Sequoia's Julien Bek published a piece called Services: The New Software. It's one of the sharpest frameworks for thinking about where AI is going, and it's worth reading slowly.
The core argument is precise: copilots sell the tool, autopilots sell the work. The best AI companies won't make professionals more productive — they'll replace the outsourcing contract entirely. Bek's example is accounting: a company might spend $10K a year on QuickBooks and $120K on an accountant to close the books. The next legendary company, he writes, will just close the books.
The framework maps an opportunity matrix: outsourced versus insourced work on one axis, intelligence versus judgement on the other. The top-right quadrant — outsourced and intelligence-heavy — is what Bek calls the Autopilot Territory. The market it contains:
| Insurance brokerage | $140–200B |
| IT managed services | $100B+ |
| Claims adjusting | $50–80B |
| Accounting & audit | $50–80B |
| Healthcare revenue cycle | $50–80B |
| KYC / AML | $30–50B |
| Tax advisory | $30–35B |
| Legal transactional | $20–25B |
| Real estate closing | $20–25B |
| Total addressable | $700B+ |
North of $700 billion in outsourced professional services that AI agents are positioned to absorb this decade. The article is right about all of it.
But there's a question it doesn't answer. And for anyone building the infrastructure layer, it's the most important question in the stack.
When the autopilot closes the books — who signs the cheque?
Bek gives the example of Rillet, building the AI-native ERP that will close the books. That's the intelligence layer — reconciliation, categorisation, period-end close. The agent does the work a CPA team used to do.
But closing the books isn't just an information-processing task. At some point, money moves. Invoices get approved and paid. Payroll runs. Tax liabilities are settled. The autopilot that replaces the accounting function doesn't just produce a spreadsheet — it needs to execute the financial output.
The same logic runs through every industry in the Autopilot Territory:
- WithCoverage is building the autopilot for insurance brokerage — selling directly to the CFO who needs coverage, not to the broker. When a commercial policy is bound and a premium is due, the autopilot doesn't hand back to a human to push the payment. It pays.
- Pace is building the autopilot for claims handling. When a standard-line claim is verified against the policy schedule, the agent settles it.
- TaxGPT is building the autopilot for multi-jurisdiction tax advisory. When the liability is calculated, the tax autopilot needs to remit.
In every category on that list, intelligence is the input. The transaction is the output. And that's where the infrastructure problem begins.
Human payment infrastructure doesn't work for agents
Every payment system in existence today was designed around one assumption: there is a human at the centre of every financial relationship. The person who signs the cheque, holds the account, consents to the transaction — the human is the financial actor. The software is the tool.
In the Autopilot Territory, that relationship inverts. The agent is the financial actor. There is no human in the loop at transaction time. And the infrastructure wasn't built for that.
Credit card tokenisation requires a human cardholder. Regulated bank accounts require a human account owner. Neither maps onto an autonomous agent operating at machine speed, executing thousands of transactions per hour across dozens of counterparties, with no single human principal.
Consider what a claims autopilot actually needs to disburse a settlement:
- Its own financial identity — a wallet tied to the agent, not borrowed credentials from a human account holder
- Programmable spending policies enforced automatically without manual sign-off — settlement caps, counterparty allow-lists, claim-type restrictions
- Agent-to-agent settlement — when the claims agent coordinates with a repair network agent, they settle directly, without routing through a human-managed payment system
- Treasury management — working capital the agent controls natively, not a delegated sub-account of a human principal
- A complete audit trail — every transaction logged against the policy, the claim, and the agent identity that executed it
You can approximate parts of this by retrofitting human-designed systems — delegated API keys, service accounts, payment rails built for business customers. But the seams show: latency, compliance gaps, counterparty risk, operational complexity that scales with human oversight requirements instead of with transaction volume.
The question every autopilot company will face
The companies building in the Autopilot Territory right now are — correctly — focused on the intelligence layer. That's where the differentiation is, and it's where the hard technical work lives. But enterprise buyers won't sign deployment contracts for autopilots that can't complete the transaction loop.
An insurance CFO buying WithCoverage's autopilot isn't buying a sophisticated recommendation engine. They're buying a system that binds the policy, processes the premium, and files the documentation — without a brokerage team in the middle. The moment you say "the agent will hand the premium payment back to your finance team," you've re-introduced most of the labour cost the autopilot was supposed to eliminate.
The payment capability isn't a feature. It's what makes the product what it claims to be.
Bek's thesis is that the next $1 trillion company will be a software company masquerading as a services firm. He's right. But behind every autopilot that sells the work, there has to be infrastructure that moves the money.
The infrastructure bet
The companies Sequoia is backing right now will be in production within 12 to 18 months. The payment infrastructure question will be on every enterprise buyer's checklist before any of them can close a deal at scale.
The Autopilot Territory is the market. Every company in it needs the same thing underneath: financial identity for agents, programmable treasury, policy-enforced settlement, and audit trails that meet enterprise compliance requirements — without putting humans back in the loop at transaction time.
That layer doesn't exist yet at production scale. Building it is the infrastructure bet of this moment in the stack.
The Autopilot Territory is real. The agents are coming. What's missing is the financial layer that lets them actually operate.
Further reading: From Automation to Autonomy — Why AI Agents Need Their Own Money · Building a Payment Policy Engine for Your AI Agent · Why Legacy Financial Rails Break Under Agent Load