Every task type has an automation mode, set per department and overridable per task type at Workflow → Automation in the app.
The Modes
- Manual: the task is yours; the AI does not touch it. Use for the work you always want to see raw.
- AI Suggest (draft): the AI analyzes and drafts a recommendation with its reasoning. Nothing sends until you approve. This is the trust-building home.
- AI Approve: you approve/reject recommendations all day; the AI never executes alone. Many carriers run here for years and are happy.
- AI Auto: recommendations that clear the autonomy gate execute on their own; the rest still wait for you. The gate weighs your org's autonomy level, the task type's eligibility, and the AI's confidence, anything uncertain falls back to a human.
Mode parity matters: in every mode, outbound communications are captured and held: the AI composes into a draft envelope, and live send happens only on approval or a cleared auto-execute. There is no mode where the AI silently free-sends.
What Auto-Executes, and What Never Does
Auto-execution is earned per task type, and a hard set is human-held by design, on every plan, at every setting:
- Receivable invoices: approving and sending an invoice always requires a human. Amounts, recipients, and the backing POD are verified server-side; the AI can draft, you press the button.
- Safety finalizations: FMCSA inspections/violations and DOT annual inspection completions are enacted by explicit human finalization with evidence rows; terminal safety states cannot be resurrected or rewritten after the fact.
- Driver settlements: payables reviews are drafted by the AI, approved by you.
- Anything the guards refuse: when the deterministic guards cannot ground a task
in your real data, the AI returns
CANNOT_RECOMMENDinstead of guessing, in any mode.
Guards Run Before the AI
Before any model call, deterministic checks already know a lot: hours-of-service math (an 11-hour FMCSA check with real trip miles), floor rates, blocked brokers, duplicate document suppression, entity grounding (does this task's load/driver/broker actually exist and match?). Recommendations you see have already passed the arithmetic.
How Autonomy Grows
Rejects and overrides feed per-task-type calibration. The practical path most carriers take:
- Run AI Suggest for a few weeks
- Promote the boring, reversible types (status updates, reminders, check calls)
- Leave money and safety human-held, that ceiling is a feature, not a limitation
"We automate the predictable. You handle the exceptional."