SaaS Is Dead

SaaS is dead.

You have read that headline before. It took off in December 2024, when Satya Nadella explained on the BG2 podcast that business applications are essentially databases with business logic on top, and that the logic is moving to AI agents. Since then, every few weeks, someone declares the end of software as we know it.

I build and sell SaaS for a living. I run Goelo, an AI meeting assistant, and I sell the code behind it as Dopamine Notetaker. So I have a stake in this question, and my answer is: the headline is right about the symptom and wrong about the patient.

SaaS is not dying. The SaaS interface is. The dashboards, the tabs, the twelve screens you click through to find one answer. Your customers will keep paying for your software. They will just open it a lot less.

The Interface Was Always a Tax

Think about what a customer actually pays you for. Not the sidebar. Not the filter dropdowns. They pay for three things:

1. The data you hold: their meetings, invoices, projects, customers.
2. The logic you run on it: the summary, the calculation, the workflow, the permissions.
3. The trust: it works every day, it is secure, someone is accountable.

The interface is how they reach those three things. It was never the value. It was the toll booth in front of the value, and every SaaS charges the same toll: learn our navigation, remember where we put things, keep one more tab open, switch context every time you need an answer.

We accepted that toll because there was no other road. Each product needed its own front door.

That is what changes now.

One Interface for Every SaaS

AI assistants are becoming the one front door. Claude, ChatGPT, Cursor, Copilot: people already spend their day in them, and they are learning to reach into other software on the user's behalf.

The plumbing for that arrived fast. Anthropic released the Model Context Protocol (MCP) in November 2024, an open standard that lets an assistant call tools exposed by any application. Within a year, OpenAI, Google and Microsoft adopted it. Today, a user can add your product to Claude or ChatGPT as a connector, log in once in the browser, and ask questions about their own data, in their own words, without ever opening your app.

Here is what that looks like with Goelo connected to Claude.

Me: I'm meeting Dinesh from Pied Piper tomorrow. What did we agree on last time, and what is still open?

Claude calls find_past_meetings with the domain piedpiper.com, then get_meeting_summary on the latest one.

Claude: Your last call with Pied Piper was on October 14. You agreed on a two-week pilot with three users, and Dinesh owed you the security questionnaire by Friday. Still open: the pilot start date and who signs on their side. Here is the meeting in Goelo if you want the full summary.

I did not open Goelo. I did not search, filter or scroll. Goelo still did all the work: it recorded the call, transcribed it, verified the names, wrote the summary, enforced who can read what. The assistant was just the interface.

What I used to do in the app	What I ask the assistant
Search the meeting, open it, scroll to the decisions	"What did we decide with Pied Piper last week?"
Filter the meeting list by attendee	"When did I last talk to someone from Hooli?"
Copy the action items into an email	"Draft the follow-up email from yesterday's call."
Read three summaries before a renewal call	"Summarize everything we discussed with this client."

Every line on the left is a screen someone designed, built, tested and maintained. Every line on the right is one sentence.

What Doesn't Die

If the interface fades, what is left of a SaaS? Almost everything that matters.

The data. An assistant cannot answer "what did Pied Piper decide?" without the meeting being recorded, transcribed and stored somewhere. Whoever holds that data holds the value.

The logic. Turning a 40-minute call into decisions and action items, with names spelled right, is hard work. In Goelo, a Deep research step checks people and companies on the web and in past meetings before the summary is written. The assistant does not redo that work. It consumes the result.

The permissions. When Claude reads Goelo, it reads only the workspaces the user ticked on a consent screen, and only what that user is allowed to see. The assistant is a guest. Your product is still the one deciding what the guest can open.

The billing. Someone still pays for the recording, the transcription and the AI. That someone is still your customer.

The interface itself, for many people. Let me temper my own headline: not everyone wants to talk to their software. Plenty of users would rather scan a list, click a meeting and read it. When you don't know exactly what you are looking for, a good screen is still the fastest way to browse. Habits change slower than technology.

The source of truth. An assistant answers with confidence, and confidence is not proof. When Claude says Dinesh owed you the security questionnaire by Friday, you want to check before you chase him. Your interface is where that check happens: the full summary, the transcript, the exact words that were said. The assistant gives the answer; your app is where the user makes sure it is right.

Your SaaS does not disappear. It becomes the backend of an assistant, and the place people come back to when they need to be sure.

Consumable Value

This is the part I want founders to take away. If your customers are going to use your product through an assistant, your job changes. You no longer ship screens people visit. You ship value an assistant can consume.

Consumable value has a few properties:

• It answers a question, not a navigation. "List the meetings with this company" is consumable. "Open the meetings page" is not.
• It is compact. An assistant reads your result inside a limited context window. Return the decisions and action items, not the whole page.
• It is trustworthy. If your summary misspells the client's name, the assistant will repeat it with confidence in an email. Accuracy becomes a feature the user can feel, even without seeing your UI.
• It links back. Every Goelo result carries an absolute link to the meeting. When the user wants the full picture, your interface is one click away.
• It is scoped. The assistant gets the user's permissions, nothing more, decided by your code and never by the model.

In practice, Goelo exposes five tools to assistants: list the meetings of a period, read a summary, find past meetings with a person or a company, find people already verified, and read a transcript. That is the whole surface. Five questions customers actually ask, answered well.

Five well-built tools are worth more than fifty screens nobody opens.

The Interface Gets Smaller, Not Zero

I am not saying you should delete your frontend. Some moments still need one:

• Onboarding and setup: connecting a calendar, inviting a team, configuring the bot.
• Consent and trust: choosing what an assistant may read, and revoking it.
• Checking and editing: opening the source behind an answer, correcting a summary, sharing it.
• Billing: plans, invoices, credits.

Notice what is missing from that list: the daily lookups, the "where was that again?". For the users who move to an assistant, the interface shrinks to the moments that need a human decision or a second look, and the assistant takes the rest.

The Uncomfortable Part

There is a real risk here, and I would rather name it.

When customers reach you through an assistant, you lose the front of the stage. Your brand appears as a tool call in someone else's chat. Your carefully designed dashboard gets fewer visits. If your only value was a nicer interface on top of commodity data, an assistant makes that very visible, very fast.

Some metrics break too. Daily active users means little when the user never logs in but asks Claude about your data ten times a day. Seat-based pricing gets strange when one assistant works for the whole team. I don't have a clean answer for pricing yet. Usage-based credits, which Goelo already uses for recording and transcription, are probably part of it.

The founders who win this shift are the ones whose data and logic are worth calling. The ones who lose built a moat out of pixels.

What to Do This Quarter

If you run a SaaS today, here is where I would start:

1. List the five questions your customers ask most. Not your features: their questions. Those are your first tools.
2. Expose them over MCP, read-only first. A tool that changes data can be triggered by a prompt injection hidden in the data the model reads. Start with reads.
3. Use OAuth, not API keys. Your customers will not copy a key into Claude. They expect a login and a consent screen, like "Sign in with GitHub".
4. Return compact results with links back to your app. The assistant answers, your interface handles the deep dive.
5. Invest in accuracy. It is now the part of your product the user experiences most, even through someone else's interface.

Where Dopamine Stands

I took my own advice. The Dopamine starter kit and Dopamine Notetaker both ship an MCP server with an OAuth 2.1 login: your users connect Claude, ChatGPT or Cursor in one click, and you expose your features as tools, one file each. In the Notetaker, the five meeting tools above are already there. I wrote down every technical decision behind it in Adding an MCP Server with OAuth 2.1 to a NestJS SaaS.

SaaS is dead. Long live SaaS.

The interface was never the product. The value was. Make it consumable.