who will claim. If a carrier prices each tenant to their exact individual risk, insurance stops working because the low-risk tenants leave, the high-risk tenants cannot afford the price, and the pool unravels. Taken literally, n-of-1 insurance is a contradiction because a pool of one is no longer a pool.

But what if we can recompose risk? Let’s use our policy as a conceptual example. Today the unit of pooling is the person. A tenant is assigned to one bucket by zip code and coverage tier, and everything about them is averaged into one premium. The alternative is to make the unit the risk. A tenant is not one risk; they are a bundle of separate exposures with separate causes. I have a high risk of losing my keys and paying a $400 emergency locksmith (I’m a forgetful person, and Manhattan locksmithing is that expensive), that risk is behavioral, and it is mine. I have a low risk of water damage because that risk is structural, and it belongs to my new building. Theft risk belongs to the doorman. Liability risk depends on how often I host.

Programmable pooling decomposes the customer into these exposures and pools each one separately, with everyone who shares it. The result is n-of-1 without a pool of one. Each pool is large and uncertain, so the cross-subsidy still survives. But no two customers hold the same combination of pools, so every product is unique.

The human perimeter: Let’s go back to the framework: The human perimeter is the ability to collect and act on context that is legal to use, trusted, and pleasant to share. Suppose an MGA launches an agentic front door to renters insurance. The agent gets to know you through a natural-language conversation, and that conversation collapses the cost of acquiring context so it learns about the bike, the doorman, the roommate, then co-designs the policy and tailors it to one person. This sounds straightforward, but is harder than it looks.

 What is legal to use? It depends where you are. As a parallel, look at national origin in auto insurance below. Some states prohibit it outright, for some states it’s strongly limited or weakly limited, or not even mentioned! A national carrier has to operationalize all these rules at once.

The obvious lines are the anti-discrimination classics like race, religion, gender, and their proxies. But these are just the surface. For example, New York prohibits underwriting on a tenant's source of income, and bars insurers from reflecting in price whether a building contains affordable housing. A conversational agent makes this test harder, not easier, because a conversation collects everything. A tenant could mention their job, their visa, their Section 8 voucher, their neighborhood, without being asked and the agent cannot un-hear that.

What earns trust? This is two-sided. From my point of view as the customer, I want to share signals that are comfortable (how frequently I travel, whether I have a dog) and that I suspect will lower my costs (that I have a doorman, that I don’t host frequently, that a renters insurance policy could wedge into a home insurance policy in a few years). But what about my relationship status? Or my health? Or job prospects? Or whether or not I’m a klutz and more likely to end up with damage claims? For the underwriter, my selectivity is the challenge. Does silence hide a risk or just guard privacy? These are ultimately as much human questions as they are technical questions. The winning platform is not one that acquires the most context but (1) earns the trust to (2) acquire the right context and (3) feeds that back across the stack to produce an n-of-1 product.

Ford's assembly line made one car a million people could afford. What comes next makes a million financial products for a million people. If it could be done for a $170 renters policy, it could be done for anything in finance.

What did I miss? Email me at hworland@nea.com to continue the conversation.