By Katari Technologies Ltd · Published
Two kinds of AI, two jobs
Machine learning looks at your history and points: a forecast of next month’s spend, an attendance pattern that looks unusual, an invoice that does not look like the others. Generative AI helps in language: it answers a question in plain words or explains how something works.
Neither should act alone. Both are best treated as advisers.
Four rules for safe use
Whichever product you use, ask four questions.
- Access: does the assistant see only what the person asking is allowed to see?
- Preview: is anything it would change shown as a preview first?
- Approval: does a person have to approve the change?
- Audit: is every action recorded against a person?
Label what the machine produced
A forecast or a flag should always be labelled as one, with the reason shown, so nobody mistakes a statistical pattern for a finding of fact. Models should be tested before they are switched on and be possible to roll back.
Where to start
Begin with questions, not actions. Let the assistant answer things people already ask each other all day: who is working tomorrow, which invoices are overdue, what is due for service this month. Check its answers against what you know, and note where it is wrong.
Once people trust the answers, consider forecasts and flags for the areas where a second look is most valuable, such as unusual invoices. Keep the final decision with a person, every time.
Bring your own assistant
Many people already use an AI assistant. The open Model Context Protocol (MCP) lets such an assistant connect to a system on the person’s behalf, with the same permissions that person already has.
Kinetic VOC includes an assistant that follows these rules, and lets people connect the assistant they already use through MCP, with the same permission checks, previews, approvals and audit trail.