The Obituary Written Too Early
On January 30, Anthropic shipped eleven vertical AI plugins for Claude, covering law, investment banking, wealth management, HR, and engineering design. Wall Street read one sentence into the launch: the core business of SaaS can now be replicated by AI. Within a week, the market value of US software and services stocks evaporated by nearly $1 trillion. Within a month, core names had lost over $730 billion in combined value, and the iShares Expanded Tech-Software ETF (IGV) fell almost 30%.
Six months later, the script flipped. CrowdStrike is up 106% year to date, Okta is up 118%, and even Salesforce — the most "legacy" name of them all — surged 22.6% in a single day after earnings. In six months, the market priced the same sector twice: first a death sentence, then a pardon.
But the story is not about a stock rebound. It is about a rewritten formula. The panic logic that opened the year ran: AI raises productivity → companies hire less → paid seats shrink → SaaS revenue declines. The darker version held that agents would simply bypass software interfaces entirely, stripping SaaS of even its value as an entry point. Six months of earnings have broken that chain everywhere it mattered: leading SaaS companies kept subscription growth between 11% and 36%, and Salesforce's net new annual contract value hit a four-year high.
Here is the thesis: AI did not kill SaaS. It killed SaaS that charges per seat. What we are watching is not a recovery — it is a forced relocation of the cash register.
The Flaw in the Panic Formula
The formula spread fast because it looked airtight. Seat-based subscriptions are the foundation of SaaS revenue, and AI's first visible effect was precisely "one person doing the work of three." Higher productivity, fewer hires, fewer seats. QED.
The fatal flaw was an unexamined assumption: that SaaS revenue is permanently chained to headcount. Refuting that requires more than proving earnings didn't fall. It requires proving that the bond between revenue and human beings is being dissolved. Over six months, the earnings reports of the sector's leaders supplied exactly that proof — in two waves, each one falsifying half of the panic logic.
The first wave came in May, led by consumption-based AI infrastructure companies. More deployed agents mean more data queries (Snowflake), more identity governance (Okta), bigger attack surfaces and thus bigger security budgets (CrowdStrike), and denser log monitoring (Datadog). These companies bill by usage, not headcount — AI inflates their demand. Snowflake jumped 48.4% in a single week, reported product revenue growth of 37% with three consecutive quarters of acceleration, and MongoDB Atlas held near 30% growth. The IGV rose 21% that month alone. The lesson: for usage-priced infrastructure, AI is not a predator. It is a demand amplifier.
The second wave, from late July through August, belonged to the traditional giants most exposed to the seat logic: Salesforce, ServiceNow, Atlassian, Workday. Salesforce posted 10.8% year-over-year revenue growth, a four-year high in net new annual contract value, and a 23% single-day stock pop. Only when the names the market had condemned most confidently delivered results did "AI kills SaaS" shift from consensus to abandoned hypothesis.
A Framework: From Selling Heads to Selling Work
Traditional SaaS prices per seat, anchoring its revenue to the customer's headcount. Its ceiling is the number of employees the customer has. The new model prices each unit of work an agent completes. Consider a customer asking for a refund at midnight. Under the old model, a human service agent logs in and processes it, and the SaaS vendor charges for that seat. Under Salesforce's Agentforce, an AI agent reads the data, checks the order, and resolves the request with no human involved — and the vendor charges for that one completed task. The ceiling is now the amount of work the customer wants done. The first ceiling gets compressed by AI; the second gets expanded by it.
This is verifiable in numbers, not just narrative. In Q2, Salesforce's 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 company revenue. More telling is the order composition: half of Agentforce's orders come from existing customers. Clients did not defect to startups; they upgraded within the Salesforce ecosystem. "AI replaces Salesforce" quietly became "AI is an add-on to Salesforce."
And the move is industry-wide, not idiosyncratic. ServiceNow's agent production deployments grew 9x in nine months, with 50% of net new business no longer priced per seat. Workday's AI SKUs hit roughly $600 million in ARR, up 200%, with usage-based Flex Credits rolled out to 200 customers. When every leader makes the same pivot in the same quarter, it is not product strategy. It is a species adapting to a new environment.
The old formula — productivity up, humans down, seats down, revenue down — has been rewritten: productivity up, output per employee up, work volume and data and interactions up, SaaS usage up. Note what did not change: the underlying logic of SaaS. Whether customers keep using, renewing, and expanding remains everything. Only the location of the cash register moved.
Divergence, Not Recovery
Under the index, the rebound is anything but uniform. The IGV is barely back to flat for the year, while the spread between its best and worst components exceeds 150 percentage points. The real story of this market is not "SaaS is back." It is "SaaS is stratifying."
Three tiers have emerged. The first tier has been repriced as AI beneficiaries and trades at new highs: CrowdStrike (+106%), Okta (+118%), Snowflake (+51%). The more AI proliferates, the more these companies earn.
The second tier — the traditional leaders — has rebounded over 50% from the lows but not reached new highs: Salesforce (-5% year to date), ServiceNow (-9%), Workday (-13%). They are stuck on two fronts. First, usage-based revenue is still a single-digit share of their totals (Salesforce about 9%, ServiceNow about 6%), so the new story remains unproven at scale. Second, selling work is a double-edged sword: every agent task consumes GPUs and inference, so costs rise faster than revenue. ServiceNow's Q2 gross margin fell nearly 7 percentage points year over year, and Salesforce's margin has declined for four consecutive quarters.
The third tier remains defined as AI losers: Intuit (-52%), HubSpot (-41%), Figma (-35%), monday.com (-36%). Their common trait is shallow, generic, feature-level software — lightweight design tools, general office suites, simple customer service, basic CRM. Without proprietary data, embedded workflows, or industry know-how, their features can be swallowed wholesale by agents.
The framework travels. Apply the same test to any software market, including China's: does the company own exclusive data, workflows an agent cannot bypass, and domain know-how? All three makes it a seller of work. Missing any one makes it a seat waiting to be compressed. Whether a company has "integrated a large model" is marketing. Where its cash register sits is destiny.
What To Do With This
- If you build SaaS: Ask where your cash register is installed before you ship another feature. Seat-priced products — however good — sit on the side of the ledger that AI compresses. Find a meterable unit of completed work: a ticket resolved, a claim underwritten, a reconciliation closed.
- If you run a traditional software company: Watch the gross-margin line as you migrate to usage pricing. Salesforce and ServiceNow show that work-based revenue has an entirely different cost structure. If inference costs are not inside your pricing model, faster growth means faster bleeding.
- If you buy software: Use "can you price by outcome?" as a supplier litmus test. Vendors willing to charge per completed task have usually wired AI into their actual workflows. Vendors who will only charge per seat are usually still treating AI as a slide-deck word.
- If you invest: Sort positions into beneficiaries, migrators, and the replaced. The migrators' upside is the pace at which usage revenue climbs; the replaced's risk is the copyability of shallow features. Neither belongs in a mental bucket labeled "SaaS index."
One sentence to keep: whoever moves the cash register from heads to finished work first holds the ticket to the next stage. The story of AI killing SaaS has reversed — but not every company makes it to shore.