AI assistants can misunderstand a request or produce incorrect information. Controls are designed into each implementation to keep people involved where it matters.
The controls listed here form part of the proposed implementation and are agreed with you. They reduce risk; they do not guarantee accuracy or regulatory compliance.
- Approved information
- Agreed sources, with a process for keeping them current.
- Limited access
- Only the information and actions the assistant needs.
- Human approval
- A person approves consequential actions before they happen.
- Clear handover
- Requests that need a person are passed to one, with context.
- Testing
- Checks for incorrect answers, missing information and failed integrations.
- Monitoring and pause
- Ongoing review, and a way to pause a workflow if needed.