Training Your Organisation

The AI acceptable use policy: what to include

An AI acceptable use policy (AUP) is the foundation of your organisational programme. It defines what employees can and cannot do with AI tools, and it provides the basis for training, monitoring, and enforcement.

What to include:

1. Approved tools and how to access them. List every approved AI tool by name, its approved use cases, and how to access it. "Use the Company AI Assistant at assistant.internal.company.com for text analysis, summarisation, and drafting. Use GitHub Copilot (enterprise licence) for code assistance. No other AI tools are approved for work data."

Be specific. "AI tools approved by IT" is not specific enough — employees do not know which tools IT has approved. Name them.

2. Data classification rules for AI. Map your AI classification framework (Module 2) to practical guidance: "You may use the AI Assistant for Internal data. You may use it for Confidential data only through the privacy gateway (the gateway is automatic — you do not need to do anything special). You may not use any AI tool for Restricted or Prohibited data without explicit approval from the Data Privacy team."

3. Prohibited actions. Be explicit: "Do not paste customer PII into any AI tool not listed as approved. Do not use personal AI accounts (free ChatGPT, personal Claude, etc.) for any work-related task. Do not upload company documents to any AI tool without checking the classification level. Do not use AI tools to process data under legal hold."

4. The rationale. Explain why, not just what. "Consumer AI products may use your input to train future models. This means confidential information you paste into a free AI tool could influence the tool's responses to other users, including competitors. The Company AI Assistant does not have this risk because it runs through our privacy gateway."

Employees who understand the reasoning are more likely to follow the policy than employees who see it as arbitrary bureaucracy.

5. How to request new tools or use cases. Provide a clear process for employees who need an AI capability that is not covered by approved tools. "If you need an AI capability not covered by the approved tools, submit a request to [email protected]. Include: the task you need to accomplish, the data involved, and why the approved tools are insufficient. Requests are reviewed within 5 business days."

If the request process is slow or opaque, employees will bypass it and use unapproved tools.

6. Consequences of violations. State the consequences clearly but proportionally. First violations of the AUP should typically result in re-training, not termination. Repeated or intentional violations (deliberately exfiltrating data via AI tools) warrant escalation. The goal is compliance, not punishment.

What NOT to include:

• Overly restrictive rules that block legitimate productivity (e.g., "AI may only be used for tasks pre-approved by management")
• Technical jargon that non-technical employees cannot understand
• Legal disclaimers longer than the actual policy
• References to specific model versions or vendor features that will be outdated within months

Different roles, different training

A one-size-fits-all AI privacy training is ineffective. A developer's AI privacy risks are different from an executive's. A customer support agent's risks are different from an HR analyst's. Role-based training addresses each audience's specific risk profile and use cases.

Executives and board members (30-minute briefing)

• The business case: AI adoption benefits and the competitive cost of inaction
• The risk landscape: regulatory exposure, reputational risk, real incidents
• The organisation's AI privacy architecture: high-level overview (the gateway exists, local inference handles sensitive data)
• Their role: champion the programme, fund the infrastructure, model compliant behaviour
• What they need to know: do not paste board materials into personal AI tools, do not discuss M&A targets with AI on personal devices

Developers and engineers (2-hour technical training)

• How the privacy pipeline works: detection, redaction, routing (technical detail)
• How to integrate with the gateway API: SDKs, endpoints, error handling
• What the gateway cannot protect against: data in code comments, data in commit messages, API keys in prompts
• Code-specific risks: AI code assistants sending surrounding code context to cloud services
• Their role: build applications that route through the gateway, flag detection gaps

Analysts and knowledge workers (1-hour interactive session)

• Which tools are approved and how to access them
• Data classification: how to determine if data can be used with AI (simplified decision tree)
• Practical demonstrations: "Here is how you summarise a report using the approved tool"
• Common mistakes: pasting spreadsheet data with customer names, uploading documents without checking classification
• Their role: use approved tools, report issues, suggest improvements

Customer support and operations (45-minute session)

• Which tools are approved for customer interaction data
• What NOT to paste into AI tools: customer account numbers, support ticket details with PII, chat transcripts
• How the approved tools protect customer data (simplified gateway explanation)
• Their role: follow the data handling procedures, escalate questions about unfamiliar data types

HR and legal (1-hour session)

• Employee data sensitivity: PII, compensation data, performance reviews, health information
• Legal privilege: AI processing of privileged documents may waive privilege in some jurisdictions
• HR-specific risk: using AI to screen candidates or make employment decisions creates AI discrimination risk
• Their role: classify HR and legal data as Restricted, use local inference for analysis

AI champions and adoption metrics

The AI champion model

The most effective organisational pattern for AI privacy adoption is the "AI champion" model: identify one person in each department who becomes the local expert on approved AI tools and privacy practices.

AI champions are not security enforcers. They are enablers — colleagues who help others use AI tools effectively within the policy framework. They:

• Know the approved tools deeply and can demonstrate them to colleagues
• Understand the data classification framework and help colleagues classify edge cases
• Serve as the first point of contact for AI privacy questions (before escalating to the Data Privacy team)
• Report common friction points and shadow AI patterns to the central team
• Test new AI capabilities and provide feedback on detection accuracy

Select AI champions who are: enthusiastic about AI (they want to be using it), respected by their peers (influence matters more than authority), and willing to invest 2-4 hours per month in champion activities.

Support AI champions with: early access to new AI features, a dedicated Slack/Teams channel for champion coordination, monthly briefings on pipeline updates and new approved tools, and recognition from leadership.

Measuring adoption and compliance

You cannot improve what you do not measure. Track these metrics monthly:

Adoption metrics:

• Active users of approved AI tools (daily active, weekly active, monthly active)
• Requests processed through the gateway (volume and trend)
• Requests processed by local inference (volume and trend)
• Percentage of the workforce using approved AI tools at least weekly

Compliance metrics:

• Shadow AI detection rate (consumer AI domain access from corporate networks)
• Policy violations reported and addressed
• Training completion rate by department
• Time to resolve AI privacy requests (from the request process described in the AUP)

Pipeline effectiveness metrics:

• PII detection rate (from red-team and production monitoring)
• False positive rate (from user feedback)
• Gateway availability and latency

Culture metrics:

• AI champion activity level (questions answered, training sessions conducted)
• Employee satisfaction with approved AI tools (quarterly survey)
• Number of new AI use case requests (indicates engagement, not just compliance)

The dashboard

Build a simple dashboard that presents these metrics to leadership monthly. The story the dashboard should tell: "AI adoption is increasing, shadow AI is decreasing, the pipeline is catching PII effectively, and employee satisfaction with approved tools is high." If any of these trends reverse, the dashboard triggers investigation.

Making people want to follow the policy

The most effective AI privacy programmes are the ones where compliance is a byproduct of good tool design, not a result of enforcement. Here are the principles:

Make the right thing the easiest thing. If the approved AI tool requires fewer clicks, less setup, and delivers comparable results to consumer ChatGPT, employees will use it. Invest in UX for your internal AI tools. Single sign-on, no separate account creation, a familiar chat interface, and near-instant responses. Every friction point drives employees toward the frictionless consumer alternative.

Make privacy invisible. The gateway should be transparent to users. They should not need to think about PII detection, redaction, or routing. They type a prompt, they get a response. The privacy machinery runs behind the scenes. If users have to manually classify their data or confirm that their prompt does not contain PII, compliance becomes a burden that degrades over time.

Celebrate AI usage, not just compliance. Share stories of employees who achieved great results using approved AI tools. "The marketing team used the AI Assistant to analyse 6 months of campaign data and identified a segment that increased conversion by 23%" is more motivating than "100% of employees completed AI privacy training."

Respond to feedback quickly. When an employee reports that the gateway blocked a legitimate query, or that the approved tool cannot do something they need, respond within 24-48 hours. Even if the resolution takes longer, the acknowledgment shows that the organisation cares about enabling AI usage, not just restricting it.

Senior leadership must use the tools visibly. If the CEO uses the approved AI tool in team meetings, sends AI-generated analysis to the board (through approved channels), and talks about AI as a productivity advantage, the cultural signal is powerful. If leadership talks about AI adoption but uses consumer tools themselves, employees notice the hypocrisy.

Address fear directly. Many employees fear that AI privacy monitoring means their employer is reading their prompts. Be transparent: "The system logs what types of PII were detected and how they were handled. It does not log the content of your prompts. Your manager cannot see what you asked the AI. The audit trail records privacy compliance, not employee activity."

Practical: training deck outline and policy template

Training deck outline (all-staff version, 45 minutes):

Slide 1-3: Why we are doing this

• AI makes us more productive (examples from our industry)
• But: real incidents at other companies (Samsung, law firm — Module 1 examples)
• Our approach: enable AI with privacy controls, not block AI

Slide 4-6: Our approved tools

• Tool name, what it does, how to access it (screenshots)
• "These tools have our privacy gateway built in — your data is protected automatically"
• What you can and cannot use each tool for (simple table)

Slide 7-9: Data classification (simplified)

• Three practical categories: "Go ahead" (Public, Internal) / "Use approved tools" (Confidential) / "Ask first" (Restricted, Prohibited)
• Examples relevant to each department
• How to determine which category your data falls into (simplified decision tree)

Slide 10-12: What not to do

• Do not use personal AI accounts for work
• Do not upload company documents to consumer AI tools
• Do not paste customer data into unapproved tools
• "If you are unsure, ask your AI champion or email [email protected]"

Slide 13-14: How our privacy architecture works (simplified)

• Simple diagram: Your prompt → Privacy gateway → Safe AI → Response
• "The gateway automatically removes sensitive information before it reaches the AI service"

Slide 15: Questions and your AI champion

• Introduce the department's AI champion
• Contact information for questions

Policy template (key sections):

AI ACCEPTABLE USE POLICY

1. PURPOSE
This policy defines acceptable use of AI tools at [Company].
It exists to enable productive AI use while protecting company
and customer data.

2. APPROVED TOOLS
- [Company AI Assistant] — [URL] — for text analysis,
  summarisation, drafting, data analysis
- [GitHub Copilot Enterprise] — for code assistance
- [Tool 3] — for [purpose]

No other AI tools are approved for work data.

3. DATA RULES
- Public and Internal data: may be used with approved tools
- Confidential data: may be used with approved tools
  (the privacy gateway protects it automatically)
- Restricted data: requires explicit approval from the
  Data Privacy team before AI processing
- Prohibited data: may not be processed by AI under
  any circumstances

4. PROHIBITED ACTIONS
- Using personal AI accounts (ChatGPT, Claude, Gemini, etc.)
  for work-related tasks
- Uploading company documents to unapproved AI tools
- Sharing AI-generated output that may contain reconstructed
  sensitive information without review
- Using AI for decisions about individuals (hiring, performance,
  compensation) without HR and Legal approval

5. REQUESTING NEW TOOLS OR USE CASES
Submit requests to ai-privacy@[company].com. Include:
the task, the data involved, why approved tools are
insufficient. Response within 5 business days.

6. REPORTING CONCERNS
If you believe sensitive data was exposed through an AI tool,
contact security@[company].com immediately. Prompt reporting
is valued and will not result in disciplinary action.

7. VIOLATIONS
First violation: retraining and discussion with manager.
Repeated violations: escalation per [Company] disciplinary
procedure.

Effective date: [Date]
Review date: [Date + 6 months]
Owner: [CISO / DPO / AI Governance Team]