Last verified 23 September 2026. Microsoft Forms can hide names, though in a small team answers may still point to someone. SharePoint pages keep a version history. Since February 2025 the EU AI Act asks organisations to build staff AI skills. Details for experts at the end of this book.

The idea in one page

A working rule is the shortest page that tells a colleague, in the middle of a task, what they may do now, when to stop, and whom to ask. It doesn't explain AI, and it doesn't replace your organisation's policy. It translates that policy into your team's own material.

Seven parts, one page.

Colours need your examples. "Yellow: internal, non-sensitive material" means nothing until a colleague sees "the group meeting agenda" next to it. People decide by matching their material to your examples. So the examples come from your team's real work, and the red ones include the cases people actually get wrong: the unpublished figure, the customer's email with the order number in it, the log file with a password.

Shadow AI tells you what's missing. Shadow AI is the use of tools the organisation hasn't approved. When people use personal accounts, it's usually because they have a real task and no legitimate route for it. A ban on its own pushes that use out of sight. Give the task a route, and keep the ban for what really must stop.

Test it like anything else you build. A rule works when a colleague who didn't write it reaches the same decision you would. So you write eight short cards, each a piece of material and a task, and write the right answer for each before anyone sees them. That's your test set with known answers. Three colleagues sort the cards using only the rule. Where they disagree with the answer, the wording is wrong, not the colleague.

Asking must be faster than guessing. An edge-case route that answers in two weeks teaches people to guess. Name a person and a deputy, and promise an acknowledgement time you can keep. The acknowledgement says who will decide, not that the answer is yes.

Where Level 3 stops. Your page covers one team and the tools already approved for it. Choosing and buying a tool, writing the rule for other teams, costing a rollout, or agreeing anything with the works council (the elected staff representatives) is Level 4: T13-L04 · Roll Out an AI Assistant to a Whole Department, With Access, Cost and Consent Settled First. So is keeping training records for the AI Act's literacy duty (T13-L05 · Keep Everyone's AI Skills Current, and Show It When Asked), though your card results are a fair start. When a question on your page turns into one of those, route it upwards and write down where it went.

Worked example: two teams, four quadrants

Both situations use the four-quadrant method from T01-L01 · From Idea to Working Result: The Four-Quadrant Method.

Lab: a research group of ten

Structure it. Mira, a postdoc in a cell-biology group: "Everyone in the group should be able to decide in a minute whether their material may go into an assistant, and which one, without asking me or the PI. The PI approves the rule. It has to work when I'm away on a beamtime trip." (Beamtime: a week of measurements at an outside facility.)

Understand it. She sends an anonymous pulse through Microsoft Forms to the eleven group members, with Record name unticked in the form's settings, three multiple-choice questions and no free-text field:

Nine people answer. Five tick Summarising papers, and six tick Don't know which one is allowed. Seven say the hardest answer to find is What to do after a mistake. None of this says who did what, which is the point. With a group this small, she reports only totals.

The answers change her plan. She'd been writing a list of what's forbidden. But the main gap is a route: people use personal accounts for abstracts because they don't know the university already offers an assistant for staff. She checks with the institute's IT page and the data protection contact. The university's assistant is approved for published and internal, non-personal material, signed in with the university account. Back to Structure it: "The rule's first line is the approved route and how to open it. The red list is short and specific. The mistake route is the second thing on the page."

Choose tools. A SharePoint page on the group's site, added as a tab in the group's Teams channel, because that's where people ask questions already, and because the page's version history shows who changed what. To add the tab, she opens the channel, clicks the + to the right of the channel name, chooses SharePoint, then Pages, picks her page and saves. Only an owner of the group's SharePoint site can add it; if that isn't you, ask whoever runs the site. The PI as the person accountable, the lab manager as deputy for edge cases. The institute's IT security contact for mistakes, because the group can't decide those alone. No new software.

Build it. She writes the page:

The last sentence of the mistake section went past the PI and the institute's data protection contact before it went on the page. A group can promise not to blame; it can't promise what the institute will do, so the wording stays with what the group decides.

Then the cards, the test set. She writes eight, each a short description of material and a task, and the right answer next to each, before anyone sees them:

Three group members who didn't write the rule sort the cards on their own, with the page open. Two get 8 of 8. The third gets 6: she calls card 4 green, because the lab manual "isn't secret", and card 8 yellow, because the names are gone. Card 4 shows that "our protocols" is too vague: people read "internal" as "confidential". Mira changes the yellow line to "Anything written for the group and not published, even if it isn't secret". Card 8 is a known trap, so it moves into the red line as an example: "sample IDs, even without names". All three rerun all eight cards and get 8 of 8.

The edge-case test: a tester posts a made-up question in the channel, marked TEST. Aylin answers the next morning, saying the question goes to the PI. The mistake test is read aloud: "You pasted a collaborator's draft into a personal account ten minutes ago." All three name the steps.

Then comes the part that makes it a group rule and not Mira's. Before her beamtime trip, Aylin runs the card test with a new master's student, alone, from the page. The student gets 7 of 8 and asks one question the page doesn't answer: "Can I use it for my thesis?" Aylin writes it down for the next review, with a link to the university's thesis rules.

What it taught her: the page took an evening. The cards found two sentences that three careful people read differently, and nothing else would have.

Company: customer operations, twelve people

Structure it. Jonas, team lead in customer operations at a lab-supplies distributor: "Everyone on the team should know whether customer material may go into the approved assistant, and what to do if they're unsure or made a mistake. It has to keep working when Sabine, who everyone asks now, is on leave."

Understand it. He reads the company's fourteen-page policy with Sabine and marks every sentence that changes what the team does on a Tuesday. There are six. Then he asks three colleagues to show him, without real data, how they use the assistant. One uses a browser extension that summarises whatever page is open, including the customer record in the CRM, the customer database. It isn't on the approved list; nobody knew it had to be. Another asks whether an order number is personal data. It isn't, on its own, but the order history attached to it names a person.

So the team's gap isn't the policy. It's the step between the policy and the customer record on the screen. Back to Structure it: "The rule names the approved assistant and says that browser extensions and connectors, the links that let an assistant read other apps, need their own approval. The red examples are the customer screens we actually use."

Choose tools. The same SharePoint page and Teams tab as the lab. The IT policy stays where it is, linked at the top. Edge cases go to Jonas, with Sabine's replacement as deputy. Mistakes go to IT's existing incident route, which the policy already names.

Build it. His page has the lab's shape. The red line reads: "Customer emails and order histories; anything from the CRM screen; contracts and prices agreed with a customer; staff records; passwords." One extra line: "Browser extensions, plug-ins and connectors that read your screen or mailbox need IT approval, even if the assistant behind them is approved."

His eight cards include "a product page from our public website, to shorten" (green) and "a customer complaint with the name removed, to categorise" (red: ask first). The first test round finds that two of three testers call the complaint yellow. He moves it into the red examples, word for word. The rerun passes.

The takeover test is the reason he started. Sabine's deputy runs the card test with a colleague who joins the week Sabine leaves. He passes, and the channel thread has two questions in its first month, both answered within the promised two days.

What goes wrong

A copy of the corporate policy. Symptom: the page is long, correct and useless in the middle of a task. Fix: link the policy, and put only the decisions that change what your team does, with your team's examples.

A ban with no route. Symptom: the rule says "no AI", and personal accounts keep being used, now quietly. Fix: give permitted tasks a route, keep the ban for what must stop, and ask in the pulse what people need.

A brand name instead of a route. Symptom: "we use Copilot" makes every Copilot, account and plug-in look approved. Fix: name the account and how to recognise it on screen. Say that extensions and connectors need their own approval.

An edge-case route nobody answers. Symptom: questions go into the channel and get opinions or silence. Fix: one person and a deputy, an acknowledgement time you can keep, and "until then, don't paste it".

Cards written after the test. Symptom: the right answers get adjusted to what testers said. Fix: write the answers before anyone sees the cards. A disagreement means you rewrite the page, then rerun all eight cards.

Two versions look current. Symptom: an old FAQ and the new page say different things. Fix: mark the old one Replaced by v1.0 with a link, and keep version, date and review date on the page itself.

The rule belongs to its writer. Symptom: when you're away, nobody knows whether it's still valid or who can change it. Fix: a named deputy who has run the card test without you, and a review date in both your calendars.

Do it yourself (3 hours, over two weeks)

1. Structure it (15 min). Write one sentence: who the rule is for, which work it covers, who approves it, and what should happen when you're away.
2. Understand it (45 min, plus a week for answers). Send the three-question pulse with names switched off. Read your organisation's policy and mark every sentence that changes a Tuesday. Ask two colleagues to show you, without real data, how they use AI today. Rewrite your sentence if the gap turned out to be different.
3. Choose tools (15 min). Pick where the page lives, the approved route, the person for edge cases and a deputy, and the mistake route. Check the route with whoever approves tools in your organisation.
4. Build it: the page (45 min). Fill in the seven parts. Use two green, two yellow and four red examples from your team's real work. Have the person accountable approve the mistake section's wording.
5. Build it: the cards and the test (45 min). Write eight cards with their right answers. Three colleagues who didn't write the page sort them alone. Fix the wording, rerun all eight. Send one test question through the edge-case route and read the mistake case aloud.
6. Build it: the takeover (15 min). Your deputy runs the card test with someone new, without you. Put the page's link where the work starts, mark older guidance as replaced, and set the review date.

Exit check

What you hand in: one AI working rule page with:

• all seven parts, approved by the person accountable for the team, with version, maintainer, deputy and review date;
• the eight test cards with the right answers you wrote before testing;
• the results of three colleagues who didn't write it, and what you changed after the first round;
• the deputy's card test with someone new, run without you.

It passes when every red card is stopped, each tester gets at least seven of eight right, the edge-case question gets an answer from the named person or deputy, and the deputy runs the card test with someone new without you.

Rule to remember

A rule works when someone who didn't write it reaches the same decision you would.

Details for experts

Checked on 23 September 2026. You don't need this to follow the book; it's for setting things up or answering a colleague.

Further reading and tools

Next books

• T12-L01 · Decide What You May Paste Into AI at Work: the green, yellow and red rule your page applies.
• T13-L02 · Show a Colleague AI on Their Own Task, in 30 Minutes: the demo that usually comes before a rule.
• T12-L03 · Give an Unattended Workflow Limits and a Stop a Colleague Can Use: the operating rule for anything that runs without a person.
• T15-L01 · Use AI in Your Research and Say Exactly How: what goes in the methods section.
• T13-L04 · Roll Out an AI Assistant to a Whole Department, With Access, Cost and Consent Settled First: Level 4, when the question becomes which tool to buy and for whom.

Tools

• Microsoft Copilot: Catalogued. Often the approved assistant in organisations on Microsoft 365; whether it's approved for your material is your organisation's decision.
• Claude, ChatGPT, Gemini: Catalogued. Business plans can be approved for internal work; personal accounts belong in the green column only.
• Langdock: Catalogued. A workspace some German organisations run under their own contract as the approved route.