Ramp and Stripe Create New Revenue Streams Around AI-Driven Tasks

Author: Simon Taylor

Compiled by Deep潮 TechFlow

DeepOcean Summary: AI brings not only efficiency gains but also entirely new customer tasks that didn't exist before: token expenditures, proxy identities, and model routing. Ramp and Stripe have built new revenue streams around these tasks, and their valuation narratives are not merely about growth, but about betting on AI-native models. Understanding this map of "new tasks" helps investors identify who is benefiting and who may be left behind.

AI has generated new customer questions

Solving these problems is where the next generation of financial products is being formed. Token spending, proxy identities, and model routing barely existed before AI.

Ramp recently raised $750 million at a $44 billion valuation. It may not need the money, but this round bought into something bigger than impressive growth—this round bought into a coherent narrative around being AI-native.

This story is credible because it is rebuilding itself around AI-generated customer tasks. For example, Router finds the lowest-cost, highest-quality model for each request, reportedly saving customers 40% in costs according to Ramp. Stripe is assembling a similar tech stack: payments at the base, Metronome handling usage-based metering and billing, and OpenRouter routing traffic across more than 400 models. It has reportedly paid over $7 billion at the top layer.

Every investor, especially growth-oriented investors, is looking for companies that benefit from the rapid rise of AI and are thus more likely to endure. They want to hold the beneficiaries and avoid the victims.

This week, I spoke with three founders. All three are nearing the completion of large funding rounds under the banner of AI-native. All three asked me similar questions:

How should we position ourselves as AI-native? What should our product be?

They're asking about a positioning issue. I believe the answer is a product issue.

This column is my response.

Quick Map:

Are you an adopter, a beneficiary, or an indigenous person?

AI created a task hierarchy: improve existing tasks; new tasks near the core; new tasks无人解决 (unresolved by anyone).

Why does your proxy need a distribution strategy?

What should you do now?

Clayton Christensen’s most useful insight for this article is: Customers don’t buy your product—they hire it to get a job done.

Financial tasks such as reimbursing my expenses, reconciling my invoices, and helping me manage fraud have existed long before AI. Modern models can now perform some of these tasks faster, cheaper, and with less human effort—but when customers wake up, they still want the same things.

AI also created a second category of tasks: monitoring my token spending; routing each inference request to the optimal model; and enabling my agent to make purchases without handing over my entire identity or bank account.

Before AI created these issues, no one had these tasks.

I think about how to become AI-native using a 2x2 matrix.

Legacy tasks have always existed (such as managing fraud), while AI-generated tasks emerge alongside technology (such as managing my token expenditures).

Company: A company or product that existed before AI, or one that exists solely because of AI.

Putting them together looks like this?

Existing Customer Task

AI-generated customer tasks

Existing company / product

Nubank / NuFormer

Ramp and Stripe

It would be impossible without modern AI.

Harvey, Hebbia, and Rogo

OpenRouter and Fal

Top left: Depth of adoption. Nubank uses NuFormer for better credit decisions, but underwriting is an ancient banking task.

The bottom left shows the locations of Harvey, Hebbia, and Rogo: these products could not have existed before modern models, yet they perform legal and financial research tasks that have long existed.

The right side represents the tailwind. Ramp and Stripe already have businesses when AI offers customers new things to buy. OpenRouter and fal matter only because inference has become an industry.

This also means a company can occupy more than one cell. Ramp’s spend product is in the top-left; its token panel and Router are in the top-right. Stripe’s payment business is in the top-left; Metronome and OpenRouter pulled it toward the right side of the grid. Things rarely fit neatly into a 2x2 matrix.

So please allow me to introduce one more concept to you.

I found three tags that perfectly describe the company's stance:

Adopters use AI to serve their existing needs—perhaps they’ve added a co-pilot. There’s real value here. They can be faster, cheaper, and leaner, but the sales pipeline structure remains unchanged. Most SaaS companies that have added an agent fall into this category.

Beneficiaries improve their core tasks while capturing new demands created by AI next door. Ramp and Stripe are the cleanest examples: expense management remains the core task, while token cost management and model routing are new revenue streams nearby.

Indigenous entities exist due to new demands. OpenRouter routes requests across more than 400 models. Fal runs generative media inference at scale. In a world without inference to sell, neither of these companies would have a business.

This chart shows you where the demand comes from.

Tasks show you where opportunities lie. Becoming AI-native is a staircase: the bottom is familiar, the top is the brilliant unsolved problems. Let’s climb it together.

AI-driven operational models are transforming company structures, governance, and systems. Teams are building external plugins around employees, integrating internal data, sharing skills, and redefining who can do what. Product managers and designers are beginning to ship code, while engineers are moving deeper into product work. Teams are building their own internal tools. (See the AI Operational Model report for more.)

This is a ticket to enter. It lets you accelerate, but by itself, it won’t change what you’re selling.

Some AI-native products serve very old tasks. This does not diminish their native nature.

Before the advent of AI, compliance screening, document and email reconciliation, legal research, and financial reporting were all tedious administrative tasks. Beacon, Sardine*, Gradient Labs, Harvey, Hebbia, and Rogo use modern models to aggregate hundreds of documents and data sources, then provide answers or manage workflows. These products simply weren’t possible a few years ago—yet the tasks themselves are ancient.

I am an advisor to Sardine.

If you're managing expenses, accounts, or helping people make payments, existing tasks for returning customers still remain, but new tasks have also appeared alongside them—available in several different variations.

Task: Help me manage AI usage. Ramp’s token dashboard is an AI-generated task embedded within an existing product. You might argue that cost dashboards aren’t new—Vantage and Datadog have been pricing cloud and compute expenses for years. The difference is the bill. Before AI gave you a token invoice, no one needed to categorize, forecast, or distinguish between COGS and OpEx for token costs. On Ramp, AI token spending increased 20.7x between June 2025 and June 2026. Metronome, now part of Stripe, sits at the other end of this same task, responsible for measuring and billing this usage.

Task: Help my agent connect to your product and complete the task. Companies like Mercury, Visa, and Ramp have all launched command-line interfaces (CLIs). Even established player Stripe released its own CLI seven years ago. But usage has surged since Claude Code’s release. A CLI is a user interface that uses command lines (terminals) rather than apps or web pages. Agents find these interfaces easier to navigate, and they consume far fewer tokens. For example, running the Ramp CLI in --agent mode returns transactions in JSON format, using about 105 tokens. Running the same transaction in --human mode requires 280 tokens to format for display. This is the kind of interface designed for “non-human customers.”

Task: Get my store discovered by AI agents. About one-third of Gen Z now uses AI instead of Google to research what to buy. If your store and SKUs aren’t visible there, you could be missing out on opportunities. Companies like Shopify and WooCommerce, along with payment service providers (PSPs) that support e-commerce merchants, are now optimizing for this.

AI has created an entirely new category of demand alongside your existing business. Your opportunity is to build interfaces that serve these needs—there’s no need to immediately transform your entire business.

But imagine a future where agents account for 80% to 90% of internet traffic, commerce, and the economy. If even a fraction of this is true, how would you reposition your company and narrative? What is the core unit of value you create, and how does it serve that customer?

In this thought experiment, we must let go of the mindset of viewing AI as an add-on to existing businesses.

Some issues did not exist before the AI boom. They are not merely from simple use of AI.

How can I trust this agent with my data or the decisions it makes? We are entering a world where third-party-developed agents may interact with your business. You need a way to ensure they are secure, reliable, manage privacy, and have a legal entity accountable behind them. A simple model is having agents embedded within a SaaS provider you already have a business agreement with. But agents are increasingly becoming the product itself—they may come from a lab like Anthropic, a startup, or an internal team. I’m seeing companies build or purchase shells or control planes (like Primitive) to wrap these agents. Another approach is AIUC, which aims to underwrite and certify agents.

How can I trust this agent to make transactions? If an agent appears in your store trying to purchase something, how do you know its reputation or who created it? Has the user authenticated that agent to make purchases? There are emerging standards such as Google’s A2A, Visa’s Trusted Agent Protocol, and FIDO’s identity standards under development—but we’re still in the very early stages. Companies like Natural Payments, Skyfire, and A-comm are establishing their presence here, managing the环节 where agents actually handle money in agent commerce workflows.

How can I achieve lower-cost inference and compute? Brex says that within five months after a company incurs its first OpenAI or Anthropic API costs, it will onboard its first open compute provider. Five months. One minute you’re paying API fees to model labs; the next, you’re comparing Together AI, Fireworks, and Baseten. You decide where to run your workloads and ask finance how to hedge those costs. AI has created a software supply chain—and is now building a capital stack beneath it. Nvidia and Wall Street are trying to mobilize $500 billion around this layer.

Source: Brex Benchmark; Brex card and bill payment data, January 2021–June 2026

These tasks are larger than features—they span product, data, identity, and funds. Once they span across product, the question becomes: who do customers trust to be in charge of them?

I don't want your agent. I want my agent to integrate into your product.

Most enterprise customers will eventually run some form of control plane or shell: a central place to orchestrate the agents they own across the tools they use. The finance agent for the CFO, the coding agent for engineers, and the procurement agent for the operations team will be budgeted, authorized, and monitored from a layer above the products.

Consumers will also have their own versions—whether it’s Grokbot or Instinct, new consumer apps from OpenAI or Google, or Apple finally making Siri less terrible, something will always emerge that puts users first across products.

Garry Tan accurately described the threats facing existing software:

The record systems already have the data, permissions, and distribution needed for the shell. Their issue is that a new orchestration layer can sit on top of them, turning each underlying product into a callable provider.

This presents three strategic options:

Own the orchestration. Become the place where customers view, authorize, and manage each agent.

Own a trusted control point. Even if others have access to the shell, your identity, reputation, routing, procurement, and settlement retain their value.

Become the most accessible product for every major interface. CLI, API, and agent interfaces are the distribution channels. (This is the correct answer for more of you than you think.)

Salesforce and core banking systems won't easily give up the sticky wedges they already have. This isn't purely about aggregation theory. Record systems can become shells; experts can own control points; products can be distributed through all of these.

This tension is precisely the point. The prize could be the control plane. It could also be the indispensable element required by every control plane.

Going back to the three founders I mentioned earlier.

The answer is always work. Position follows the product.

Find the work that AI creates alongside the value you've already delivered. Build an interface for it.

Then decide how it spreads. You can have a shell with a trusted control point inside, or become a product that can be elegantly invoked from all shells.

This is what it means to reinvent yourself. You start by using AI to run your own operations. You use it to do your old jobs better. Then you climb toward roles that didn’t exist before the model caused problems for customers.

Ramp reached $44 billion not because it managed fees the best. It got there because investors believed AI could continuously create new demand alongside its core offerings—and Ramp moved quickly to capture it.

Select job. Select interface. Assign it.

(Ramp’s growth, by financial company standards, is extraordinary: TPV (total payment volume) increased 170% year-over-year as of March, the fastest pace in three years, while its business has expanded twentyfold.)