Big software wins AI by selling to itself

EIGHTEEN MONTHS into the enterprise-AI boom, every big software company is doing a version of the same thing: shipping an AI-powered update to the product it already sells. Salesforce has Agentforce, ServiceNow has Now Assist, SAP has Joule, Microsoft has Copilot. Each promises an AI agent that will live inside the system a company already uses, understand its data, and take action on the user's behalf — a sales record updated, a customer ticket resolved, a report written. At roughly the same time, a very different kind of company has been growing. Harvey, Cursor, Sierra, and Glean did not bolt AI onto an existing product. They started with a foundation model and built a workflow around it. The industry calls the first motion "wrapping" and the second "native." The reader's natural question is which one wins. Eighteen months in, the honest answer is that both do, but they are winning very different things.

Base instinct

Salesforce's Agentforce is, by any normal measure, a runaway success. In the eighteen months since Marc Benioff unveiled it on stage, the product has signed tens of thousands of paid deals and reached nearly $1bn in recurring revenue — the fastest product ramp in the company's history. ServiceNow's Now Assist is on a similar trajectory. The trouble is that most of that money is coming from customers who were already paying Salesforce for something else. In the most recent quarter, more than 60% of Agentforce bookings came from existing customers expanding existing contracts. The native upstarts, by contrast, are selling to customers with no prior relationship. Cursor, the AI-native code editor, reportedly crossed $200m in annual revenue before it hired its first enterprise sales rep. Harvey, which rebuilt legal research and drafting around a foundation model, has gone from nothing to roughly $200m in recurring revenue in three years, with clients at most of America's largest law firms. The incumbents are cross-selling to a captive base; the upstarts are reaching customers the incumbents never had.

That distinction matters because cross-sell is a lease, not a moat. A March survey of SAP's own customer community found that only 3% of SAP users run the company's flagship AI product in production — the vast majority had quietly chosen Microsoft's Copilot instead. SAP has shipped dozens of specialised AI agents, but the product is only available to customers on a specific kind of cloud contract, and most of the base has declined to upgrade. When an incumbent's own customers pass on its AI wrap, the wrap has nowhere else to sell.

The divide between the two motions, at the eighteen-month mark, maps onto something simpler than revenue numbers: where the work actually happens. If the valuable information lives inside an existing piece of software — a sales opportunity inside Salesforce, a support ticket inside ServiceNow, a pull request inside GitHub — the incumbent is hard to dislodge, because prying the AI away from the data is more trouble than it is worth. If the valuable information lives in the work itself — a legal brief, a customer conversation, a chunk of working code — a native company can build a better workflow from scratch and deliver it faster than the incumbent can update its own. Harvey's rollout at Allen & Overy involved four thousand lawyers and was fully operational before the largest legal-software incumbent reached feature parity. Sierra, a customer-service platform only two years old, now handles support for roughly half of the Fortune 50 — a domain the big CRM vendors were supposed to own.

A quieter, more structural shift is happening underneath both motions. Both sides are gradually abandoning the model that made enterprise software profitable for two decades: charging per user per month. Salesforce now sells flexible AI credits. ServiceNow meters usage through something it calls assist packs. Sierra charges around a dollar fifty per successful customer-service resolution. Harvey is piloting deals where it takes a share of the revenue its software helps produce. The reason is simple enough: software priced per human makes little sense when the human is the very thing the software is trying to remove. What comes next is an accounting problem with philosophical stakes. Once every agent call has a measurable cost and a measurable return, customers will be able to compare the incumbent's AI and the upstart's AI on the same ledger, for the first time.

For now, the market is letting both bets run. Salesforce's stock has drifted down by roughly a quarter this year despite the record AI numbers, because investors suspect its AI boom is refilling a leaky bucket more than tapping a new well. Cursor and Harvey, meanwhile, trade privately at prices that assume the upstart approach eventually wins everywhere. Neither view is quite right. What is actually happening is a bifurcation — a market that looks like a winner-take-all race in every press release but is quietly turning into two parallel markets, each defensible on its own terms. That is an awkward outcome for anyone trying to pick a side, and a clarifying one for anyone willing to watch.