11People, roles and skills
Roles first, expert time as a management commitment, one team across both sides, and the skills to build.
The second principle stated in chapter 01 has a practical consequence: the business is not consulted about this system. It runs part of it.
Roles come first
A product owner from the business sets priorities, triages feedback and accepts each increment. Each source has an owner who decides what it means, what is validated and who may see it. Experts define the vocabulary, correct drafts, fill gaps and judge responses. Technology builds and runs the instruments that make their decisions hold (Figure 11). The CTO and the CIO share that work as the first article described, but neither decides in the business's place.
Expert time is a management commitment
Validating knowledge, correcting interview drafts and growing the test set fall on a small number of people who already have full-time jobs. If the work is simply added on top, it slips, and the project stalls without anyone deciding to stop it. Expert time must be budgeted and released by the experts' management, like any other resource, and it is often the scarcest one. Where experts are about to retire, it is also running out. Their engagement is earned the same way as sponsors': by seeing, cycle after cycle, that their corrections make the system better.
One team across both sides
Separating an "AI team" from the people who own the data and the source systems recreates the drift between the two sides that chapter 02 warned against. The walking skeleton needs one team, product owner included, building both sides together.
Own, buy, or bootstrap
The make-or-buy logic applies to people too. What the company will keep for years is owned: the description of its knowledge, the test set, access rights, product ownership. What is commodity or temporary can be bought: specialised document processing, the initial hosting setup, a security review. What is needed early but must be owned later is bootstrapped with outside help and a planned transfer, written into the roadmap rather than assumed. One skill is easily forgotten: interviewing experts well, on which Crawl depends. Hiring often takes longer than several cycles, so a missing skill can block a stage as surely as missing infrastructure.
Engineers need new skills, on top of the old ones
Most tools for preparing knowledge and running evaluations come from the Python world, which many enterprise teams have never used (Figure 12). More fundamentally, these systems are not deterministic: the same question can produce different responses, so "correct" becomes a rate measured on a test set rather than a test that passes or fails. Meanwhile, the established disciplines matter more than ever, because they are what separates a production system from a demo. Learning takes time, and that time belongs in the roadmap. A team asked to build a production system while learning its tools will deliver a demo.