Skip to content
Prototype request form with a document that couldn't be verified, next to the one-time passcode step.

Lead UX Designer · 2026

Designing agentic document ingestion

Designing how an AI agent that requests missing documents fits into an existing platform, and where people step in to complete the job.

01

Overview

Freight shipments run on documents. When a document is missing, someone has to notice, chase it, and get it into the right place before anything moves.

An AI agent could do the noticing and the chasing. That part was easy to describe. The harder design question was everything around it. Where does the agent's request reach a person, how do they respond, and how much should they trust what it does on their behalf.

I worked on the systems side of that question and on the screens people would actually touch. I built the prototypes with Claude Code.

02

Context

The agent's job was simple. It ran as a workflow task that watched a job for documents that should be there, such as a customs declaration. When something was missing, it triggered a request for information, or RFI, for the person who could supply it. One RFI could hold several information requests, so a single request could ask for more than one thing.

The platform it had to live in was not simple. CargoWise already had established portals, navigation patterns and platform services. A new AI feature couldn't sit beside all of that as a separate thing. It had to enter through the existing structure and use the services that were already there.

Diagram showing a workflow AI task watching a job and triggering an RFI that holds several information requests
How a request comes about. A workflow AI task watches the job and triggers an RFI, which can hold several information requests.

03

Mapping the system

The team planned the flows together in working sessions, sketching them out on whiteboards before anything was built. My focus was the systems thinking behind those flows.

An RFI reaches people through a link, and that link had to work for two very different recipients. Some used CargoWise every day and were already signed in. Others had no account and just needed to send a document. So a request could go out as a public link or as a link for signed-in users only. The public version verified people when they submitted, using a one-time passcode sent to their email. The signed-in version used the existing sign in, and anyone already signed in went straight to the request.

I mapped how each of those entry points connected to the existing portal information architecture. I also traced which platform services the feature would depend on, so the design matched how the platform already worked rather than inventing a parallel path.

Diagram of the two ways a request for information reaches a person, through a public link or a signed-in link
How a request reaches someone. A public link verifies with a one-time passcode. A signed-in link uses the existing sign in.

04

Designing the human touchpoints

Two moments needed real interface design. The first was getting someone into the request, either with a one-time passcode or through the existing sign in. It had to feel trustworthy and take as little effort as possible. The second was the form for supplying the missing document itself.

I prototyped both with Claude Code, so the flow could be clicked through end to end. The aim was that a person who received a request could understand what was being asked and finish the task without needing to know an AI agent was behind it.

The form also had to handle a document the agent didn't trust. Each upload was checked by the agent. If it failed verification, the person could still upload it anyway, as long as they gave their email so they could be identified. The agent raised the flag, and the person kept the final say.

Prototype of the one-time passcode entry step
Getting into the request.
Prototype of the request form with a document that couldn't be verified, offering a clearer file or upload anyway with an email
A document that couldn't be verified. The person can choose a clearer file or upload anyway and give their email.
Diagram of the upload check, showing a failed verification that a person can override by giving their email
What happens after an upload. A verified document goes straight through. A failed one can still be uploaded, but the person has to identify themselves.

05

Building trust in stages

Handing a task to an agent only works if people believe it will get things right. That trust has to be earned, and a design can help or get in the way.

The path we mapped started with a manual upload step. The person stayed in control of supplying the document. From there the agent took on more, one step at a time. First it sent requests for a person to approve. Then it checked uploads and flagged problems while the person decided. After that it sent reminders automatically, and then filed documents that passed on its own, with people reviewing only the exceptions. The last step was full automation, with people stepping in only when something looked wrong. Trust set the pace of automation instead of the other way round.

Diagram of six stages from manual upload to full automation, with the agent taking on more as trust in it grows
Six stages, each giving the agent a little more responsibility as trust in it grows.

06

Outcome

  • Mapped the entry points and how they connect to the existing portal structure
  • Defined how the feature relies on existing platform services rather than adding new ones
  • Built working prototypes of both touchpoints with Claude Code
  • Mapped a staged path from manual upload to full automation, with the agent taking on more as trust grows

07

Reflection

The agent was the easy part to explain. The design work was deciding where people stay in the loop, and that turned out to be a systems question as much as an interface one.

This was a team effort, and my part was the structure and the touchpoints. I built the prototypes with Claude Code, which made AI both the subject of the design and part of how I worked on it. It changed how I approach AI features. I now start by asking where a person needs to be, and how the design can earn the right to take that step away.