AI is coming for software companies

Enterprise software stocks crashed in early February, and specialist software ETFs fell deep into correction territory. The trigger was the sudden commercial launch of advanced agentic AI tools for developers, which left investors doubting the whole software-as-a-service (SaaS) business model. The logic of the sell-off pointed at the application vendors, the companies that sell a login to each employee, rather than at the suppliers of the computing behind the new tools.

Fewer seats to sell

For two decades software companies were valued on one comfortable measure: recurring revenue from per-seat licences, earned at eighty per cent gross margins. Autonomous AI agents break that model. If one software agent can do the customer service, sales calls or code upkeep of ten employees, customers need far fewer seats. The model took hold at the turn of the century, when Salesforce began selling customer-management software as a subscription delivered over the internet, and by the 2010s investors valued the sector on revenue multiples that assumed customers would buy more seats every year.

The threat is no longer hypothetical. In February 2024 Klarna, the Swedish payments company, said its AI assistant was handling two-thirds of its customer-service chats and doing the work of 700 full-time agents. Whether or not that figure holds up over time, it is the kind of claim that makes a procurement department count the logins it pays for, and every seat not renewed comes straight off a vendor's recurring revenue.

How the valuation unwinds

The damage travels through one number above all: net revenue retention, the share of last year's revenue that existing customers pay again this year, including upgrades. A healthy SaaS firm counts on retention above 100 per cent, because customers hire more staff and add seats without any extra sales effort. If agents let a customer-service department shrink, the seat count shrinks with it, retention falls below that line, and a company that looked like a compounding annuity starts to look like a business that must win new customers just to stand still. Multiples built on the first assumption do not survive the second.

The sector has been through one derating already. When interest rates rose in 2022, software stocks fell hard because their value rested on profits far in the future, and a higher discount rate shrank them. That episode hit the multiple while leaving the business model intact. This one questions the growth itself, which is harder to recover from, because a cut in rates cannot restore seats that customers no longer need.

A costly change of model

Software groups trying to switch to charging by usage or by results will see margins squeezed for years while they make the change. Corporate customers are already cutting the number of vendors they use and cancelling overlapping SaaS subscriptions to free up money for AI computing. Usage pricing also brings a cost that seat licences never had: every task an agent performs consumes inference computing, which the vendor must buy from a cloud provider or model developer, so gross margins on agent revenue are structurally lower than on a login.

There is a precedent for such a transition, and it is not comforting for the short term. When Adobe stopped selling boxed copies of its creative software in 2013 and moved customers to monthly subscriptions, reported revenue and profits dipped for a period because upfront licence sales were replaced by smaller recurring payments. The change paid off in the end, but it took several years, and Adobe had the advantage of dominant products and no rival offering the same work more cheaply.

The move from installed software to the cloud offers a harsher lesson. Siebel Systems dominated customer-management software around 2000, then lost ground to cheaper subscription rivals and was sold to Oracle in 2006. Incumbency protected it for a while, but it did not protect its pricing, and the new entrants won by charging less for the same job.

Winners and the exposed

The likely winners are the model developers and cloud providers that sell the computing, and the software firms that own a system of record, such as the databases and enterprise resource planning suites where a company's core data lives, since agents still need somewhere authoritative to read and write. The most exposed are tools priced per head for functions that agents can take over, such as customer support desks, sales-development software and basic coding aids. Lenders are exposed too, because private-equity firms bought many mid-sized software companies with heavy debt on the assumption that their subscription revenue was as dependable as rent.

That last group deserves more attention than it gets. Many of those buyouts were financed by private-credit funds that lent against recurring revenue rather than current profit, on the theory that customers almost never cancel mission-critical software. If retention slips below 100 per cent, the collateral behind those loans shrinks at the same time as the borrower's cash flow, and the losses would surface first in the valuations of private-credit funds, which are marked less often than listed shares.

The incumbents' defence

The strongest counterargument is that incumbents control what agents need: customer data, integrations, security approvals and procurement relationships that take years to build. Salesforce began charging per conversation for its Agentforce agents in 2024, and others are experimenting with similar pricing, so the leaders may capture the new revenue rather than lose it. Most vendors' reported results have yet to show a collapse in seats. That defence is real, but it points to a smaller and lower-margin industry rather than a return to the old valuations.

Contract structure also slows the damage. Enterprise licences typically run for one to three years, so a customer that needs fewer seats can cut only at renewal, and the effect reaches reported revenue gradually. That buys incumbents time to reprice, but it also means the market will be judging them for several years on forward indicators rather than on revenue that has already been booked.

What to watch in results

The figures to watch in coming earnings are net revenue retention, seat counts where companies disclose them, remaining performance obligations, and gross margins on newly launched agent products. AI is starting to eat the software industry. Subscription businesses that once commanded rich multiples will probably be priced more like suppliers of cheap digital labour, and the rerating has further to run.