Purpose & audience

• Tell users what the AI can/can’t do, when to escalate, and how we handle their data.
• Audience splits: general users, admins/operators, and auditors/regulators.

Reusable template (copy/paste)

• Intended use: <1-2 sentences> including exclusions.
• How to use: step list (max 7), include examples and safe prompts.
• Known limits: hallucination risk, data freshness, domain gaps, escalation triggers.
• Privacy: what is logged, retention, opt-out, sensitive data cautions.
• Human support: how to reach a person; response times.
• Last updated: version and date; link to change log.

Do/Don't guidance

• Do: provide context, verify critical outputs, cite sources where available.
• Don’t: paste secrets; rely on AI for legal/medical/financial decisions without review.

Escalation to humans

• Define thresholds (confidence low, data missing, policy-sensitive topics); provide one-click handoff.

Privacy & data entry cautions

• Mark fields likely to capture personal or confidential data; add inline warnings and consent links.

Localisation & accessibility

• Readability at B1-B2; alt-text; high-contrast options; translations reviewed by domain experts.

Versioning & records

• Archive prior instruction sets; track comprehension test results; link to release versions.

Placement patterns

• Onboarding interstitial; inline hints; “learn more” drawer; downloadable PDF for auditors.

User comprehension tests

• Multiple-choice checks; misinterpretation logs; A/B testing of copy; pass thresholds for release.

Channel samples

• Web widget; email auto-responses; IVR scripts; mobile push notice; CRM bot quick-reply text.

Common pitfalls

• Vague “limitations”; missing human contact; no copy versioning; not localised for key markets.

Authoring checklist

• Purpose clear; steps actionable; limits explicit; escalation easy; privacy stated; records kept.

© Zen AI Governance UK Ltd • Regulatory Knowledge • v1 05 Nov 2025 • This page is general guidance, not legal advice.