Ops assistants that must not invent fleet numbers
Why AI that sounds right about usage and SLAs is worse than silence—and how tool-grounded assistants keep chat honest.
Failure modes of ungrounded chat
- Calendar vs. cycle confusion — month totals that don’t match metered windows
- Wrong tenant scope — fluent answers about someone else’s fleet
- Stale memory — yesterday’s outage narrated as current
- Silent interpolation — filling gaps instead of saying “unknown”
The tool-first pattern
| Step | Responsibility |
|---|---|
| Understand the ask | Model |
| Choose a scoped tool | Model + allowlist |
| Fetch canonical data | Your platform APIs / helpers |
| Explain the result | Model |
| Refuse when tools fail | Product UX |
If the model can answer usage without a tool call, you have a design bug.
Evaluation questions for buyers
- Does a wrong service line ID hard-fail?
- Do billing answers cite a defined window?
- Are write actions confirmable and audited?
- When the provider is busy, does the UI say so?
Where Vigil stands
Vigil’s product thesis is that chat is another client of the same truth as Analytics and reports—not a second source of truth. That is why tools and allowlists matter more than clever prompts.