A Copilot is not a Strategy

(Unless you're an incumbent)

Microsoft’s announcement about its AI copilot was interesting for two reasons - (1) Copilots (yes, plural) are a critical expansion strategy for them, and (2) The rumors of consumption-based pricing’s demise may have been exaggerated. For this post I’ll focus on (1), with more to come in a future post on (2).

Copilots are still in their early days, but it’s clear that companies of all sizes are pursuing this strategy. The real question is - will they work? Looking over the landscape, I believe a few things:

1. A true copilot is one that’s in the flow, helping you complete your core tasks faster, with a non-interruptive UX. It’s Almost Invisible AI¹

2. Copilots are better as an incumbent strategy

3. Copilots can work as a wedge for startups, but only if they sit on top of some complex data acquisition or synthesis

Let’s tackle them in order.

---

I’d argue most of the copilots today aren’t true copilots. They’re helpful for answering questions, but not for taking action. And they exist separately rather than in the flow. They’re like a map you can get at a rest stop on the side of the road rather than a unified autopilot + GPS experience.

Which answers why copilots will be more powerful for incumbents: they help companies compensate for products with a large footprint, limited interoperability, and poor UX.

Think about products that have a delightful experience; for those products, will it be faster to type in a long question / set of commands in sentence format or to navigate it yourself? Now think about products where the information is split across dozens if not hundreds of pages (AWS Console, iOS, documentation, etc.) - having the right copilot experience would save you a lot of time! Accordingly, I believe copilots will generally be a sustaining innovation rather than a disruptive one.

That does reveal a compelling approach for startups to use copilots as a wedge, but only if they can find areas with the following characteristics:

1. Requires solving real infrastructure or integration challenges to build a great experience²

2. Large, latent user base ready to get activated

3. Users see an immediate ROI from your product even when not using the copilot

4. Incumbents are too focused on increasing lock-in to build an interconnected product

---

So what are some examples of great copilots?

I’d love to see copilots in the enterprise take inspiration from Perplexity*, where the copilot is in the flow of a question you have in progress and is personalized rather than generic. They’ve solved difficult problems with their own index and web crawling, the product is amazing even without the copilot, etc.

Glean* is another great example with their Chat feature, which is effectively the internal knowledge copilot to Perplexity’s global knowledge copilot. They’re solving truly difficult problems with indexing and RAG, every knowledge worker in the world would benefit from their product, and there is an immediate ROI on the search product even independent of Chat.

And I’m excited for a future where both products can take action on my behalf.³

On the flip side, there are a lot of “data copilots” today. They address a really critical need with a large latent user base and incumbents that are more motivated by increasing lock-in. But the risk is that they all have access to standard APIs for querying data sources and that the copilot is the whole product instead of a part. To stand out in this space, I think a data copilot would need to inherit security permissions from the various source systems while seamlessly connecting to both new and legacy data stores (without extracting the data).

Of course, many will go a step further and suggest copilots are a bandaid on our way to an agent-driven future. While that may prove to be correct, I think enterprises are years from trusting agentic solutions. Copilots are real and here to stay, and I hope they graduate from chatbots on the bottom right to teammates that help us accomplish our tasks faster than ever before.

1. AIAI just rolls off the tongue, doesn't it?

2. Some examples - data is trapped in silos, sources lack standardized APIs, performance expectations from the end user are high, intricate security requirements, etc.

3. Both of those examples are horizontal products. Copilots are also quite powerful as a wedge for vertical startups