Every logistics company has a version of this problem. Shipment requests flood in via email. Each one is formatted differently. Someone has to read every single one, extract the critical details, enter them into a system, and make sure nothing gets missed.
For United Transportation — a regional air freight and logistics company — this manual process was consuming their dispatch team's entire day. As volume grew, so did the errors. And hiring more people to do the same repetitive work wasn't a sustainable answer.
We built a five-agent AI system that changed that. Here's exactly what we did, how it works, and what it took.
The Problem: Messy, Unstructured Emails at Scale
The dispatch team received a high volume of inbound shipment requests via email every day. Each required a coordinator to manually read the email, extract seven critical fields — customer, tracking number, shipment type, airline, pickup date, screening status, and any special notes — then enter everything into a Notion dispatch board and assign a driver.
The challenge wasn't just the volume. It was the variability. Every customer sends information differently:
- Air waybill numbers buried in PDF attachments
- Pickup times mentioned casually in the body text
- One customer says "tender export," another says "deliver to airport" — same thing
- Some submit via shared Excel spreadsheets with inconsistent formatting
Traditional automation tools can't handle this. They need structured, predictable inputs. Rule-based systems would require constant maintenance and would still miss edge cases. What United Transportation needed was something that could actually understand what it was reading.
"Dispatchers were spending their entire day reading emails and copy-pasting data. Now the system handles the routine work and our team focuses on what actually needs a human."
Why This Was "Too Messy to Automate" — Until Now
United Transportation had tried to automate before. Every attempt ran into the same wall: the moment a customer sent something slightly different, the automation broke. A rules-based system couldn't distinguish between a purchase order number and an air waybill number. It couldn't infer that "TSA screening required" means this is an export. It couldn't open a PDF attachment and find the tracking number three pages in.
Modern large language models can do all of this. They bring contextual understanding to unstructured data — the same way a human reads and interprets, but at machine speed and without errors from fatigue.
This is exactly the kind of workflow AI agents were built for.
The Solution: Five Specialized AI Agents
We built a system of five agents, each handling a specific part of the dispatch workflow. Together they cover the process end-to-end.
Email Administrator
Monitors the shared inbox and classifies every incoming email — dispatch request, status inquiry, billing question, or other. Only true dispatch requests move forward.
Spreadsheet Coordinator
Handles customers who submit via Excel. Extracts relevant rows, normalizes the data — even from messy workbooks with merged cells — and passes clean records downstream.
Dispatch Intake Agent
The core engine. Reads email content, opens PDF attachments when needed, extracts all seven required fields, and creates a structured record in the Notion dispatch board.
Driver Tracker
A mobile app for drivers to check in and out at pickup locations. Updates driver status in Notion in real time and alerts dispatchers of delays automatically.
Status Reporting Agent
Pulls live data from the dispatch board and generates professional status reports for clients and internal stakeholders. Formatted, timestamped, and distributed automatically — no manual assembly required.
How the Dispatch Intake Agent Actually Works
The Dispatch Intake Agent is the most complex piece and worth understanding in detail, because it illustrates what AI agents can do that traditional automation cannot.
The agent follows a clear source hierarchy: email subject and body first, then attachments only when necessary. For export shipments, if the AWB number isn't in the email text, it knows to open the transfer order PDF and find it there.
Each field has detailed extraction logic. For example:
- Customer name: Extracted from the sender's display name or email signature — not from shipper or consignee details in the shipment itself. Output is standardized to first word, all caps.
- Shipment type: Determined by keywords and context. "Tender export," "deliver to airport," and "TSA screening" all indicate an export. The agent can even override an apparent IMPORT classification to LOCAL if it detects both a pickup date and a delivery destination that isn't an airport.
- Tracking number: Strict token rules prevent it from confusing a phone number, zip code, or purchase order with an AWB number.
Why prompt engineering matters
The quality of any AI agent depends entirely on the quality of its instructions. We invested significant time upfront writing detailed extraction rules, priority hierarchies, edge case examples, and validation logic. That upfront investment is what delivers consistent, production-quality results — and it's where most DIY attempts fall short.
Before and After
Before
- Dispatcher reads every email manually
- AWB data extracted field by field by hand
- Notion board updated after each email
- Driver status tracked via phone and text
- Status reports written by staff
- Errors discovered at delivery
After
- Email Administrator classifies automatically
- Dispatch Intake extracts all fields instantly
- Notion board updates in real time
- Driver App captures check-in/out at each stop
- Status reports generated and sent automatically
- Exceptions flagged before they become problems
The Results
Eight hours of manual dispatch work eliminated daily. Zero missed details from email overload. The dispatch team now spends their time on what actually requires human judgment — solving complex customer problems, handling exceptions, and building relationships.
Perhaps the most telling sign of success: the team asked us what else could be automated.
It's also worth noting what we didn't do: we didn't ask United Transportation to change any of their tools. The system connects Microsoft Outlook, Excel, Google Sheets, and Notion — the environment they already had. We built around their world, not ours.
What This Means for Your Business
United Transportation is a logistics company, but the pattern applies everywhere. If your team is spending hours every day reading emails and entering data — in any industry — that workflow is a candidate for AI agents.
The key question isn't "can this be automated?" It's "is this too messy for traditional automation?" If the answer is yes, that's exactly where AI agents excel.
The technology that was previously only accessible to enterprise companies is now practical for SMBs. The only question is where to start.
Read the Full White Paper
Deep dive into the technical architecture, implementation lessons, and broader implications for SMB logistics.