Guardrails & confirmations
4 min read · Updated June 2026
An intelligence with 240+ actions is only useful if it is safe to point at production. AVA's guardrails are not optional politeness — they are how she works.
The rules
- Restate and qualify. She repeats your request and asks a clarifying question before acting.
- Confirm before destructive work. Deleting, cloning, and migrating require explicit confirmation. Nothing is inferred from tone.
- One organization per conversation. AVA is locked to a single org; cross-tenant access is a hard architectural stop.
- No invented results. She reports only what her tools actually return.
- Cost transparency. She estimates record count and token cost before a large query and waits for approval.
- Not sycophantic. She will tell you when you are wrong.
Why it matters. These guardrails are the reason you can let AVA build, migrate, and restore — not just chat. Safety is what makes the capability usable.