What if an AI was your best customer?

The paradigm shift of MCP. A lot of ink has been spilled on the growing capabilities of AI. One of those capabilities — the ability to select and invoke tools — may ultimately yield an entirely new way of doing business. Back in November of 2024, Anthropic announced the launch of Model Context Protocol (MCP), an open framework designed to make it easier for AI models to interact with external data sources and tools. Earlier this month, Alexandre Pesant of Lovable tweeted about how Lovable used MCP to easily build their Supabase integration.

The critical impact of MCP was that Supabase, not Lovable, bore the brunt of the integration work: defining and exposing how third-party services like Lovable could best make use of Supabase’s API. MCP let Supabase handle the heavy lifting, defining how its API could feed data and actions to Lovable’s AI via an MCP server. This hinges on three parts: ‘resources’ deliver raw data like file contents or database rows; ‘prompt factories’ package that data into AI-ready formats, like a Supabase schema summary; and ‘toolkits’ let the AI trigger actions, such as adding a table column — all boosting efficiency without custom code. While direct attribution is impossible to an outside observer, the high growth that Supabase has recently experienced (and publicly reported on social) media seems closely related to the fact that MCP has made it very easy for third-party AI tools like Lovable, Bolt, and V0 to choose and use Supabase as their default backend database.

Building for AI customers. A few weeks ago, a CEO we work with said that he planned to ensure that his infrastructure service (I am afraid I can’t yet say which one) would be very easy for AIs to choose and use. This was the first time I’ve heard a founder explicitly say this, and it seems — to me — like a sort of watershed moment. We are entering an era where AI-powered tools (agents?) will autonomously select 3rd-party services, instantiate them, integrate with them, and begin to consume them. In many cases, they will be significant drivers of growth and revenue. In other words, for many companies that are building infrastructure services today — one of the principal decision-makers that will be driving revenue will be an autonomous system built by another company making intelligent optimization decisions based on a set of attributes that are published and consumed using MCP or some similar open standard.

Open questions. We’ve only just begun to wrestle with the implications of this for entrepreneurship — and as we do that — I want to turn to the community of founders and technologists who are reading this with some questions:

1. How widespread is this already? So far, I can only find a few examples of tools that are being selected, instantiated, and consumed at scale by AI. Some examples of tools that are being selected by AI already include Supabase, Github, Postgres, and Huggingface. Most of these are open-source and probably being used for free. It’s going to get really interesting if and when money starts flowing. Most of the examples of tools that are performing the “choose and use” (select and consume) are coding assistants (Bolt, Cursor, Lovable, and V0) and, perhaps, tools like Langchain. If you are aware of other documented examples, please let me know.
2. What does it mean to build a tool to optimize its selection by an AI? I assume that best practices would include an open-source version, clear documentation (including MCP), very well-specified APIs, and some clear benchmarking on performance and efficiency so that an autonomous agent could make a high-confidence selection. I suspect this will be an area of growing attention for infrastructure founders over the coming two years. Are you building for selection by AI? If so — how are you thinking about that?
3. Is this a new business model or just a new target customer? It is tempting to declare that we are on the cusp of a new type of business model (B2AI?) in which an AI is making purchase decisions — but I am not entirely sure that this rises to that level. It’s entirely possible that pricing models remain essentially the same — but that protocols like MCP become a new and necessary type of documentation for a new and crucial audience — AI agents of various types. Much like you might translate your existing API documentation into another language for certain humans, you will translate them into MCP for certain agent consumers. On the other hand, it could be that B2AI really does emerge as a totally new way of doing business and thinking about business — and it could unleash a new wave of infrastructure companies — built very leanly to serve an entirely new function in an entirely new way. A third possibility, however, is that this is not a stable point. Any AI that can automatically select and consume some infrastructure tooling might just be able to automatically author its own — completely optimized — version of that. How do you think this will play out?

I am far from any answers on these questions — but am grappling with them and very eager to hear from others who have begun to think about them as well. If you are building a tool that is designed to be consumed in a meaningful way by autonomous systems, I would love to hear from you.