Is SaaS dead? What happens to software when AI can generate it

Is SaaS really dead, or is that a headline?

SaaS as a delivery model is not dead: renting software over the internet is still how every business runs. What is dying is a specific product shape, the static workflow sold per seat, because a model can now generate that workflow on demand.

The debate started in December 2024, when Satya Nadella said on the BG2 podcast that business applications are essentially databases with business logic on top and that "that's probably where they'll all collapse, right, in the Agent Era", as Cloud Wars reported with the timestamps. He never said "SaaS is dead". He said the logic moves to the agent tier.

I agree with the precise version. Nobody is regenerating a general ledger at runtime, and nobody should. But the invoicing tool that only exists because someone packaged 12 steps into screens 5 years ago is now competing against a version of itself that the customer can rewrite on the spot.

What did the February 2026 SaaS sell-off actually show?

The sell-off showed that investors now separate software that holds durable state from software that packages knowledge work, and that the second category can lose 25% of its value in 6 weeks.

The sequence was fast. Anthropic introduced Claude Cowork on 12 January 2026 and released 11 open-source plugins on 30 January. According to DeepLearning.AI's The Batch, the S&P Software and Services index lost 25% of its value between 12 January and 23 February; on 4 February alone the index fell 4%, wiping out $285 billion in market capitalisation. LegalZoom fell nearly 20% and Thomson Reuters 16%.

Notice what fell. Legal templates, financial data terminals, task-level knowledge work. Notice also that the rebound on 24 February came when Anthropic announced integrations with Docusign, FactSet, Intuit and Salesforce, not replacements. The market's message was not "software is worthless". It was "software that an agent can do from the outside is worth less, software that an agent needs to connect to is worth the same or more".

Which SaaS survives: what makes software durable?

Software is durable when it holds state that matters, is shared across an organisation, has been hardened by years of edge cases and sits in the path of money, data or compliance. Payments, data warehouses, observability, identity, source control and communication rails fit that definition.

The reason is simple: an agent needs somewhere to keep the truth. It can generate a report, but the numbers have to live in a system that many people and many agents trust at the same time. Regenerating that system for each user would destroy the only thing that makes it valuable, which is that everyone sees the same state.

There is a second durable category that is easy to miss: legacy systems. The 20-year-old ERP with millions of undocumented lines is not getting rewritten by a model. Legacy becomes infrastructure, and agents reach it through whatever door is open: an API, a command line or the browser.

Which SaaS dies: why predefined workflows lose?

Predefined workflows lose because their whole value was deciding the steps in advance for you, and a model can now decide those steps at the moment of use, for your exact case, at almost no marginal cost.

Think of the mid-market tools most companies pay for: a form builder, a reporting layer over a database, a template library, a project tracker with a fixed methodology. That bet paid for 15 years and stops paying when the customer can describe the sequence they want and get working software in a minute.

Gartner quantified the exposure in July 2026: $234 billion of enterprise application software spending is at risk from agentic AI between now and 2030, and price adjustments will account for approximately 20% of enterprise SaaS spending by 2030, according to CIO Dive. The Gartner analyst's phrase is the one to remember: agents break the link between user growth and revenue growth. If your revenue only grows when the customer adds a seat, and the customer's agent does the work of 5 seats, no feature release fixes that.

What is ephemeral software and why is it rising?

Ephemeral software is software created at runtime, for 1 person, for 1 task, and discarded afterwards: a view, a transform, a one-off tool, a workflow that exists for the length of a single job. It is the second half of the split, and it is where most new software will live.

The economics come from the Jevons paradox. When the cost of producing a tool collapses, the world does not end up with fewer tools. It ends up with orders of magnitude more, most of them made for 1 person and 1 afternoon. TechCrunch reported in March 2026 that Lovable users were building about 200,000 projects a day, with peaks above 500,000. Most of those will never be products. They are software as a consumable.

The clearest signal that this is not a fringe idea is who is shipping it. Cloudflare open-sourced Cloudflare OS under Apache 2.0 in August 2026, the platform its own staff had used since May. Its unit of software is a "Gadget", a private instance of an app created for 1 user, who can change the code with an agent instead of filing a feature request. The repository states that in the AI era "the centralized model of software stops making sense". A company with one of the clearest views of how the web is used built per-person software as its answer, and gave away the source.

How is SaaS pricing changing?

SaaS pricing is moving from seats to usage and outcomes, because agents do the work that seats used to measure. IDC predicts that by 2028 pure seat-based pricing will be obsolete and that 70% of software vendors will refactor their pricing around consumption, outcomes or organisational capability.

The logic is unavoidable. If a support tool charged per human agent and an AI agent now resolves 2 out of 3 tickets, the vendor's revenue falls while the value it delivers rises. The only way to capture that value is to charge for the resolution, the transaction or the outcome.

In my companies we made this change over the last 18 months. Every new product prices on the unit of value the customer actually receives, an invoice processed, a lead qualified, a report delivered, with a floor that covers our costs.

What should a SaaS founder do now?

A SaaS founder should move the product's value into state, data and outcomes, price on results, expose the product to agents through an MCP server and let customers reshape the workflow, instead of defending the fixed one.

I run +10 companies through this transition and I am building +100 more with Aurum VOS, my lifelong project, so this is the list I actually use. It requires admitting which half of your product was infrastructure and which half was a workflow someone else can now generate.

Software is not dead. The predefined workflow is. The founders who accept that this quarter will own the infrastructure that every agent, and every ephemeral tool, needs to connect to.

• Find the state: make the data and records only you hold the core of the product.
• Price on outcomes: resolutions, transactions or results, with a cost floor.
• Ship an MCP server: let agents read and act on your product.
• Let customers reshape the workflow: expose building blocks, do not lock the sequence.
• Stop investing in screens that describe steps; invest in the systems that hold truth.