Banking for Vertical Agents

The next customer of a bank won't be a company.

It will be an agent acting on behalf of one.

That line still sounds strange to most people. It won't for much longer. A new category of buyer is showing up in the financial system — not a person, not a business in the traditional sense, but a vertical AI agent running the operational spine of an industry. These agents are already here. They're moving contracts, quoting materials, scheduling deliveries, reconciling invoices. What they can't do — yet — is pay.

In the previous two memos, we argued that the next great fintech won't be a bank, and that the bank of the future has no interface — it's a ledger, a memory layer, and a set of agents. This piece is the third turn of that argument: the customer of that bank isn't going to look anything like today's SME or enterprise client. It's a vertical agent. And that agent needs a different kind of bank.

Three waves of agents

The AI agent market has gone through three waves in about three years, and most of the industry has only noticed the first one.

General-purpose agents. ChatGPT, Claude, Gemini. Broad, horizontal productivity. You can ask them anything. They're useful across every industry and expert in none. This is the wave everyone talks about.

Specialist agents. Claude Code for engineering. Harvey for law. These plug into custom workflows and specialized tools, with fine-tuning and integrations that make them meaningfully better than general agents at a narrow set of tasks. Still horizontal in the sense that they serve a function — coding, legal research — across any industry.

Vertical agents. This is where things get interesting. A vertical agent automates the workflow of a specific industry — not a job function. It comes with industry-specific integrations, playbooks, fine-tunes, and data. It understands the domain the way a ten-year operator understands it. Sierra for support operations. Surfaice for construction. Abridge for clinical conversations. ClaimSorted for insurance claims.

The analogy worth holding is search. Google was horizontal. Then vertical search appeared — eBay for goods, Zillow for homes, Indeed for jobs, Avito for everything local. The horizontal index didn't go away. But the valuable, high-intent commerce moved into the vertical surfaces, because specialization compounded. Same dynamic is playing out with agents. The generalist remains the front door. The money gets made — and the workflows get transformed — by the vertical.

What vertical business agents actually do

Most vertical agents for business are a layer on top of the ERP.

They don't replace the ERP, and they mostly aren't trying to. ERP systems hold decades of process, data, and compliance scaffolding that isn't going anywhere. What agents do is sit on top and automate the operational layer that used to require human hands — the parts of the workflow where a person was reading emails, opening PDFs, updating tabs, typing into fields, chasing a vendor for a quote, reconciling a mismatch. That work is being eaten by agents, one business process at a time.

The practical shape of a vertical agent today looks like this: it understands the domain deeply, it knows which systems to read from and write to, it runs workflows end-to-end with human oversight at key decision points, and it plugs into whatever ERP, accounting, or back-office stack the customer already runs.

Which is great. Until you hit payment.

The workflow stops at the payment

Here's the specific thing that's broken: current AI agent workflows automate everything up to the moment money needs to move, and then they hand the task back to a human.

A construction-ops agent can negotiate with a cement supplier, evaluate contractor quotes, verify delivery, match the invoice to the PO, flag a discrepancy, and propose an approval — and then someone from finance has to log into a bank portal and actually send the wire. For a construction company with hundreds of counterparties buying concrete, steel, and labor, the speed and reliability of those payments is a direct input to margin. Late payments mean bad terms. Bad terms mean compressed margin. Payment execution is not a back-office footnote — it's one of the most operationally important levers the business has. And today it's the step the agent can't touch.

Multiply that across every industry where the same dynamic holds. Logistics agents that can't settle freight. Property management agents that can't pay contractors. Procurement agents that can't finalize vendor payments. Healthcare back-office agents that can't pay suppliers. The workflow is 90% automated and 10% human — but the 10% is the step that determines whether the automation is actually worth anything.

This is the cliff edge every vertical agent today walks up to and stops.

The ten-year picture: automated production talking to automated production

Zoom out for a second. The economy is moving toward more automated, more robotized production. Not in some utopian sci-fi sense — in the very concrete sense that factories, warehouses, logistics networks, and increasingly service operations are becoming programmatic. Machines talk to machines. Systems read from systems. Humans are pulled up into strategic and exception-handling roles.

If you take that trajectory and extend it, you end up somewhere specific: one automated production facility interacting directly with another automated production facility. One AI-operated business ordering from another AI-operated business. The logistics, the procurement, the payment — all happening without a person in the loop, because a person in the loop is the slow part.

This sounds speculative until you look at it from the inside. A manufacturer knows — programmatically, in real time — which components it needs, in what quantity, by when. A supplier knows — programmatically, in real time — its capacity, its pricing, its lead times. There is no good reason these two systems should route through a purchasing manager, an account manager, an email thread, and a wire transfer. They should transact directly. Faster, cheaper, with full audit trail, no manual reconciliation.

That's the ten-year picture. Maybe sooner in some industries, later in others. But the direction is one-way.

What's actually needed for that world to exist

A few things have to be true for agent-to-agent commerce to work in practice. Most of them weren't true five years ago. Most of them are true today.

Payment rails fast enough to be part of the workflow. Agents operate at machine speed. An ACH that settles in two days or a wire that takes a day isn't infrastructure — it's a bottleneck that forces the whole workflow back into human time. Blockchain rails clear in sub-second to seconds. That's not a marginal improvement; it's a category change that makes payment a normal step in an API call instead of an out-of-band process.

A stable, usable unit of account. Stablecoins are no longer the Wild West of a few years ago. They're a solid, regulated, production-grade part of the financial system. USDC alone clears more volume than most card networks on a monthly basis. The GENIUS Act in the US and MiCA in Europe put formal frameworks around them. This is no longer an experimental rail; it's a real one.

Programmatic guardrails. Agents can't be handed a checking account and told to have fun. They need permissioned contracts, spending limits, policy enforcement, audit trails, and real-time risk controls. Smart contracts provide part of this natively. Payment protocols built for agents — x402, the MPP that Tempo and Coinbase launched recently — add the rest: standardized ways for an agent to authorize, authenticate, and execute a payment with the guardrails baked in.

All three exist today. The infrastructure is in place. What's missing is the integration layer that turns these pieces into a usable product.

What has to get built

Three layers, roughly.

A unified payment stack. Something that spans every blockchain, every stablecoin, every existing fiat rail (ACH, SEPA, wires, cards), and presents it as a single interface to an agent. The agent shouldn't care whether the counterparty takes USDC on Base, a wire in EUR, or an ACH push in USD. It should ask for "the cheapest, fastest rail that clears in policy" and get it.

Deep accounting and ERP integration. Every transaction that flows through that stack has to land correctly in QuickBooks, NetSuite, SAP, Xero, whatever the company runs. Without this, the agent is creating a reconciliation problem rather than solving one. The bookkeeping has to happen automatically, in the right accounts, with the right metadata, the right tax treatment, the right audit trail. This is the unsexy part that most crypto-native payment companies underestimate and most traditional fintechs are too slow to build.

Financial agents as the operating layer. A bookkeeping agent that verifies every transaction has the right supporting documentation — invoices, contracts, POs, receipts — and that the accounting treatment is correct. A treasury agent that understands when to execute, which account to pay from, when to move balances between entities and rails, how to manage float and FX exposure. These aren't "nice to have." They're the substitute for the finance team that used to do this work by hand, and they make the whole stack usable at scale.

And a new kind of marketplace

There's one more piece, adjacent to the payment stack but not the same thing.

If you're a large manufacturer today, you don't buy from suppliers by browsing a public marketplace. You run a private procurement platform — often built on enterprise software like Ariba or Coupa — where approved vendors can see orders, submit bids, manage POs. That's the marketplace model for serious B2B commerce.

The agent-native version of this is going to look different. It'll be built on smart contracts and new payment protocols. The interface will be programmatic, not visual. The discovery, negotiation, contracting, and settlement will all collapse into a single transactional surface that two agents can operate against directly. Whoever builds that — and it's a separate company from the vertical agents themselves — owns a valuable layer.

But this layer is orthogonal to vertical agents. A vertical agent's core competence is understanding the industry and running the workflow. Building a marketplace, a payment stack, and a financial control layer is a completely different product. It's horizontal work that serves every vertical.

Why vertical agents won't build this themselves

This is the key structural point. The task of a vertical agent is to be the best possible operator of a specific industry workflow. That's where the domain expertise lives. That's where the defensibility lives. That's where they spend their engineering cycles.

Payment infrastructure is not their core business. It's an enormous tax on their roadmap if they have to build it. They'd need payment expertise, compliance expertise, banking relationships, blockchain infrastructure, ERP integrations, treasury logic — none of which they have today and all of which take years to build well. AI makes all of it faster to implement, but faster isn't the same as trivial. There are failure modes — regulatory, operational, reputational — that you don't want to hit while also trying to be the best vertical agent in your category.

And crucially, the same payment and treasury stack is reusable across verticals. A construction agent, a logistics agent, a property management agent, and a procurement agent all need the same underlying financial rails, the same ERP integrations, the same bookkeeping and treasury agents. The product composes horizontally. It's a classic platform play.

So vertical agents will buy, not build. And someone needs to be the seller.

The bank of the future is the bank for agents

Put it all together and you arrive at a specific architecture.

A vertical agent sits at the top — the interface between strategic decision-makers and the business. It's the multiplier, the operator, the system of action. Underneath it, legacy ERPs remain, because they hold the historical truth of the business. Next to the ERP, a new financial stack: unified payment rails across blockchain and fiat, compliance and KYB/KYC, accounting integrations, treasury and bookkeeping agents, an agent-native marketplace layer. All of it packaged not as a traditional banking product with a UI and a portal — because we've already established that the bank of the future doesn't have an interface — but as a set of APIs, protocols, and agents that other agents talk to.

This is the bank for vertical agents. A B2B2B business. Embedded finance, but at an entirely different level of abstraction than the embedded finance of the 2010s — which was really just "put a debit card in a consumer app." This is embedded finance for a world where the primary economic actor is a machine acting on behalf of a business.

The customer is the vertical agent. The vertical agent's customer is the operating company. The operating company's customers are its actual end users. Three layers of abstraction, and the bank sits at the bottom of the stack, invisible to everyone except the agent it talks to.

"Can't you just build this on existing rails?"

This is the obvious pushback, so worth addressing directly. No — you can't, for several reasons.

Speed. Existing rails were designed decades ago for human-speed workflows. They assume a human is waiting a day or two for a payment to clear and is fine with that. Agent workflows break if payment isn't part of the same execution loop as the rest of the operation. The architecture of ACH, SEPA, and card networks isn't something you can patch into sub-second finality.

Programmability. Existing rails don't have smart contracts, don't have policy-aware execution, don't have granular agent permissions, don't have the kind of audit trail that a regulator will want for agent-initiated payments. Banks can bolt on APIs, but the underlying settlement model isn't programmatic in the way agent workflows require.

Interface. The existing stack is designed for humans to log in, click, and confirm. Agents don't do that. The whole interaction pattern is wrong.

Integration depth. Low-level banking infrastructure exists — Column, Bridge, Circle, Lead, a handful of others. These are infrastructure providers, not products. A vertical agent builder trying to assemble a usable system from those primitives has to write a small bank of their own. That's possible — AI makes it faster than it used to be — but it's still a multi-year distraction from the actual product.

Repeatable, horizontal, not owned

This is the interesting structural property of the opportunity: it's repeatable across industries in a way that most fintech products aren't.

A construction-agent bank, a logistics-agent bank, a healthcare-agent bank, a procurement-agent bank — they all need the same core stack. The surface integrations differ. The vertical knowledge lives in the agent, not in the bank. The bank's value proposition is that it composes cleanly with any vertical agent and handles the financial plane so the agent doesn't have to.

That's the shape of the company that wins this category. Horizontal financial infrastructure, agent-native, embedded into every vertical workflow, with the unit economics and network effects that come from being the shared substrate underneath a large fragmented market.

The emerging category

Agents are already everywhere in the obvious places — support, sales, coding, legal research. They're just starting to show up in the heavy, high-stakes parts of the real economy: manufacturing, construction, logistics, healthcare operations, procurement, energy. These industries are where the automation opportunity is biggest, and where the payment problem bites hardest.

The US in particular has a strong policy and capital tailwind here. Stablecoin regulation is in place. Blockchain infrastructure is production-grade. Vertical agent companies are being funded at real scale. Every one of those companies will hit the payment wall at some point in the next eighteen months.

This is the emerging category. A new type of customer — the vertical agent — and a new type of financial product built for it. The infrastructure exists. The need exists. The integration layer — the bank for agents — is what we're building.