Skip to content

Approvals: Your Review Queues

Everything VMXpert and its agents want to do to your data comes to you first. There are two queues, both at the top of Admin → Toolbox → VMXpert → Agents, and they hold different kinds of requests.

📥 Review Inbox — proposed data changes

Every proposed change to your POS data lands here — whether it came from an agent's scheduled run or from a chat conversation. Each item shows:

  • Where it came from (Agent · name, or Chat) and whose it is — filter by source or by owner
  • When it was proposed
  • A before/after preview of exactly what would change
  • The agent's own summary of why

Approve applies the change through the normal audit trail — the change log, permission checks, and website sync all fire exactly as if the owner had made the change by hand. Reject discards it; nothing happens.

Chat proposals don't vanish

If someone starts a change in chat and never clicks Confirm, the proposal doesn't disappear — it waits here in the Review Inbox until someone approves or rejects it.

🛠️ Tool Approvals — actions awaiting an OK

When a capability is set to Approve on the Managing AI Access page (say, sending outside email, or database lookups for a particular person), each use becomes a request in this second queue. Every request leads with what's being asked in plain terms — the capability, the action, and the key details — with the full technical detail available behind a disclosure if you want it. Approve lets that one action run; Deny stops it.

Where else approvals appear

  • In chat — the person chatting sees the same requests inline: a "Review before applying" card for data changes, an "Approval required" card for gated actions. Chat and the admin queues are two views of the same thing.
  • On each agent's dashboard — its own pending proposals, so you can review one agent's work in context.
  • Resolved items stay visible in the agent's Activity feed with their outcome.

The rules that hold it all together

  1. Nothing is applied without a person's OK. Agents and routines cannot write data on their own — they can only propose.
  2. Approving acts as the owner. A change is applied with the owner's permissions and identity, exactly as if they'd made it themselves — so someone's agent can never do what they themselves couldn't.
  3. There's no undo. The review is the safety step — read the before/after preview before approving.