Why bolting AI onto legacy funds misses the revolution

Richard Rumelt, who spent a career studying strategy, argues that the way to find undefended high ground is to ride a wave of change: a shift in technology or cost or rules that no single firm controls, that levels old advantages and creates new ones. His warning is that most people watch the wave's main effect and miss the higher-order effects, which are often larger. When the microprocessor arrived, everyone could see cheap computing. Fewer saw that once every component in a computer had a processor of its own, the work of integrating them would become trivial, and the vertically integrated computer company, which had built its advantage on that integration, would deconstruct into a horizontal industry of specialists.

Finance is a higher-order effect of the AI wave. The main effect is on cognitive work in general, and the work of finance is cognitive work that was computerized decades ago: reading, comparing, modelling, writing, monitoring, reconciling. It is the work AI is best at, and it is already going. Anyone can now get a term sheet explained, a cap table modelled or a market sized in seconds, for nothing, which a year ago cost a retainer. People hand agents their credit cards and let them book their travel. Inside the industry, agents source, model and write memos. CVC sold a company to Blackstone this summer with an AI in place of its bankers, and an agent called Boardy closed its own seed round. This is not only cheaper. AI is already better than people at the grunt work and the analytics. Judgment comes next, and it will arrive through a feedback loop between AI and people.

To see what the wave does to the fund, use a distinction from the study of innovation. Henderson and Clark showed that the changes incumbents miss are not usually new components but new architectures: the components of a product stay the same and the way they are linked changes, and the incumbent cannot see it because its knowledge of how the pieces fit together has become tacit, embedded in routines and org charts and the habits of people who have done it one way for years. A venture fund has five components: sourcing, diligence, the decision, ownership, and the exit. Nothing in the wave removes any of them. What changes is the architecture. The work moves to an agent. The decision moves to a market of the owners. Ownership moves onto a public ledger, where it is a token that anyone can hold and sell. The components are linked in a way that has no manager in the middle, and the incumbents, whose expertise is in being the middle, will add AI to the old architecture and call it progress. That is what Rumelt calls the predictable bias of every transition: the forecast is a battle of the titans, and the titans prepare to grapple while the floor beneath them gives way.

Industries also have what Rumelt calls an attractor state, the arrangement they are pulled toward because it serves buyers most efficiently, and they reach it through disruption, which shakes an industry loose from arrangements that no longer serve anyone but their incumbents. For private capital the attractor is the arrangement with no manager's margin in it, where the people whose money it is do the steering. Demonstrations speed the pull. Napster was the demonstration for music. What follows is the demonstration for capital.

The work becomes an agent

Doing the work of investing is not the same as holding the money, and two things stand in the way of handing it to an agent.

The first is that an agent has an enormous attack surface, and the record of agents given wallets makes the point better than any argument. In February an agent three days old tried to tip someone four dollars and sent its entire balance, about $250,000 by its own account, by mistake. Freysa was built with a single rule, never transfer its funds, and on the 482nd attempt someone talked it into transferring about $47,000. Aixbt, one of the best-known agents in crypto, was drained of 55 ETH through a hack. These were three different failures, a state bug, social engineering and hostile instructions, and the variety is the lesson: the attack surface is the whole interface between the model and the world, and it grows with every tool an agent is given. The second is that an agent does not know what it does not know. It can price a deal with complete confidence over whatever reached its context, and the most dangerous gaps are the ones no checklist names: the founder's history that lives in a whisper network, the structure that failed in an earlier deal, the market that turned after the model was trained. Good investors know what they don't know. An agent's map has no edge. Neither problem is fixed with a better prompt. They are fixed with an institution, and that institution is a market.