AI Document Audit for Freight Forwarders

Background

GoFreight is a SaaS platform used by 1,000+ freight forwarders worldwide, the companies that handle the documents and logistics behind ocean and air cargo. Their operators, called OPs, spend the day turning shipping paperwork into structured shipment records. Document Intelligence is the feature that reads an uploaded PDF and fills those records for them.

TL;DR

The problem

What I did

22existing companies asked to join the Early Access.

The behavior users never mentioned

An AI feature was already live, but users were quietly re-checking everything it filled in.

GoFreight's Document Intelligence had shipped before I joined. After a user uploaded a PDF, a modal opened: the document on the left, the AI-extracted fields on the right. The user confirmed each field, fixed anything that looked wrong, and pressed Save. It worked.

But while I was mapping the PM's customer interviews onto a journey map, one pattern kept surfacing. After saving, users would reopen the same shipment and read back through the data they had just confirmed. Once we noticed it, the team started asking about it directly in later interviews, and most users said it was a normal part of how they worked.

We named it Double Review.

To check Double Review wasn't isolated to a handful of customers, I led a cross-functional workshop with 10+ teammates across product, design, and engineering. The session had two jobs: share the customer observations I had gathered, and walk the group through grouping them into themes and weighing each pain point by impact and feasibility. By the end, the team had aligned on accuracy and Double Review as the core pain points for this iteration, and the PM used that alignment to adjust the Scrum schedule for the coming sprints.

The workshop board, where we grouped interview findings and voted on which pain points mattered most.

What the modal couldn't do

The modal had been built as a deliberately simple surface, and that turned out to be the deeper problem.

A lot of what the Shipment Entry page could do just didn't work inside the modal, so users couldn't finish or verify their work there. Three gaps came up the most.

The bet: move the review onto the Entry page

Build the review where the work already lived, and run it alongside the old flow instead of replacing it.

So we made the call to build the review experience on the Shipment Entry page itself, where all of this already worked. The data would be verified in the same place it was going to be used.

We didn't replace the modal outright. v1 stayed live for customers already using it, and v2 shipped alongside it on the same product. Running both against real workflows was the only honest way to see whether the new flow actually helped, for new and existing customers alike, before committing to a long-term direction.

Original

New

Many ways in, one review

A document could reach the review through several different paths. No matter which one, it had to end in the same place.

A document reached the audit through two routes that lived on different pages. Drag-and-drop and upload happened in the List view, under Document Intelligence. An email forwarded in surfaced as a task in the Action Center, the area built for managers to triage and assign work, so it opened an Edit Task step before anything else. Same audit at the end, two unrelated starting points, each with its own pre-steps.

The hardest case was an invoice on its own. An invoice is always an update to an existing shipment, so it has to attach to a real HBL or MBL. When the AI matched nothing, the design had to decide where to stop the user and how. A document with no match at all is held with a prompt to match it manually before it can enter; a document that matched only some of its pages lets the user in and flags the rest at the bottom to assign by hand. Most of the work here was not the screen, it was choosing which step each prompt belonged to.

All of these paths had to land in one consistent review, and the Review Guide panel was the surface that absorbed them. The old modal's actions, update existing and create new, could not move back onto the Shipment and Invoice Entry pages without overloading screens that were already dense, so they were pulled into one panel that kept everything the AI touched in a single place. Reusing the same entry point the modal had used carried the user's existing mental model with it. They recognized where they were and knew the actions they could finish here, instead of relearning a new surface.

The new design, piece by piece

With every AI action in one panel, the rest was making the guidance feel like part of the Entry page the user already knew.

The review now happens on the Entry page itself, with a Review Guide panel docked at the bottom right.

The Shipment Entry page is where forwarders already work, every field, every tab, every cross-reference they trust. A single Shipment can carry a dozen HBLs and several Invoices, and not all of them need changes. The panel had to guide a review across all of that without pulling attention off the page itself.

So the rest of the work was in the details, each one keeping the AI visible without asking the user to stop and learn it. How a filled field reads, what happens when AI overwrites an existing value, how the review moves from one tab to the next, and where the source documents sit.

Making existing users switch

The new panel only succeeds if existing users actually move into it.

After the initial prototype, two extensions came out of further use-case review, both carrying capabilities from the old modal into the new panel so existing users could move over without losing the moves their workflow depended on.

(1) Copy From an existing Shipment

When creating a new Shipment from a PDF, users sometimes want to start from an existing Shipment's data. Copy From lets them select any past Shipment by file number; the new Shipment opens pre-filled, and the AI-extracted values land on top. It came out of recurring Shipments users wanted to clone, where starting from a blank form meant retyping known information.

(2) Manual select Update Shipment

Update Shipment originally relied on the system to detect which Shipment a PDF belonged to. Detection worked in most cases, but not all. The panel now lets users select the target Shipment manually when the system's match is wrong.

Lowering the cost of switching kept the migration on track. We didn't roll the new flow out to everyone at once; we offered it to the customers who regularly help us trial experimental features. 22 of them asked to turn it on through Early Access.

What the data changed my mind about

We set up Clarity to watch how the review played out in real sessions, and one finding cut against the panel's core assumption.

The step-by-step was built for those busy Shipments, on the assumption that users would want to be walked through every item.

In the session recordings, the same move kept showing up. An experienced OP would open the review, glance at the panel once, then skip the order it suggested. Instead of starting at Basic Information and working down, they went straight to the tab that mattered for this Shipment, often Container, checked the AI's values there, and finished without walking through the rest.

A Clarity session recording of the review on the Entry page.

By the time they had uploaded the PDF and let the AI extract it, they already knew which Invoices and HBLs to check and which to leave alone. The sequence we laid out had nothing new to tell them.

For an OP who knows the Shipment, the walkthrough was optional. It earns its place somewhere else, in a handover. When someone picks up a Shipment they didn't set up, they don't yet know what has been done or what still needs auditing, and the panel gives them that map.

So instead of forcing the sequence, we let users collapse or close the panel when they don't need it, and open it back up when they do.

The guidance an expert skips is the same guidance that helps whoever inherits the work.