When Building Gets Cheap, Your Roadmap Stops Working

For years your roadmap ran on one hidden rule: protect engineering time. Every framework you used, RICE, value-versus-effort, all of it, put build cost in the denominator. Now a working prototype shows up before the estimate does. That one change breaks the math your whole planning process rests on. Let me catch you up on what to do instead.

The denominator just went to zero

When a capable engineer can build a feature in an afternoon, effort stops being the thing that sorts your list. Jeff Gothelf makes the point plainly: divide impact by effort, and when effort shrinks toward nothing, everything looks worth doing. Which means nothing is actually prioritized. Your roadmap turns into a to-do list with a nicer name.

The reflex is to celebrate. And you should, some. Being able to build almost anything is real. But "can we build it?" was never the hard part. Deciding what's worth building, and living with it when you're wrong, still is. Cheap building doesn't answer that. It makes the question more dangerous, because nothing slows you down now except your own judgment.

Sort by what you learn and what you can undo

Gothelf offers two inputs that still mean something. Learning value: how much building the thing teaches you about whether it was worth building. A feature that settles a live argument or tests a strategic bet scores high. A feature you're already sure about scores low, even if customers end up liking it. Reversibility: how easily you walk it back if you're wrong. A config toggle is one thing. A migrated data model or a public promise is another. Bezos called those two-way and one-way doors.

Run the two against your backlog and you get four boxes. High learning, easy to reverse: build it now and watch what users actually do. High learning, hard to reverse: prototype first, use fake doors and small releases to capture the insight without locking in the consequence. Low learning, easy to reverse: make a small bet, put a date on the calendar, move on. Low learning, hard to reverse: stay out. It's the worst use of cheap capacity you have.

The trap is over-building the easy stuff

The danger in this era isn't under-building. It's pouring effort into things precisely because they're easy to produce. When something is cheap and reversible, the instinct is to polish it, hold a meeting about it, ship five versions. Don't. Ship one, check whether it moved a real number, and get back to the work that teaches you something.

Gothelf gives you a gut check worth doing Monday. Run the exercise once and count how many of your "quick wins" land in the small-bet box. That number is roughly how much of your roadmap was being sorted by cheapness instead of by learning. It usually stings.

The programs shipping decks and nothing else

The cost of building fell for everyone, including the people who never learned to build in small increments. Patrick Neeman points at the MIT NANDA study: roughly 95 percent of enterprise generative AI pilots delivered no measurable impact. The models work. The programs don't. Same death as the dot-com era, a thing nobody needed badly enough to pay for.

The fast movers weren't the big enterprises running the most pilots. They were mid-market teams that scoped narrow and shipped, often buying from a vendor before building custom. Ambition, measured in pilot count and headcount, correlated with getting stuck. Speed came from narrow scope and a real problem, not from attempting more at once.

And cheap output cuts both ways. AI can generate a fifty-page strategy deck and a governance framework before lunch, none of which moves a product one inch. Apply one test to every document: does it help someone make a decision or do the work? If it exists to prove you were busy, it's waste with a nice cover.

The habit that survives all of this

If output is nearly free and judgment is the expensive thing, then staying close to customers is the skill that keeps paying. Teresa Torres has been teaching this for years: you find real opportunities by staying near real customers, not by admiring your roadmap. Her ideation work makes a related point worth stealing. Generate 15 to 20 ideas for one opportunity, individually first, then compare them against each other instead of asking yes-or-no about each one. Cheap building tempts you to marry your first idea and go build it. More options, compared honestly, is the counterweight.

Watch what the data science role is turning into for a preview of your own. KDnuggets lays it out: the wage premium isn't for people who train models, it's for people who plug them in, keep them honest, and answer for the output. Prompt engineering roles grew 135.8 percent in 2025. When agentic accuracy runs 75 to 90 percent per step and compounds to about 50 percent over three steps, a human who knows the domain becomes the reliability layer. Judgment is the job.

The deep cut

Here's the part that actually changes your Monday. Swap the effort column in your scoring sheet for two new ones: "what will this teach us?" and "how hard is it to undo?" Don't announce a new framework or rename the meeting. Gothelf is right that you shouldn't. Just run the top 15 to 20 items through those two questions and watch where the arguments happen. That's where your real prioritization was hiding all along, buried under effort estimates that no longer mean anything.

One concrete guardrail while you speed up: put evaluation in front of the ship button, not after. A loop that runs faster also fails faster. Automated checks plus a human on anything a customer sees, with a definition of good set before you generate. That's what lets you keep your foot down without driving into a wall.

Three questions for your team

1. Pull this quarter's roadmap and sort every item into the four boxes. How many of our "quick wins" are low-learning small bets we're over-investing in?
2. For the top three items, what specifically will building each one teach us that we don't already know? If the honest answer is "not much," why is it near the top?
3. Where in our current AI work are we shipping documents and demos instead of something a customer used on a Tuesday, and what would we cut if the deck stopped counting as progress?