Pruning My Pothos
← StoryboardsStoryboard · 13 frames · Stage 08, Readiness

When Is an AI System Ready for Real Use?

Tested, approved and deployed are not the answer. Thirteen frames on readiness as a judgment about one exposure, for one use.

Download PDFRead the article →

01 / 13

  1. A small bridge with a sign reading tested with: bicycles, a lorry marked PROD driving onto it, and an engineer on the far bank saluting: ship it. Deployed, delighted, disowned. Ready for what, exactly? Tested, approved and deployed are each worth something, and none of them is the answer. Readiness means the evidence and controls justify this exposure: these users, doing this job, with these consequences when it is wrong.

  2. The same system is fine for a team reviewing its output internally, and not yet for customers acting on it. Same system, different exposure. Naming the exposure precisely, which users doing what with what riding on it, turns an unanswerable question into one with an answer.

  3. Tested from a chair. Deployed from a plane. A system is not ready in general. It is ready, or not, for a particular exposure to a particular use.

  4. The demo: one user, one happy path, little at stake. Real use: malformed input, a slow upstream system, a goal nobody anticipated, and failures that are expensive. Readiness asks about conditions the demo may never have exercised, and silence in the evidence isn't a pass.

  5. Twelve rehearsals. One user. One path. A demonstration compresses the system: one user, one goal, a selected input, one path through, little at stake. It shows the thing is possible, which is a narrower question than whether it is ready.

  6. A release bar set first, and a lower one moved after seeing the results, which is crossed out: that is fitting the rule to the results. Define the release criterion against an evaluation set that fits the use and its failure costs, before exposure.

  7. Bullseye. Target painted afterwards. Define an explicit release criterion against an evaluation set that fits the intended use and its failure consequences, before exposure rather than after.

  8. Fallback: a queue. Reader: nobody. Name the safe alternative that actually exists when the system should not continue: stop and say so, decline, keep the prior state, queue it, restrict it, or hand it to a person who can act. A queue nobody can act on doesn't count.

  9. Some effects don't come back: a sent message, a payment made. Ask what can be stopped before more happens, reversed, corrected after the fact, or contained to limit how far it travels, and at what cost.

  10. The undo button. Connected to nothing. Before exposure, name what can really be reversed and how. In this repository, when checks that run after a draft is placed fail, the previous state of the affected files is restored.

  11. Everyone points. Nobody's name is on it. Before exposure, somebody owns the decision to expose, and somebody owns the response when it misbehaves. If nobody can be named for either, that is itself a finding about readiness.

  12. Approved. Best before the next model update. A change to the model, the instructions, the retrieved corpus, a tool, the workflow or the intended users can undermine part of the basis the release decision rested on.

  13. Three answers, in writing: what evidence exists against the failures that matter, what limits the damage when that evidence is incomplete, and who decides and who answers, named rather than assumed. If they can't be answered, nobody can say yet, which is more fixable than unready.