The SaaSpocalypse, sorted: who AI is actually disrupting

Disclaimer: Not investment advice. I may hold positions in some of the companies mentioned. Figures are from company filings and daily closes to 25 September 2026; returns are price-only.

A word, then a selloff

On 30 January 2026 Anthropic released eleven plugins for Claude Cowork aimed at ordinary office work: reviewing contracts, triaging NDAs, building financial models. Four days later a Jefferies trader, quoted by Bloomberg, gave the reaction a name. The "SaaSpocalypse" was born on 3 February, a day when the software ETF IGV fell 4.6% on the heaviest volume in its 25-year history. Thomson Reuters fell 16% and LegalZoom about a fifth.

It kept going. Claude Opus 4.6 on 5 February, Claude Code Security on 20 February, a wave of agent headlines on 9 April. By the April low, IGV was 37% below its September 2025 peak. By StockCharts' count, January to March was its worst quarter since the end of 2008.

Fig. 1The year the stack split in two

The chart shows what didn't happen. Investors didn't leave AI; semiconductors are up about two-thirds over the same stretch. They moved down the stack, from the software that uses AI to the silicon it runs on. The implicit verdict was that AI is great news for whoever sells the picks and shovels and bad news for whoever sells the applications.

That verdict treated software as one thing. It isn't. A tax-preparation app, a fleet-telematics network, a database, a freelance marketplace and a language-learning game all got the same label and, for a while, the same treatment. Some of them are genuinely being disrupted. Some of them are being helped. And some sit exactly where they were, with the same customers paying for the same things, while their share prices were marked down as if they were next.

Three questions I ask of any software company

I sort software businesses with three questions. None needs a view on how smart models will get, only on what the business sells and to whom.

1. What does the customer pay for? A seat (a license per human employee), a unit of the real world (a truck, a site, a patient record, a warehouse), a volume of usage (queries, messages, data, compute), or a finished piece of work (an answer, a lesson, a tax return, a logo)? Seats shrink if companies employ fewer people. Finished work gets cheaper if a model can produce it. Units of the real world and usage don't obviously do either.

2. Could a general-purpose model produce what you sell, or does it need what you own? A model can write a homework answer from scratch. It can't know where your trucks are, what your hospital's regulatory filings say, or which version of the code shipped last Tuesday unless someone who owns that data, hardware or system of record lets it in. The first kind of business competes with the model. The second kind supplies it.

3. When AI usage goes up, does more flow through you, or less? Every agent that runs a task makes API calls, reads and writes data, logs events, sends messages and pushes code. If those land on your platform and you charge by volume, AI adoption is your growth driver. If AI adoption means fewer people log into your tool, it's the opposite.

The answers put companies into three groups:

Group	What they sell	What AI does to them	Examples
Accelerated	Usage, infrastructure, the layer that connects models to action	More traffic, more data, more workloads	Palantir, Datadog, Cloudflare, Snowflake, CrowdStrike, Twilio
Resilient	Records and operations the model needs but can't replace	Little direct effect; AI becomes a feature on top	Samsara, Veeva, Manhattan Associates, Autodesk, Procore, Tyler
Disrupted	Finished work that a model can now produce	Direct substitution or price collapse	Chegg, Fiverr, Upwork, EPAM, Duolingo, Intuit's TurboTax

The groups aren't moral judgments or stock picks. Plenty of disrupted businesses will survive, and accelerated ones can still be overpriced. They are a way to ask whether the market's reaction matches what's happening in the revenue.

The scoreboard

To test it, I took 36 US-listed software companies across the three groups. For each, I measured how much its revenue growth rate changed over the last four reported quarters and compared that with its 2026 share-price return.

Fig. 2What the market actually paid for

There's a relationship, but it's looser than the narrative suggests. Ranked, the correlation between acceleration and return is about 0.5; as a straight line, about 0.35. Acceleration explains perhaps a tenth to a fifth of the variation in returns. What stands out is the asymmetry. The quarter of companies whose growth slowed the most had a median return of about −43% this year. The quarter whose growth sped up the most had a median of about +7%, dragged down by names like Oracle (−30% despite growth accelerating by 17 points) and Palantir (+7% despite accelerating by 45). The market was far more certain about punishing deceleration than about paying for acceleration, partly because some accelerators were already priced for it.

Margins didn't matter at all: the change in operating margin had no relationship with returns.

Fig. 3Durable was priced like doomed

This is the finding I think matters most. The median resilient company grows its revenue about 17% a year, faster than the median disrupted company, and I can find no evidence in its numbers of AI replacing it. Its median share price is down about 20% this year, against about 25% for the disrupted group. The market sorted software by momentum, not by durability. Durable got priced like doomed.

Accelerated: when AI runs through you

Palantir is the clearest case of AI as a demand driver rather than a threat.

Fig. 4Twelve quarters, every one faster

Palantir's growth bottomed at under 13% in the quarter ending June 2023, two months after it launched its AI Platform. It has accelerated every quarter since, twelve in a row, to 93% in the June 2026 quarter, with US commercial revenue up 149%. The reason, in my framework, is question three. Palantir sells the layer that turns an organisation's messy data into a model of its real operations, on which AI agents can act. Every new agent a customer deploys needs that layer, so better models mean more demand for it, not less.

The same logic runs through the rest of the accelerated group. Datadog's revenue growth has accelerated for five quarters to 36% as AI companies and agent workloads generate more telemetry to monitor; its stock rose 31% on its May report. Snowflake's product revenue growth went from 30% to 37% over three quarters, with management crediting its AI coding tools; the stock jumped 36% on 28 May. Cloudflare re-accelerated to 36% in the same year that machine traffic overtook human traffic on the web it carries. CrowdStrike's net new recurring revenue grew 51% in its latest quarter as AI multiplied the attack surface (a subject for another piece). Twilio's organic growth has accelerated for four quarters on voice AI.

I'll go deeper into why agents create this kind of demand in the next essay. The point here is that for this group, "AI is eating software" has it backwards.

Resilient: the model needs you

Samsara sells cameras, gateways and sensors to fleets, construction firms, utilities and factories, plus the software that turns their data into routing, safety and maintenance decisions. Its customers pay per vehicle and per site, not per office worker. Its latest quarter had revenue up 30%, recurring revenue of $2.1 billion, also up 30% for the sixth quarter running, and a record 20 new customers paying over a million dollars a year. It posted its first GAAP operating profit.

In early February, during the first SaaSpocalypse wave, the stock hit $24, about a third below where it ended 2025. Nothing in the business had changed.

Fig. 5Nothing broke; half of them fell anyway

Run Samsara through the three questions. It sells units of the real world, not seats. A model can't produce what it sells; a model needs what it owns, the data from millions of vehicles and sites, to be useful to its customers at all. And AI makes the product better: Samsara is building agents on top of its own data that no outside model can replicate without it. The same holds, in different ways, for Veeva's regulated life-sciences records, Manhattan Associates' warehouse execution, Autodesk's engineering files, Procore's construction workflows and Tyler's local-government systems.

The chart shows how the market handled them. Veeva and Manhattan Associates are up with the market. Autodesk, Procore and Tyler are each down around 30% with no deterioration in their numbers. Autodesk is still below where it traded at the April low. For these companies, the risk the market priced in isn't showing up anywhere I can measure.

Disrupted: where it's real

There's a group where the disruption shows up in revenue, not just in valuations, and it's worth being specific about who is in it.

Fig. 6Where it is real

Chegg sold answers to homework questions; its revenue halved in a year and its headcount went from about 3,200 in 2022 to under 600. Fiverr and Upwork sell small, well-defined tasks done by freelancers (a logo, a translation, a landing page) and both are now shrinking. EPAM sells engineering hours, the kind of work coding agents are increasingly doing. What these businesses share is question one: the customer was paying for a finished piece of work, and a general-purpose model now produces something close to that piece of work for a few cents.

Duolingo and Intuit are the harder cases, because the evidence is more mixed than their share prices suggest.

Fig. 7Duolingo: the users slowed first

Duolingo's daily active user growth fell from 49% in early 2025 to about 21% a year later, just as free AI tutors appeared in ChatGPT, Gemini and Google Translate. The moment that crystallised it came in August 2025, when OpenAI demonstrated GPT-5 building a French-learning app live on stage. Then came the worst day in the stock's history: −25% on 6 November 2025, after bookings guidance disappointed. By April 2026 the shares were 83% below their peak. Management says part of the slowdown was deliberate, trading bookings for free-user growth, and I haven't found any source that proves AI caused it. Users have re-accelerated slightly since. My read: Duolingo sells a finished product (a lesson) that general models can now approximate, and its defence is habit and gamification rather than anything a model needs from it. That's a real exposure, even if the timing of the damage is arguable.

Fig. 8Intuit: priced for what the model might do

Intuit is the case that shows how far ahead of the numbers the market can move. On 21 May 2026 it reported a quarter that beat estimates and raised its full-year revenue guidance. The stock fell 20% that day, because it trimmed the TurboTax outlook, announced a 17% cut in staff, and did so the same month OpenAI launched a personal-finance experience in ChatGPT, which it says gets money questions from more than 200 million people a month. One analyst estimate puts the model cost of preparing a simple return at around 12 cents, against an average TurboTax charge of about $162. Intuit's revenue is still growing 10 to 14% a year, and its operating margin is expanding. The market is pricing the first two questions: TurboTax sells a finished piece of work, and a general-purpose assistant can increasingly produce it.

Where I'd push back is on treating Intuit as all answer and no infrastructure. Filing a return is also authorisation to file, liability, audit support and a decade of a household's financial history. Those are closer to the resilient group. How much of Intuit's value sits in the answer versus the rails is exactly the question the next two tax seasons will settle.

What about the seats?

The loudest version of the SaaSpocalypse was about seats: if AI agents do the work of office employees, companies need fewer licences for tools priced per employee. It's a sound argument, but the evidence of it happening in revenue is still thin. The cleanest disclosures I found:

• Salesforce cut its own support staff from about 9,000 to 5,000 using AI agents, which is the kind of seat reduction it now has to worry about at its customers.
• ZoomInfo is moving from a third to half of its contract value off seat-based pricing within two years, specifically to protect against seat downsizing.
• GitLab reported that seat contraction among its small-business customers was worse than planned, while total seats still grew.

What's much more common is vendors moving their pricing away from seats before seat losses show up. More than half of ServiceNow's new business is now on non-seat pricing. HubSpot is moving to credits and outcome-based pricing for agents. Salesforce is selling Agentforce by consumption. Gartner estimates that agentic AI puts about $234 billion, around a fifth of enterprise SaaS spending, at risk by 2030. That's a forecast, not a measurement.

The companies priced as seat casualties are, for now, mostly still growing: ServiceNow at 24%, HubSpot at 20%, Adobe at 13% with record revenue. HubSpot is down 47% this year and Adobe 33%. Either the revenue will eventually confirm those prices, or those prices will turn out to have been the overreaction. I don't think it's possible to know which yet. What I do notice is that the market reached its verdict before the evidence arrived.

So, is it real?

Yes, but narrowly. AI is genuinely destroying value in businesses that sell a finished piece of cognitive work that a model can now produce: answers, small freelance tasks, outsourced engineering hours, maybe lessons and simple tax returns. The revenue lines prove it for the first three.

For most of the rest of software, 2026 has been a repricing more than a disruption. The market lumped together businesses that AI replaces, businesses that AI needs and businesses that AI accelerates, and then only slowly started telling them apart. Salesforce's report on 27 August, when the stock rose 23% and the software ETF had its best day of the year, marked the moment the sorting became visible.

The mispricing I find most interesting is in the middle group. Companies that own the physical data, the regulated records or the operational systems that AI has to work through are growing steadily and, in several cases, trading as if they were on the losing side. The accelerated group has already been rewarded, and the disrupted group is being repriced correctly, sometimes brutally. The resilient group is where the market's reaction and the revenue line disagree most. The three questions are how I decide which side of that disagreement to be on.

GitLab is my favourite edge case: its reported revenue growth has slowed to 21%, but the net new recurring revenue it adds each quarter grew 42% last quarter, just as its usage-priced agent platform started to sell. The next essay picks up from there, with the agents themselves.