ADR-002 — Product Shape: Supervised Operations

Status: Accepted
Date: 2026-05-21

Context

AI agents can operate in two modes: fully autonomous (act without human review) or supervised (draft for human approval before executing irreversible actions).

Decision

Supervised operations. Agents draft all customer-facing and irreversible actions. Humans approve before execution.

Rationale

E-commerce operators cannot afford mistakes on refunds, cancellations, or replies to unhappy customers. Trust is built gradually. Supervised mode lets the operator stay in control while the agent handles the cognitive load of drafting.

The confidence threshold (≥80%) determines whether the agent drafts at all. Below threshold, the conversation is escalated to a human queue with an explanatory note.

Approval-gated actions

Action	Gated?
Send reply	Yes
Issue refund	Yes
Cancel order	Yes
Resolve conversation	Yes
Add internal note	Yes
Assign conversation	No — internal, reversible

Consequences

• Every agent run produces either an approval_request (pending human action) or an escalation note
• The approval queue UI is a core product feature, not an afterthought
• Filament is used for the approval dashboard (minimal build cost)