In freight forwarding, "AI customer service" is an overloaded phrase. It can mean an assistant that drafts replies for your agents, a bot that answers rate and schedule questions on its own, or simply a knowledge base your team can search. Those are three different builds with three different price tags, and lumping them together is the fastest way to buy the wrong one.
Two forwarders called us the same week asking for a price on an AI customer service desk. Both move similar volumes. Both are tired of answering the same rate and schedule questions all day. The first opened with a single line: how much? We quoted nothing, because there was nothing yet to quote. The second spent thirty minutes walking through where the pain actually sat, who the tool would replace, and what number he wanted to see move in six months. By the end he had cut his own request in half and knew which part to build first.
The difference was not budget. It was whether the right questions get asked. Here are the three we go through before any figure is offered.
Question one: which part is actually hurting
"Customer service can't keep up" is a feeling, not a task. It has to be broken into parts, and each part needs a rough share.
The usual four are these. Volume is too high and one or two people cannot answer fast enough. Replies are too slow, so the customer waits half an hour and asks a competitor. The same questions repeat daily while the answers sit in three different heads. Or complex enquiries eat too much time, and a single rate conversation takes half an hour.
Each pain points to a different fix. For volume, start with triage and auto-replies. For speed, attack response time. For scattered knowledge, consolidate answers into one searchable source first, a step that has little to do with AI. For expensive complex cases, that is where AI earns its place, pulling information together and drafting a judgement.
Ask the owner for three proportions: of a week's enquiries, how many are simple and repetitive, how many need a rate or order check, and how many truly require a person to decide. We worked with one US-trade forwarder where thirty percent of enquiries were schedule and tracking questions, half of the rest needed pricing, and fewer than ten percent needed a human call. Seeing that split, the owner moved the budget from "AI for everything" to "triage first."
Question two: is it replacing people or helping them
Two very different builds sit under the same phrase "AI customer service."
In the first, the AI answers simple questions on its own, handling rates and schedules and tracking without a human. It saves labour, but the margin for error is thin. One wrong answer can lose a booking.
In the second, the AI works beside the agent. It assembles the facts and drafts a few possible replies, and a person confirms the final line before it goes out. Slower by a beat, far safer, and a better fit for high-value cargo that cannot afford a mistake.
Most forwarders start with the second and hand over part of the simple volume only once the knowledge base is complete and accuracy holds. Either way, say up front whether the tool replaces certain tasks or simply helps certain people move faster. Without that, the two sides will disagree on delivery day.
Question three: which number should change in six months
"Faster" and "smarter" cannot be written into an acceptance test. Replace them with something countable. Response time from how many seconds down to how many. The share of enquiries that need a human from one level to a lower one. The completion rate from first contact to a sent answer.
Fix those numbers first, and six months later you have a basis for judging whether the build was worth it and where to tune it. Set targets against real volume, not against ambition. One owner announced he would cut his service team in half, then discovered there was not enough enquiry volume to keep either the machine or the people busy.
Where AI does not belong, and where humans must step in
Clearly ruled, fixed-step work does not need an AI. Checking bill numbers, container numbers and declaration elements has a definite right and wrong answer. A hard-coded workflow runs it more reliably and more cheaply than asking a model to "understand" it.
AI earns its keep where the work needs several rounds of back-and-forth and pulls data across systems. A customer asking whether a shipment can arrive two days earlier forces a check of schedules, space and past correspondence to shape a conditional reply. A fixed workflow cannot hold that, and people find it exhausting.
Decide one more thing before kickoff: which situations must go straight to a human. Angry customers, amounts over a threshold, and any change or cancellation that moves money or liability should always go back to a person. Draw that line early and it is insurance. Draw it late and it becomes an incident.
FAQ
Does every customer service enquiry go to AI?
No. Fixed-step work is better served by a plain workflow. Angry customers, large amounts, and rebooking or cancellation go to a human.
How much does an AI customer service desk cost?
There is no single price. It depends on enquiry volume, how many steps you automate, and whether the tool answers directly or assists an agent. Answer the three questions first and a workable budget appears.
How soon can we see results?
It depends on the stage. Consolidating a knowledge base and starting triage usually moves quickly. Overall metrics depend on business volume, so fix response time, human-handoff ratio and completion rate and review quarterly.
How do we judge a vendor?
Look at whether they ask about the business before quoting, and whether they offer measurable targets and a human-handoff rule. A vendor who promises results up front deserves more caution, not less.
We are Shanghai Naili (上海奈李信息技术有限公司), a freight-forwarding marketing partner with ten years in the sector. We build AI customer service and AI digital employees for forwarders, alongside full-service marketing department outsourcing. Our habit is simple: ask the business questions first, then quote.