SaaS is dead: five claims, checked

Five claims, one headline

Most arguments about whether SaaS is dead go nowhere because the two sides are discussing different propositions. Pulled apart, the headline contains at least five.

1. 1Seat-based pricing is finished, because agents do the work a seat used to represent.
2. 2AI agents will replace the application layer, leaving databases and a model tier.
3. 3Companies will build their own tools instead of buying, because building got cheap.
4. 4The market has already priced the collapse in, so the collapse is real.
5. 5This has happened before, and the incumbents lost that time too.

Where the phrase came from

The death of SaaS is not a new idea, but it acquired a date, a price, and a nickname between December 2024 and February 2026. The sequence matters, because the strongest version of the argument is about pricing models and the loudest version is about a stock chart.

What the market actually did

Start with the number, because the number is where this argument gets its authority. You have probably seen the selloff quoted as a trillion dollars, or two. Here is the awkward part: those figures describe the same event, measured four different ways, and they do not agree.

The two Jefferies figures are the ones with a stated index and a stated window, so they are the ones worth quoting. The trillion-dollar numbers circulate without a named index, and the larger of the two appears mainly on aggregator sites citing each other. A reader who repeats "$2tn" is usually repeating a chain with no original at the end of it.

What a repricing is, and is not

A share price is not a reading of how a business performed. It is a bet on how it will perform, expressed as a multiple of what it earns now. When a sector falls 25% while its revenue rises, the market has not observed customers leaving. It has revised what it will pay for a dollar of subscription revenue in a world where agents exist.

That is a real and important event. It is also a forecast, and forecasts by large numbers of people are wrong regularly enough that the distinction is worth holding on to.

The forecasts underneath it

• ~70%, share of software vendors expected to move off pure seat pricing (by 2028). IDC
• $318bn to $576bn, global SaaS spending, forecast to keep rising through the selloff (2025 to 2029). Gartner

Those two forecasts sit oddly together, and the tension is the most useful thing on this page. One says the way software is billed is about to change completely. The other says the amount spent on it nearly doubles. Both can be true. Neither is "SaaS is dead".

Claim 1: seat-based pricing is finished

This is the strongest version of the argument and it has serious analysts behind it. The logic is clean: a seat prices access for one human. If an agent does the work, the number of humans stops tracking the value delivered, so the meter is measuring the wrong thing. IDC has forecast that per-seat pricing will be structurally obsolete by 2028, with about 70% of vendors moved to consumption, outcome, or capability pricing.

We think that is directionally right. We also think it is a claim about 2028, and that almost nobody making it in the present tense has checked what vendors charge today. We tried to check, using our own CRM pricing dataset, and could not publish the result. Why is in Method, and what we withdrew, and it is worth reading before you trust anyone else's number on this.

Claim 2: AI agents replace the application layer

If agents were dissolving applications, the applications would be turning into agents. That is broadly what is happening, and it is the part of the argument that looks most right. Nearly every CRM now ships something described as an assistant, and a smaller group ships something described as an agent.

The interesting question is how it is sold. Across the vendors we track, AI capability tends to arrive above the entry plan rather than replacing it: it is positioned as a reason to move up a tier, not as a reason to need fewer seats. That is close to the opposite of a technology eating its host, and it is what you would expect at this stage of a transition, when the new capability is a premium rather than a default.

We have deliberately not put a percentage on that here. Our own dataset was not clean enough to carry one, and a number we cannot stand behind is worse than a claim we can describe.

Claim 3: companies will build their own instead of buying

This is the practical conclusion most people draw. If a model can write the code, why rent the software? It is the claim with the most anecdotes and the least data behind it, and the anecdote it always rests on is Klarna.

Worth reading that story to the end. Klarna announced it was dropping Salesforce and Workday in early 2025. By May the CEO had said the cost-cutting went too far and human staff came back. Reporting from CX Today indicates that what replaced Salesforce and Workday was substantially other vendors' software, alongside in-house work, rather than a pure build. The most-cited proof of leaving SaaS is partly a story about switching SaaS.

Claim 4: the market has already priced the collapse in

It is the argument that feels most like evidence, because it comes with a chart. HubSpot is the clean test, because both numbers are public and they point in opposite directions.

Claim 5: this has happened before

It has, several times, and the pattern is consistent enough to be useful. Each wave was directionally right about a layer and wrong about the category.

The lesson is not that incumbents always survive. Plenty did not. It is that the layer that moves is the interface, the layer that stays is the record, and the billing model changes last, usually several years after everyone has declared it already changed.

What is actually dying

Dismissing the whole argument would be as lazy as accepting it. Three things in the SaaS model are genuinely in trouble, and none of them is the subscription.

Where we stand, since we have a stake in this

Lumenbase is a CRM, sold by subscription, priced per seat. We are not a neutral party and it would be dishonest to write this page as if we were. Our position is that the seat is a fair way to bill for access and a poor way to describe value, that the record should maintain itself rather than wait for someone to type into it, and that a vendor whose main defence is the cost of leaving has already lost the argument. You can check the last one: our migration tooling is built to move data in, and the API and MCP surface is built so an agent can read and write the record without a human in the form.

What to do if you buy software

The useful output of this argument is not a prediction, it is a shorter list of questions to ask a vendor. These five separate the tools that are adapting from the ones repeating the word AI.

1. 1Can an agent do what a seat does? If the product only works through its own interface, an agent cannot use it. Ask for the API and MCP surface, and ask what share of the product it covers.
2. 2What happens to the bill when headcount falls? If the answer is "you pay less", the vendor has accepted that seats price access, not value. If the answer is evasive, you are the hedge against their revenue model.
3. 3What does the record cost to maintain? Count the hours your team spends typing into it. That number, not the licence, is usually the larger line and the one AI actually changes.
4. 4How much does leaving cost? Ask for a full export including custom fields, history, and attachments. A vendor confident in the product will quote it in hours.
5. 5Which AI features are real, and on which plan? AI capability is usually sold above the entry tier. Ask which plan you actually need, and ask to see the feature working on your own data before you upgrade for it.

Our CRM evaluation guide covers the rest of the shortlist process, and the cost calculator runs the seat maths across the CRMs we track.

Method, and what we withdrew

Everything numeric on this page is someone else's reporting, attributed in the line where it appears. We have not independently verified any of it, and where sources disagree, as they do about the size of the selloff, we have shown the disagreement rather than picking the most dramatic figure.

The study that is not here

This page was originally built around our own measurement: we track public CRM pricing, and the plan was to test "the seat is dead" against what vendors actually charge. We audited that dataset before publishing and it did not hold up. Three problems, each of which would have put a false statement on this page:

• A field we read as "publishes no price" actually means "no price resolved for the band we compared". Four vendors we would have named as secretive publish openly, and our own database holds their prices.
• The read date was the newest date across all vendors, not a date they shared. Some rows were five weeks older than the date we would have printed.
• Our count of vendors using consumption pricing came from a field that can only hold three values, none of which is consumption. The zero was guaranteed by the schema, not found in the market.

One further problem we claimed in an earlier draft turned out not to be real: we thought the per-seat price column mixed units in a way that spoiled the median, and it does mix units, but the code already normalises the affected vendors, so that figure was sound. An audit that overstates its findings is its own kind of failure, so it is recorded here rather than quietly dropped. The three above stand, and the first and third are disqualifying on their own.

We are saying this out loud for two reasons. It is the honest explanation for why a page about pricing evidence contains no pricing statistics from us. And it is a fair warning about everyone else's: the confident percentages circulating about seat pricing are drawn from datasets with the same failure modes, assembled by people who did not publish their method. When the cleanup is done we will publish the study, with the per-vendor read dates attached.

Questions