With AI

§Every chapter in this whitepaper described a step that existed before 2023 and exists now. This conclusion is about what changed in between, and what did not.

§ 17.1What got cheap#

§Desk research. Chapter 2 used to be a week. A good model with browsing collapses the first pass into an afternoon: the market as it is, the competitors and the adjacent tools, the post-mortems. It does not read the graveyard for you with judgment, and it will confidently cite a company that does not exist, so the afternoon is spent verifying rather than searching. But the search is no longer the cost.

§Synthesis, in the room. Chapter 6's first moment, the same-day page, is still yours to write, because the face and the pause are not in the transcript. The second moment, the weekly pattern pass, is now assisted: transcripts clustered, quotes pulled, contradictions surfaced, in the time it takes to make coffee. The learning wall updates itself. What does not change is that the team has to read it.

§The prototype. This is the change everyone sees. Two navigable prototypes in two days, by four people with no training, from a written brief and eleven prompts, on tools that did not exist three years earlier. In 2022 that was two weeks for a designer and a developer. Tom Chi's line about building a prototype in the time it takes to ask a question stopped being a metaphor.

§The documents. Every artifact in the template pack that follows can be drafted from one operating brief: the research plan from your assumptions, the conversation guide from your questions, the brief for the machine from your rules, the tension map from your synthesis pages. The job becomes maintaining the brief and correcting the drafts, which is a different job from writing, and rewards the people who write the brief precisely.

§ 17.2What that did to the sprint#

§Fidelity stopped being the conversation. For twenty years the first hour of every prototype review was about the prototype: the colors, the alignment, whether it looked real. Now the design pile says clean, simple, premium, and the whole argument moves to scope, which is where it always belonged and which is harder, because scope is where the disagreements live.

§The failure surface moved. It used to be "we could not build it in time." Now it is "we built it and nobody asked." When a feature costs one prompt, features arrive from nowhere, and every one of them is untestable because no question produced it. Chapter 11's rule, only build what came from a question, exists because of this, and it is the most important sentence in the whitepaper.

§ 17.3What did not get cheap#

§Recruiting. A stranger who fits your segment has a life. Getting an hour of it takes days of asking and no model can ask for you. Recruiting is now the critical path of the whole sprint, and the team that starts it on day one is the team that ends with a pattern rather than a finding.

§The person in the room. Interviewer skill became the quality ceiling. In the field sprint the customer interviews were strong and the operations interviews were weak, and the prototype was the same in both. The difference was who was asking, where, and how long they could stay silent. No tool changes that.

§Same-day synthesis. The transcript is instant. The decision about what you saw is not, and it can only be made by the person who was there, on the day. Memory keeps the words and drops the evidence, and the evidence was the point.

§Courage. In the field sprint a room of ten that had ranked payments first chose to put payments last, because twelve strangers said so and the product lead said it aloud. No model did that. The tool made the prototype in two days; the decision to let the prototype's failure overrule the team's vote took one person and a sentence.

§ 17.4What AI does not do in this sprint#

§It does not interview. A model can draft the guide and cluster the transcript; it cannot sit across from a restaurant owner and notice that she reached for her coat when asked for a second meeting. It does not vote. The written silence before a discussion is a human discipline, and a summary of five people's notes is not five people's notes. It does not decide. The tension map is filled by a team looking at quotes, and the memo is proposed by someone willing to be wrong in writing.

§Ethan Mollick's frame is the useful one: treat the model as a tireless, occasionally confident-and-wrong collaborator, and give it the parts of the work where tirelessness matters and judgment does not. In the sprint that is the desk, the drafting, the clustering and the build. Everything else is the field, and the field is yours.

§ 17.5What you leave with#

§The written position from chapter 13, in hand, on your way to the fourth whitepaper. And one question to carry with you into every prompt for the rest of the company's life: which question did this come from? If you cannot answer, do not build it.