Anonymized field case · Cross-border logistics
Freight inquiry AI workflow: from intake to reviewable quote material
This is not a claim that autonomous quoting is complete. It is a process design for selecting a measurable entry point and placing rules, exceptions, handoff, and ownership in one workflow.
No client identity, price, route, contact, or internal document is included. The public material is the process-selection and system-design method.
The business starting point
An inquiry connects customer information, cargo fields, route, timing, pricing sources, and exception handling. Asking an agent to “complete the quote” compresses several decisions into an unobservable black box.
The first step is to separate receipt, extraction, missing-field detection, rule matching, exception detection, human confirmation, and write-back.
Why the first phase prepares a decision
While rules are still being verified, the system prepares a review package: extracted fields, missing information, rule source and validity, conflicts, and a draft. A commercial operator keeps authority over price and the external message.
Four design layers
- Intake: preserve inquiries from email, chat, forms, and attachments.
- Rules: retrieve sources with version, scope, and validity.
- Handoff: route missing information, conflicts, and high-risk cases to a named owner.
- Write-back: return the decision, correction, and new exception to the process record.
Six testable workflow nodes
The six nodes are not six labels around one model call. Every node has an observable input, output, acceptance condition, and stop condition.
| Node | Required output | Acceptance | Stop condition |
|---|---|---|---|
| 1. Intake record | Task ID, source, timestamp, input hash | Traceable to the original message or attachment | Untrusted source or unreadable attachment |
| 2. Field extraction | Structured fields with field-level source | No guessing; every value traces to the request | Ambiguous, conflicting, or unsupported field |
| 3. Completeness check | Missing fields and next question | Each required field passes or becomes a clarification | Destination, dimensions, cargo type, or another required value is absent |
| 4. Rule retrieval | Rule ID, version, scope, validity, and citation | One active source supports each consequential decision | Conflicting active versions or no authorized rule |
| 5. Review material | Candidate option, assumptions, risks, and open decisions | No unapproved price or external commitment | High-impact exception, unauthorized action, or irreversible risk |
| 6. Human decision and write-back | Decision, rationale, status, and resume point | The task and exception queue receive the result | Unknown owner or failed write-back |
Pre-launch acceptance criteria
- Every extracted field can be located in the original input; missing values are never guessed.
- Each pricing-relevant rule carries a source, version, scope, and validity period.
- Missing fields, rule conflict, unauthorized actions, and irreversible risk produce a defined stop.
- The handoff package includes the request, fields, rule citations, actions taken, stop reason, and next owner.
- No price, timing, or customer commitment is sent without business-owner approval.
- The human decision and correction return to the original task and exception review.
Three synthetic failure samples
These samples test the stop path. They do not represent a real customer, route, or price.
| Input change | Expected state | Prohibited action | Next owner |
|---|---|---|---|
| Origin and weight are present; destination and dimensions are absent | ASK_FOR_CLARIFICATION | Retrieve a price or draft a quote | Requester / business operator |
| Retrieval returns two rules still marked active | HANDOFF_RULE_CONFLICT | Select the newer or cheaper rule | Process owner |
| Fields are complete but the requested action sends a customer price directly | READY_FOR_HUMAN_REVIEW | Send, write a final price, or change permissions | Business owner |
The first public validation target is not “quote accuracy.” It is whether the workflow reliably distinguishes continue, clarify, and handoff while preserving enough evidence for a person to resume.
What the first pilot should measure
Track field completeness, follow-up questions, review time, rule conflicts, exception recovery, and write-back. Only discuss deeper automation after this bounded loop is stable.
Download the freight workflow evidence plan (CSV) to record input, acceptance, handoff, and evidence status by node.
What is known and what is not yet proven
The current evidence supports a clear workflow boundary: the inquiry can be separated into fields, rule sources, conflicts, human review, and write-back, with an input, basis, exit condition, and owner for each node.
It does not yet support a stable operating-outcome claim: the public material does not prove sustained review-time reduction, higher quote conversion, or revenue impact. Those remain hypotheses rather than reported results.
The next evidence should come from consecutive representative samples: field completeness, review time, rule conflict, handoff, and recovery. Customer and operating outcomes should only be discussed after the process evidence is stable.
Continue
Read the complete freight inquiry workflow breakdown, define the evidence window with the AI agent productivity scorecard, see how FDE connects field delivery and adoption, and place the case inside the AI agent work-ownership path.