The SaaS Counter-Revolution

Something important happened this week. Intercom announced they'd built their own model - one that outperforms every external model they'd previously used, at a fraction of the price. Most people scrolled past it. They shouldn't have. It turns out the frontier model providers don't have as much of a moat as we thought.

This is a signal of where this is all heading. But to understand why, we need to challenge three assumptions baked into the "everything app" narrative.

OpenAI, Anthropic with Cowork, and others are racing to build an AI agent that sits above all your tools and orchestrates work across Figma, Notion, HubSpot, Google Workspace, and the rest. But look carefully at how these agents actually work, and you'll notice something telling: they still need apps to get things done.

That's not a footnote. It's an admission. If AI could truly replace applications, the agents wouldn't route through them. They'd bypass them. Instead, the most capable agents are leaning into existing tools - using spreadsheets to model financials, design tools to produce assets, CRMs to manage records. Why? Because applications aren't just containers for data. They're optimized experiences built around specific kinds of work, and they function as guardrails that help AI find its way to accurate outputs. When an agent builds a financial model in a spreadsheet, the formulas either work or they don't. The application provides feedback. It creates structure. It catches errors that a free-floating agent would never encounter. Hallucination and drift are very real problems for AI today - and well-designed software was always built to guide workflows. Now it's doing the same for agents. The more guidance the application provides, the better the output. The app is the necessary scaffolding and infrastructure needed to make agents accurate enough to be useful.

Even Slack - which is perhaps one of the most horizontal applications out there is experiencing a boom of agents. Despite endless talk of AI replacing everything, neither Anthropic nor OpenAI has replaced it - and their own teams use it wall to wall. Because Slack isn't just a messaging app. It's a million small product decisions, made in the right order, stacking up into something exceptional. It's not just for human to human communication - it's also a proven orchestration layer for bridging the communication gap between machines and humans - and that power is now helping agents.

Counter narrative: software brings deep domain specific guardrails to agents. Without them, the agents are lost.

Here's a misconception worth challenging: that the primary friction in modern work is switching between user interfaces. It isn't.

Moving from Figma to Linear to Slack is not hard. Those tools represent three logically distinct aspects of work - design, task management, communication - and most people navigate between them without much friction at all. The interface is rarely the bottleneck.

There's a deeper problem with the everything-app vision that doesn't get talked about enough: it reduces the richness of human interaction to a text box. This reminds me of something I was told at school in the late 1980s - don't bother learning to type, because soon we'd be able to talk to computers and they'd record everything for us. Typing, we were assured, was a skill with an expiry date.

That prediction aged poorly. I type on my iPhone, my laptop, my iPad - everywhere. Because it turns out that our hands and our voice together are far more powerful than our voice alone. Try editing a complex document using only voice commands and you'll understand immediately why. Human communication is multidimensional. Great user interfaces are built around that richness - visual, spatial, tactile, contextual. Collapsing all of that into a single text prompt isn't progress. It's a step backwards dressed up as simplicity.

What this means is that the application providers - the Intercoms, the Figmas, the Salesforces - are actually well-positioned to win. People already spend most of their working hours inside these tools. And because these vendors built the product, they understand the domain deeply. Their models are trained on the right data, with access to everything under the hood and will get the most meaningful feedback to drive successful execution. Those important guardrails aren't accessible through the external API.

If I need to resolve a customer support issue, I'm going to trust Intercom's purpose-built model over a general-purpose agent. If I'm updating designs, I'll trust Figma's optimized model. If I'm working inside Salesforce, I'll trust Agentforce. Not out of loyalty - out of competence. General-purpose agents will always be locked out of the most valuable system feedback and guardrails, and application vendors have very little incentive to let them too far in.

Counter narrative: everything apps consistently provide a worse user experience and now they provide worse intelligence and automation too.

So if context-switching between apps isn't the problem, what is?

It's fragmented data. When I need to understand what's happening with a customer, and the answer is spread across Gong, HubSpot, Salesforce, and Zendesk - that's where the real pain lives. Not because I had to open four tabs, but because no single place holds the full picture. I have to piece together a story that should already be told.

Developers understood this intuitively long before the AI agent conversation started. Think about how a modern engineering team actually works: they use a dozen different tools across a single release cycle - GitHub, Jira, CI/CD pipelines, monitoring platforms, on-call systems. Each one sends its own alerts, its own notifications, its own updates. Left unconnected, it's noise. So what do developers do? Almost universally, the first thing they do when setting up a new workflow is connect all of those tools to Slack. Not to replace them - but to bring their signals into one place, so the whole story of a release is visible in a single thread. Nowhere is this more valuable than in incident response: when something goes wrong in production, the fastest path to resolution is a unified, real-time narrative of what happened, when, and what's been tried.

Here's where the AI platforms are getting it badly wrong. They've looked at this problem and concluded that it's fundamentally about the UI and application layer - that applications are just dumb databases sitting underneath, waiting to be orchestrated by something smarter.

If you believe that, you literally don't understand what business software does.

Applications aren't wrappers around data. They encode process logic, enforce constraints, generate feedback loops, and embed decades of domain knowledge into every interaction. Treating them as passive storage is a category error, and building strategy on top of it compounds the mistake.

The result is that the AI platforms are trying to solve a $1m data layer problem with a $15m application layer solution. The real pain - fragmented, inaccessible context - is a data connectivity and visibility problem. It's hard, but it's a targeted problem with targeted solutions. Instead, the everything-app vision substitutes that work with tokens: paying, at scale, for an agent to manually stitch together what should already be connected. It's replacing human error with AI-powered inefficiency, and calling it automation.

The oldest insight in enterprise software still holds: the best workflows are built on top of unified data and context, not bolted on top of a constellation of siloed APIs. The BPM vendors knew it. The CRM vendors built entire empires on it. The defining question now isn't "which agent can control the most tools?" - it's "who owns the shared data layer that makes context meaningful?" Get that right, and the coordination largely takes care of itself.

Counter narrative: Build the context layer, and you won't need $15m in agent token consumption to paper over the cracks.

Intercom just fired the starting gun on the domain-specific model era. The applications are coming back - and this time, they're bringing their own intelligence with them.

--

If you've read this far...

We built Noded AI to connect your customer data and to consolidate your customer signals in one place - inspired by my work rebuilding the Slack platform. Our goal is to give you the complete story on two very important questions: are you delivering on the commitments to your customers? And, are they going to grow?