July 2026 · 7 min read
A Roadmap Compounds. A Pilot Resets.
Most companies play AI like a county-fair booth: pay per throw, keep nothing, and call the pile of prizes a strategy. The fix isn't to stop throwing, because a cheap scouting throw is often the smartest money you spend. It's to build the foundation that catches what a throw wins, so the spend compounds instead of resetting. That foundation, funded on purpose and sequenced, is a roadmap.
You know the game. Five dollars for three ping pong balls, and all you have to do is land one in a milk bottle. The rim is barely wider than the ball. The bottles are packed so tight there’s no bounce that helps you. You throw all three, you miss all three, and the guy running the booth is already looking past you. If you win, you carry off a stuffed animal that cost a dollar to make and comes apart before you reach the car.
Most companies play AI like this. Not because they run pilots. Because they run pilots the way you play a booth: pay per throw, keep nothing, and call the pile of prizes a strategy.
I want to be precise about the mistake, because the easy version of this argument is wrong. The mistake is not the throw. A cheap throw can be the smartest money you spend. The mistake is a booth where winning and losing leave you in the same place, and three years of throwing that never adds up to anything you own.
A throw can be scouting. A toll buys nothing.
Some throws are worth every dollar. You step up once to see whether the game is winnable at all, watch how the bottles are stacked, and decide whether to keep playing or walk. That throw’s value was never the prize. It was what it told you. In AI that’s an honest use of a pilot: spend a little to learn whether the thing is possible, whether the data supports it, whether anyone would use it. A pilot that kills a bad idea before you fund the build just saved you the build. That is a win, and the spreadsheet should book it as one.
The toll is different. A toll is the throw that was never going to teach you anything, because it was scoped to look good instead of to decide anything. A proof of concept built to demo well. Hand-picked data, a controlled slice, the messy parts of production kept out of frame. It clears the bar of “look, it works,” which is a different bar from “this survives contact with the business,” and the gap between those two bars is the width of the bottle mouth. Nobody has to be acting in bad faith for this to happen. The demo satisfies everyone in the room, and the foundation satisfies no one’s calendar, so the demo gets built and the foundation doesn’t.
MIT put a number on where that ends. In 2025 it found that 95 percent of enterprise generative AI pilots deliver no measurable P&L impact. Read that as the booth’s payout rate. Not because the models can’t win. Because almost nothing was built to catch what a win would have paid out.
Name what the throw is for
The trick of the booth is that you’re playing before you’ve asked what winning would get you. You see the game, the reflex fires, the money is gone before you priced the prize. Most AI spend starts the same way. The board wants to see AI. A peer stood up a thread. A vendor ran a good demo. None of those tells you what you’re trying to win.
So name it before you pay. Not the whole prize in perfect detail, because sometimes the honest answer is that you’re throwing to find out. Name what the throw is for. Either it is meant to produce a capability you will keep, in which case say which decision it changes and what changing it is worth, or it is meant to answer a question, in which case say which question and what you will do with each answer. Both are legitimate. The toll is the third kind, the throw that is for nothing, that can’t tell you it failed because it was never told what winning meant. Nobody agrees on what a customer is until someone does that work, and nobody agrees on what you’re trying to win until you say it out loud, before the money moves.
A toll is metered. Investment is committed.
Here is where the analogy earns its keep, and where it stops. Five dollars for three balls is metered spend. You pay per attempt, the meter resets every time, and nothing you spent on the last throw lowers the price of the next. That’s fine for scouting, where the information is the whole return. It’s ruinous as a way to build, because building is the one thing a meter can’t do.
Most pilots are metered when they should be invested. Each is a fresh withdrawal, scoped as a one-off. The data model built for the pilot was built for the pilot. The integration was a demo integration. The next initiative starts from roughly zero and pays its own toll. A company can run this for three years, spend real money, and hold no asset at the end, because it was never buying an asset. It was renting demos.
Investment is committed to a return and sized to the outcome instead of the attempt. It compounds, but only if there is something for it to compound on. That something is not another pilot. It’s the foundation underneath all of them: the governed data layer you build once, the entities you define once so the third initiative doesn’t relitigate what a customer is, the evaluation and the guardrails that every later bet inherits instead of rebuilding. Build that, and the throw stops resetting, because now there is a shelf at home for what it wins. This is the whole of what I mean by tying every dollar to a number the board already watches. The dollar is an investment when a return is named and something is built to keep it. It’s a toll when it isn’t. Same dollar. The difference is whether anything is behind it.
The roadmap doesn’t ban the booth. It gives the throws a floor.
A roadmap is not the opposite of a pilot, and it is not one giant bet placed up front. That version fails worse than the booth, because a rigged game at least caps your loss at five dollars, and a three-year platform committed before anyone tested the prize can lose the whole budget at once. A roadmap is the other thing. It’s a foundation, and then a sequence of bets that ride on it, each one gated, each one cheaper and more certain than it would have been alone.
Which means the roadmap has pilots in it. It is supposed to. A serious AI portfolio budgets for the exploratory bets whose only job is to tell you something, alongside the proven work that pays the bills. The difference is that on a roadmap those bets land on the foundation. What they learn is kept. What they build is reused. The exploration that I’ve argued is the right instinct under uncertainty is still right here. The half that was missing is the floor under it, and the discipline to converge onto the few bets worth hardening. Explore on the foundation, converge onto it, and the same throws that were tolls become the way the asset gets built.
That is the tell of a roadmap against a midway. On a midway, last quarter’s five dollars is a stuffed animal in a landfill. On a roadmap, last quarter’s spend is still working this quarter, because it went into something that is still standing.
Where ROI actually comes from
ROI is not a thing you bolt on at the end. It’s what falls out when the rest is real. You named what the throw was for, you built a foundation for what it returned, and you sequenced the bets so each one started higher than the last. A return was the plan, not the hope. The reason most AI spend can’t show ROI is not that the models are bad. It’s that no prize was named, nothing was built to keep the winnings, and every throw started over. A pile of demos is not an asset, however much it cost.
Don’t stop throwing. Stop paying tolls.
So don’t get better at throwing. Better aim doesn’t fix a game where winning and losing leave you in the same place. And don’t stop throwing either, because the scouting throw is how you find out what’s worth building. Stop paying tolls. Stop funding throws that land nowhere, scoped to demo instead of decide, with no foundation to catch what they win.
Build the floor first. Name what each bet is for. Let the bets ride on the foundation and converge onto the few that pay. Strategy plus investment, sequenced so it compounds, is a roadmap, and a roadmap is the only version of this where the money you spent is still there in the morning.
Keep the stuffed animals if you want them. Just stop calling the pile a strategy.