Tools Multiply. Seats Collapse - Why the software that survives AI is the software agents can use

Per-seat software pricing is a twenty-five-year artifact of a specific condition: the human user was the bottleneck. AI agents break that condition. When an agent can invoke a capability ten thousand times in an hour, the software market sorts into two categories: the tools whose usage multiplies, and coordination surfaces whose seat economics hollow out. The sorting has already started and the market is pricing it before the contract data catches up.

What was named and what wasn't

In April 2026, Nvidia CEO Jensen Huang sat down with Dwarkesh Patel and made a distinction nobody had asked him to make. Most software companies, he said, are tool makers. He named the tools: Excel, PowerPoint, Cadence, Synopsys. He said the number of agents would grow exponentially, and the number of instances of those tools would skyrocket.

But he did not name Salesforce or Workday, or ServiceNow, or Slack, or any other company whose economics depend on charging per human seat.

That omission matters more than the list he gave. Jensen was describing which category of software still makes sense when the human is no longer the scarce operating unit. And once you hear the distinction, a large part of enterprise software sorts itself very quickly.

The twenty-five-year artifact

Per-seat software pricing feels permanent because most people working in software today have never known anything else. But this is a business model that became dominant under very specific conditions.

Before Salesforce, enterprise software was usually sold as a perpetual license. You bought the software, installed it on your own hardware, and paid maintenance on top. The emergence of Salesforce and NetSuite helped normalize something different: pay per user, per month, with the vendor hosting the product in the cloud. Over time that model stopped looking like a choice and started looking like the natural shape of software.

Recurring revenue was easier for public markets to value than lumpy license deals. Cloud delivery made user-based metering enforceable in ways on-premises deployment never could. Per-seat pricing won because it aligned vendors, investors, and revenue visibility. Because for a while, the human user really was the bottleneck.

That is the condition now changing. Once it does, the software market sorts into two broad classes.

Software where the value is in the function. Software where the focus is on the human as the operating unit.

The tool side

Tool-makers benefit from agents because agents multiply usage of the capability inside their applications.

A tool is a discrete capability inside a larger suite. A crop function, a render engine, a design-rule check, a merge operation. Today those capabilities are bundled behind a single interface and sold as one subscription. But when an agent can invoke a specific function, automate it, chain it with others, or run it in parallel, the number of times that capability gets called can explode independent of the suite it sits in.

And the charge follows the function. It is done per crop, per render, per merge, per simulation run.

HubSpot's new Prospecting Agent is priced per lead recommended, not per user receiving the recommendation. Cadence is building agent systems that call its design tools autonomously and claims major gains in design-space exploration. Synopsys is doing the same from the other side of the EDA duopoly. GitHub is embedding Copilot across more of the software lifecycle, turning version control and code operations into agent-callable infrastructure. Adobe's subscription revenue is stable but the faster growth is not in seats. Firefly Enterprise, its API-based generation service, saw credit revenue grow 75% quarter over quarter, and new enterprise customers grew 50% year over year. That is what the early stage of the shift looks like inside a tool-maker: the old meter still works, but the new meter is already outrunning it.

In each case, what is growing is not the number of humans opening the application but the number of times a specific capability gets invoked.

The tool survives because the capability survives. In many cases it becomes more valuable, because the number of invocations is no longer capped by the pace of human attention. And its usage can be fragmented to the point where the individual actions themselves become the billable unit.

The seat side

Now look at the other category.

Salesforce, Workday, ServiceNow, Jira, Slack, and large parts of the seat-based SaaS universe are not discrete capabilities in the same way. Their real product is a coordination surface for human work. They organize approvals, handoffs, queues, status changes, escalations, ownership, dashboards, and process visibility for teams of people.

That was enormously valuable when the institution needed lots of humans to move work from one stage to another. But when the work inside the system begins to be performed by agents, that coordination layer starts to hollow out.

If the prospecting is done by an agent, the scheduling is done by an agent, and the CRM updates are logged by an agent, the shared dashboard is no longer where the work happens. It is where humans occasionally check on work happening elsewhere. The software does not disappear overnight but the thing that justified charging every employee a recurring fee begins to contract.

That is why the current market repricing is not just a general tech wobble. A large share of the punishment has concentrated in software companies whose economics depend on seat expansion. The market is beginning to ask whether the old meter still maps to the work.

The obvious objection is that if this were already true, customers would be renegotiating licenses down immediately.

Not necessarily. Enterprise software moves at renewal speed, not at technical speed. Contracts are multi-year and budget cycles lag operational reality. The first sign is not always falling contract value. Often it is weaker seat growth, flat expansion, or slower net-new adoption. The market, meanwhile, starts pricing the shift before contract data catches up.

In early February 2026, roughly $830 billion in global software stock value was erased in a single week — the sharpest repricing the sector has seen. Atlassian is down roughly 75% from its late 2024 peak. The company is still growing seats, its Teamwork Collection sold over a million in under nine months. But the market is no longer pricing seat growth. It is pricing the question of whether seat growth still means what it used to. That is what a leading indicator looks like: not the seat count falling, but the market deciding the seat count no longer matters.

There is another serious objection here. In regulated sectors (financial reporting, procurement, healthcare) humans cannot simply vanish from the process. The audit trail and legal sign-off remain and the institution still needs a record, an approval event, and someone accountable.

But preserving that compliance layer is not the same thing as preserving the full economics of a coordination product. What disappears is the revenue logic built on charging large numbers of people to manually move work through the system.

The escape attempt

Incumbents understand this. That is why so many are trying to move one layer up.

Salesforce is no longer trying to sound like a CRM vendor. It wants to be an agent platform. Workday wants to be a gateway for workforce-aware agents. ServiceNow wants to be the grounding layer for enterprise AI. SAP wants to host agent development on top of enterprise data.

The strategy is obvious. If the old seat-priced surface weakens, become the platform above it. Sell the orchestration, the agent access, or the layer the new automation runs through.

And to be fair, this does produce revenue. But so far it looks more like extension revenue from already-captive customers than a new economic foundation. Salesforce's Agentforce reached $800 million in ARR by early 2026 but that is roughly 2% of the company's total revenue, and 60% of Q4 bookings came from existing customers buying more, not from new market captured. These numbers sound large until you put them next to the business they are meant to replace.

But the deeper problem is that the escape layer is unstable too.

The same forces that weaken seat pricing are also weakening proprietary orchestration layers. Open coordination standards, model-agnostic tooling, and agent protocols are reducing the switching costs that make those layers sticky. If enough of the orchestration logic becomes standardized, then the new moat starts to look less like a moat and more like a temporary tollbooth.

What remains

Not everything in the old stack disappears at the same pace.

Three things tend to last longer than the seat model itself: the underlying data, the compliance shell, and the narrow points where institutions still need a human to own the consequence. Things like records, audit trails and exceptions still matter. In regulated environments, those layers may persist for years after the seat-priced coordination layer begins to weaken.

What remains is the residue that cannot yet be automated away: the ledger, the proof, the permission boundary, the accountable sign-off. Those layers may still support real businesses but they do not justify assuming the old seat-based software stack remains economically whole.

The real question is: where is the actual capability, and where is the institutional residue left behind when the humans stop doing most of the operating?

The sorting test

So here is the test.

Ask of any software product:

If you remove the humans, is there still a capability an agent can invoke? Or was the human coordination surface the product all along?

Figma: survives. The design engine remains usable through APIs and automation.

GitHub: survives. Version control, CI, and repository operations remain agent-callable.

Adobe Creative Cloud: survives. The rendering and media engines are the product.

Synopsys: survives. The EDA tooling still does the work.

Salesforce: hollows out toward collapse. The data may remain. The seat-priced coordination layer weakens.

Slack: hollows out toward collapse. Agents do not need chat as their native coordination surface. They coordinate through protocols and structured data. A channel is a UI built for human eyes.

Workday: hollows out toward collapse. Approval-heavy human process is the core product.

ServiceNow: hollows out toward collapse. Ticketing and escalation logic built for human bureaucracy loses density when agents do the operating.

Then come the ambiguous cases.

Notion is genuinely mixed. Part tool, part workspace.

Jira looks mixed, but resolves more toward coordination than capability.

So the question is not whether a vendor can bolt AI onto the product. At this stage it would be difficult to find an enterprise product without it. The real question is whether the product still makes sense when the human is no longer the scarce operating unit. And a 500-seat license built to move work from one human queue to the next does not.