The reason was that coding agents had stopped being a demonstration. Through 2025 they had gone from finishing your sentences to writing and shipping working software, with a person supervising rather than typing. In late January, Anthropic announced Claude Cowork, a desktop tool that took on multi-step office work: drafting documents, running compliance checks, handling the administrative substance of a job rather than a fragment of one. Software shares fell hard and immediately.
From there the reasoning ran in two directions, and both were bad for the industry. If an agent can do a knowledge worker’s job, companies need fewer knowledge workers, and business software is sold by the person. And if an agent can write software, then any product that was expensive to build has just become cheap for somebody else to copy.
The market seized on the fears and wasted no time repricing. By the end of February the main software index had lost almost a quarter of its value. It bottomed on 10 April, down 31% from where it started the year, two days after Anthropic began selling a hosted service for running agents at scale. Somebody called the episode the SaaSpocalypse and the name stuck.
Then it unwound. The index closed August slightly above where it opened the year. Nothing about software’s prospects had been settled by the argument itself. What changed in between is that the companies reported.
How software actually gets paid
Before getting into the earnings, it is worth being clear about how software companies actually make money. There are two ways to charge for software.
The first is to charge per person. A company with 500 staff who use the product buys 500 licences. This is how most business software is sold, because it is easy to explain to a finance department and it grows quietly as the customer hires.
The second is to charge per unit of work. Every query answered, every gigabyte moved, every message delivered. This is how infrastructure is sold, and how your electricity is sold.
Only the second is a meter. The first is closer to a lease. Companies sign annual or multi-year contracts for a fixed number of licences agreed in advance, and they cannot hand a few back in March because a team got smaller. The bill arrives whether the seat is used or not.
That distinction does a great deal of work later in this story.
The health of a per-person vendor shows up in two numbers, both reported quarterly. Whether existing customers spend more this year than last. And whether the order book, meaning contracted revenue not yet delivered, is still growing.
One claim, in two parts
Take the first of those two fears, the one about headcount, because it is the one the market actually traded. It has two legs, and they are worth separating because they met very different fates.
The first leg is about people. Agents do the work of employees, so companies employ fewer of them, so they renew fewer licences. Per-person vendors lose revenue from customers they have not lost. Retention drops, discounts appear, and the damage arrives first where price is bound most tightly to headcount, which means payroll and human-resources software before anything else.
The second leg is about machines. An agent works constantly and never sleeps, so it generates far more queries, traffic and storage than the person it stands in for. Everything sold by the unit gets busier.
Both legs describe the same event from opposite ends. Work moves from people to machines, so the leases shrink and the meters spin. That is one coherent forecast, and it is auditable.
By August the machine leg had been confirmed emphatically, and the people leg had not shown up at all. January’s other fear, the one about software becoming cheap to copy, dropped out of the argument after the spring. It is the one that did the damage.
What the summer actually showed
Take the machine leg first. Software sold by the unit accelerated across the companies we track that charge for consumption rather than headcount, revenue growth picked up from roughly 26% at the start of 2025 to about 35% by the middle of this year, and the gap between them and the per-person vendors is the widest it has been since 2022. The clearest single piece of evidence came from Cloudflare, which disclosed that for the first time in its history more than half the traffic crossing its network was not human. Every one of those requests is metered.
The share prices agree. Sorted by what the product charges for, businesses selling agent security and identity were up 84% at the median this year, and infrastructure sold by the unit up 49%. On the opposite end, per-person application software was down 20% and sat more than a quarter below its highs. A hundred points of dispersion inside one sector is not a market panicking about software. It is a market that has already sorted it.
Now the people leg. Workday sells human-resources software priced per employee. If agents shrink payrolls, this company feels it before anyone. Subscription revenue grew 14%. Margins went up rather than down. The order book grew faster than revenue, which is what a company looks like when demand is running ahead of what it has already delivered. And agents arrived on the revenue line: more than a quarter of the new business Workday signed that quarter was AI product.
Salesforce built the case against itself. Its Agentforce product automates customer-service work, which is work Salesforce charges for by the seat, and its chief executive cut several thousand of his own support staff to prove the product worked. Salesforce then booked new business at its fastest rate in four years. Customer losses ran near record lows. Seat counts were still rising.
Five9 charges per call-centre agent seat, and AI voice agents do exactly that job, which makes it the least ambiguous test in the sector. Its AI revenue grew 78%. Its seat count grew as well, and the revenue earned per seat went up.
Paycom and Paylocity charge per employee per month, which makes their filings close to a census of how many people their customers employ. Those counts grew. Retention improved at both. Paycom’s growth accelerated and its shares finished August at a one-year high.
Four companies in the people leg where seat destruction should have been visible, and it was absent in all four. The nearest thing to an exception is Asana, whose existing customers now spend slightly less than they did a year ago. Even that does not survive inspection. The figure is improving rather than deteriorating, Asana publishes no seat count at all, and the company attributes the pressure to renewal decisions in a weak economy rather than to agents. A small product losing share in a crowded market is an ordinary competitive problem wearing an agentic label.
One crack is worth recording, however. Workday guided its next-quarter order book below what analysts expected and framed the following year slower still. That is a real step-down, and it would be convenient to call it ordinary maturity in a ten-billion-dollar business. We cannot tell. Workday does not publish seat counts, so nothing in the filing separates a slower order book from customers renewing on fewer licences. It is the one place where the people leg crisis survives.
Why the people leg did not land
The work did move to machines. The forecast was right about that. What it got wrong was the timing of the money, for four reasons.
The first is the lease. A seat is contracted a year or more in advance, so a company can cut staff in May and still be paying for those licences the following spring. Cloudflare cut about a fifth of its workforce in May. Intuit cut 17% of its staff. Salesforce removed several thousand support roles. None of their software vendors saw anything, because nobody hands back licences mid-term. The gap between a headcount decision and the renewal conversation runs twelve to twenty-four months, and in mid-2026 the industry was still inside it.
The second is that the vendors sold the agents themselves. When a company decided to automate a workflow, the obvious supplier was whoever already held its data and ran that workflow. So the customer bought agents from its existing vendor. Workday booked agent products as more than a quarter of new business. Salesforce sold Agentforce to the same customers whose service teams it was shrinking. The job was automated and the vendor was paid for automating it.
The third is that the work did not disappear, it multiplied. An agent handling a task a person used to handle makes far more calls to far more systems, at machine speed, around the clock. That is the machine leg again, and it lands on the same industry.
The fourth is where the money came from. Companies did not fund agent projects by cutting the software budget. They funded them out of labour budgets, and the software line grew.
Put those together and 2026 makes sense. The industry did not defeat the forecast. It absorbed it, on terms set by contracts signed before anyone had heard of an agent, and by vendors quick enough to sell the replacement for their own product.
The company that shows the real risk
None of which explains Intuit, whose shares fell 47% this year, the worst drawdown in this study, while its actual quarter beat on both revenue and earnings.
This is January’s other fear, the one that dropped out of the headlines.
Intuit is usually described as protected because it owns a ledger and a tax filing workflow. That is the wrong reading of its moat. What Intuit really owns is thirty years of engineering that encoded the tax code and accounting rules into working software. Tax law is enormous, fiddly and changes every year. Turning it into a product that millions of people can use without understanding any of it took decades and an army of developers. The barrier protecting Intuit was never the customer relationship. It was the cost of building that software.
Coding agents attack precisely that barrier. The cost of writing large amounts of careful, tedious, rule-following software has fallen faster than the cost of almost anything else in this cycle. A competitor that would once have needed a decade and a fortune to encode the same rules can now do it in a fraction of both.
The results say this is already happening. Intuit guided next year’s profits about 16% below what analysts expected, said tax unit growth would run at 2–3%, and told investors it had lost do-it-yourself customers to cheaper providers because its pricing had become hard to defend. Its answer was to launch a free tier of QuickBooks and to partner with OpenAI. Both are price responses.
So customers did leave, but not to an agentic replacement. They left for a rival whose product only exists because agents made it cheap to build. Any company whose defensibility rests on how hard its product was to build is exposed to the same erosion, and a great deal of business software was defended on exactly that basis.
What would change this quickly
Nothing here is settled, and the things that would move it are visible.
The lease runs out. The headcount cuts of 2025 and 2026 reach their renewal dates through 2027, and that is the first moment the people leg gets a fair test. If retention at the per-person vendors still holds a year from now, that argument is finished. If it slips, everything above was a timing effect.
The build barrier keeps falling. Intuit’s problem was a competitor made viable by cheap code. Every quarter that coding agents improve, the same door opens for someone else’s incumbent.
The infrastructure layer gets bundled. The per-unit vendors are winning because agents run on their pipes, and the companies selling the agents would rather sell the pipes too. Anthropic’s hosted runtime is the first move in that direction, and the businesses that fell hardest in April were precisely the ones whose function it absorbs.
Agents get write access. Today an agent drafts and a person approves, inside the vendor’s product. The systems-of-record moat holds because of that last step. It ends the day agents are trusted to file the return and post the entry themselves.
And the crossover arrives. A per-person vendor stops being exposed to the people leg entirely once its metered revenue passes its seat revenue, because at that point it is paid for the work rather than the worker. ServiceNow is furthest along, with roughly half its new business already arriving on a non-seat basis. Salesforce is further back: hold its current growth rates, with the seat core compounding at 8% and the data business at 20%, and the two lines cross in about six or seven years. That turns an existential question into an arithmetic one.
The market spent the first four months of 2026 pricing the perceived disaster and the next four unwinding it. As shown in the latest earnings, agents are not taking the software. They are taking the difficulty in developing sophisticated software, and a large part of this industry has spent thirty years charging for difficulty.
This piece extends the framework in “Where Does the AI Build-Out Top First”, which argued that a build-out this large tops in stages rather than on a day.