Is AI a SaaS killer? The 2026 numbers

"SaaS is dead" has been said at almost every software conference for two years now.

It is worth checking, because people are making real decisions on the back of it. Founders are changing roadmaps. Buyers are delaying purchases. Investors are marking software down.

The short answer: AI is killing one kind of software business and making another kind stronger. Which one you are in is something you can work out yourself, today, in about an hour.

What this article covers

1. Why people say AI is killing SaaS. The three arguments, and who is making them.
2. Why it is not the full story. What the numbers actually show.
3. Six questions to ask about your own product, so AI works for you instead of replacing you.

Part 1. Why people say AI is killing SaaS

Three separate arguments get mixed into one sentence. They are not the same argument, and it helps to keep them apart.

The three arguments

• Pricing. Most software is sold per person. If AI means a company needs fewer people, it buys fewer licences. Same customer, less money.
• The app layer. Satya Nadella put the argument on the BG2 podcast in December 2024. Business applications, he said, are "essentially CRUD databases with a bunch of business logic", and in the agent era that logic moves up into the AI layer. His point: a CRM is a database with rules on top. If an agent can read the database and apply the rules, you may not need the app.
• Copying. Building software got very cheap. What took a small team three months can now take one good developer a weekend. So what stops anyone copying your product?

What people point to

• Klarna. In 2024 its chief executive said the company had shut down Salesforce and would shut down Workday.
• Software valuations fell hard. On Aventis Advisors' running measure of public software multiples, companies were valued at 18 to 19 times revenue through most of 2021. As of March 2026 the median was 3.4 times, a fall of about 82%.
• HubSpot grew 20% in its second quarter of 2026 and still trimmed its outlook for the year.

Read that list quickly and the case looks closed. It is not.

Part 2. Why it is not the full story

The pricing argument is right

This one holds up. Selling per seat assumes a company's work grows with its headcount. AI breaks that assumption.

If a support team of forty handles the same work with twelve people, the vendor loses twenty-eight licences without losing the customer and without doing anything wrong.

So vendors are changing how they charge, to usage, to credits, sometimes to results. HubSpot's trimmed outlook is partly this. The company pointed to its own move to trial-led, outcome-based pricing, which lengthens the buying process, alongside tighter budgets and longer sales cycles.

But note what that is. A company changing its price list is not a company dying. It is a repricing project and a difficult quarter.

The "apps will collapse" argument is early, not wrong

For an agent to replace an app, two things must be true.

• It has to reach the data. Most company data sits inside vendors who have no reason to make leaving easy. That is improving, but "an agent can read it" and "an agent can read it with the permissions, audit trail and reliability a finance team accepts" are years apart.
• Someone has to be accountable. This is the part the argument skips. An app is not just screens. It is where responsibility lives: who approved this, who changed that, when, and under what rule. Regulators ask those questions of people, not of models.

Klarna is the example everyone quotes and nobody finishes

Klarna did drop Salesforce and Workday. What it moved to is the part that rarely gets repeated.

It did not replace them with an agent working on raw data. It moved to other software vendors, with Deel taking over the HR work, alongside its own consolidated internal data layer. Siemiatkowski said so himself in March 2025: "we did not replace SaaS with an LLM".

Then, in May 2025, Klarna reversed its customer service automation and started recruiting people back, after deciding quality had suffered.

The lesson is not that AI does not work. Klarna's assistant handled enormous volume and still does. The lesson is narrower, and it is the most useful sentence in this article: automating a task and replacing a system of record are different projects. The company held up as proof that agents replace software ended up buying different software.

The copying argument has the most surprising answer

Building did get cheap. If cheap building were enough, the new AI products flooding the market would be taking customers and keeping them.

They are not keeping them. ChartMogul measured retention across roughly 3,500 software companies through 2025:

• AI-native products kept 48% of their revenue base over a year. Established B2B software kept 82%. Both figures are net revenue retention, so they are measured the same way.
• The cheapest AI products did worst by a distance. Those under $50 a month held on to 23% of their revenue.
• The more expensive ones did not. AI products above $250 a month performed roughly in line with established business software.

That middle number is the whole story. Cheap AI products attract people who are trying something out, and trying something out ends.

It is worth adding the part that cuts the other way, because it is the more interesting trend. AI-native retention improved through 2025, from 27% in January to 40% by September on a gross basis. The early triallists churned out and the committed customers stayed. Some of that is which companies were being counted rather than the same companies getting better, so do not lean on it too hard. But it is a market maturing, not a market collapsing.

Building the product was never the hard part. The hard parts were distribution, trust, support, integrations, compliance, and slowly becoming the place where the data lives. AI has made none of those easier. It removed the one moat that was never really a moat and left every other moat standing.

What the large software companies actually reported in 2026

All of these are from the companies' own quarterly results, not from anyone's summary of them.

• ServiceNow. Subscription revenue up about 25%, and its AI products passed $1 billion in annual contract value.
• SAP. Current cloud backlog up 26% at constant currency.
• Adobe. AI-first recurring revenue more than tripled year on year and passed $500 million, with full-year targets raised.
• Workday. Subscription revenue still growing, with the profit margin outlook raised.
• HubSpot. 20% growth, outlook trimmed during its pricing change.

One more thing about that 82% fall in valuations. Most of it happened before agents were a serious topic. Interest rates rose in 2022 and repriced every loss-making growth company in the world, and software growth had already slowed. AI fear is the newest layer on a repricing that was four years old before it arrived.

So what is actually dying

Not SaaS. A part of it.

At real risk

• Thin wrappers around a model, meaning a prompt, a form and a subscription. The model provider will ship that as a feature.
• Products whose value was the interface. If the job is "make this data easier to look at", an agent can look at the data directly.
• Per seat software sold to roles that are being consolidated.
• Single-feature tools bought on a card and forgotten. This is where the 23% lives.

Getting stronger

• Systems of record. Whoever holds the true version of the data is not replaced by something that reads it.
• Software with audit or regulatory weight. Someone must be accountable.
• Deeply integrated workflow software. The cost of leaving is not the licence. It is fourteen integrations, trained staff and four years of history.
• Industry-specific software with data or process a general model does not have.

Part 3. Six questions to ask about your own product

Answer these honestly and you will know which side of the line you are on. Most can be answered from data you already have.

1. If a customer's own AI agent had full access to your database, what would they still need you for?

If the honest answer is "the screens", there is a problem. If it is "the integrations, the audit trail, the fact that the data is only correct because we keep it correct", the position is solid.

2. What share of your revenue is per seat, and who sits in those seats?

Count the licences held by roles most exposed to automation. That is the size of your exposure, and you can measure it this week.

3. Are you the record, or a view of the record?

Views get replaced. Records get inherited.

4. How long does a customer's data have to live in your system before leaving becomes painful?

If leaving is never painful, your pricing power will keep sliding, with or without AI.

5. Is your churn concentrated in your cheapest plan?

If yes, the problem is who you are selling to, not the product. The cheapest tier is often a marketing cost wearing a revenue costume.

6. What does AI in your product let a customer stop doing?

"AI-powered" on a feature list does very little. One specific task the customer no longer performs does a lot. If nobody can name the task, it is marketing.

Using AI as leverage instead of waiting to be replaced

The pattern in the numbers is simple. The companies doing well did not add AI as a feature. They used AI to do more of their customer's job.

Three practical moves follow from that:

• Go deeper, not wider. Own more of one workflow rather than adding another feature. Depth is what makes leaving expensive.
• Become the record. Every piece of data a customer trusts you to hold is worth more than a feature you ship.
• Charge for the result, not the seat. If your customer needs fewer people because of you, your price should not fall when that happens.

The short version

• Seat-based pricing is in real trouble. That part of the story is true.
• Systems of record are safer than they were, not less safe.
• Building software got cheap. Being trusted did not.
• Most products dying loudly in the data were cheap experiments that were never going to last. AI did not kill them. It just made it easy to launch thousands of them.

AI is not a SaaS killer. It is a very good test of whether a piece of software was ever worth paying for.

Where we come into this

We build and maintain software products for companies in Europe, Japan and the Gulf, and we build the AI systems that sit inside them. A good part of our work right now is helping founders answer question one on that list honestly.

If you want a straight second opinion on your own product, book thirty minutes. No presentation.