When Change Costs Almost Nothing
Upfront planning also has an economic value: the more expensive a correction becomes later, the more sense it makes to resolve as many questions as possible before implementation begins.
In knowledge- and software-intensive work, AI is starting to shift that relationship. Drafts can be produced faster, variants explored more easily, and results tested earlier. Discarding an approach and rebuilding it can also require significantly less effort than before.
That does not make preparation obsolete. But it changes the question: When is more upfront planning more valuable—and when do we learn more by starting to build and test?
When Even Perfect Plans Fall Apart
Robert C. Martin describes exactly this tension from his work with AI agents. In a conversation with Matt Pocock, he is asked how much planning makes sense before an agent starts implementing. Martin says that he, too, experimented with specifying as much as possible upfront and turning that specification into detailed plans. His experience was sobering: those plans repeatedly fell apart because it simply was not possible to think through everything in advance. (Robert C. Martin in conversation with Matt Pocock: “LIVE: Uncle Bob on Software Fundamentals in the Age of AI”, from approx. 36:15; YouTube.)
He later captures the point in one remarkably simple sentence:
“The agents love to write plans.”
According to Martin, agents will happily expand those plans and enrich them with ever more detail. The result can look exceptionally convincing—and still fall apart during implementation. His alternative is much smaller-grained: build something, get feedback, continue, check again, and reorganize when necessary.
AI agents, too, can produce an impressively detailed plan without anyone having thought through everything that matters.
Martin then illustrates the point with a house. If every change to the house under construction cost only one dollar, the economic case for perfecting the entire blueprint before the first piece of work began would look very different. For AI-assisted software development, Martin’s conclusion is that the cost of change has fallen so dramatically that experimentation and early feedback can become more valuable relative to extensive upfront planning. (Ibid., from approx. 40:04.)
When Planning Shifts Its Focus
If AI can make operational work faster and cheaper, planning does not automatically become less important. But the point at which it creates the most value may shift.
The less effort it takes to create variants, build initial versions, or make later adjustments, the less attractive it becomes to describe every possible step in exhaustive detail upfront. More weight may move to defining the outcome, making the assumptions behind a direction explicit, and identifying the feedback that will tell us early whether that direction is working.
For Project and Product Management, that would shift the emphasis: less detailed description of an assumed future, more deliberate design of a framework in which teams can learn and adapt quickly.
AI does not protect us from planning the wrong thing. It may simply help us execute it much faster.
Faster experimentation is not the same as working without a plan. The central question remains: what are we trying to learn from the experiment?
What Still Matters in Planning
A reliable project plan shows what is expected to happen—and what that expectation is based on and where the limits of our knowledge lie.
As AI increasingly accelerates analysis, elaboration, and implementation, that capability may become even more important. Speed does not make the wrong direction any more correct. One larger question remains open: If AI takes over a growing share of tactical work, what will remain at the core of Project and Product Management?
I intend to return to that question in a future article.
