AI will not kill SaaS. It will sort it.

OpenAI runs Slack heavily across the organisation. Its hiring flow goes through Ashby. It has posted roles specifically for SaaS enablement and governance -- the kind of headcount investment you only make if you expect SaaS to be load-bearing for years, not quarters. Anthropic uses Stripe for payments and billing. Greenhouse runs its recruiting. It has worked deeply with Salesforce and Slack.

Sit with that for a moment. These are not companies short of engineers. If any organisation on earth could rebuild its internal tooling from scratch using AI -- and justify the sprint cycles on strategic grounds -- it would be one of these two. They build frontier models for a living. The infrastructure cost of rebuilding a CRM, an ATS, or a billing platform is trivially within their engineering reach.

They are not doing that.

Why they buy

The build-versus-buy question is not fundamentally a capability question. It is a time, trust, and workflow-gravity question. When you are racing to ship the next version of a frontier model, every engineering cycle spent rebuilding recruiting workflows is a cycle not spent on the thing that actually differentiates you. The opportunity cost is not a small number, and it does not get smaller as your product ambitions grow.

But there is a deeper reason than opportunity cost. The SaaS tools that serious companies use are not thin databases with a UI layered on top. They are years of accumulated workflow logic -- edge cases, compliance handling, audit trails, integrations with dozens of adjacent systems, support infrastructure for failure modes you haven't encountered yet. When you choose Greenhouse over a custom ATS, you are not just buying software. You are buying accumulated knowledge about how hiring actually works across thousands of organisations, wrapped in tooling that has been stress-tested against real recruitment operations at scale.

That trust takes time to build. It is not replicable in a single sprint, even with AI writing the code. And the organisations that have learned this -- the ones that burned two years rebuilding something that already existed, only to find the rebuilt version had none of the edge-case handling -- tend not to repeat the experiment.

What AI actually kills

There is a real version of the AI-kills-SaaS argument. It just applies to a specific category. AI is a genuine threat to products that are essentially capability wrappers -- tooling that extracts a margin by sitting between a user and something a modern language model now does without a subscription. If your value proposition is that you do something slightly more conveniently than a generic model capability, and that capability is now table stakes in any enterprise LLM deployment, the economics are brutal.

It also threatens products built on retrieval and summarisation alone. If your SaaS surfaces data that users can now access conversationally through a general-purpose agent with the right integrations, the product layer above the data becomes fragile. The question to ask about any SaaS product is always: what is the user actually paying for? The surface, or the system underneath? If the honest answer is the surface, you have a problem.

AI may kill weak SaaS. It will not kill workflow gravity.

What survives -- and may get stronger

The SaaS that owns a real business workflow is a different animal. Consider what Salesforce actually is to the organisations that run on it: not just a database of customer records, but the audit trail for compliance, the source of truth for commissions, the system of record for revenue forecasting, and the integration point for fifty other tools. Unplugging it is not a product decision. It is an organisational archaeology project.

The more interesting observation is that deep SaaS may become more valuable as AI adoption grows, not less. If your product is the authoritative data source -- the place where the truth lives for a given business process -- then every AI layer that gets built on top of your domain needs you. The integration surface becomes a moat. The AI companies know this, which is why the large SaaS vendors have been adding AI capabilities rather than ceding the workflow layer to something else. They are fortifying, not retreating.

Anthropic's use of Stripe is a clean illustration. As AI billing complexity increases -- metered usage, multi-tier access, dynamic pricing at scale -- the last thing a fast-moving team wants is to own payment infrastructure. Stripe handles it and keeps handling it as the complexity grows. The dependency deepens as the product evolves. That is the direction of travel for workflow SaaS that gets this right.

The sorting function

I've spent a long time evaluating and deploying enterprise software in large organisations. The products that survive disruption cycles are almost never the cheapest or the cleverest. They are the ones where the workflow gravity is high enough that the cost of replacing them -- in time, in risk, in organisational disruption -- exceeds the benefit of what replaces them. That calculus is not going away. If anything, AI makes it sharper: the more AI-augmented tools get built on top of existing SaaS data, the more expensive the migration becomes.

The AI-kills-SaaS thesis is roughly right about a portion of the market. The thin-wrapper products, the simple-retrieval tools, the features that were always a bit too small to be products -- those are under genuine pressure. Not from AI replacing them exactly, but from the raised expectations that AI has set for what software should do without charging separately.

But the SaaS that owns workflow, data governance, integrations, and the kind of trust that accrues through audit trails and compliance coverage -- that is not dying. It is being sorted for. The best signal I know for which side of that line a product sits on is straightforward: watch which SaaS the AI companies themselves renew. They are doing the analysis for you.