The one-page ground rule

The situation

A department of sixty receives the new AI policy. Twelve pages, reviewed by legal, genuinely thorough. Everyone clicks the box that says they have read and understood it. Attestation rate: one hundred per cent.

Six weeks later a sales coordinator pastes a customer list into a personal account to clean up the formatting. The policy covers this — page seven, clause 4.2, second bullet. It was correct, current, and completely absent from the four seconds in which the decision was made.

The same sixty people can all tell you where to assemble in a fire. That instruction is one sentence long.

The principle

A rule only changes behaviour if it is short enough to recall at the moment of the decision, which is why one page beats twelve every time.

This is not an argument against the long policy. The long policy has a job: contractual cover, audit evidence, a defensible answer when a regulator asks. It is a reference document. What it is not is an operating instruction, and organisations get into trouble by publishing the first and assuming they have issued the second.

The operating instruction is one page, written in your team's own nouns, and it passes three tests. The recall test: someone can state the red line from memory a month later. The new-starter test: someone on their third day can apply it to a real document without asking. The edge test: it says what to do when the answer is unclear and names the person who decides, with a time by which they answer.

Categories are what break the recall test. "Confidential information" makes everyone guess. "The pricing sheet, the customer list, the interview notes" makes it decidable in four seconds, which is all the time the decision ever gets.

This is the page the tool map from Unit 3 points at. The map says where work goes; this says what never goes, and when the two disagree, this one wins.

Worked example

Shown here in Gemini.

The one page is derived from the twelve, not written alongside it. Run it in one pass: the policy and a list of what your team actually uses AI for go in, one page comes out. Shown here for a commercial team at a mid-size distributor.

Job: draft an operating rule for my team, not a policy.

Inputs: our AI policy (attached) and the list of eight tasks my team
actually uses AI for (attached).

Output: the six sections of the template below, filled in. Every line
names a real document type this team handles — no generic categories.
200 words maximum.

Standard: nothing in RED may contradict the attached policy. Where the
policy is silent, write the line and mark it "not covered — decision
needed from the owner". Do not soften anything into guidance.

The template it fills:

AI GROUND RULE — <team> — owner <role> — reviewed <date>

1. WHAT WE USE IT FOR
   <three named tasks, one line each>

2. GREEN — paste without asking
   <named document types>

3. RED — never paste, no exceptions
   <named document types>

4. AMBER — ask first
   Ask <named role>. Answer within one working day.

5. CHECK BEFORE IT LEAVES
   <what must be traced to a source, by whom, before what>

6. SAY SO
   <where AI-assisted work is declared, and to whom>

Sections one to four came back filled:

1. WHAT WE USE IT FOR
   Drafting renewal and quote-follow-up emails
   Summarising call notes into the CRM
   Rewriting supplier terms into plain language for internal review

2. GREEN — paste without asking
   Published price list. Marketing copy. Public tender documents.
   Product spec sheets. Anything already on the website.

3. RED — never paste, no exceptions
   Customer lists. Signed contracts. Anything from the HR folder.
   Unreleased pricing. Incident reports.

4. AMBER — ask first
   Ask the commercial director. Answer within one working day.
   Draft contracts, named-account call notes, anything with a
   customer's own document attached.

Five red items, all of them things a person on that team touches by name. Section four names one person and a deadline, which is what stops amber becoming a silent green. Budget an hour for the page, and expect most of it to go on the argument about section three — that argument is the valuable part, and writing down where it landed is what stops it recurring.

In your tool

The page has to sit where the work happens. Every one of these keeps it in front of people; none of them enforces it.

Product surfaces checked 2026-08-04.

Try it

Write sections two and three only — green and red. Use your own document names, not categories, and cap each line at eight words.

Then test it. Read the two lists aloud to a colleague once, put them away, and name three real documents that crossed your desk this week. Ask which pile each one goes in.

You are done when they sort all three the way you would, from one hearing, without asking a clarifying question. If they hesitate on one, the line covering it is written as a category. Name the document instead and test again.

Common failures

• The comprehensive draft. Every clause is defensible on its own and the total is unreadable. Coverage and adoption trade against each other, and the one page is the side of that trade nobody chooses by accident.
• Amber with no owner. "Check with your manager if you are unsure" names nobody and sets no deadline, so the question is never asked and the paste happens anyway.
• Written once, never dated. A rule with no owner and no review date describes the tools you had last year. Put both in the header, where the first thing anyone reads is how old the page is.