What is the difference between a roadmap and a bet portfolio for AI products?

A roadmap and a bet portfolio can look identical on the page, right up until one of them is wrong. Deep Current Fisheries Cooperative runs TideCast, a tool that predicts where fish will be, zone by zone, so boats get sent somewhere worth the fuel. Ilse Marrant is the AI PM who has to explain why her planning document was never really a roadmap at all.

How to answer this, stage by stage

Nobody is scoring whether you know the word "portfolio." They're scoring whether you can tell a real commitment from a hopeful one.

Let's learn

Here is what happens when a plan built on an uncertain model gets written up in the same language as a plan for something you already know how to do.

Before TideCast, a dock master decided where boats went each week from tide charts, radio chatter with other crews, and a lifetime of noticing patterns nobody wrote down. With TideCast, a per-zone catch prediction comes back overnight, built from satellite temperature data and years of logged hauls. In the zones it knew well, it beat the dock master's own guesses by a wide margin.

Both are plans. Only one of them should ever get a promised date attached.

Here's the turn: Deep Current's planning document listed "sub-zone prediction live in all 12 zones by Q3" the same way it listed "boat maintenance completed by June," one line, one date, no distinction between a thing that was scheduled and a thing that was hoped for. The zones TideCast knew well got the same confident language as the zones it barely understood yet.

At its worst, a plan that promises the same date for a proven capability and a hopeful one doesn't fail everywhere at once. It fails exactly where confidence was lowest, and it fails loudest, because that's where a captain trusted the date most.

What I would leave alone: south banks prediction, at 84 percent accuracy on two straight seasons of its own eval set, earns a real roadmap date. Not everything on the plan needs a confidence tier bolted on; some of it has genuinely graduated.

The lesson: the words "roadmap" and "bet portfolio" aren't a branding choice. They're a promise about how sure you actually are, and using the wrong word is how a hopeful guess ends up believed like a certainty.

Now here is the same thing as a story

The short version above is what you'd say defending the plan to the cooperative's board. Read this one for how one sentence in a meeting gave the whole problem away.

Casimir Norlund has run dock scheduling at Deep Current for nine years. He can look at a week's forecast and know, before anyone tells him, which two boats will fight over the same good zone.

When TideCast's roadmap came out, Casimir built his whole quarter's crew rotation around the promised dates, same as he always had for maintenance and dry-dock schedules. South banks, open channel, reef shallows, north banks, all live by Q3, all with the same weight on the page.

One box is small on purpose. The other one is the box that actually costs something.

Reef shallows and north banks both slipped past Q3, not because anyone missed a deadline, but because the model's own accuracy on those zones had never actually cleared a bar worth promising. Nobody had written that bar down. The roadmap just said "live," the same word it used for the zones that were genuinely ready.

In a fleet-ops meeting, one of the boat captains looked at the roadmap and asked, plainly: "Wait, so north banks prediction is as sure as the maintenance schedule being done by June?" Nobody had a good answer. It wasn't. It had never been.

Deep Current split the document that quarter: a real roadmap for maintenance, staffing, and anything already proven, and a bet portfolio, with a confidence tier and a cost-if-wrong number, for every zone prediction still short of its bar. Casimir stopped scheduling crews around north banks entirely, and started planning that zone week to week instead, the way he always had before TideCast existed.

North banks is the one zone that should never have shared a column with boat maintenance.

PICK, the commitment that admits what it doesn't knowNot a definition to recite. PICK is what stops a hopeful guess from being believed like a certainty.

The recap, one line per letter: position is running the plan as a bet portfolio with a roadmap-shaped summary layered on top, impact is naming captains and a regulator as the ones who absorb a broken promise, cost asymmetry is that the false-certainty error is the expensive, hidden one, and kill criteria is a zone graduating to a real date only after clearing its bar twice.

And if you want to be sure it really works, try it somewhere elseSame four letters, a regional airline's turbulence-prediction tool instead of a fishing fleet. A different sort of promise breaks this time.

Vasterlund Regional Air runs a model that predicts turbulence risk along each route, feeding into how crews plan service timing mid-flight. Its planning document listed "smooth-service window prediction, all routes, by spring" with the same confidence as "new galley carts installed by spring." Mapped onto PICK: position is treating each route's prediction as its own bet, not a fleet-wide roadmap line. Impact lands on flight attendants who plan service timing around a window that sometimes closes without warning, and on passengers who get a drink knocked into their lap when it does. Cost asymmetry is the same shape: a route wrongly roadmapped as reliable costs a real injury report; a route wrongly bet-portfolio'd that turns out solid just costs an attendant double-checking a window that didn't need it. Kill criteria is a route graduating once its prediction holds across two full seasons of weather variation. The reversal here isn't a missing confidence column, it's a packaging choice: Vasterlund had bundled all routes into one fleet-wide accuracy number for its investor updates, which hid the fact that mountain-corridor routes were carrying nearly all the error while flatland routes were already reliable enough for a real date.

The same four branches decide a fishing zone and a flight route alike.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "a roadmap promises a date for something known, a bet portfolio prices something uncertain, and mixing the two into one column is how a hopeful guess gets believed like a promise," and stop.
Cost: no time to build a formal confidence-tier system this quarter. Say so honestly, and start by just splitting the document into two columns, since that alone stops the worst failure.
The model gets better, for real: if a zone's prediction genuinely clears its bar two seasons running, that's the honest trigger to move it into the roadmap, not a reason to keep hedging out of habit.

None of these four things belong on a maintenance schedule. All four belong on a model prediction.

Where people run it wrong.
They let one column hold both proven capabilities and hopeful bets, with the same confident word attached to each.
They treat "bet portfolio" as a permanent hedge instead of a temporary label with a real graduation test.
They publish an external, regulator-facing promise before the internal eval bar has actually been cleared.

How to use it live. The moment someone asks you this question, ask yourself: which items on this plan could reasonably lose, and does the document actually admit that. If it doesn't, it's not a roadmap. It's a bet wearing one.

Any one of these four signs is reason enough to move the item off the roadmap and into the portfolio.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0