How human approval works in freight automation
Automate collection, checking, follow-up, and data entry first. Keep money, customer promises, compliance decisions, and uncertain exceptions with people. The agent should attach the source, explain why it stopped, and make the review faster.
Approval should follow the cost of being wrong
Sending a routine document request is easy to reverse. Approving a margin, promising a delivery date, accepting a customs classification, or paying an invoice can create a financial or legal problem. Those actions should not share the same autonomy setting.
NIST's AI Risk Management Framework recommends defining human roles, testing before deployment, and monitoring systems while they operate. For a freight team, that means writing down who decides, what the agent may do, and when it must stop.
Use four levels of authority
- Read. The agent collects emails, documents, portal data, and TMS records. It cannot change anything.
- Draft. It prepares the update, quote, dispute, or follow-up. A person reviews every action.
- Act within rules. It may complete a short list of tested actions. Exceptions still go to a person.
- Expand carefully. More actions become automatic only after the workflow proves accuracy and the team understands the failure cases.
Autonomy is not a switch. It is permission for a specific action under specific conditions.
Where the human gate belongs
| Workflow | Agent handles | Person approves |
|---|---|---|
| Milestone chasing | Checks missing events, sends routine follow-ups, files replies | Ambiguous replies and customer commitments |
| Exception monitoring | Tracks dates, calculates countdowns, gathers evidence | Operational response and commercial decision |
| Quote drafting | Extracts the RFQ, collects rates, applies approved rules | Final price, margin, and unusual routing |
| Invoice audit | Matches lines, finds differences, drafts a dispute | Payment holds, dispute submission, and settlement |
Shadow mode comes before production
In shadow mode, the agent works on live traffic but cannot send a message or write to the system of record. The team sees what it would have done, compares that with the human action, and records the difference.
This catches the failures a polished demo misses: vague replies, renamed attachments, duplicate booking numbers, branch-specific rules, stale rate sheets, and customers who omit required details.
Production permission should depend on evidence from this period. If the workflow still needs frequent correction, keep it in draft mode.
Every review needs the source
A useful approval screen shows the proposed action beside the email, rate sheet, invoice line, tracking event, or TMS record that supports it. The reviewer should not have to search three systems to confirm one draft.
Each action should also retain the input, rule used, output, reviewer, timestamp, and final result. That record makes an error diagnosable instead of mysterious.
Low confidence is a routing rule
Confidence should decide where work goes, not whether the agent hides its uncertainty. A missing field should trigger a clarification request. Conflicting sources should go to a person. A portal failure should create a visible exception instead of a silent gap.
The right threshold depends on the action. Filing a copy of a document may tolerate more uncertainty than changing a billed amount.
Monitor the workflow after launch
Document layouts, carrier portals, customer habits, and internal rules change. A workflow that worked last month can drift. Track accuracy, human overrides, low-confidence cases, failed writes, and incidents. Track the business result as well: staff hours returned, response time, cost flagged, or extra work handled.
Someone must own the review. Monitoring cannot be another dashboard that nobody checks.
Questions to ask before giving an agent access
- Which systems and records can it read?
- Which actions can it take without approval?
- What information appears with each review request?
- How are failed actions and low-confidence cases surfaced?
- Who can pause the workflow?
- What is logged, and how long is the record retained?
- Who fixes the workflow when a source or rule changes?
Frequently asked questions
Keep final pricing, customer promises, disputes, payments, compliance decisions, and low-confidence exceptions behind human approval. Agents can gather evidence and prepare the action.
In shadow mode, the agent processes live work and drafts the next action without sending messages or changing the system of record. The team compares each draft with the action a person would take.
Yes, for bounded and reversible work that has passed testing, such as filing a document or sending an approved follow-up. The allowed actions should be explicit, logged, and monitored.
Track accuracy, overrides, low-confidence cases, failed writes, response time, and the business result. Review incidents and changes to document formats, carrier portals, and workflow rules.
Sources
Map the approval points
Bring one live workflow. We will identify what an agent can handle, what stays with your team, and how to measure the first build.