The Seat Was the Business

Salesforce reported its Q1 FY26 earnings last month. Buried about a third of the way through the deck, past the revenue headline and the RPO chart, was a number the company framed as a success and which is also, read honestly, the most interesting admission the enterprise software industry has made this year. Agentforce — Salesforce’s AI agent platform, the product the company has staked its AI narrative on — reached approximately $100 million in annualized revenue across more than 4,000 paying customers. About 30% of the quarter’s bookings came from existing Salesforce customers expanding into Agentforce.

The standard read is that this is good news. Agentforce is working, customers are buying, the existing install base is converting. The more revealing read is what the underlying pricing model says about the math. Agentforce has been repriced twice in less than a year. The current recommended model, Flex Credits, charges approximately $0.10 per action — or $500 per 100,000 credits. This is not a seat license. It is, in the specific technical sense that enterprise software has spent twenty-five years avoiding, a consumption bill.

For the previous quarter-century, Salesforce has priced software by the seat. At $150 per month per user for the rough mid-tier, a customer bought access for one human to a deep body of functionality, some small fraction of which any individual user would touch in a given month. The gross margin on that seat was somewhere around 80%, because the marginal cost of a seat was nearly zero. This is the architecture that enabled the entire SaaS era. Not the feature set. The pricing.

What the seat model was actually selling was access — priced against headcount, calibrated to what enterprises had decided they could afford, indexed to a linear relationship between what employees did and what their employer paid for. This worked beautifully for as long as software usage scaled with headcount. For most of the last two decades, in most enterprise workflows, it did. More salespeople meant more CRM seats. More engineers meant more Jira seats. More customer service reps meant more ServiceNow seats. The SaaS business model was a toll on the number of employees using the product, and the toll was a function of what the software did for those employees, not what it cost to run.

Foundation models broke this assumption. Not slowly, and not in the future — structurally, and already. The work that five customer service agents used to do is now being attempted, with mixed success, by one agent and an AI that handles the first-tier triage. A sales team that used to have twenty account executives with their own CRM seats is now being restructured around ten AEs and a pipeline of AI-drafted outreach. A marketing team that needed fifteen Hubspot seats now needs eight, because the agents are drafting half the campaigns. The seat is not a useful unit of value anymore because the work is no longer indexed to the headcount.

The vendors that notice this first are the ones repricing by the unit of work rather than the unit of human. Agentforce at $0.10 per action. Intercom’s Fin agent at $0.99 per resolution. Zendesk’s AI agents at seat-plus-resolved-interaction fees. GitHub’s Copilot Enterprise at a blended per-seat-plus-usage tier that Microsoft’s finance team reportedly rebuilt at least once. Each of these is a different kind of promise. The vendor is committing to deliver a specific quantum of work, and the customer is committing to pay for that quantum rather than for access.

The consequence the market has not fully priced is that per-seat ARR, as a metric, has become untethered from the underlying business. For twenty years, “annual recurring revenue” at a SaaS company meant a specific, durable thing: the number of seats, times the annual price per seat, reliably renewing. It was the cleanest number in enterprise finance. A company with $1 billion of ARR growing at 30% and 120% net revenue retention was a machine whose mechanics the investment community understood.

A company transitioning significant product surface area to consumption pricing does not have that kind of ARR anymore. What it has is a forward revenue estimate that depends on usage patterns the company cannot control and customers cannot always predict. Salesforce’s Q1 Agentforce figure is an “annualized” revenue number — the company took the most recent period’s run-rate, multiplied it out, and presented it as though it will recur at the same level for the next twelve months. This is the standard SaaS convention. It was built for seats. Applied to consumption, it is a projection, not a contract.

The multiples the market currently assigns to enterprise software companies are calibrated against the seat-era framework. Revenue of the sort these companies are increasingly generating is not calibrated the same way. The gap is already showing up in the sell-side models — analysts are quietly segmenting “committed” and “consumption” ARR in their own footnotes, because the same dollar of revenue does not behave the same in both buckets.

This is where the argument turns uncomfortable. If consumption pricing is the correct answer — and the large vendors’ move toward it is a signal that they believe it is — then a significant portion of the enterprise SaaS commentariat has spent the last three years looking at the wrong number. NRR against a seat base is not a durable benchmark when seats are being actively reduced as a product strategy by the vendors themselves. Gross margin calculations that assumed near-zero marginal cost per unit of value delivered do not carry over to a model where the vendor is actually paying an inference bill every time the product does what the customer is paying for.

The unit economics that made the SaaS decade what it was — 80%+ gross margins, predictable expansion revenue, CAC payback periods that shrank as cohorts seasoned — were a consequence of a specific pricing architecture. That architecture is being dismantled by the vendors themselves, on their own earnings calls, in pricing memos that are quietly republished every few months as the economics clarify. The gross margin on an inference-mediated product is not 80%. It is, depending on model routing and the quality tier the customer is paying for, somewhere between 40% and 65%. The ceiling on that number is bounded by what foundation model providers charge and, more importantly, by how fast they cut each other’s prices.

The argument against this framing, offered by the large vendors, is that bundled offerings — seat plus consumption — preserve the old unit economics on the seat portion and add the new consumption revenue on top. This is the story Salesforce is telling. It may work. It has a history worth noticing, though. The last time an enterprise software company tried to bundle a premium per-seat product with a new consumption layer to preserve margins, it was Oracle in the 2010s, and the subscription wave rolled through anyway. Defensive bundling buys time. It does not change the direction the pricing model is moving.

Nobody has a robust framework yet for what a healthy SaaS business looks like when agent-priced revenue becomes the majority of the book. The models everyone is using are still the seat-era models with consumption revenue bolted on as an asterisk. That is not a durable analytical posture. Within the next two to three years, somebody is going to publish the framework that replaces it. The companies that will have the easiest time explaining their business in that framework are the ones whose cost structure was already consumption-aware, whose gross margins were never the artifact of the seat fiction, and whose growth was not underwritten by a seat-expansion playbook.

The vendors that priced seats the longest have the most to explain. The vendors that priced consumption from the beginning have the quietest spring.