Assist before automating
Useful early applications include policy retrieval, summarising exceptions and preparing a draft analysis. These reduce searching without allowing a model to make employment or payroll decisions.
- Cite the underlying record or policy
- Label generated content
- Require human confirmation
Respect the tenant and role boundary
AI access should be no broader than the authenticated user’s existing permission. Prompts, retrieval and output need tenant-aware controls and protected logs.
- Never pool customer records by default
- Minimise personal data in prompts
- Keep model and retention choices explicit
Measure useful outcomes
Evaluate whether assistance reduces resolution time or improves completeness, while monitoring incorrect suggestions and user overrides.
- Start with reversible workflows
- Capture feedback and corrections
- Do not present planned AI as a current capability
Use this as an operating guide
Adapt the checklist to your policies, legal advice, workforce and implementation scope. A product demonstration should prove the representative workflows you identify here.