THE DECISION-RULE CARD From "AI for Your Business — Advanced", chapter 3 ================================================================ A finding in a document has no trigger, no owner and no moment at which a named person must act. It sits outside the working week, and the working week always wins. That is why good analysis changes nothing. This card fixes that for exactly ONE recurring decision. Written down, a decision improves. In someone's head, it just ages. ⚠️ ONE RULE, ONE QUARTER. You will spot five decisions that need this. Five rules written at once means five unrevised rules in December, and the revision is where all the value is. ⚠️ NOTHING HERE IS A BUILD. No wiring, no agent, no workflow. You are writing a card and four names. Every firm that skipped straight to automating a decision it never wrote down ended up automating its worst assumptions, at speed, with nobody able to say why. 0 - CHOOSE THE DECISION ---------------------------------------------------------------- Two tests, and both must pass. [ ] DOES IT CROSS A BOUNDARY between two people? If the same person spots the situation, decides, and lives with the consequence, nothing gets codified -- they just keep doing what they do with a document nearby. [ ] DOES IT CURRENTLY REQUIRE A SPECIFIC PERSON'S HEAD? If yes, that's your one. If anyone could do it, it's a process, not a judgment, and it doesn't need a rule. [ ] Does it happen at least monthly? (It must, or you'll never accumulate enough cases to revise on.) The decision: ________________________________________________ THE RULE CARD ================================================================ 1 PATTERN IT CAME FROM (from your chapter-2 record -- how many judged outcomes, and what they cluster on. A rule with no evidence behind it is just a policy.) ______________________________________________________________ ______________________________________________________________ 2 TRIGGER (observable -- something a person notices without interpreting. "A client who seems likely to be confused" is a mood, not a trigger.) ______________________________________________________________ 3 DEFAULT ACTION (what happens unless someone stops it. A default, not a command.) ______________________________________________________________ 4 WHAT THE AI DOES At the moment of the decision it produces three lines for a person to read: Trigger matched: yes / no, and on what Recommended: the default action Because: the past cases from OUR OWN record it is matching ⚠️ THE CITATION RULE, in one line: NO CITED CASE FROM OUR OWN RECORD, NO RECOMMENDATION. If it can't cite, the flag doesn't appear and the decision is made the old way. That single line of policy is what stops a confident- sounding machine inventing grounds that were never in your business -- and it is a management instruction, not a technical control. Read the third line before acting on the second. Then it stops. It does not decide, does not send, does not act. If your rule can't be run on paper by a person who has read it, it isn't a rule yet. 5 PROPOSER ______________________________________________ (spots the trigger, puts up the default -- usually whoever's doing the work) 6 REVIEWER, WHO DECIDES _________________________________________ (a DIFFERENT person -- that's the boundary) 7 MAY OVERRIDE ON (name the KIND of reason, not just the person. "Client-specific knowledge -- something true of that client the trigger can't see." NOT "being busy.") ______________________________________________________________ A rule with no override gets quietly ignored within a month. A rule with unlimited override was never a rule. 8 EXCEPTION AUTHORITY ___________________________________________ (a THIRD person, for cases outside the stated override grounds. If the reviewer also owns every exception, "exception" quietly becomes "whatever the reviewer prefers.") 9 OUTCOME LOGGER ________________________________________________ (often the recorder from chapter 2 -- and it can be. Say so rather than assuming.) 10 WHAT GETS LOGGED BOTH the followed cases and the overrides, with a reason for the override in one sentence. This is the part everyone drops, and dropping it is what stops the rule ever getting better. ⚠️ OVERRIDES ARE NOT FAILURES. They are the most valuable entries you have -- each one is an experienced person telling you exactly where your rule is too crude, in the only language that carries weight: a real case. 11 REVISION DATE _________________________________________________ (a quarter out, in the diary now. On that date the overrides get read and the rule either changes or is confirmed. Confirming counts -- a rule reaffirmed on evidence is stronger than one nobody has examined.) ⚠️ FOUR ROLES, FOUR DESKS YOU ALREADY HAVE. This does not mean four new hires. Name what people already do, then say which of them is allowed to decide. DECIDE NOW WHAT HAPPENS ON THE FIRST OVERRIDE ================================================================ Write this while nothing is at stake, so that when it happens in front of the whole team you are reading a line rather than making a ruling: An override INSIDE the stated grounds is the reviewer's call. It gets logged with its reason and goes to the revision. An override OUTSIDE those grounds goes to the exception authority. Blur those two and either the reviewer starts asking permission for everything, or every awkward case quietly becomes an "exception" that the reviewer decides anyway. There are two easy answers when a senior person overrides in public, and both are wrong. OVERRULE THEM and the message lands that the rule outranks judgment -- so within a month people follow a rule they can see is wrong, which is worse than having none. WAVE IT THROUGH with a shrug and the rule is decorative by Thursday. THE OVERRIDE LOG ================================================================ DATE | CASE | FOLLOWED / OVERRIDDEN | REASON (one sentence) | OUTCOME -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | -----+------+-----------------------+-----------------------+--------- | | | | (Copy this block as often as you need. A rule that records only its failures teaches you nothing; one that records only its successes is a marketing document.) AT THE REVISION ================================================================ Read the quarter's overrides together, not one at a time. What you are looking for is a SHAPE -- four overrides that share a reason are the rule telling you what it's blind to. Overrides this quarter: ______ Shape they share: ________________ [ ] The rule CHANGES. New wording: ______________________________________________________________ [ ] The rule is CONFIRMED, on the evidence. (Confirmed twice running? That's a flag, not a comfort -- either it's genuinely settled or nobody's reading. Ask the proposer, not the reviewer. They'll know.) [ ] The RECORD needs a new field, because the rule is blind to something the record never captured. (This is the loop closing in the other direction, and it's often the most valuable outcome of a revision.) Next revision date: __________________________________________ TWO TRAPS ================================================================ THE RULE THAT'S REALLY AN AVERAGE. Write "do whatever usually works" and you have encoded the middle of your record and thrown away the edges -- and the edges are frequently where your advantage lives. A good rule handles the common case and NAMES what it doesn't cover. MORE THAN ONE RULE. See the top of this page. One. Really. ePORTS — your ecom source. eports.global