Acceptable use for staff
- Lesson 11
- 30 min
- Module 4: Policies people can follow
Where we are
In Module 3, we listed Larkfield’s eight systems, gave each one a tier and assessed the riskiest change. That work covers the systems that teams build or buy. It does not tell a member of staff what they may paste into an AI tool this afternoon. In this lesson, we write rules that a new employee can follow alone, as a swimming pool does with the sign at its gate.
Lesson 11 of 19 · AI Governance: Policies, Risk and AccountabilityAcceptable use for staff
1. What governance is for
2. Laws and standards in plain words
3. Inventory, risk tiers and impact assessment
4. Policies people can follow
5. Controls, documentation and evidence
What you will be able to do
Write a one-page policy for every member of staff. It says which data may go into which AI tool, what a person must check, and what is never allowed. A new employee must be able to answer “may I…?” from the page alone.
Before you startWhat to know and to have ready
What you need to know
- Larkfield’s eight systems and their owners, from Meet Larkfield’s AI portfolio.
- The tiers, and the controls that every tier has, from Risk tiers and proportionate controls. “Acceptable-use rules apply” is one of them.
What you need
- The draft acceptable-use policy, with its three “may I…?” questions, copied into your document tool.
- Larkfield’s one-page AI policy, as Farida found it, for comparison.
The sign at the swimming pool
Section titled “The sign at the swimming pool”A public swimming pool has a sign at its gate. It does not say “swim safely”. It says: no running on the side, and no diving in the shallow end. Children under 8 swim with an adult. Call the lifeguard if someone is in trouble.
Each line answers a question that a visitor really has, and a child can apply it alone. The sign does not need a lifeguard to explain it. And the pool makes the right choice easy: there is a deep end for diving and a shallow pool for small children.
A pool with a sign that says “be careful” still has rules. But nobody can follow them, because nobody knows what they are.
The problem: a policy that nobody can apply
Section titled “The problem: a policy that nobody can apply”Every member of Larkfield’s staff can open an AI chatbot in a browser today. So Tomás wrote a draft acceptable-use policy in one afternoon (constructed, like all of Larkfield). It has seven sections, and it sounds firm. Farida tests it with three questions that staff really ask:
- May I paste a customer’s complaint, with her name and order number, into a free chatbot to help me write a polite reply?
- May I use the meeting-notes tool in a call with a supplier?
- May I let the reply suggester in the helpdesk tool write a reply and send it without reading it, when the queue is long?
She asks each question of the draft alone, as a new employee would. Not one gets a clear answer.
Text description of the diagram
The draft is on the left and Farida’s notes are on the right. Section 2 says staff should use approved tools where possible, and that IT keeps the list. The note: “should” and “where possible” allow any tool, and nobody can see the list. Section 3 says never to put confidential or sensitive data into any AI tool. It allows customer data in tools the company provides, and asks for care with personal data. The note: no word is defined. Is a complaint with a name sensitive? Is a free chatbot “provided”? Question 1 fails here. Section 4 says always check AI output. The note: check how, and who checks? A glance counts. Question 3 is answered only weakly. Section 5 forbids anything illegal and decisions about people. The note: a new employee cannot know what is illegal, and every reply is about a person. Section 6 says to ask IT for a new tool and not to use unapproved tools. The note: there is no form and no answer time, and nothing about a new AI feature in a tool Larkfield already has. Question 2 fails here.
The draft fails in a second way, and Larkfield has already seen it. In The AI inventory, the platform team found the weekly digest script. Each week it sends the customer messages of the week, with names and order numbers, to an AI service that Larkfield has no contract with. Somebody needed a weekly summary of complaints, and there was no approved way to get one. In one month the script sent 1,150 customer messages, and its bill line was 87% of Larkfield’s AI spend (constructed). The old policy already said “do not put confidential information into AI tools”. A rule that says only “do not” does not stop the work. It moves the work out of sight.
The idea: an acceptable-use policy
Section titled “The idea: an acceptable-use policy”An acceptable-use policy is the short set of rules that tells every member of staff how they may use AI tools at work. “Acceptable use” is an old name from IT: the rules for company laptops and e-mail. For AI, a useful one answers four questions, each with an example:
- Which data may go into which tool? A short list of approved tools, and for each one the kinds of data that it may receive.
- What must a person check? Who reads an output in full before it reaches a customer or a supplier, and who approves an action.
- What is never allowed? A few lines with no exceptions, in words that a new employee recognises.
- What is the approved path? How to ask for a new tool, and how fast the answer comes. And how to report a mistake without fear.
The policy needs data classes: named kinds of data, each with examples, so that “sensitive” means the same to everyone. Four are enough for Larkfield. Public is anything already published. Internal covers plans and prices before they are published. Customer covers names, messages, order numbers and addresses. Sensitive covers payment details, health, children’s data, passwords and keys.
For organisations in the scope of the EU AI Act, the policy also helps with one duty. The Act asks providers and deployers to “take measures to support the development of” AI literacy among their staff (Digital Omnibus on AI, Regulation (EU) 2026/1744, new Article 4, official source, checked 22 October 2026). A short policy with examples is one such measure. Whether it is enough for your organisation is a question for a lawyer.
A tiny example: fix question 1
Section titled “A tiny example: fix question 1”Farida fixes one question first, so that she can see each step. She writes every intermediate result down.
Step 1: find the lines that answer it. Section 2 says “should use approved tools where possible”. Section 3 says “never confidential or sensitive data”, and allows customer data in “tools that the company provides”.
Step 2: write what a new employee would conclude. A complaint with a name is not called “sensitive” anywhere. A free chatbot is “provided” by its maker, if not by Larkfield. And “where possible” allows it when the approved tool is slow. So the result is: unclear, two readings.
Step 3: write the missing parts. Name the data classes (above). List the approved tools, each with the data classes it may receive:
| Approved tool | Public | Internal | Customer | Sensitive |
|---|---|---|---|---|
| Document assistant | yes | yes | no | never |
| Reply suggester in the helpdesk tool (once switched on) | yes | yes | yes | never |
| Meeting-notes tool (internal meetings only) | yes | yes | no | never |
| Any tool not on this list | yes | no | no | never |
Then add one “never” line. Never put customer or sensitive data into a tool that is not on the list. Never put sensitive data into any AI tool.
Step 4: ask the question again. A complaint with a name and an order number is customer data. A free chatbot is not on the list. So the answer is no, from the “never” line and the list, in under a minute. If the new employee still needs help to write the reply, the list sends them to the helpdesk tool, where the work belongs.
The same four steps work for the other two questions. Question 2 needs the list to say where the meeting-notes tool may be used. Question 3 needs section 4 to say who reads, and how much.
Make the approved path the easy one
Section titled “Make the approved path the easy one”The weekly digest script shows the real test of a policy: is the approved path faster than the workaround? Three things make it so.
- A list that people can find, with a use for each tool. The list sits on one intranet page, and the policy names that page.
- A short form with a fast answer. “New AI use” asks two questions: what do you want to do, and what data does it need? For a low-tier use, the owner and the risk lead answer within 2 working days. In the board’s first quarter, the fast path’s median was 2 working days (constructed).
- No blame for reporting. A member of staff who pasted the wrong text must say so the same day. That happens only if nobody is punished for saying it.
Text description of the diagram
The flow starts with one question: is the tool on the AI tools list? If it is not, the member of staff may use only public data. For anything else, they do not paste it, and they ask for the tool with the form “New AI use”. A low-tier use gets an answer within 2 working days. If the tool is on the list, the next question is the data class. Sensitive data (payment details, health, children’s data, passwords and keys) never goes into any tool. For customer or internal data, the list must allow that class for that tool; if it does not, they ask. Then: will the output reach a customer or a supplier, or act on one? If yes, the person who sends it reads it in full, and approves only what they have opened. If no, they use it. A note at the side: if something went wrong, tell your team lead and the risk lead the same day.
Discuss: how strict should it be?
Section titled “Discuss: how strict should it be?”Farida shows the four-step fix to the team.
- Tomás: “Make it simple. Ban every AI tool that IT has not approved, and add disciplinary action. People will follow it.”
- Omar: “Long lists slow everyone down. Trust people to be sensible, and keep one line: use your judgement.”
- Mei: “I want the data classes in it. Without them, ‘sensitive’ means something different in every team.”
- Sam: “I joined last month. Where would I even find the list?”
Before you read on, answer these:
- Which position would have stopped the weekly digest script?
- What does each of the others get right?
- What would you test the final page with?
The course’s answer. Neither Tomás’s ban nor Omar’s single line would have stopped the script. The draft already forbade unapproved tools, and the script ran anyway. A ban with no approved path hides use; it does not end it. “Use your judgement” fails question 1, because a sensible person can still think a complaint is not sensitive.
Mei is right: the data classes turn arguments into look-ups. Tomás is right that some lines need no exceptions: “never sensitive data in any tool” is one. Omar is right about speed, so the answer is a fast path, not fewer rules. Sam’s question is the test: the page must name where the list is. Then Farida tests the page with the three questions, and a fourth from a new employee.
What would change the answer: if Larkfield worked with health or children’s data every day, “sensitive” would cover most of its work. Then a list of approved uses would need more detail, and fewer tools would be on it.
When to write one, and what it costs
Section titled “When to write one, and what it costs”Write it as soon as staff can reach an AI tool, which today means now. A first page takes a day: the data classes, the list and the checks are short, but the examples need care. Keep it to one page and six lines or so, and review it every 6 months, because the list changes.
The cost is in keeping the list true. Every new tool or feature must reach the list, or staff go back to guessing. And a policy is only read if somebody teaches it: in the board’s first quarter, 42% of staff had completed the AI training (constructed).
Do not use the acceptable-use policy for things that are not staff choices. Whether the triage service may route tickets alone is a decision about a system. Its tier and its owner settle it, not a line for every member of staff.
Common mistake
A policy that only lawyers can read
Symptom: the policy is signed, but staff still ask their manager for every case. The questions are the same every week, and the answers differ between managers.
Cause: the policy uses words that need a judgement (“sensitive”, “appropriate”, “where possible”) and gives no examples. So each manager applies it in their own way.
Fix: test it with real “may I…?” questions from staff. Every line that leads to “ask someone” needs a defined word, an example or a place on the list. Ask a person who joined last month to read it, not the person who wrote it.
Guided practice: lines 1 to 3
Section titled “Guided practice: lines 1 to 3”-
Open your copy of the draft acceptable-use policy. Start a new page called “Use of AI tools at Larkfield”.
-
Write the owner (Farida), who approves it and when it is reviewed.
-
Line 1: write where the approved tools list is. Copy the three tools and their uses from the table above.
-
Line 2: write the four data classes, each with two or three examples.
-
Line 3: write the “never” lines. Use words that a new employee recognises: “a password”, not “credentials”.
Check: question 1 now gets a clear “no”, with the line numbers.
Your turn: lines 4 to 6, and the test
Section titled “Your turn: lines 4 to 6, and the test”Write the rest of the page. Line 4 says who checks an output before it reaches a customer or a supplier, and how. Line 5 says how to ask for a new tool or a new AI feature. Line 6 says what to do when something goes wrong. Then answer the three “may I…?” questions from your page alone. Write the line number beside each answer.
Show one possible answer
Use of AI tools at Larkfield. Owner: Farida. Approved by: the governance board. Review: every 6 months. (Constructed.)
- Use the approved tools. The list is on the intranet page “AI tools” and says what data each tool may receive. Today: the document assistant (internal and staff-only documents), the helpdesk reply suggester (customer messages, once switched on), the meeting-notes tool (internal meetings only).
- Data classes. Public: anything published. Internal: plans, prices before publication. Customer: names, messages, order numbers, addresses. Sensitive: payment details, health, children’s data, passwords and keys.
- Never put customer or sensitive data into a tool that is not on the list, and never put sensitive data into any AI tool. Never paste a password or an API key into any AI tool.
- A person checks before it goes out. Anything AI writes to a customer or a supplier is read in full by the person who sends it. An AI action on a customer (refund, replacement) needs the approval the tool asks for; never approve what you have not opened.
- New tool or new AI feature? Ask the risk lead with the form “New AI use” (two questions). Low tier: an answer within 2 working days. An AI feature that appears in a tool we already buy counts as new.
- Something went wrong? Tell your team lead and the risk lead the same day. Nobody is blamed for reporting a mistake.
The three questions: (1) no, lines 1 and 3; (2) not yet: internal meetings only, so ask the risk lead, lines 1 and 5; (3) no: the person who sends reads it in full, line 4.
Knowledge check
Section titled “Knowledge check”Check your understanding
A new employee asks: 'May I paste our unpublished price list into the document assistant to check my summary of it?' Answer from the rewritten policy alone, and say which lines you used.
Show the answer
Yes. Prices before publication are internal data (line 2). The document assistant is on the list and may receive internal documents (line 1). Nothing in line 3 forbids it, because the data is neither customer nor sensitive data. If the summary goes to a customer or a supplier, line 4 applies too: the sender reads it in full.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Check your understanding
Predict: Larkfield adds the line 'Use your judgement with AI tools' and removes the data classes, to keep the page short. Which of the three 'may I…?' questions stops getting a clear answer first, and why?
Show the answer
Question 1, the complaint in a free chatbot. Without the data classes, “customer data” is a judgement again: a member of staff can decide that a complaint is not sensitive and paste it. The weekly digest script shows that somebody already decided this way. Questions 2 and 3 still rest on the list and on line 4, so they last a little longer.
Does your answer have the main points of ours?
Sign in and enroll in the course to save your answers. Every check counts toward the certificate.
Recap
- An acceptable-use policy tells every member of staff which data may go into which tool, what a person checks, what is never allowed and how to ask.
- Data classes with examples make “sensitive” mean the same to everyone. A list of approved tools says which classes each tool may receive.
- Test the page with real “may I…?” questions, asked of the page alone. A line that leads to “ask someone” needs a word defined, an example or a place on the list.
- Make the approved path the easy one: a list people can find, a two-question form with a fast answer, and no blame for reporting. A ban with no approved path hides the work.
You can now write the rules for what staff do with AI tools. But staff choices are not the only way AI reaches Larkfield. Teams build new systems, and a tool that Larkfield already pays for can add an AI feature overnight. Next, you set the gates that each of them must pass.
Next: Building and buying AI