It is about who owns the data, who owns the relationship, and who is left standing, accountable, when something breaks.
By Arpita Saxena, Frontier Signal
On 29 January 2026, enterprise software had its worst day since the COVID crash of March 2020.
The trigger was not a recession or a rate rise. It was a fear simple enough that a child could follow it: if a machine can do in seconds what a person used to do all afternoon inside a $150-a-seat piece of software, why keep paying for the seat? Wall Street gave the moment a name, half joking and half not. The SaaSpocalypse.
It kept going. By early April the software ETF tracking the sector was down more than 21 percent for the year. HubSpot, Atlassian and Figma had fallen 70 to 80 percent from their highs. For a few months it looked like the future Satya Nadella had warned about in December 2024, when he questioned whether the software business model survives agents at all, was arriving exactly on schedule.
And then it did not. By late July, Salesforce rose 7 percent in a single session, ServiceNow 8 percent, Workday 10 percent. The companies investors had written off were posting the best agent numbers in the market.
The panic and the rebound are not opposites. They are two readings of the same shift, and the market spent 2026 lurching between them because almost everyone was handing down a single verdict on a category that splits cleanly in half. Here is where the line actually falls.
Here is what happened
The sell-off was indiscriminate, which is the tell. In the April trough Cloudflare dropped 12 percent in a session, Snowflake 9 percent, ServiceNow 7 percent, Salesforce 4 percent, and Palantir fell 17 percent across three trading days. Trade commentary put the total value erased across enterprise software at around two trillion dollars. Companies with almost nothing in common were repriced on one idea.
The proof text was a product launch. Anthropic's release of Claude Cowork in January 2026, able to run multi-step software workflows on its own, was treated as the demonstration everyone had been waiting for. ServiceNow's own guidance a week later, quietly acknowledging the risk, was read the same way.
One casualty was unusually clean. The commercial company behind Tailwind CSS saw revenue fall roughly 80 percent once AI models began writing its output for free inside anyone's code editor. Smaller than the rest, and the most legible: what it sold was the convenience of producing something a model now produces at no cost.
Then the incumbents reported. ServiceNow's second-quarter revenue grew 24 percent year over year, crossed one billion dollars in AI annual contract value, and showed agentic deployments up ninefold in nine months. Salesforce's Agentforce, the product investors assumed would eat the company from inside, reached 1.2 billion dollars in annual recurring revenue, up 205 percent.
Buyers had already said this. A January 2026 survey by Avenir found 63 percent of enterprise buyers expected their existing vendors to benefit from AI. Only 8 percent expected them to lose. The market spent five months pricing the opposite.
How it works
Moat one: data an agent cannot get anywhere else. A model is only as sharp as what it is allowed to see, and the most valuable enterprise data was never sitting in public. Veeva Systems, which sells to life sciences and competes with Salesforce in that market, argued on its fiscal Q3 2026 earnings call that it is not building a smaller version of what Salesforce is attempting but something resting on its Commercial Content and Crossix assets: decades of regulator-facing data no general model has a route to replicate. The same logic shelters Bloomberg and ZoomInfo. A horizontal AI platform can rent all the compute in the world and ship a capable agent into any industry by Tuesday. It cannot manufacture twenty years of structured data it never collected.
Moat two: being the thing nobody can safely remove. Andreessen Horowitz has a phrase for companies like NetSuite and Workday: they have hostages rather than just customers. Systems wired so deeply into finance, payroll and compliance that switching is not a preference but a multi-year risk. An agent can route around a clunky dashboard. It cannot route around the fact that pulling out the general ledger mid-financial-year is its own kind of disaster.
Moat three: someone to blame when it breaks. This is the layer nobody discusses and it may matter most. An agent can draft the clause, flag the anomaly, route the ticket. It cannot be the named, insured, auditable party a bank's compliance office or a hospital's risk committee points at when something goes wrong. That is not a gap a smarter model closes. It is a role only an accountable vendor with an audit trail and a support contract can occupy.
The incumbents are not resisting agents. They are metering them. SAP began routing third-party AI agents through a metered gateway in June 2026. Workday built an Agent Gateway that charges Flex Credits for autonomous work rather than for human seats. ServiceNow disclosed in April 2026 that half its net-new business no longer comes from seats at all. SAP's chief executive Christian Klein put it plainly in March 2026: it would be foolish to keep charging a flat subscription once AI can automate away much of what that subscription used to pay for.
Which turns the threat into a toll road. If every agent, yours or a rival's, has to pass through a governed gateway to reach the system of record, the gateway owner is not being disrupted. It is collecting on the disruption.
Why this is a frontier story
The casualty list fits the test almost too neatly. The companies hurt worst were, underneath everything, interfaces wrapped around commodity work: generating code, organising tickets, arranging shared documents. That is precisely what agents are built to be good at, stitching together visible, well-documented, structured tasks quickly and cheaply. When what you sold was mostly the interface for doing what an agent now does by calling an API, the reason to exist thins out fast.
The survivors are not claiming immunity. They are claiming a position. Veeva, ServiceNow, Salesforce, Workday and SAP each hold at least one moat: data nobody else has, a system nobody dares remove, or the gateway agents must pass through. None of them says AI cannot do the work. They say the work has to come through them.
Most of the market is in neither pile. The uncomfortable middle has some data and some switching cost and not quite enough of either. For those companies survival is not a coin flip, it is a race: get real proprietary data and real governance into the product before an AI-native challenger, unburdened by legacy pricing, ships something 80 percent as good for a fraction of the cost.
Pricing is where the shift becomes visible first. Agent gateways and AI control towers appeared in vendor roadmaps that had no such thing eighteen months ago. When a vendor stops selling seats and starts selling consumption, it has read the same memo the short sellers read and reached the opposite conclusion.
The lesson generalises past software. In any industry where AI can perform the visible task, value migrates to whoever holds the private data, the embedded position, or the liability. That is a template, and software is simply the first sector to be repriced against it in public.
The honest catch
This is an argued thesis rather than a measured finding, and the biggest complication runs the other way from the panic.
The agents doing the disrupting are failing at a rate that should give anyone pause. Gartner forecasts that more than 40 percent of agentic AI projects inside companies will be cancelled by the end of 2027, on runaway costs, unclear returns and governance nobody thought through. Its 2026 Hype Cycle places agentic AI at the Peak of Inflated Expectations, with only 17 percent of organisations having actually deployed agents against more than 60 percent who say they intend to within two years. That gap between intention and execution is exactly the condition that produces cancellations.
Much of what is being bought is not what it says on the label. Gartner estimates only around 130 of the thousands of vendors selling agentic AI have anything close to real agentic capability. The rest is what its analysts call agent washing: older chatbots and automation tools rebranded to ride the wave without the autonomy to match. Researchers at RAND Corporation studying AI project failure more broadly found the overall failure rate tops 80 percent, more than double that of ordinary technology projects, largely because organisations chase a striking pilot instead of doing the unglamorous data engineering and governance first.
The forecasts are forecasts. The 40 percent figure is a prediction, not a result, and Gartner is explicit that the technology still represents a genuine advance. Read it as a ceiling on how fast disruption can move through a real enterprise, not as a verdict on whether it moves.
The numbers behind the panic are softer than the panic was. The two trillion dollar headline comes from trade commentary rather than an exchange or index provider, market moves quoted here are point-in-time, and the revenue and ARR figures are as reported by the companies themselves. All of it is directionally consistent and none of it is audited for this purpose.
Three moats is a frame, not a law. It fits the 2026 casualty list well, which is a reason to take it seriously and also a reason to be careful: frameworks built to explain a past year tend to fit it better than the next one.
EDITOR'S TAKE
Everyone picked a side in 2026 and both sides were arguing about the wrong variable. The question was never whether agents are capable. It is what they are allowed to touch, and how many of them survive contact with production while everyone works that out. Those are two different clocks, and the 2026 crash and rebound are best read as two bets riding in the same share price: one on how fast agents commoditise the interface, the other on how slowly and expensively enterprises actually get agents into production. Do not bet on the category. Bet on the moat, which means building toward the governed layer rather than around it, or picking off the categories where the moat was always thinner than the market cap implied. The uncomfortable part of this thesis is that it makes the incumbents the toll collectors of their own disruption, and nothing about that outcome is obviously good for anyone buying software.
Quick questions
What was the SaaSpocalypse, in plain terms?
It is the half-joking name Wall Street gave to a sell-off in enterprise software shares that began on 29 January 2026, when the sector had its worst day since the COVID crash of March 2020, and deepened through early April. The reasoning behind it was simple. Most business software is sold by the seat, meaning a company pays a monthly fee for every employee with a login. If an AI agent can complete in seconds the work a person used to do all afternoon inside that software, the number of seats a company needs falls, and so does the revenue. Anthropic's launch of Claude Cowork in January 2026, which could run multi-step software workflows on its own, was treated as evidence the moment had arrived. By early April the sector ETF was down more than 21 percent for the year, with HubSpot, Atlassian and Figma 70 to 80 percent below their highs. Then the same companies reported strong agent revenue and the shares recovered sharply through July, which is why 2026 looks less like a verdict and more like an argument nobody has finished.
Why did agents damage some software companies and not others?
Because agents are good at the visible, well-documented, structured part of work and poor at everything surrounding it. If a product was essentially a convenient interface for doing something a model can now do by calling an API directly, the product's reason to exist gets thin, which is what happened to the business behind Tailwind CSS and, less severely, to parts of HubSpot, Atlassian and Figma. Three things protect a company from that. The first is proprietary data an agent has no route to obtain, which is Veeva's argument about decades of regulator-facing life sciences data. The second is being embedded so deeply in finance, payroll or compliance that removal is a multi-year risk rather than a preference, which is the position NetSuite and Workday hold. The third, and the most underrated, is accountability: an agent cannot be the named, insured, auditable party a compliance officer points at when something fails. None of the three depends on the vendor's AI being better than anyone else's, which is why the market's original test was measuring the wrong thing.
If agents are so disruptive, why do most agent projects fail?
Because capability and deployment are different problems, and enterprises are bad at the second one. Gartner forecasts that more than 40 percent of corporate agentic AI projects will be cancelled by the end of 2027, citing runaway costs, unclear returns and governance that was never designed. Its 2026 Hype Cycle places agentic AI at the Peak of Inflated Expectations, with roughly 17 percent of organisations having actually deployed agents against more than 60 percent intending to within two years, and that gap is where cancellations come from. Part of the problem is what buyers are sold: Gartner estimates only around 130 of the thousands of vendors marketing agentic AI have genuine agentic capability, and calls the rest agent washing. Broader work by RAND Corporation puts the failure rate for AI projects above 80 percent, more than double ordinary technology projects, largely because organisations chase an impressive pilot rather than doing the data engineering and governance first. None of this means agents are hype. It means the disruption has a speed limit set by enterprise execution, not by model quality.
Sources
Gartner: the forecast that more than 40 percent of agentic AI projects will be cancelled by the end of 2027, with Forbes and IHL Services on the 2026 Hype Cycle placement, the 17 percent deployment figure and agent washing.
Smart Supply Hub: secondary write-up of the RAND Corporation research placing AI project failure above 80 percent.
Yahoo Finance: the late-July session in which Salesforce rose 7 percent, ServiceNow 8 percent and Workday 10 percent.
Morningstar and Global Banking & Finance: what software earnings actually showed about the AI threat, against the share price reaction.
The Founders Report: Veeva's earnings-call argument that the vertical moat is proprietary data rather than agent features.
Rob Saker: the Andreessen Horowitz hostages rather than customers framing for deeply embedded systems.
Crown Point Advisory Group and GSPANN: the rewrite of seat-based pricing, and the agent gateway as a tollbooth on enterprise interoperability.
Chargebee: the Avenir survey finding that 63 percent of enterprise buyers expected incumbents to benefit from AI and 8 percent expected them to lose.
Long Angle, Digital Applied, Tech Insider and Big Ideas DB: trade coverage of the sell-off, the two trillion dollar figure and the moat debate. Treat the headline value as commentary rather than an index measurement.
Hashbyt: the Claude Cowork launch timing as the trigger event in the January narrative.
Frontier Signal explains frontier technology in plain English. This piece is an argued thesis, and company figures, market moves and analyst forecasts should be independently verified. Nothing here is investment advice.