The SaaS-pocalypse is overhyped. Nonetheless some SaaS companies will go the way of the dinosaurs.

“AI is killing SaaS.” Everyone is saying it.

The argument goes like this. AI can build software now. So why pay Salesforce for a CRM when you can ask Claude to build one? Why pay Asana for project management tools when you vibe-code a tracking tool? Why pay for any of it when the substrate of every business app is now a couple of prompts away?

We think the SaaS-pocalypse is overhyped. There is a kernel of truth in it. Some thin applications and one-off internal tools really will get vibe-coded away. The SaaS that dies will mostly be the SaaS that was not earning its keep. But systems of record are not going anywhere. Workday, Salesforce, Gmail’s backend, each one took a decade of corner cases to make reliable. You do not reproduce that in a weekend, or even in a year, with Claude Code. Production-grade software still requires real engineering. Most people would rather pay for it than maintain it.

The interesting question is not whether SaaS dies. It is which SaaS dies, what replaces it, and what the rest of it looks like when AI gets done with it.

Software has been here before

In the early years of Microsoft Office, Access was a core component. Every company had a power user who built the company’s internal tools in Access: an inventory tracker, a customer database, a quote builder, an expense report system. Access could do almost anything a database could do. If you were willing to learn it, you could build the app your job actually needed.

Then specialized SaaS arrived. Salesforce won the CRM. Asana won project management. Shopify won commerce. Workday won HR. NetSuite won the back office. Even though some people use Excel as a database, Access did not lose to Excel. Excel kept its niche of grids and formulas. Access lost, one domain at a time, to a constellation of purpose-built products that each nailed one job better than a general-purpose database tool could. Today Access is a back-office and developer product. The companies that ran their business on Access in 1998 do not run their business on Access in 2026.

Two things drove this. The first is that most people would rather buy a solution than build one. Buying gets you support, updates, security review, integrations, and it’s someone else’s problem when it breaks. Building gets you the job your weekend Access app does today and a maintenance burden forever. For internal tools used by a small number of domain experts, rolling your own is fine. But for widely used or complex tools, buying is preferable.

The second is that generic UI incurs a productivity tax. Generic tools are great if you do the job once a week. They are a tax you pay daily if you do the job constantly. Excel was fine for occasional sales tracking. Salesforce won because doing it every day across a team made the tax of Excel unbearable.

You can see the same pattern outside software. Cars used to have buttons and knobs for everything. Then touchscreens arrived and replaced them. Then drivers noticed that adjusting volume or climate by tapping through menus while driving felt worse than the knob did. Now the buttons are coming back. Mercedes, Hyundai, and Volkswagen have all announced returns to physical controls for high-frequency interactions. Nobody is going back to all-buttons and thankfully no one offered touch-screen steering. But the most used buttons are coming back, probably for good.

The chat box is the AI version of the touchscreen. Great for the long tail of occasional things. Tax-bearing for what you do every day.

AI is at the same point

The first wave of consumer AI is chat. Claude, ChatGPT, Gemini, Microsoft Copilot, and a growing crowd of general-purpose assistants. They are extraordinary. You can ask them anything and get a useful answer. You can paste a contract in and get a summary. You can hand them a screenshot and get a critique.

Between the chat box and what comes next is a middle layer that is getting most of the recent press. OpenClaw, Town, Hyperagent, Manus: polished personal agents that wrap general-purpose models in routines, multi-surface presence, approval modes, and a memory of you. OpenClaw is the dominant one by far, with more traction than the others combined. Call this Wave 1.5. These products are closer to fit-for-purpose than vanilla chat, but they are still chat-shaped at the core. You still go to them and ask. They are a polished Wave 1.

Wave 2 is different. Wave 2 is fit-for-purpose AI that lives inside the work surface itself. Not a chat box you visit, but the right surface for the work: an inbox, a sprint board, a CRM, a support queue, a code editor. The AI is where the work is, not waiting for you to come over and prompt it. If you are an engineer, that is Claude inside your IDE. If you are a designer, it is an AI designer inside Figma. If you are running sales, it is an AI inside the inbox doing summaries and drafts automatically, not ChatGPT you paste threads into. If you are running customer support, it is an AI inside the support tool acting on tickets as they arrive.

What Wave 2 is not

Wave 2 is not toolkits. Crew AI, LangChain, and Letta are infrastructure for builders. They are great if you are constructing a multi-agent system, but they are the AI equivalent of the database engines Access could connect to. Useful, for sure, but not what most people want to buy.

Wave 2 is not plumbing. Zapier and n8n are workflow automation tools. They connect arbitrary services with arbitrary triggers. They automate connections, not jobs.

Wave 2 is not chat with a domain skin. Many products take ChatGPT or Claude, bolt on a custom prompt and a different visual, and call it a sales agent or a customer-support agent. That is Wave 1 in a costume. Wave 2 products are the ones where the UI, the data model, the workflow, and the agent are designed together because the whole product is shaped around one job.

Token economics and the race to avoid commoditization

Today, when most people “use AI,” they are typing into a chat app — ChatGPT, Gemini, or Claude — or reading an AI Overview at the top of a Google search. The chat box and the search box are the product, and the model providers own that customer relationship directly.

As Wave 2 products proliferate, that ratio will shift. The lawyer who used to paste contracts into Claude gets a contract-review tool that does it inline. The salesperson who used to ask ChatGPT for follow-ups gets an inbox that drafts them automatically. The PM who used to summarize standups in chat gets a sprint board that summarizes them in place. Each of those products spends tokens. The point is not how many. The point is where. Consumption that used to flow through a chat interface owned by the model provider now flows through a third-party product the user trusts to pick the best model for the job.

For the model providers, this is the risk they are racing to avoid. If most consumption goes through third-party Wave 2 products, those products pick whichever model is best or cheapest. The model providers become OEM suppliers, the way chip makers supply PCs. OpenAI’s ChatGPT Agent, Custom GPTs, and Memory, Anthropic’s Claude Code, Cowork, and Projects, and Google’s Gemini in Workspace, AI Overviews, and NotebookLM are all defensive moves to build sticky consumer surfaces so the model providers do not get commoditized by their own customers.

The cleanest path out for them is to skip the build and acquire. Buying a Wave 2 layer is faster than building one. We would be surprised if some of the obvious acquisitions did not happen, and perhaps soon, since both OpenAI and Anthropic have filed to IPO.

What this means for SaaS

The SaaS-pocalypse narrative says AI is the asteroid and SaaS products are the dinosaurs. We think a good analogy is the arrival of television. Cinema did not die when TV arrived. The category transformed. Theaters had to adapt to a world where most people watched moving pictures at home, and many did not make it. The business model changed first to advertising-supported broadcast, then to subscription streaming. New formats became cultural phenomena. Prestige TV drama is not a movie with episodes; it is its own art form. Movies survived but reshaped around the event end of the spectrum. The category got bigger, not smaller.

The rise of AI may cause the same sort of shift. The SaaS that survives is the SaaS that adapts: incumbents that put AI inside their products and rebuild for the Wave 2 shape, and new Wave 2 companies that start fit-for-purpose from day one. The SaaS that dies is the SaaS that was slow, or that bet on the wrong model and got locked in, or that was never really earning its keep to begin with.

AI is not an asteroid. Nonetheless some SaaS companies are dinosaurs.

Where we fit

this+that is one of the first products built for Wave 2 of messages, tasks, and team workflow. The agent does not sit in a chat box waiting for you to ask. It reads every message you receive: Gmail, Outlook, Slack, Teams, Google Chat, Telegram. It reads them as they come, extracts what needs doing, surfaces tasks for the whole team, and runs workflows on the actions that should happen automatically.

The chat UI is wrong for this job. You do not want to paste an email into Claude to get a follow-up draft, then paste the follow-up back into your inbox to send. You do not want to summarize Slack threads by copying them into ChatGPT. You do not want to remember to ask the chat to check on the customer who has been waiting two weeks. You want the AI to live where the messages live and act on them while you are doing something else. And you need tasks to track what’s been done, what needs doing and who’s on it.

A note on threes

There’s an adage that things come in threes. Our earlier essay argued that agentic AI is the third wave of productivity applications, after desktop suites and cloud collaboration. This piece argues something narrower: that AI itself has its own waves, and so far we count two. Different framings, both useful. We could not figure out how to make this one come in threes. Wave 3 of AI is not visible yet from where we sit. Whatever it ends up being, we hope to be part of it.