Supervised autonomy, explained
Every Odoo Employee follows the same mechanics — read state, propose one action, get checked, and leave a trail you can undo.
How one action happens
Same mechanics for every department, every action, every time.
- Step 1
Reads what's true right now
No memory between runs. Every action starts from the current Odoo state, the tenant's playbook, and whatever triggered it — nothing stale, nothing assumed.
- Step 2
Proposes exactly one action
The agent doesn't chain steps or improvise. It emits a single structured action — the same message a human would see and approve.
- Step 3
A separate executor validates it
The agent never touches Odoo directly. A separate executor checks the action against policy before anything happens.
- Step 4
Every action leaves an undo
The executor records the inverse of what it did. Undo isn't a special case — it's built into every write.
Autonomy is a dial, not a switch
Every action type is configured individually — not a single on/off setting for the whole agent.
Ask first
The agent proposes the action and waits. Nothing happens in Odoo until a human approves it in Slack, Teams, or the dashboard.
Act now, reversible
The agent acts immediately, and the action carries a working undo — like Gmail's undo send, for a whole team's work.
Some actions never get a choice
Refunds, apologies, and telling a customer they've been escalated always require a human's sign-off. That's hard-coded into the platform — no operator setting can turn it off.
Trust is earned by data, not a toggle
An action type only becomes autonomous once a human operator sees it clear five rolling metrics — promotion always needs a person to make the call.
Demotion doesn't need a person. The moment those metrics regress, the action automatically reverts to needing approval, and the team gets notified — tightening the leash never waits on someone noticing.
A second opinion, every time
No action ships on one AI's judgment alone.
A second AI reviews it
Every customer-facing action is checked by a separate AI acting as judge.
A confidence floor
Actions below the confidence threshold never ship without a human first.
Human spot-checks
10% of completed actions are reviewed afterward by a person and rated correct, partially correct, wrong, or severely wrong.
Ready to hire your first Odoo employee?
Book a demo and see Helpdesk handle a real ticket queue — live, supervised, undoable.
Book a demo