Explain how you would use scenario planning for a two-year AI product strategy.

Picture a two-year plan built on one guess about how fast a computer-vision model will improve, and a warehouse that has already spent the money before anyone checks if the guess was right. Cascade Recycling runs BinVision, a camera system that sorts recyclable material on the line. Bianca Ferraz is the product lead who stopped the company from betting two years of capital on a single number.

How to answer this, stage by stage

Nobody is scoring whether you can predict the future. They're scoring whether you built three answers for the futures you can't predict.

Let's learn

Here is what a two-year AI strategy needs, so a company doesn't spend two years of capital on a guess nobody wrote down.

Before BinVision, workers on Cascade's sorting line pulled recyclable material off a belt by eye, roughly forty items a minute, sorting by material type into separate bins. BinVision adds a camera and a classifier above the belt that flags material type before it reaches the worker, cutting missed items by more than half. Cascade's leadership wanted a two-year plan for how much further to automate the line, and the first draft assumed one number: the model's accuracy would keep improving at roughly the same steady pace it had for the last year.

This is the process the two-year plan is actually about speeding up, one truck at a time.

Here's the turn: the danger wasn't picking the wrong pace of improvement. It was writing a plan that only had one pace in it at all. When a vendor's upgrade looked like it might be arriving faster than expected, the team had no pre-agreed answer for what that would mean, so the temptation was to treat any promising number as proof the fast future had already arrived, capital and all.

At its worst, betting on the wrong trajectory doesn't just waste one quarter's planning meeting. Cascade almost placed an order for a second automated sorting line, a six-figure commitment, on the strength of a benchmark number that didn't hold on the material the line actually processes most.

What I would leave alone: the worker training schedule for the sorting line doesn't need a scenario plan at all. Whether the model improves quickly or slowly, workers still need the same onboarding for how to work alongside the camera system, so that part of the roadmap stays fixed regardless of which track the company ends up in.

The lesson: a two-year AI plan doesn't fail because the forecast was wrong. Forecasts are always a little wrong. It fails because nobody wrote down, in advance, what to actually do about each way it could be wrong.

Now here is the same thing as a story

The short version above is what you'd present to a capital-planning committee. Read this one for how a wall-mounted terminal on the sort-line floor almost triggered a purchase nobody had actually agreed to make.

The terminal that shows BinVision's live accuracy numbers is bolted to a post near the start of the sorting line, where Bianca Ferraz can see it from the catwalk above. Cascade hired her as its only product person eighteen months before this happened, reporting straight to the plant's general manager.

Three named tracks and one fixed check. This is the whole plan, not a summary of it.

For five straight quarters, the terminal told a boring, reassuring story. Accuracy climbed a little each quarter, roughly the pace everyone had assumed. Bianca's scenario plan sat mostly unused, a document nobody had needed to open.

Then, in the sixth quarter, the vendor pushed a model upgrade. The published benchmark jumped. In the plant's weekly ops meeting, someone floated approving the second sorting line early, since "the model's clearly hitting the breakthrough numbers now." It was the kind of moment that, without a written plan, becomes a decision made on excitement rather than evidence.

The anchor didn't prevent the near miss. It caught it before the order went in.

Bianca's scenario plan had a rule for exactly this moment: no track shift counts until the trigger number is tested against Cascade's own held-out material, not the vendor's general score. The test came back the next morning. Film plastics accuracy, the category driving most of the line's actual sorting decisions, had dropped 11 points even as the overall benchmark rose. The breakthrough track's trigger hadn't been met at all.

Four branches, not three. The fourth one, hold and recheck, is what stopped the near miss from becoming a purchase.

So here is the decision I would take back: writing a two-year plan around one assumed pace of improvement in the first place. It made sense when the goal was just getting a plan approved fast. It stopped making sense the moment a single vendor headline could have redirected two years of spending on its own.

The near miss lands right where a fixed quarterly check was already scheduled to look.

With the scenario plan in place, the replay is almost boring. The vendor's benchmark jumps in quarter six. Someone in the ops meeting gets excited. The terminal's numbers get checked against Cascade's own material the next morning, same as every quarter. The plan holds to the Slow track, the sorting-line order stays paused, and two quarters later, once film plastics accuracy genuinely clears the trigger, the Breakthrough track activates on purpose, not on a headline. And the thing I'd tell myself, if I could go back: the plan was never supposed to predict which quarter the future would arrive. It only needed to say, in advance, how we'd know when it had.

SPARK, in one screen

The recap, one line per letter: situation is workers sorting by eye off a belt, payoff is the habit of checking a written trigger instead of reacting to headlines, anchor is the three named tracks with pre-agreed spending shifts, risk is a benchmark hiding a real regression on the material that matters most, and keep out is capping it at three tracks with one fixed quarterly check.

Three tracks and a fixed date beat ten tracks nobody checks under real pressure.

And if you want to be sure it really works, try it somewhere else

Meridian Pharmacy Group runs AuthFlow, a tool that drafts prior-authorization requests for prescriptions, deciding which cases route straight to an insurer and which need a pharmacist's review first. Deshawn Pruitt, Meridian's pharmacy operations manager, faced the same two-year question: bet the roadmap on a steadily improving classifier, or hold room for a frontier model jump that could handle far more cases without a pharmacist touching them at all. Mapped onto SPARK, situation is pharmacists reviewing every borderline case by hand today, payoff is the habit of checking a written trigger before expanding automation, and the anchor is the same three-track structure: steady, slow, breakthrough, each with a number tested against Meridian's own denial-and-appeal data, not a vendor's general accuracy claim. The risk plays out differently here than at Cascade: a new model looked ready to handle appeals unsupervised, clearing a general benchmark easily, but it was quietly worse on the specific insurers whose rules change most often, exactly the cases pharmacists were already spending the most time on. Keep out is the same discipline: three tracks, one quarterly check, no attempt to forecast the exact model six months out.

The breakthrough track always carries the most to lose if the trigger check is skipped, in a pharmacy or on a sorting line alike.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "three tracks, one trigger number each, tested on our own data, checked every quarter," and stop.
Cost: no budget for a formal quarterly review process. Say so honestly, and start with just the three written triggers and a calendar reminder, since writing them down costs nothing and stops the worst version of this mistake.
The model gets better, for real: if the breakthrough trigger genuinely clears on real data, that's the plan working as intended, and the honest move is to activate that track on purpose, not treat a headline as proof it already happened.

Where people run it wrong.
They build the scenario tracks but never write the trigger number down in advance, so the check becomes a debate in the moment anyway.
They trust the vendor's own benchmark instead of testing against their own data, and miss exactly the regression that matters most.
They build ten tracks to feel thorough, and then nobody actually checks any of them once real work gets busy.

How to use it live. The moment someone asks about a multi-year AI strategy, ask yourself: what three futures am I actually planning for, and what number would prove which one we're in? Name both, and the rest of the answer follows.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0