The Death of SaaS

“SaaS is dead” is a headline covering three different claims: that the per-seat pricing model breaks, that the software gets commoditized by agents that can rebuild it, and that the incumbents lose. The first is mechanical and largely true. The other two are contested. This note separates them.

The seat was the meter

SaaS pricing attached revenue to a proxy for value: a human logging in. More employees using the tool meant more seats, and net revenue retention above 100% came from seat expansion inside existing accounts. The proxy held because work was done by people at keyboards.

An agent breaks the proxy. It does not log in, does not occupy a named seat, and can execute thousands of tasks under one credential or none. When one person supervising agents does the work that took five, seat count stops tracking value, and it starts tracking in the wrong direction. IDC projects that by 2028 pure seat-based pricing will be obsolete, with 70% of software vendors refactoring pricing around consumption or outcomes. Gartner’s version: by 2030, 35% of point-product SaaS tools are replaced by agents or absorbed into larger agent ecosystems, and at least 40% of enterprise SaaS spend shifts to usage, agent, or outcome pricing.

The threat is not only that agents replace human seats. It is that agents lower the cost of building software at all. A non-technical user can now assemble a workable internal tool by describing it, which is the same mechanism as Async Coding Agents applied to line-of-business apps. When rebuilding a single-purpose tool costs an afternoon, the thin point solution loses its pricing power. This is Jevons Paradox pointed at software supply: make creation cheap and the volume of bespoke software explodes, which is bad for the vendor selling one frozen version of it.

What February 2026 priced in

The repricing was fast. In early February 2026 the S&P North American Software Index fell roughly 15% while the broad market held, and analysts tagged the episode the “SaaSpocalypse” after about $285 billion in SaaS market cap evaporated. Anthropic’s Claude Cowork launch was widely tagged as the catalyst that crystallized the market’s conclusion that agents could absorb whole categories of knowledge work. (This note is Anthropic-produced, so weight that attribution accordingly.)

The dispersion inside the selloff is suggestive. Per-seat, workflow-of-record names were hit hardest. The confound: high-multiple growth software de-rates hardest in any software selloff, and Atlassian and Salesforce are also high-beta names, so pricing model and valuation multiple are partly entangled here. Atlassian dropped about 35% after a Q3 report showing enterprise seat count declining for the first time in company history, and it cut roughly 1,600 jobs (10% of staff) the following month to self-fund an AI and enterprise-sales pivot. Salesforce fell around 28%. The market was not pricing “software is worthless.” It was pricing “revenue indexed to headcount is now a liability,” which is a narrower and more defensible claim.

It is worth holding the counter-numbers next to these. Forrester still projects SaaS spending rising from $318 billion in 2025 to $576 billion by 2029. A model transition and a category collapse produce different stock charts but can produce the same spending line. “Death” is doing rhetorical work that “repricing” does more honestly.

Outcome pricing inverts the risk

The replacement models split into consumption (pay per token, API call, or agent run-minute) and outcome (pay per resolved ticket, closed deal, completed task). Outcome pricing is the sharper break because it moves risk onto the vendor. Under a seat license the customer pays whether or not the tool produced value. Under per-resolved-ticket the vendor eats the cost when the agent fails.

That inversion makes the vendor’s margin depend on something SaaS never had to think about: the marginal cost of doing the work. Every resolved ticket burns inference. If the price per outcome is fixed and the model call to hit it is expensive or variable, gross margin is now a function of AI Inference Unit Economics rather than the near-zero marginal cost of serving one more login. It also exposes vendors to the AI Price Wars running underneath, where falling token prices help outcome-based margins and rising reasoning-token consumption per task hurts them. Salesforce and ServiceNow have both shipped agent-native tiers that charge for work rather than for workers, which is an admission that the old model is at risk from people who know its economics best.

What survives

The strongest survival argument is not about features. It is about who owns the data and the source of truth. AI agents need somewhere authoritative to read state and write results, and Systems of Record (the CRM, the ERP, the ledger) hold unique data structures, long activity histories, and embedded regulatory logic that is expensive to replicate. Palantir, Bloomberg, and Veeva sit on proprietary datasets that an agent must pay to consult regardless of whether the consumer is a human or a bot. Owning the system of record means charging for access to ground truth even in a world with no seats.

The more interesting claim is that the disruption relocates the moat rather than removing it. Bain frames a $100 billion opportunity in cross-system labor: the highest-value automation is exactly where no single system of record owns the outcome, where a decision spans ERP, CRM, billing, and support at once. The new advantage is cross-workflow decision context, the ability to see and act across systems that each own only a slice. If that holds, agents do not kill SaaS so much as demote the single-workflow tool and reward whoever governs the connective layer between tools. That is a shift in where the rent sits, not the end of rent.

This returns to the third claim, that the incumbents lose. The same facts cut the other way. A broken pricing model is not a broken company if the vendor owns the customer relationship and the data and can swap meters in a quarter, which is exactly what Salesforce and ServiceNow shipping agent-native tiers demonstrates. Incumbents can reprice faster than a startup can win distribution. The load-bearing question for “who survives” is switching cost, not pricing mechanics. Note that the survival framing here leans on Bain, a consultancy paid to find the transformation work it forecasts, and on analyst houses rewarded for bold predictions. Discount accordingly.

Try it

Reprice a real SaaS tool as an agent would (an afternoon, spreadsheet plus an LLM API). Take a tool your team pays per-seat for. Estimate the volume of discrete outcomes it produces per month (tickets resolved, records updated, reports generated). Then rebuild one narrow slice of it as an agent: an API loop that takes an input and emits the same outcome, and measure the token cost per run against a provider’s current pricing. Compare three numbers: current per-seat cost, a plausible per-outcome price, and your measured inference cost per outcome. The gap between the last two is the vendor’s gross margin under outcome pricing, and whether it is positive tells you if the model even works for that use case. Watching your per-outcome inference cost move as you swap models makes the AI Inference Unit Economics dependency concrete.

Rebuild a point tool by description (1-2 hours, any vibe-coding agent). Pick a single-purpose internal tool (a form-to-tracker, a status board with an email integration). Describe it to a coding agent and see how close you get in an hour. The point is not to replace the tool. It is to feel where the difficulty actually lives: usually not the UI, but the data, the permissions, and the integration state the Systems of Record argument is about.

Sources

Bain, Will Agentic AI Disrupt SaaS? — the disruption-vs-survival framing and the data-moat argument
Bain, The $100-Billion SaaS Opportunity in Cross-System Labor — cross-workflow decision context as the relocated moat
Forrester, SaaS As We Know It Is Dead — the case that “death” is overstated, with the spending projection
IDC, Is SaaS Dead? — the seat-obsolescence and pricing-refactor predictions