AI created new problems for customers, and fintechs are racing to solve them

Ramp has just closed a $750 million round at a $44 billion valuation. The article argues the company probably didn't need the money: what it bought with the round was a coherent narrative about becoming 'AI-native', and that narrative is credible because Ramp is rebuilding itself around jobs its customers only have because AI exists. The central example is Router, Ramp's tool that finds the lowest-cost, highest-quality model for each request and which, according to the company itself, saves customers 40%. Stripe is said to be assembling a similar stack: payments at the base, Metronome measuring and billing usage, and OpenRouter routing traffic across more than 400 models; Stripe reportedly paid more than $7 billion for that upper layer.

The underlying idea, borrowed from Clayton Christensen's 'Jobs to be Done' framework, is that customers don't buy a product but rather 'hire' it to do a job. Jobs such as managing expenses, reconciling invoices or preventing fraud predate AI: modern models can perform them faster and more cheaply, but the customer already wanted that beforehand. AI, however, created a second category of jobs that didn't exist before: monitoring token spend, routing each inference request to the best model, or letting an agent buy something without handing over the user's entire identity and bank account. The article thus distinguishes between 'AI-native' (which describes the product) and 'AI-created' (which describes the demand).

To organize this, the author proposes a 2x2 matrix crossing the type of job (old vs. created by AI) with the type of company (pre-existing vs. impossible without modern AI). Nubank, with NuFormer, exemplifies a pre-existing company using AI to do an old job better (credit underwriting). Harvey, Hebbia and Rogo are products that could not have existed before today's models but that solve old jobs (legal and financial research). Ramp and Stripe are pre-existing companies that now also capture demand created by AI (token management, model routing). And OpenRouter and fal are companies that only make sense because inference became an industry in its own right. The article itself notes that a single company can occupy several boxes at once: Ramp's expense business is 'old work', while its token dashboard and Router are 'new work'.

From there, the author proposes three possible postures toward AI. 'Adopters' use AI to serve the demand they already had (by adding a copilot, for example), which makes them faster and cheaper but doesn't change their sales funnel. 'Beneficiaries' —the case of Ramp or Stripe— keep serving the old jobs, but also capture the new demand AI generated right next to their core business. 'Natives' —OpenRouter, fal— exist only because of that new demand; they have no business in a world without inference to sell.

The article lays out a ladder of progression toward the 'AI-native'. The prerequisite is running the company itself with AI (teams already using AI internally, product managers and designers writing code, internal tools built by the teams themselves), something that speeds up operations but does not by itself change what is being sold. The first rung is old jobs solved better: regulatory compliance screening, reconciliation of documents and emails, legal research or financial reporting, tasks that companies such as Beacon, Sardine, Gradient Labs, Harvey, Hebbia and Rogo tackle by combining hundreds of documents and data sources with today's models; the product would have been impossible a few years ago, even though the job itself is old.

The second rung is the new jobs that appear alongside the core business. One is helping manage AI spend: Ramp's token dashboard lives inside a product that already existed, and according to the company its customers' AI token spend grew 20.7x between June 2025 and June 2026; Metronome, now part of Stripe, sits on the other side of that same job, measuring and billing that usage. Another is letting agents connect to the product and get to work: companies such as Mercury, Visa and Ramp are launching command-line interfaces (CLIs), and Stripe itself launched its own seven years ago, noting a sharp jump in usage since the launch of Claude Code. CLIs are easier for agents to navigate and consume far fewer tokens: according to the article's example, a transaction in the '--agent' mode of Ramp's CLI costs about 105 tokens in JSON, versus the 280 tokens the same transaction costs pretty-printed in '--human' mode. A third new job is getting your own store or brand discovered by AI agents: roughly a third of Gen Z already turns to AI instead of Google to research what to buy, which is leading platforms such as Shopify and WooCommerce, and the payment providers already working with e-commerce, to optimize for that channel too.

The third and final rung is the jobs that flat out did not exist before the recent AI boom and that nobody has fully solved yet. One is trust: how to know whether an agent can be trusted with your own data or decisions, especially when that agent doesn't live inside a software vendor you already have an enterprise agreement with, but is the product itself (from a lab such as Anthropic, a startup or an internal department). The article mentions companies building or buying 'harnesses' or control planes to wrap those agents, such as Primitive, and approaches like AIUC, which seeks to certify and underwrite agents. Another new job is the trust needed for an agent to transact: if an agent shows up at a store trying to buy something, how to know its reputation or its creator, and whether the user actually authorized it. Standards are emerging here such as Google's A2A, Visa's Trusted Agent Protocol and a FIDO identity standard, along with companies like Natural Payments, Skyfire and A-comm, which are positioning themselves at the part of the agentic commerce flow where agents actually move money. A third new job is obtaining cheaper inference and compute: according to Brex data, companies add their first open compute provider within five months of their first OpenAI or Anthropic API charge, moving from paying a model lab to comparing options such as Together AI, Fireworks and Baseten. The article notes that AI first created a software supply chain and is now generating a capital structure beneath it, with Nvidia and Wall Street trying to mobilize $500 billion around that layer.

A separate section is devoted to agent distribution. The idea is that no enterprise customer wants each vendor's 'agent' separately, but rather to have its own agent operating inside every product it uses. Most companies will end up running some kind of central control plane or 'harness' from which the CFO's, engineering's or operations' agents are budgeted, permissioned and observed, sitting above the individual tools; and consumers will have their own versions, whether through assistants from OpenAI, Google or, eventually, Apple. The article cites a warning from Garry Tan about how that orchestration layer threatens incumbent software: systems of record already own the data, permissions and distribution a 'harness' needs, but they run the risk of a new orchestration layer sitting on top and turning every underlying product into a mere callable provider. From this, according to the author, three non-mutually-exclusive strategic options arise: own the orchestration (become the place where the customer sees and manages all their agents), own a trust control point within that orchestration (identity, reputation, routing, provisioning, settlement), or become the easiest product for any major 'harness' to call, via CLIs, APIs and agent-oriented surfaces.

The article closes with practical advice for founders seeking to raise big rounds by claiming to be 'AI-native': the answer is not about positioning but about product. Find the job AI created right next to the value you already deliver, build the surface for that job, and then decide how to distribute it (by owning the orchestration, a trust control point, or being easy to call). The text recalls that Ramp did not reach a $44 billion valuation for being the best at expense management, but because investors believe AI will keep creating new demand alongside that core and that Ramp has the speed to capture it: its total payment volume (TPV) grew 170% year over year through March, its fastest pace in three years, after the business had already grown 20x.