I’ve tested about a dozen AI copilots over the past six months. Customer service copilots, coding assistants, sales enablement tools, HR support bots. Different vendors, different industries, different price points. And here’s what struck me: they all feel weirdly similar.

Not just the interface—though yes, they all have that chat window and suggested prompts thing going on. It’s deeper than that. They make the same kinds of mistakes. They have the same conversational quirks. They struggle with the same edge cases. It’s like test-driving cars that all have different paint jobs but identical engines.

There’s a reason for this, and it’s not what most people think.

The Foundation Model Bottleneck

Here’s the thing: nearly every enterprise copilot on the market is built on one of maybe four foundation models. OpenAI’s GPT-4, Anthropic’s Claude, Google’s Gemini, or occasionally Meta’s Llama if they’re going the open-source route.

The vendors layer on some RAG (retrieval-augmented generation), connect it to your company’s knowledge base, slap a brand on it, and call it proprietary. But under the hood? It’s the same reasoning engine everyone else is using.

This isn’t necessarily bad. These foundation models are genuinely impressive. But it does explain why your “industry-specific” copilot sounds a lot like your mate’s “completely different” one. They’re reading from the same script, just with different costume changes.

The Customisation Mirage

Most copilot vendors will tell you their secret sauce is in the fine-tuning or the prompt engineering or the proprietary workflows. And look, some of that’s real. A well-engineered system with good retrieval can genuinely outperform a poorly built one.

But the performance ceiling is still set by the underlying model. You can optimise a Honda Civic beautifully—better tyres, tuned suspension, the works—but it’s still not going to perform like a purpose-built race car.

The teams doing interesting work in this space—like the Team400 team—are starting to look beyond just wrapping existing models. They’re exploring model routing, ensemble approaches, and actually training domain-specific components where it matters.

What Differentiation Actually Looks Like

So if most copilots are functionally similar today, what would genuine differentiation look like?

First, data moats. A copilot that’s trained on genuinely unique, high-quality data from your specific domain will perform differently than one that’s just accessing a generic knowledge base. The law firms that are feeding years of case outcomes into their systems? Those tools will actually know things that generic models don’t.

Second, workflow integration. A copilot that lives inside your actual work tools—not just as a sidebar, but genuinely woven into how you work—creates different value than another chat interface. The ones that can trigger actions, update systems, and orchestrate across platforms aren’t just answering questions; they’re doing work.

Third, specialised reasoning. This is the frontier that’s opening up in 2026. Instead of one general-purpose model trying to be okay at everything, you’re seeing systems that route different types of questions to different specialised models. Medical reasoning goes to a medically-trained model. Code generation goes to a code-specific one. This architectural shift could finally break the “everything feels the same” problem.

The Vertical Model Wave

The biggest shift I’m watching is the emergence of domain-specific foundation models. Not just fine-tuned versions of GPT-4, but models trained from scratch on specialised corpuses.

Medical AI company Hippocratic AI released a healthcare-specific model late last year that absolutely destroys general-purpose models on clinical reasoning tasks. We’re seeing similar efforts in legal, financial services, and scientific research.

These aren’t copilots in the traditional sense—they’re the engines that next-generation copilots will run on. And they’ll feel genuinely different because they’ve learned different things, not just memorised different facts.

What This Means for Buyers

If you’re evaluating AI copilots today, here’s what to actually look for:

Don’t get distracted by the demo. Every copilot demos well on prepared scenarios. Ask about failure modes. Ask what it doesn’t do well. Ask about the underlying model and what actual customisation they’ve done beyond prompt engineering.

Look at the data strategy. How is this tool learning from your specific context? Is it just searching your documents, or is there actual model improvement happening based on your domain?

Evaluate the integration depth. Can this thing actually do work, or is it just an expensive search interface?

And honestly? For a lot of use cases right now, the boring answer might be the right one: just use ChatGPT or Claude with a decent knowledge base. The premium you’d pay for a “specialised” copilot that’s using the same underlying model often isn’t worth it.

The Next Twelve Months

Here’s what I think we’ll see by this time next year: the current crop of wrapper copilots will start struggling. The ones that are just a chat interface over GPT-4 with some RAG bolted on won’t justify their enterprise pricing anymore.

The winners will be the ones that either go deep on vertical-specific models or go wide on orchestration—tools that coordinate multiple AI systems to handle complex workflows. The middle ground of “generic copilot with industry flavour” is going to get squeezed.

We’re at the end of the first wave, where just having AI was differentiated. The second wave requires actually building something that couldn’t exist by just calling an API.

That’s harder. But it’s also where things get interesting.