Automating logistics customer service with AI

Automating logistics customer service means a system reads, looks up and answers customer questions using your own data. The same questions arrive in the shared inbox every day, from clients and from their end customers. Where Friday’s shipment has got to, how much of an item is left, whether the delivery can move a day and what happens to the pallet that arrived damaged. Someone in 3PL customer service looks up the answer across three screens and types a reply. A system can do much of that looking up, as long as you agree beforehand which part.

How is this different from a chatbot?

The system does not make up an answer. It fetches a fact that is already in your WMS, your TMS or your carriers’ tracking and puts it in an email.

A chatbot on a website handles general questions. In a warehouse, a fulfilment centre or a 3PL almost every question is about one order, one shipment or one item for one client. The answer is in your own data, so that is where the system needs access. Without it you get a polite email that says nothing.

Which questions does a 3PL receive?

Most emails fall into a few fixed groups. How big each group is differs per warehouse and per client. You count that yourself, and further down is how.

Where is my shipment?

Track and trace queries are the best-known example. An end customer emails the web shop, the web shop forwards it, and you find out whether the order was picked, whether it is on a route and what the carrier shows as the last scan. The answer is spread out. The WMS knows whether the order shipped. The TMS or your shipping software knows which carrier and which tracking number. The carrier knows where the box is now.

Stock and availability

A client wants to know how much of an item is in stock, whether yesterday’s inbound is already available or when a backorder can go. That is in the WMS, but per location and per status, and rarely in the shape the client asks for.

Changes to an order

A different delivery address, a later delivery day, a line removed. Whether that is still possible depends on where the order is now. If it has not been released for picking, it is a change in the WMS. If it is already packed on the dispatch floor, someone on the floor has to pull the box back.

Damage, loss and complaints

A damaged pallet, a box that never arrived, the wrong item in the parcel. This is about liability and money.

How does AI customer service work in a warehouse?

An AI system reads the question, looks up the answer in your own systems and prepares a draft that a member of staff approves. In practice that runs in four steps.

1. The system reads the email and the attachment. A language model works out which client the question comes from, which order or tracking number is in it and what kind of question it is. If there is no number, it searches on the end customer’s name, postcode or reference.
2. It looks up the answer: the status in the WMS, the route in the TMS, the last scan at the carrier through its integration or portal. It only reads and changes nothing.
3. It writes a draft in the language of the question, with the fact in it and where it came from: which status, from which system, at what time.
4. A member of staff reads the draft, edits it where needed and sends it. If the question should not have gone to the system, they set it aside for a colleague.

Here the system chains several steps on its own and chooses which system to consult. That is called an agent, and what an agent can and cannot do is in what is an AI agent. How the system reaches your WMS and TMS without replacing anything is in connecting WMS, TMS and ERP.

Orders often land in the same inbox. They take a different route, into the WMS instead of back to the sender. That part is in orders from email into your WMS.

Which customer questions can you answer automatically?

Answering customer questions automatically, without anyone reading the draft, only suits questions whose answer is a fact from your system and where a mistake costs little. A shipment status with a tracking link is that kind of question. So is a stock level, as long as the sender is allowed to see it.

So you do not start by sending automatically. For the first few weeks every draft goes past a member of staff, and you count how often it goes out unchanged. If that stays consistently high for one kind of question, you switch on sending for that kind. Everything else stays a draft.

A few rules keep it safe. The system only replies to a sender who belongs to the client of that order. It never mentions data of another client or another end customer. If it is unsure about the order number or about who is writing, the email goes to a person.

Where must a person always stay involved?

Anything involving money, liability or a complaint is decided by a person. The system may recognise those questions, sort them and prepare the file. Towards the customer it promises nothing and admits nothing. A credit note, a compensation or an exception to the agreements always comes from a member of staff.

With damage there is a legal side as well. For international road transport under the CMR Convention, the consignee has to report damage that was not apparent at delivery to the carrier in writing within seven days, Sundays and public holidays excepted. If that does not happen, taking delivery counts as prima facie evidence that the goods arrived in the condition described in the consignment note. A claim arising from such carriage is generally time barred after one year. The system helps by recognising a damage report straight away and putting the deadline on the file. Who holds the carrier liable and what you promise the client remains work for people.

If the system sends replies itself, the reader has to know they are dealing with an AI system. Article 50 of the AI Act requires an AI system that interacts directly with people to be built so that they know, unless that is already obvious to an attentive reader. That obligation has applied since 2 August 2026. How to handle the personal data of end customers in those emails is in GDPR and AI in logistics.

How do you measure whether inbox automation works?

You measure it against a baseline of how your inbox runs today, taken before anything is built. Inbox automation without that baseline leaves you unable to say afterwards what changed. Over a few ordinary weeks, count the following.

• How many emails arrive per day, per client and per kind of question.
• How long it takes from arrival to reply, on a Monday morning and on a quiet Wednesday.
• How often a question needs a second email because the first answer was wrong or incomplete.
• How many minutes a member of staff spends per question looking things up in the WMS, the TMS and carrier portals.

After go-live you measure the same. Two numbers are added: how often a draft is sent unchanged and how often a member of staff corrects the system. Those corrections are the most useful. They show which question the system misreads, and sometimes which system gives a status that is wrong.

That is also how our way of working describes it: one process, measured against a baseline, and the next only once this one works. For the inbox that usually means one kind of question for one client. Only when that runs does the next kind get its turn.

How to count one week of inbox

1. Export an ordinary week of email from the shared inbox.
2. Mark each message by kind: shipment, stock, change or damage.
3. Note in which system the answer was.

From that, pick a kind of question that comes up often and whose answer is a fact in one system. Bring the count to an introductory call and we pick it together.