AI Didn't Kill SaaS — It Killed Per-Seat Pricing
In January, Wall Street delivered a death sentence to software. On January 30, Anthropic released 11 vertical AI plugins for Claude covering law, investment banking, HR, and engineering. Within a week, US software and services stocks lost nearly $1 trillion in market value; the IGV software ETF fell almost 30% in under a month. The subtext among traders was blunt: AI can do what SaaS does, as a side project.
Six months later, the story flipped. CrowdStrike is up 106% year to date, Okta up 118%, and even Salesforce — the "legacy" vendor everyone loved to bury — jumped 22.6% in a single day after earnings.
Most commentary stops at "the market overreacted." That misses the point. Inside this round trip sits a rewritten business formula, and it applies far beyond software.
Here is the thesis in one sentence: AI didn't kill SaaS. It killed per-seat pricing. The reversal happened because the cash register is moving from seats to work units.
The Panic Formula and Its Hidden Assumption
The bear case earlier this year looked airtight: AI raises productivity → companies hire fewer people → paid seats shrink → SaaS revenue declines. A darker version added that agents would bypass software UIs entirely, stripping SaaS of its value as the point of entry.
The chain's fatal weakness was a premise nobody questioned: that SaaS revenue is permanently locked to headcount. Under per-seat pricing, the customer's employee count is the revenue ceiling — so any technology that compresses headcount compresses the vendor's income. If that premise holds, AI is an existential threat.
Six months of earnings say it doesn't hold. No leading SaaS company reported a collapse in seats; subscription growth across the sector's leaders held at 11%–36%. Salesforce's net new annual contract value hit a four-year high.
Note what this actually shows. It's not merely that revenue didn't fall — new bookings accelerated to a record. If AI were simply "not yet" eroding SaaS, orders should be flat, not better. Better orders point to only one explanation: AI is creating new reasons to pay.
Two Waves of Reversal, Each Falsifying Half the Panic
The SaaS recovery didn't happen all at once. It came in two waves, and each falsified one half of the doom logic.
The first wave, in May, belonged to consumption-based AI infrastructure vendors, marked by Snowflake's 48.4% single-week surge. These companies charge by usage, not headcount, and AI directly inflates demand: more deployed agents mean more data queries (Snowflake), more identity governance (Okta), bigger attack surfaces and thus security budgets (CrowdStrike), denser log monitoring (Datadog). In late May, this group delivered a cluster of earnings beats — Snowflake's product revenue grew 37% year over year, accelerating for a third straight quarter, while MongoDB Atlas held near 30% growth — driving the IGV up 21% for the month. Lesson one: for metered infrastructure, AI is a demand amplifier, not a predator.
The second wave, from late July into August, featured the incumbents most exposed to seat-based logic: Salesforce, ServiceNow, Workday. The bellwether Salesforce posted 10.8% revenue growth, a four-year high in net new annual contract value, and a 23% single-day pop on the print. Legacy leaders broadly rebounded more than 50% from their lows. Only after this wave did "AI kills SaaS" shift from consensus to abandoned hypothesis.
From Selling Heads to Selling Work: A Pricing-Migration Framework
Put the two waves together and a framework emerges: SaaS billing is migrating from seats to work units, and each vendor's position in that migration decides who wins and who exits.
Salesforce's revenue mix shows the mechanics. In its fiscal second quarter, AI-related revenue — Agentforce plus Data 360 — reached roughly $3.9 billion in annual recurring revenue, up about 210% year over year, approaching 9% of total revenue. Agentforce's typical scenario: a customer asks about a return at midnight. In the old model, a service agent logs in and handles it, and Salesforce bills per seat. In the Agentforce model, an AI agent reads the data, checks the order, and resolves the issue with no human involved — and the billing unit becomes "that one unit of completed work."
The ceiling for seats is the customer's headcount. The ceiling for work units is however much work the customer wants AI to do. The first gets compressed by AI; the second gets expanded by it.
Even more telling is the order composition: half of Agentforce's bookings come from existing customers. They didn't defect to startups — they upgraded within the system. "AI replaces Salesforce" became "AI is an add-on to Salesforce." And this is not one company's product strategy. ServiceNow's AI agent production deployments grew 9x in nine months, with 50% of net new business no longer priced per seat. Workday's AI SKUs reached roughly $600 million in ARR, up 200%, with usage-based Flex Credits rolled out to 200 customers. A collective pivot means this is the industry's new survival mode.
Within this framework, SaaS is splitting into three tiers:
- Metered infrastructure — repriced as AI beneficiaries, many at record highs (CrowdStrike +106%, Okta +118%, Snowflake +51%);
- Application incumbents mid-migration — up over 50% from lows but below records (Salesforce -5%, ServiceNow -9%, Workday -13% year to date), held back by two problems: work-based revenue is still a single-digit share (Salesforce ~9%, ServiceNow ~6%), and every unit of agent work burns GPU and inference costs, so margin is the pressure valve — ServiceNow's Q2 gross margin fell nearly 7 points year over year, and Salesforce's has declined for four consecutive quarters;
- Shallow, generic point software — still defined as AI losers (Intuit -52%, HubSpot -41%, Figma -35%, monday.com -36%), sharing a common trait: little accumulated data, workflow depth, or industry know-how, which makes their features easy for agents to absorb wholesale.
The Framework Applied: Every Head-Priced Industry Is in Line
The seat-to-work migration is not a software story. Any industry whose revenue is locked to headcount is standing on the same conveyor belt.
Consulting and outsourcing are the nearest analogues. Traditional firms bill by person-day; in an AI world, clients will buy the delivered analysis, not the consultant-days consumed. The firms that reprice from time to deliverables will replicate Salesforce's second curve. The rest will reenact the shallow-SaaS script.
Corporate procurement changes too. When software bills per completed ticket, per processed lead, or per closed refund, IT budgets shift from "buying tools for employees" to "buying output." That single change rewrites the sales pitch, the pricing page, and the renewal negotiation for the entire enterprise-software industry.
Even the labor market is being re-ranked by the same logic: value is measured less by your job title and more by how much agent-driven work output you direct. The January panic — "AI takes your job" — turns out to have a truer formulation: "AI takes your pricing unit."
What to Do About It
- If you build SaaS: audit how much of your revenue formula is locked to headcount. If you own data, workflows, or industry know-how, package AI capability as a usage- or outcome-priced SKU fast — the Salesforce path proves existing customers will buy add-ons in place. If you sell shallow, generic features, the window is closing.
- If you buy software: demand usage-based options in renewal negotiations and push idle-seat waste back onto vendors. But watch the margin trap: the vendor's inference costs will eventually return to you as price increases.
- If you invest: separate the three tiers. Metered infrastructure is fully priced; chasing carries risk. For incumbents, the key metric is work-based revenue share moving from single digits to double digits — tracking that beats tracking the stock. For shallow point software, the discount may not be finished widening.
- If you're an operator: tilt your personal value toward defining and QA-ing agent work output, not toward executing repeatable tasks — the latter is precisely what's being repriced.
One line to close: in January, Wall Street sentenced SaaS to death; six months later, it commuted the sentence to manslaughter. But what was spared is the species, not every way of living in it. Whoever moves the cash register from heads to work first holds the ticket to the next stage.
Sources: company earnings and market data as compiled in "曾被AI误杀的赛道，大反攻了" (36Kr, Sept 17, 2026).