Funding Your AI Bets. James March – 1991.

Two projects hit the same budget line this quarter. Same approval process, same ROI template, same three-year forecast demanded from both. That is already the problem.

They are fundamentally not the same kind of bet:

• The first bet starts from a pain somebody in the business already feels, already works around, already pays to avoid.
• The second bet starts from a capability nobody asked for, aimed at a market that does not exist yet.

Finance cannot tell them apart, so it treats them the same way, and both suffer for it.

March, 1991: Exploration vs. Exploitation

James March wrote the foundational reason in 1991, in one of the most cited papers in management theory:

Exploitation, refining what you already do, produces returns that are positive, proximate and predictable.

Exploration, experimenting with new alternatives, produces returns that are uncertain, distant and often negative.

Those are his words: often negative. That is the expected statistical outcome of genuine exploration, not a sign that somebody ran the project badly.

The Safe Bet Has the Smaller Ceiling

Which bet has the bigger upside? Not the safe one.

A bet that starts from an articulated need lands more often and pays back faster, but your competitors can see the same need. It gets copied inside a year and drifts into a commodity where only price separates you.

A bet that starts from a capability fails most of the time. The few that work do not return an incremental percentage; they create a market that did not exist. You will never get that from asking customers, because customers can only describe needs they can already picture.

They Need Different Money

So the two need different money:

Fifteen Pilots, No Capability

March described where the first error ends. Organizations that explore without exploiting suffer the costs of experimentation without the benefits, and end up with too many undeveloped ideas and too little distinctive competence:

Fifteen pilots, no capability. He wrote that in 1991 and it describes most enterprise AI programmes today.

He was just as blunt about the reverse! Exploit without exploring and you end up trapped in a suboptimal equilibrium—still profitable, until somebody takes the bet you refused to fund.

Sort Your AI Portfolio Into Two Piles

Run the check on Monday. Take your entire AI portfolio and sort every item into two piles:

1. Pile 1: This one is buying an option on a market we cannot measure yet.
2. Pile 2: This one is fixing a pain we can already price.

Ask the counterfactual test: Would this pain be on the list if GenAI did not exist? If not, it is an option, whatever the business case claims.

Next time someone brings you a three-year revenue forecast for an experiment, ask what it would cost to find out whether they are wrong. That number actually exists.