Stop Asking If SaaS Dies. Ask Which.

An expert in M&A outside the software industry had a theory I had reservations about.

His ten-year bet on software was that companies would stop buying SaaS. AI had made building in-house cheap enough to become the new default. Buyers would become builders. The market would shrink from the inside out.

I consult with companies trying to adopt AI, and that theory does not survive what I see.

AI made parts of software cheaper. Writing functions, scaffolding apps, and spinning up integrations are easier than they were two years ago. But the expensive part of software was never only the code. It was the organisational debt around it: the structure that dictates processes, the processes that dictate tooling, and the ownership questions that make every adoption curve slower than the demo suggests.

So the question is not whether SaaS survives AI. It is which SaaS survives AI. That is a more useful question, because it points to the categories, moats, and buying decisions that still compound when software itself becomes easier to produce.

”SaaS is dying” is a category error

Narratives beat data when the narrative is cleaner.

“SaaS is dying” is clean. It is short, confident, and familiar enough to sound plausible. The problem is that it reads a compositional shift as a collapse.

Technology markets rarely die in the clean way the headline suggests. Retail was supposed to die when e-commerce scaled. Amazon would hollow out every mall. Main Street would disappear. What happened instead was recomposition. E-commerce took share from generic mid-market retailers. Marketplaces took share from catalogue businesses. DTC brands took share from commodity shelves. Logistics, fulfilment, and payments became larger categories underneath the visible shopping experience.

The headline category did not collapse. The stack underneath it changed.

The same pattern is showing up in SaaS. Seat-based pricing is softening. Generic horizontal tools are facing substitution pressure from AI-native alternatives. Buyers are more sceptical than they were in 2020. All of that is real. But it describes a thinner middle, not a dead category.

The categories under pressure are the ones that sit lightly on top of an existing workflow: a simple form builder, a lightweight task tracker, a generic slide-maker, a CRM with no domain-specific model underneath it. If a competent solo builder can produce a credible replacement in an afternoon with AI, the squeeze is real.

The categories that compound are different. They are the ones where a ten-minute replacement falls apart the moment the problem becomes real: the tax calendar inside accounting software, the eight-step approval chain in procurement, the retention model embedded in analytics, the compliance audit trail inside a healthcare workflow.

Those are not features. They are accumulated judgement encoded over time.

Linear is the useful edge case. On the surface, it is a task tracker, which should put it in the squeezed column. It is exactly the kind of product a Saturday afternoon in Claude Code seems able to clone. But Linear’s original moat was not “tasks in a database.” It was an opinionated model of issue tracking, keyboard-first execution, and speed for a specific kind of product team. Now the question is whether that model can recompose around agents doing parts of the work.

That is the real lesson. Surviving SaaS products are not static moats. They are teams that notice when the moat is moving.

A recomposing category can be healthy. Value is migrating somewhere new, not disappearing. The useful question for founders, buyers, and investors is where that value is migrating.

The data tells the same story

The market has cooled. The market is still hot. Both are true.

Cooling from the 2021 zero-rate peak is not the same thing as dying. The 2021 market was not pricing software at its long-term value. It was pricing software during a global lockdown, when every company had to digitise at once and procurement friction broke under pressure.

The 2024-2026 reset is a different comparison set. Mature non-tech sectors trade at 1-3x revenue. SaaS in 2026 still trades at 5-20x revenue with double-digit growth. That is normalisation, not collapse.

The aggregate is still growing. Gartner’s narrower public-cloud SaaS cut grew from $251B to $299B in twelve months. Broader analyst slices project the category between $819B and $1.48T by 2030.

The headline number depends on the definition. The direction does not.

The distribution is more important than the total.

The category split shows where the signal is:

• AI SaaS is forecast to grow from $22B in 2025 to $368B by 2034, a 36.6% CAGR (Fortune Business Insights).
• PaaS is growing at 21.8% CAGR through 2030 (Grand View).
• Vertical software, where compliance, domain data models, and regulated workflows matter, is growing at 12.5% CAGR through 2033 (Grand View).

M&A shows the same sorting at finer resolution. Of the 2,698 SaaS M&A deals SEG tracked in 2025, 72% referenced AI. But the multiples sort by workflow depth, not AI branding. Security trades at 6.3x revenue. ERP and supply chain trade at 6.7x. The broader vertical SaaS bucket sits at 4.6x, slightly below SEG’s 4.8x median.

The premium is not going to “vertical” as a label. It is going to the categories where workflow depth has actually accumulated.

Public SaaS tells the same story. The BVP Public Cloud Index, with $1.7T in market cap, trades at 5.8x revenue on 18.6% average revenue growth. That is mature growth pricing. It is not death-of-category pricing.

The AI divergence appears one tier higher. Bessemer’s Cloud 100, which covers top private cloud and AI companies more broadly than pure SaaS, averaged 23x revenue in 2024 and 20x in 2025, while aggregate value grew 36% in the same window. Revenue growth outran multiple compression. Inside that average, AI-in-stack companies trade at 24x while non-AI sits at 19x.

“SaaS valuations are up” is too broad. “SaaS valuations are down” is also too broad. Valuations are diverging by category, and the shape of the divergence is the useful signal.

Pricing models point in the same direction. IDC expects 70% of software vendors to refactor pricing around consumption, outcomes, or capability metrics by 2028. Microsoft 365 Copilot already meters agent usage alongside its $30/user seat. Salesforce Agentforce charges $0.10 per action.

The seat is no longer the only useful unit of account. Vendors moving fastest toward consumption and outcome pricing are adapting to the recomposition earlier than the ones still pricing every product as if human seats were the only source of value.

Organisational debt is the adoption brake

The “SaaS is dying” narrative has a second leg, and it is the more dangerous one to believe.

It goes like this: AI made software cheap to build, so companies will stop paying for other people’s software and build their own.

I understand the intuition. I write software with AI every day. I can move faster than I could two years ago. If I extrapolate my own workflow onto a 500-person company, I can see why someone might conclude that every organisation now has a latent build team inside it.

The extrapolation breaks when it reaches the organisation.

Most companies carry organisational debt. I mean something specific by that term: structure dictates processes, processes dictate tooling, and once those layers have settled into a shape, innovation becomes hard even when everyone agrees the current process is bad.

I have worked with mid-sized companies that wanted to introduce tools for workflows eating their teams’ time. The tool was picked. The ROI was obvious. The decision still sat in a bottleneck for months. It was not budget. It was not one person’s veto. The decision required a conversation across several roles, each with its own model of the workflow, and no one owned the meeting where the decision had to happen.

That is the pattern. Not dramatic failure. A slow, diffuse stall.

AI amplifies this instead of resolving it. Adopting an AI agent inside an enterprise creates questions no existing role owns by default. Who owns the prompts? Who updates them when a new model ships? Who decides what the agent can do: sign contracts, send emails, issue refunds? Who signs off on failure cases? Who writes the eval suite that confirms behaviour after a model upgrade?

None of these questions has an obvious home in a 2023 org chart. Every one of them either needs an owner or forces the structure to change.

Companies that can reshape quickly have a real DIY opportunity. Most companies cannot.

Covid is the closest recent precedent. Companies reshaped overnight because not reshaping meant not operating. Remote work, asynchronous collaboration, and digital procurement became possible at a speed most executives would have called unrealistic in 2019. Five years later, return-to-office mandates and the walk-back of remote policies show the structural pull reasserting itself once the pressure lifted.

Forced reshaping does not dissolve organisational debt. It suspends it.

AI is not existential pressure for most enterprises today, which is why the default brake still holds. Building your own tool with AI is not only engineering time. It is ownership. Who maintains the tool after it ships? Which team answers when it breaks? Who upgrades the underlying model? Who retires it when the internal tool is worse than the vendor product the company could have bought?

AI made the code cheaper. It did not make the org chart cheaper. It did not make ownership cheaper. It did not make the decision to commit to a long-lived internal product cheaper.

The “everyone will build in-house” thesis imagines a company with almost no organisational debt. That company exists, but it is rare. Most companies will keep buying SaaS because SaaS is a delivery mechanism for capability that sidesteps the org-chart rewrite in-house software would require.

SaaS is not winning a head-to-head against AI tools. SaaS is winning because it is often the path of least structural resistance.

I am the counter-test

The DIY thesis should apply to me if it applies to anyone.

I am a one-person agency. I build products with AI as my core workflow. I can stand up a full-stack app in a weekend if the problem is contained enough. If anyone should replace paid SaaS with internal tools built in Claude, it is me.

And I still pay for different categories of SaaS every month.

I pay for accounting and invoicing software. I could write a tax-compliant invoice generator in a day. I would not trust my own version to survive year three of running a business under German tax rules. Vertical and compliance-deep SaaS wins because domain knowledge is not only a generation problem. It is a maintenance and judgement problem.

I pay for an AI presentation service. I could assemble a prompt chain that generates decent decks. The reason I do not is not that it would be technically difficult. It is that I do not want to own the templates, brand logic, export flow, editing model, and edge cases. Someone else has put product attention into a problem I only want solved.

I pay for agentic workflow tooling. I could glue something together. The product I pay for does the gluing better than I would, and it improves while I am doing something else.

Those categories map closely to the SaaS categories the data says are compounding: vertical-compliance-deep, AI-native, and platform/integration. I am not buying them because I cannot build them. I am buying them because my time is worth more than the build, and because the teams building them care about the problem more than I do.

That distinction matters. AI changes what I can build. It does not automatically change what is worth owning.

The average company is a less favourable environment for DIY than I am, not a more favourable one.

The moats AI did not lower

If I were evaluating what to build, buy, or back in 2026, I would not start with “AI features.” I would start with the moats AI did not lower.

Domain depth is the first one. Tax codes, healthcare compliance, insurance underwriting, supply-chain reconciliation, and regulated manufacturing workflows are hard because the knowledge is accumulated in conflicting rules, audit edge cases, and the gap between how a domain says it works and how it actually works. AI can summarise these domains. It does not replace the accumulated judgement required to build a product that survives contact with a new regulation, a failed audit, or a customer edge case.

Customer data gravity is the moat AI strengthens. A product becomes more valuable when the customer generates proprietary data inside it: workflow history, edits, corrections, structured decisions, approvals, exceptions. An accounting product with years of categorised transactions gives an AI classifier a better operating environment than a fresh clone with a clean schema. A CRM with a decade of pipeline activity gives an agent more leverage than a weekend-built replacement with twelve integrations and no history.

Generic horizontal tools that store only a thin layer of metadata do not accumulate much data gravity. Workflow-deep tools that own the entity model do, and the gap widens every quarter the customer keeps using the product.

Time is the moat people usually see last.

Someone has to care about the software after it ships. Someone has to watch telemetry when a model provider changes behaviour. Someone has to rewrite prompts when a better model is released. Someone has to deprecate old features, update the UI, file the security review, and respond to the customer who found the edge case.

AI has not removed that work. In some cases, it has increased it, because the underlying stack now changes faster.

This is why software retention economics have not cracked. It is why 60% of new ARR at $50M+ ARR companies comes from existing customers rather than new ones. It is why SaaS Capital’s bootstrapped $3-20M ARR cohort sustains 103% net revenue retention on 91% gross revenue retention.

Caring about software over time is a job. SaaS companies are, at their core, teams paid to keep caring about one problem so their customers do not have to.

Sovereignty-by-design is the newer moat, especially for European workflows. The EU Data Act, the ENISA cloud certification framework, and the European Commission’s EUR180M sovereign-cloud contracts in April 2026 are turning data residency, portability, and exportability from procurement details into product architecture questions.

For regulated European buyers, sovereignty is no longer a footnote. It is becoming part of the product surface.

These moats are not equally important in every category. But if none of them are present, AI substitution pressure should be the default assumption.

How to pick

The category-level answer is useful for strategy: vertical, AI-native, platform, workflow-deep. But that altitude is too high for a procurement decision, a vendor pitch, or a build-vs-buy call.

For those, use the four moats above as the passive screen. Then check the leading indicator: pricing.

Pure seat-based pricing is the early warning sign of the squeeze. It assumes value scales with human users. In AI-era software, value increasingly scales with actions, consumption, outcomes, or capability. Vendors moving toward those models are adjusting to the new unit of value before the multiples fully reflect it.

Finally, run three questions against the specific product:

1. Could a competent solo builder produce a credible replacement in a Saturday afternoon with AI?
2. Does the product compound on proprietary data the user keeps generating?
3. Would owning the internal version create maintenance, compliance, or decision costs the company is underestimating?

If the answer to the first question is yes and the answer to the next two is no, the product is exposed.

If the answer to the first question is yes, but the next two are also yes, the build decision is more complicated than the demo suggests.

If the answer to the first question is no, the product is probably not being protected by code. It is being protected by domain depth, data gravity, time, or regulation.

That is the difference the “SaaS is dying” narrative flattens.

Stop asking if. Ask which.

The next ten years of SaaS are not a shrinking pie fought over by defensive incumbents. They are a recomposition, and the useful work is learning to see which parts of the stack are thinning and which parts are getting richer.

If that outside perspective changes, I do not think it will be because the market narrative gets better. It will be because companies fail to absorb what they thought AI had made cheap enough to build themselves.

Narratives watch pricing. Patterns watch structure. Structure is where the answer usually sits.