From AI Tool to AI Firm

Most companies asking "how do we adopt AI" are asking the 2023 question. The firms pulling ahead stopped treating AI as something you add to a workflow and started treating it as the thing the workflow is built around. The distinction sounds semantic. It is the whole game.

Buying an AI tool is a procurement decision. Becoming an AI firm is an organisational one. The first changes a line item; the second changes how decisions get made, who makes them, and how fast the loop between observation and action closes. We invest on the assumption that within a decade, the second kind of company will out-compete the first the way software companies out-competed their analog incumbents — not by being marginally better, but by operating on a different cost and speed curve entirely.

Features sit on top. Intelligence sits underneath.

A feature is a chatbot bolted onto a support page, a summariser in the inbox, a copilot in the IDE. Useful, measurable, and almost entirely surface. It improves the experience of an existing process without changing the process. When the novelty fades, you are left with a modest productivity gain and a recurring bill.

Embedded intelligence is different in kind. It means the underwriting decision, the pricing decision, the routing decision is made by a system that learns from every prior decision and its outcome. The human moves from making the decision to designing, supervising, and overriding the system that makes it. That shift compounds. Each decision improves the next. A bolt-on feature does not compound; it depreciates.

A feature makes a task faster once. Embedded intelligence makes every future task better than the last.

The data architecture is the strategy

You cannot build embedded intelligence on top of a data estate designed for quarterly reporting. The AI-native firm treats its operational data — every decision, every outcome, every correction — as the raw material of its core asset. That requires capturing decisions and their results in a structured, queryable, feedback-ready form, which most organisations simply do not do today.

This is why the cheapest-looking AI adoption is often the most expensive long-term mistake. A surface feature requires no data discipline and delivers a quick win, which is precisely why it postpones the work that actually matters. The firms we back are willing to spend a year rebuilding how they capture and structure their own decisions before they ever ship a model into production. It is unglamorous, and it is the moat.

Talent: from operators to designers of operators

The AI-native organisation does not employ more people doing the same work faster. It employs fewer people designing the systems that do the work, plus a layer of judgment for the cases the system should not decide alone. This is a genuine restructuring of the org chart, and it is where most transformations quietly fail — not on the technology, but on the unwillingness to change who does what.

For investors and operators alike, the diagnostic question is blunt: is this company using AI to do its existing job slightly faster, or to do a fundamentally different job that was not possible before? The first is a feature. The second is a firm. We are interested almost exclusively in the second.