Brokers already run a delegation ladder for people: a new dispatcher can update internal notes on day one, can message carriers after a few weeks, and can commit margin only with a manager's sign-off. AI freight agents need the same ladder, enforced by the system rather than by habit.
The freight action safety ladder
- 1Free: read and interpretReading load state, tracking events, documents, and carrier data is safe by construction. So is flagging exceptions and scoring risk. Let agents do this everywhere, immediately.
- 2Free with review: internal draftsDrafting next steps, internal notes, and recommended actions changes nothing in the outside world. The draft itself is the safety mechanism — humans see the agent's judgment before it ever matters.
- 3Approval-gated: external commitmentsTendering freight, sending customer-facing messages, requesting detention, changing rates, committing to a carrier. These bind the company, so they route through human approval with the agent's evidence attached.
- 4Blocked: policy overridesOverriding carrier risk holds, service policies, or financial limits should be impossible for the agent regardless of confidence — these escalate to a person, full stop.
What makes the ladder enforceable
- Explicit actions: the agent chooses from named operations, not free-form access — you cannot gate what you cannot name.
- Policy answers per request: every action returns allowed, approval required, blocked, or missing context, so the agent always knows its standing.
- Approval packets: when a human decision is needed, the agent delivers the context, evidence, and proposed step in one reviewable unit.
- Audit records: every action — taken, approved, or blocked — is written down with its reasoning, so trust is verifiable rather than assumed.
- Idempotency: retries and duplicates are detected by the system, because an agent that double-tenders a load is worse than no agent at all.
Haulbase enforces this ladder in the platform: the Haulbase Agent works read-and-draft by default with approval routing on sensitive actions, and Headless Haulbase exposes the same permission engine to agents you build yourself.