Making AI customer service work across a global freight forwarding business

Challenge

The company wanted an AI agent to take the routine shipment questions off its customer service team, mostly people asking where their package was and whether it was going to be late. Nobody thought of it as a hard project. The plan was to give the agent access to the tracking data and let it answer, so the humans could spend their time on the calls that needed a human.

It got complicated the moment we started asking what a late delivery meant inside the company. To the customer it was a late delivery, but to the person answering the phone it was often a signal that something else was going on, and they'd learned over the years which kind of late was which. If a lane was backing up, sales needed to know before they promised another customer the same transit time. If a truck had broken down or a customs hold was involved, operations needed to get on it. And if the shipment was for one of the accounts with a service guarantee in the contract, the delay was about to become a conversation with legal and finance. None of this was documented anywhere and, historically, it didn't need to be, as long as the same experienced people kept answering the phones.

The agent couldn't do what they did, which was decide in a few seconds whether this was a tracking question or an early warning, and if it was a warning, whom to tell. Before it could go live, the company had to work out which situations the agent should handle alone, what it was allowed to say to the customer about why a shipment was late, what it could pass along to sales, and who had the authority to act when it flagged a problem. That turned out to be most of the project. The agent itself was the easy part.

‍

Approach

In the first two weeks, I traced how shipment information moved through the business, from carrier systems to customer service to sales and operations, and where along the way someone decided. That meant working with five functions in five countries. The United States, Canada, Germany, France, and Spain each had their own privacy regime, their own customs realities, and their own way of doing things, and the answer to a question as basic as which customer data the agent could see changed depending on which border a shipment was crossing.

I then spent four weeks getting the right people in the same room to settle the questions the agent had forced into the open. When a delivery looked likely to slip, who needed to hear about it, and how early? Could the agent give the customer a revised estimate on its own, or did someone have to sign off? When the carrier's system said one thing and the customer's contract said another, whose problem was that? And how did the company add automation without adding a pile of new work and a compliance exposure in three jurisdictions? None of these belonged to customer service or IT alone, which is why no one had answered them.

For each, we documented the data-sharing rules, the escalation path, who had authority to act, and where a person needed to stay in the loop. I also mapped the agent into the systems and workflows the teams already used, rather than asking them to work around the technology, which is the usual way these projects go sideways.

The most valuable part of the work was connecting carrier scheduling and operational data to the people who could act on it. Once that connection existed, a shipment delay stopped being bad news to pass along to the customer and became an early warning that sales could use.

Outcomes

Earlier visibility into delivery risk. Potential shipment problems now reach sales and operations before the customer calls, not after.

Clear responsibilities across five functions. Customer service, sales, operations, technology, and legal agreed on who handles what information, who escalates, and who decides, for the first time in writing.

An agent with defined limits. The company drew the line between what the agent can report, when it can trigger a workflow, and which decisions still require a person. The agent operates inside those lines in all five countries.

A better business case. The value of the project is no longer measured only in the cost of answering inquiries. It includes customer relationships protected, commercial concessions avoided, and the time sales gains to respond to a constraint before selling into it.

In eleven weeks, what began as a customer service automation project ended up changing how information and decisions move between departments. Making an AI agent faster at answering questions is useful. Making sure the business can act on what the agent finds is where the real return is.

‍