Perspective

Nov 9, 2026 · Nugawi Intelligence · Perspective · 5 min read

Shipment to Settlement: A Freight Forwarder's Workflow Under the Evaluation Harness

Shipment to Settlement: A Freight Forwarder's Workflow Under the Evaluation Harness

A workflow on the bench, not a slide

Most logistics AI content stops at the demo. This piece walks a shipment-to-settlement workflow for a leading freight forwarder, stage by stage, through the evaluation harness Nugawi ships with every production build. The outcome numbers live in the published case study, linked at the end. What follows is the anatomy: which stages the agents run, which stay human, and how each stage is graded so the workflow holds on any given Tuesday with no vendor in the room.

The workflow, end to end

Shipment-to-settlement is the forwarder’s spine. Five stages:

Rate request and quote. A customer asks for a price on a lane; the request lands by email or portal with weights, volumes, incoterms and dates buried in prose and attachments.

Booking and dispatch. The approved quote becomes a booking with a carrier, and dispatch documents follow it out the door.

Tracking milestones. Carriers report departures, arrivals, transhipments and customs events across emails and portals, each in its own format.

Proof of delivery. The POD arrives as a scan or a photo, and its contents must be matched to the shipment record before anyone can bill.

Invoice and settlement. The settlement file is assembled from the quote, the booking and what was actually delivered, then invoiced, reconciled and closed.

Four of the five stages are mostly reading, retyping and checking: precisely the work agents are good at, provided somebody keeps score.

Where the agents run

Three stages carry agent traffic.

Quote drafting. The agent reads the inbound request, identifies the lane, pulls historical rates and surcharges from the operator’s own quote archive, checks them against the pricing matrix and drafts the quote for approval. It drafts from what this operator actually charged on these lanes, not a generic market feed.

Milestone capture. The agent watches carrier emails and portals, extracts each milestone, timestamps it and writes it to the system of record. Exceptions, a missed connection, a container that failed to load, a customs hold, are flagged to a person rather than silently absorbed.

Settlement document assembly. When the POD lands, the agent assembles the settlement pack: commercial invoice, freight invoice, credit notes for the variances between quoted and delivered. Data is keyed once, at the source, and the documents follow.

Where humans stay

Three decisions never leave the desk.

Exception handling. When a shipment goes sideways, a person owns the recovery. The agent surfaces the exception with the full history attached; it does not negotiate with carriers or customers.

Credit and pricing approvals. Anything outside the pricing matrix, a new customer on credit hold, a lane priced below floor, goes to a human with the agent’s draft attached. The agent can prepare the decision; it cannot be the decision.

Final sign-off. Nothing leaves the building unbound to a name. Every settlement file and every non-standard quote ships under a human signature, and the audit trail records whose.

How the harness grades every stage

This is the part most vendors skip. Each stage is graded by an evaluation harness: the test suite for agents, described in how production AI agents stay honest.

Golden cases come from the client’s real history: last year’s actual rate requests on its actual lanes, with the quotes the pricing team actually approved; the carrier emails that carried real milestones; the settlement files finance actually issued. The agent is graded against decisions this operator already made, not invented examples.

The regression set grows with every caught error. A misread incoterm, a milestone attributed to the wrong shipment, a credit note that missed a variance: each becomes a case that fails the release if it ever returns. The set only grows, which is the point.

Quality gates sit in the release path. When a model changes, a prompt is rewritten or a carrier portal redesigns its emails, the change ships only when the evals pass. A failing eval is a routine Tuesday, not an incident.

The controls an auditor asks for

Freight runs under audit: customs declarations, AEO status and customer SLAs all demand evidence. So every stage ships with the controls that evidence depends on: an access matrix per workflow step, so the quote-drafting agent cannot approve credit and the settlement agent cannot touch pricing; an audit trail of who requested what and who approved it; UK/EU data residency; and a rollback rehearsed as an operation, because the first rollback should never be the first incident. None of it is bolted on afterwards; it ships in the first release.

What to ask a vendor before signing

Before any logistics AI contract, four questions. Where do the golden cases come from: our lanes and our quotes, or your demo dataset? Who owns the evaluation suite, and does it live in our repository, versioned beside the workflow? Show me a release your own evals blocked; a vendor that has never blocked one is not evaluating anything. When the agent mis-captures a milestone, who sees it first: my operations team, or your dashboard?

The walk, in one line

Map this walkthrough onto your own operation with AI agents for logistics, read the harness anatomy in how production AI agents stay honest, and if agents are already live and ungraded, the Agent Audit closes the gap. For measured outcomes, the published case study has the numbers; this piece has the method.

← Back to Blog