Takeaways

What the article asks of each of its three readers.

If you are the CTO or CIO

If you are the CTO or CIO

  • Run your own gates before you commit an organisation: give the AI a problem you know well, have it build several versions of the same thing, and judge them on evidence. Here, three implementations in 48 hours settled an architecture choice (prologue).
  • Put a method around the AI, and add deliberately what the method leaves out, as you would for Scrum. BMAD gave roles, a chain from product brief to tested story and a three-way review meant for every story and recorded for 149 of the 172; the quality system and the roadmap had to be added (chapters 04, 06 and 10).
  • Build the walking skeleton first, and draw the boundaries with Domain-Driven Design before the AI writes across them. Cheap features pile up quickly on an assumption nobody has tested (chapters 03 and 08).
  • Ask whether the AI's tests would catch a mistake, as well as whether they run. Mutation testing on the safety-relevant code showed which mistakes the tests would let through, and a defect that no test could have caught (Appendix B).
  • Keep the fundamental decisions with the team, and move the rules that matter into checks the build cannot skip. Here instructions held most of the time, and the build checks every time (chapter 07).
  • Use cheap building to try options instead of debating them, and plan around the new bottleneck: the time people need to decide and to verify (chapters 03 and 10).

If you lead quality or regulatory affairs

If you lead quality or regulatory affairs

  • AI can produce the evidence of a regulated process with the code, provided the quality records live where it works: plain text, next to the code, under version control, with an identifier on every requirement, risk, decision and gap (chapter 06).
  • Take nothing on trust. Ask for proof that each AI workflow receives the rules: here none did for ten weeks, and audits and corrective actions run by a person are what caught it (chapter 07). The plugin used here is published under an open licence, with its tests: start from it, or from its pattern (Appendix F).
  • Make every approval show the difference being approved, keep validation separate, with an environment of its own, and watch for process that is cheap to build and costly to go through (chapter 09).
  • Decide what the AI may see, and keep the signatures with people competent to give them. Plan their time: the AI increases what they are asked to review (chapters 04 and 06).

If you are an executive or a team lead

If you are an executive or a team lead

  • Expect the cost of writing code to keep falling, under conditions: a method, a quality system and people who can judge the result. That changes your teams more than your tools (chapters 01 and 02).
  • The work moves to reading, judging and verifying. Keep, and develop, the people who can judge code across the whole stack, and start the conversation about roles early, with the people concerned (chapter 11).
  • Fund AI-built projects in gates, like any uncertain project, with conditions under which they stop. Cheap building does not make a product worth having (chapters 02 and 10).
  • Keep domain experts and a product owner close: no AI persona replaces the people who do the work (chapters 05 and 11).

In one line

In one line

Generative AI can bring critical software within reach of a very small team, provided a method surrounds it, a quality system backs it and experienced people hold the helm: here, one experienced engineer working part-time, under a Class C process, with a €20-a-month AI subscription whose flat fee hid a far higher consumption. The harder change is for teams: the work moves to deciding and verifying, and that work stays with people.