When SaaS can be rebuilt

For twenty years the bargain was simple. A company rented software it could never have afforded to build, paid for every person who used it, and adapted its work to the product. AI has changed the first part of that bargain. An internal tool, a client portal or a workflow that once took a team months can now be specified, generated and running in weeks, and it can fit one industry, or one company, more closely than a product built for thousands of customers.

That has started a loud debate about the end of SaaS. Here is what was said, what the markets did, what is changing in pricing, and what remains hard to copy.

The argument

The sharpest version came from a company that sells business software. In December 2024, on the BG2 podcast, Microsoft’s Satya Nadella described business applications as “essentially CRUD databases with a bunch of business logic”, software that creates, reads, updates and deletes records. “The business logic is all going to these agents,” he said, and the applications would collapse “in the Agent Era”.

Klarna’s chief executive, Sebastian Siemiatkowski, had spoken on an investor call about dropping Salesforce, and many read that as AI replacing software. In March 2025 he corrected the reading: “So no, we did not replace SaaS with an LLM, and storing CRM data in an LLM would have its limitations.” Klarna had built its own stack around a graph database and used AI coding tools to put new interfaces on it quickly.

The idea has since become a sales pitch. Ema, which raised 77 million dollars in September, first wraps itself around a company’s existing applications. Its chief executive, Surojit Chatterjee, told TechCrunch that for some customers the large SaaS products “are mostly becoming like a database”.

What the markets did

Investors took the argument seriously this year. On 12 January Anthropic released Claude Cowork, an agent for office work. By 20 January a basket of SaaS stocks tracked by Morgan Stanley had fallen 15 percent, after 11 percent in 2025, and traded at 18 times expected earnings, the cheapest on record. When Anthropic added plugins for legal work, sales, finance and other functions at the end of January, the selling spread. The S&P Software and Services index lost 25 percent between 12 January and 23 February, and the Jefferies strategist Jeffrey Favuzza gave the episode its name: the SaaSpocalypse.

Then the mood turned. On 24 February Anthropic announced integrations with Salesforce, Intuit, Docusign and FactSet, connecting its agents to their products rather than going around them. By October, JPMorgan’s Samik Chatterjee counted the software group and the IGV index up 21 percent since the end of June, against 4 percent for the S&P 500, with Workday, Salesforce and Docusign among the leaders. He said the core businesses range from “modestly decelerating to stable growth”.

Neither swing is a verdict. The markets first priced the end of software, then its survival. What they show is that nobody yet knows which products agents will replace and which they will run on.

Seats under pressure

The clearest change is in pricing. A price per person makes less sense when an agent does part of the work those people did. Bain put it plainly in September 2025: “Seat-based pricing may not fit when AI is doing the work.” Its advice to vendors was to price for outcomes rather than for log ins.

Vendors now sell both. Salesforce offers its Agentforce agents at 2 dollars per conversation, through Flex Credits at 500 dollars per 100'000 credits, with a standard action costing 20 credits, or as add ons from 125 dollars per user and month. Intercom charges 0.99 dollars per outcome for its Fin agent, next to human seats from 29 dollars. Ema charges neither by seat nor by tokens, but by completed tasks and business outcomes.

For a buyer this is mixed news. Paying for results can be fairer than paying for log ins. It is also harder to forecast, and every vendor counts a result in its own way.

What is hard to rebuild

Code is now the cheap part. What a product has gathered around its code is not. Bain locates the edge of the systems of record in “unique data structures, long histories of activity, and built-in regulatory logic” that would be costly for others to copy.

Data is the first moat. A CRM or an ERP holds years of history and links to many other systems, and moving all that safely is a project of its own. Network effects are the second: a marketplace, a payments network or a platform that partners build on is valuable because others are on it, and no agent can generate the other side.

Compliance is the third. Payroll across countries, tax rules that change every year, audits, certifications and liability for errors are maintained work, not features, and a company that rebuilds such software takes that work on. That is why even Siemiatkowski doubted that all companies would do what Klarna did, and expected fewer SaaS vendors to consolidate the market instead.

What it means for buyers

For a company the question is no longer build or buy, but where to do which. In our view it helps to sort the software you pay for into three groups.

The first is generic tools: forms, simple workflows, internal dashboards and portals that are, in Nadella’s terms, a database with some logic, paid per seat. These are the easiest to rebuild to measure, and often the cheapest to replace.

The second is the systems of record, with your data, your compliance and your counterparties inside. Keep them, connect agents to them, and renegotiate the pricing as agents take over work. The third is where your company differs from everyone else in its industry. That is where a product built for thousands of customers fits worst, and where software of your own pays back most.

Two rules hold in every group. Own your data and make sure it can leave. And software you build is software you run: someone has to secure it, update it and answer for it.

This is the work Werk was made for. It is the platform we build our own products on, with the parts every app needs made once and tested, and we build clients’ apps and processes on it with them. AI writes most of the code, people decide what to build and review every release, and the code belongs to the client.