I’ve spent 15 years in the tech industry. I’ve seen waves come and go… cloud, mobile, SaaS, crypto, and now AI. Every single time, the noise drowns out the signal. So let me skip the hype and offer a perspective I think is worth sitting with.

If I were building software today, I wouldn’t be building a product.
I’d be building an operating system.

I know how it sounds and no, I’m not talking about Windows or Linux but something closer to what Anthropic has been quietly assembling… lt me explain exactly what I mean.

To understand where the industry is going, ignore the headlines and watch what the big players are actually building.

Claude didn’t stay a chatbot. It’s now an extensible desktop application with agents, skills, third-party integrations, MCP support and so on...

Perplexity recently introduced the concept of “computers”, a system that “acts” on information.

OpenAI made the same move (of course, louder): ChatGPT launched Operator in early 2025, an agent that browses the web, fills out forms, and completes tasks end to end.

Google instead executed one quietly. In March 2026, Gemini went live across the entire Google product suite. Google Search added an AI Mode

Four companies. Same structural bet. The future isn’t a model you talk to. It’s a platform that acts… and that’s not breaking news… but it highlights the path IMHO.

Here’s the problem with these big players right now: they’re built for everyone.
Which means, in practice, they’re optimised for no one in particular.

A tax accountant doesn’t need a general-purpose assistant. They need something that already knows tax codes, speaks the language of their workflows, and connects directly to the tools sitting in their existing stack. Same goes for an insurance underwriter, an e-commerce operator, a compliance officer and so on and so forth…

A general-purpose agent isn’t going to own those spaces unless you heavily fine-tune it. The depth isn’t there. And depth - in a specific domain - is exactly what enterprises will pay for.

That’s the gap. And it’s wide open.

Here’s what I also find interesting: you don’t need to reinvent the experience: Anthropic already built it.

Look at Claude Desktop. Left sidebar with task history. A persistent chat interface in the centre. A dynamic right sidebar for tools, context, and outputs.
That three-panel structure is essentially an operating system shell - and it’s the right model if you ask me.

The play is to take that architecture and make it vertical.

Three things separate a generic AI platform from a vertical AI OS.
Get these right and you have a defensible product.
Get them wrong and you have a chatbot with a logo.

1. A custom knowledge layer

It goes without saying… you need a fine-tuned RAG system bolted on top.
A knowledge layer tailored on the actual documents that matter to that profession. Tax codes. Insurance policies. Product catalogues. Legal precedents… etc.
The agent needs to speak the language of that industry natively.

This is the foundation. Without it, you’re just reskinning ChatGPT.

2. Profession-specific tooling

Not generic integrations like Google Drive or Notion. The exact tools that specific professional uses every day, already there… already connected, already contextualised. An accountant’s stack looks nothing like an underwriter’s stack.
The tools need to reflect that.

The more tightly you match the tooling to the profession, the less onboarding friction you have. The product should feel like it was built for them… because it was.

3. Generative UI

This is the most underrated component IMHO.
You give the agent the ability to build its own interface for the tools it’s working with. It doesn’t just use the tools, it shapes how they’re presented to the user based on the task at hand.

Instead of the interface being fixed, the agent decides in real time what to render based on what the user is trying to do. Google Research described it as “an AI model generating not only content but an entire user experience” - and they’ve already shipped it inside Gemini 3. The UI isn’t loaded from a template. It’s composed on demand.

For a vertical AI OS, this matters more than it does for a generic platform. An accountant reconciling invoices needs a different interface than the same accountant filing a quarterly return. You don’t want to build both screens manually and maintain them forever. You want the agent to figure out what components need to be used.

That’s a fundamentally different product architecture - and it’s what separates a smart assistant from something that actually feels like an operating system.

The vertical you pick matters more than the technology you use. Pick something with a clear document-heavy workflow, a professional audience that pays for software, and enough complexity that a generic tool will always fall short.

Accounting. Legal. Insurance. Healthcare administration. These aren’t sexy. That’s the point. The horizontal AI platforms will spend years trying to reach them. A focused vertical product gets there in months.

The architecture is replicable. The domain expertise isn’t.