The roll-up bank

The fastest way to win a business that borrows isn't to build a better bank. It's to buy the one it already has.

A business that borrows almost never leaves the bank that lends to it. You can build a better product, and the customers you need still won't come.

In The P&L was the moat, we argued that AI lowers the cost of running a bank enough to make building one possible again. We also left a warning unanswered: "A challenger could win the interface and still lose the customer." Building wins new customers. It doesn't win the ones already locked into a lending relationship. This memo is about those customers, and our answer comes from somewhere unexpected - the roll-up.

Some markets can't be entered by building. They have to be bought.

The rest of this is when buying beats building, why specialist business lenders fit that pattern, why rolling up banks used to destroy their value, and what we haven't worked out yet.

When buying beats building

A friend of ours builds computer-use agents for legacy software: decades-old systems with no API, which much of the economy still runs on. One of his biggest customer segments is AI roll-ups, companies that buy traditional businesses and rebuild them with AI. His insight was that many businesses used to be impossible to roll up profitably, and AI has changed that. We started testing the idea against two cases we know well, and a pattern emerged.

Lettings. Renting a flat in Europe is a poor experience, and plenty of startups have tried to fix it. They kept failing for the same reason. Demand was never the problem; finding people who want flats is cheap. Supply was. Landlords sign multi-year contracts with their agencies, and the window to switch is tiny. You can't build a marketplace without supply.

Dwelly solved it by buying the agencies. Buy an agency and you buy its landlords. Then AI rebuilds the back office. According to its backer General Catalyst, Dwelly acquired 15+ agencies managing 10,000 properties and reached the UK top 15 in under two years. Finding a tenant went from three weeks to under two days. Manager productivity doubled.

Wealth management. Advice is heavily regulated. Advisors need long certification, firms need licences, and few new advisors enter the market. Clients stay with their advisor for years and never switch on impulse. Many firm owners are approaching retirement with no succession plan, and are open to selling. Buy the firm, keep the advisors, give them AI.

Both cases have the same shape. Building was possible but pointless, because the customers were locked up somewhere else. General Catalyst puts it plainly:

Acquisitions are simply the fastest path to customers.

From those two cases, a roll-up makes sense when six things are true:

1. It's hard to build from scratch. Licences, regulation, expertise.
2. Customers are sticky. Switching is rare, slow and painful.
3. Acquisition is expensive. You can't buy your way in with ads.
4. The whole is worth more than the parts. Something you can only do once you own many of them.
5. Owners want to sell. Succession, fatigue, rising costs.
6. The market is fragmented. No monopoly, lots of targets.

Banking, segment by segment

So we pointed the six conditions at banking. Most of it fails.

Retail banking. We covered the economics in The P&L was the moat: small revenue per customer, acquisition by ad auction. Add low switching costs, since plenty of people hold half a dozen accounts, and there's nothing to lock in and nothing to roll up.

Micro-businesses. Most businesses are tiny, borrow little beyond a credit card, and behave like consumers. Every neobank has chased them for a decade. Same story as retail.

Startups. High-growth companies bank with fintechs. They were never at a legacy bank, so there's nothing to buy.

Businesses that borrow. Here the economics flip. The UK competition regulator found that only 4% of small businesses switched their main account in a year, and 70% of businesses older than ten years had been with the same bank for at least a decade.

What holds them isn't the app. Digital experience is nice to have. Fees matter to a price-sensitive minority who talk about them more than they act on them. Onboarding mattered once, when neobanks won by making accounts easy to open. But once you've banked somewhere for three years, you don't leave until it really hurts. And nobody leaves the bank that holds their credit line.

The target is mature small and mid-sized businesses that depend on credit. The lending book is both the barrier to entry and the lock-in.

The book is the moat.

Why there are so many business banks

There are few retail banks and a huge number of banks for businesses. The reason is underwriting.

Underwriting a consumer is relatively easy: a credit history and some cash-flow data tell you most of what you need. Underwriting a business is hard. You need to understand what the company does, what will happen to it, and every document it has.

So banks specialise. A 2021 New York Fed study found banks concentrate lending in particular industries, and that specialisation lets them offer better terms, get better loan performance and keep lending to small, hard-to-assess firms that other banks walk away from. The FDIC sorts US community banks into specialty groups (commercial real estate, business lending, agriculture) and finds they hold 15% of all bank loans but 36% of small-business loans and 70% of farm loans. In commercial real estate alone, community banks held $690 billion of loans at the end of 2019, almost a third of the US total, on just 12% of the industry's assets.

One bank knows construction. Another knows healthcare. Their edge is people who can read a P&L in that sector and know what credit to extend, why, and how to get it back. A construction lender's real skill is reading a contractor's job-level numbers: which jobs are on budget, which change orders will get paid, where the margin is quietly fading. It's exactly the kind of knowledge a model can now read at scale. Thousands of small specialists, each with a sticky, credit-dependent book that's impossible to win from outside. That's what a roll-up needs.

The trap

The obvious move is to buy a dozen specialists and merge them. It's also the move that destroys them.

The classic research here is Berger, Rajan, Stein and colleagues: large banks lend less to firms without formal financial statements. Small banks are better at using soft information (the judgement, the relationship, the local knowledge) because it doesn't survive being passed up a hierarchy. Merge the banks, centralise the credit committee, and the specialist's edge is gone.

Merge the specialists and you buy their expertise, then make it unusable.

Plain scale doesn't rescue it either. The FDIC finds community banks capture most economies of scale by around $600 million in assets. A roll-up that only shares overheads isn't worth the trouble. Bank roll-ups weren't impossible. They just destroyed the value they were trying to capture.

Consolidate without centralising

AI changes the trade-off. Soft information was never lost because it didn't exist. It was lost because it couldn't travel. Loan officers' notes, call transcripts, emails, contracts and cash-flow histories were unreadable to anyone outside the room. Models can read all of it.

Picture how a roll-up could work on those terms.

Specialists keep the decision. Each bank keeps its bankers, its brand and its authority to make credit decisions. Nothing gets centralised that shouldn't be.

One platform underneath. One core, one compliance and KYC stack, one AI layer for reading documents, analysing cash flow, monitoring covenants and answering customers. Construction expertise doesn't help underwrite a dental clinic, but the model that reads a contractor's P&L also reads a clinic's.

Agents on the old cores. Every bank you buy comes with its own core system, often decades old and without a usable API. Replacing ten cores before you see any benefit would sink the roll-up. Computer-use agents, the kind our friend builds, can operate those systems the way a person does: pulling data out, keying decisions in, running the workflows nobody has documented. Integration starts on day one, and the cores can be replaced when it makes sense, not before.

One balance sheet. A bank can only lend so much to one borrower: for US national banks, broadly 15% of capital. Small specialists turn away their best customers' biggest deals. Inside a group, they don't have to. Regulators also watch sector concentration, so a portfolio of specialists can grow where a single one would hit its ceiling.

Shared data. No single niche lender has enough defaults to learn from. A group of them does. Every specialist gets better because the others exist.

That's the fourth condition met. The whole is worth more than the parts, not because of shared overheads, but because AI lets you scale specialisation instead of replacing it.

Who sells, and where

The sellers look real. Between 2011 and 2019, voluntary mergers, not failures, became the main reason the number of US banks fell, and the FDIC cites succession as a motive: family-owned banks where the next generation doesn't want the job. Meanwhile, building from scratch nearly stopped. The US chartered an average of 183 new banks a year from 1985 to 2011, and four a year from 2012 to 2019.

Geography decides the rest. The US had 4,750 community banks at the end of 2019, about as fragmented as banking gets. The UK is the opposite: four banks held 83% of business current accounts. This is a US thesis, or one for markets that look like it.

Why now

Private equity has rolled up businesses for decades. The playbook was financial: buy, cut costs, add leverage, sell in five years. That improves a business. It doesn't change what the business is. And in banking, it kills the specialist edge it pays for.

The new playbook inverts it. General Catalyst describes its roll-ups as investing $100M+ into technology, holding for the long term, and aiming for outcomes that reshape a sector. It's not cost-cutting. It's rebuilding the operating model.

And the money is already moving. Dwelly has raised a $170M Series B, backed by General Catalyst and EQT Growth. General Catalyst is backing AI roll-ups across accounting, insurance, IT services and property management. Banking is harder to buy than a letting agency, and worth far more.

What's new isn't the roll-up. It's AI that lets you keep what made each business valuable while running them as one.

What remains hard

None of this is easy. Here's what we're still working through.

1. Regulators approve the buyer. Every acquisition is a change of control, and a tech-led owner will be scrutinised harder than a bank buying a bank.

2. Banks cost capital. A lettings agency is a team and a client list. A bank is a balance sheet. Each acquisition needs real equity, and the returns depend on credit and margins as much as on operating costs.

3. AI in credit decisions is regulated. Models have to be explainable and fair-lending compliant. Compliance and operations will move first; underwriting will follow.

4. Credit cycles bite. A model trained on good years hasn't seen a bad one. The specialists have.

Not built, assembled

Look for markets where building is too slow, acquiring customers is too expensive, customers never leave, owners want out, and AI makes the whole worth more than the parts.

Specialist lenders to businesses check every box. The next great SMB bank may not be built. It may be assembled, one specialist at a time.