5. What AI changed, honestly

§Each of the six whitepapers ends with a section on what AI changes at that stage. Read together, they form one honest inventory, and this chapter collects it. The pattern is consistent enough to state once: AI made the middle of every stage nearly free, and the middle was never where the stage was won.

§ 5.1What got cheap, stage by stage#

§In discovery and the sprint, desk research collapsed from a week to an afternoon, synthesis of transcripts became a task done in the room, and the prototype fell from a two-week specialist job to a two-day one for anybody. The middle of the loop, the part between an idea and a stranger's hands, went nearly to zero.

§In strategy, the strawman ladder that a facilitator once drafted over an evening is now generated from the evidence in minutes, with citations, and the four handoff documents regenerate together from one brief instead of drifting apart across four evenings.

§In business design, the four value proposition canvases, the model and the monetization sheet draft from the synthesis and the strategy rows, and, more importantly, stay consistent with each other when a row changes.

§In experimentation, variants generate in minutes, analysis runs in the room, and hypotheses draft from the funnel, so experiment velocity, the number of honest loops per week, rises by a large factor.

§Across all of it, one genuinely new thing: every document can be generated from a single operating brief, which the next chapter is about.

§ 5.2The three consequences#

§The collapse of the middle did three things, the same three at every stage.

§Fidelity stopped being the conversation. For twenty years the first hour of every review went to the artifact: the prototype's rough edges, the deck's polish, the canvas's neatness. Now the artifact is clean by default, and the argument moves to what it should contain. In the field sprint, a hundred demo-day comments, none about design, all about scope. This is a gift, and it is uncomfortable, because scope is where the real disagreements live and pixels were where teams hid from them.

§The failure surface moved. When a feature cost a week, cost filtered features. When it costs a prompt, nothing does, and teams add features that trace to no question. The failure surface used to be "we could not build it in time." Now it is "we built it and nobody asked." Three features, three failures, one cause, in the field sprint. This is the single most important change in the whole inventory, and it is why the rule "only build what came from a question" is the spine of this whitepaper.

§The bottleneck moved to people. This is the one that matters, and it is the same at every stage. Recruiting became the critical path, because a stranger who fits your segment still has a life and no model can ask for an hour of it. Interviewer skill became the quality ceiling, because the person in the chair sets what real people are willing to say. Synthesis speed became the memory of the work, because what is not written the same day is gone. And courage became the deciding factor, because evidence only beats a vote if someone lets it.

§ 5.3Additive by nature, and product is subtraction#

§There is a structural reason AI tempts teams toward the wrong response, and it is worth naming plainly. Generation is additive. Asked for options, a model gives more options. Asked for a strategy, it gives a ladder that includes every good thing the evidence supports, which is Rumelt's bad strategy, a list of goals with adjectives. Asked for a business model, it gives a canvas where every box is full. Asked for experiments, it gives a backlog longer than any team could run.

§Product work, at every gate, is mostly subtraction. Strategy is what remains after most of the good ideas are excluded. A first version is defined by what it leaves out. An experiment worth running is the one chosen from the many that could be. The tool is fluent at the half of the work that adds, and silent on the half that cuts, and the half that cuts is the half that was always the point.

§ 5.4The margin question, briefly#

§One change is not about the process but about the product, and it belongs in the honest inventory. Casado and Bornstein observed that AI products carry a cost per user that does not vanish with scale, and the fifth whitepaper carried that into unit economics as a line item. The teams building on cheap generation should remember that cheap to build is not cheap to run, and that a product which calls a model on every interaction has a bottom line the free prototype never hinted at. Building got cheap. Serving did not, necessarily, and the spreadsheet still has to know the difference.

§ 5.5Ask where, not if#

§Vince Law makes the same point from the other direction, and it is worth stating because most teams get the question wrong. The question is not whether to be AI-native; it is where. His AI adoption matrix sorts work along two axes, whether the outcome is definitive or divergent, and whether the organization actively or passively prioritizes it, and gives each quadrant a different operating model: automate the definitive work nobody guards, sharpen the definitive work that matters, settle for good enough on the divergent work nobody guards, and deliberate, with humans owning the outcome, on the divergent work that matters most. His conclusion is this whitepaper's, arrived at independently: value migrates toward the divergent, actively-prioritized decisions, the ones that need judgment and accumulated experience. That is exactly the corner AI cannot take, and the corner the two human ends of every stage occupy.

§Law's warning pairs with it. AI eats whatever the organization treats as good enough, so a team that quietly settles on the decision of what to build, the most divergent and consequential work it has, will let the model make it by default, and get a fluent answer to a question nobody chose. Being AI-native is not adopting AI everywhere. It is knowing which corner is yours and refusing to hand it over.