Logistics runs on communication, and most of that communication is repetitive. Where is my shipment. Can you quote Rotterdam to Belgrade for two pallets. Why was it not picked up yesterday. Send me the CMR. Every one of those questions has an answer already sitting in a system, and a person is being paid to look it up and retype it. That gap is where AI agents pay for themselves in this sector, faster than in almost any other, and without touching a single truck.
§ 01Where agents earn their cost
- Quote intake and response. Requests arrive as unstructured email with the details scattered across a paragraph. An agent extracts origin, destination, dimensions, weight, and required date, checks them against your rate structure, drafts a quote, and escalates anything outside normal parameters. The bottleneck in quoting is rarely the pricing, it is the typing.
- Track and trace. The single highest volume question in any forwarder or carrier office. An agent pulls current status from the TMS or carrier data, answers the customer in their own language, and only involves a person when the status is bad news or the data is missing.
- Exception detection. A shipment that has not moved for eighteen hours, a delivery window that will be missed, a document that never arrived. Agents are well suited to watching for conditions continuously and alerting a dispatcher before the customer calls, which changes the conversation from apology to notification.
- Documentation checks. CMR, packing lists, customs paperwork, and delivery confirmations have to match the order. An agent compares fields across documents and flags mismatches, which is tedious for a person and instant for a machine.
- Carrier and subcontractor chasing. Status requests, confirmations, and unanswered emails are pure follow-up work with no judgment attached until something goes wrong. Agents chase, log, and escalate.
The pattern across all five: read from systems, prepare an output, and hand anything unusual to a human. That is a deliberately narrow definition of what an agent does, and it is why these deployments survive. The architecture behind that split is documented in our engineering notes.
§ 02What does not work
- Autonomous pricing on non-standard lanes. Rates on established lanes are rules and agents handle them well. Rates that depend on a relationship, on knowing which carrier owes you a favour this month, or on a judgment about a customer's future volume are not in any system the agent can read.
- Replacing dispatch judgment. Deciding which of three late shipments gets the available truck is a decision with commercial consequences and a phone call attached. Agents prepare the picture. People decide.
- Working around missing data. If carrier status only exists in a portal a person logs into manually, the agent needs that data pipeline built first. That is an integration project, and it should be quoted as one rather than hidden inside an agent project.
- Claims and disputes. High stakes, legally sensitive, low volume. Wrong shape for automation and the savings are not there anyway.
§ 03The integration reality
This is the part that separates a realistic quote from an optimistic one. A modern cloud TMS with a documented API is a few days of work. A forwarding system from 2011, three carriers who only publish status through a web portal, and a customs process that runs on emailed spreadsheets is where the schedule actually goes. Integration is typically 60 to 70 percent of the build effort in this sector, and the discovery call exists largely to find out which situation you are in before anyone commits to a number.
One consequence worth planning around: start read-only. An agent that reads shipment data, drafts responses, and writes nothing back carries almost no risk while still removing the majority of the manual work, because looking things up is the slow part. Write access comes later, process by process, with a person approving until the numbers justify removing the gate.
§ 04How multilingual actually helps here
International logistics is multilingual by definition, and this is one of the few places where language models produce an operational advantage rather than a demo. An agent handling status enquiries can read a message in one language, check systems that record data in another, and answer in the customer's language with consistent terminology. For teams that currently route by who speaks what, that changes the shape of the working day more than any dashboard.
§ 05Scoping a first deployment
Choose track and trace or quote intake. Both are high volume, both fail visibly and cheaply, and both are measurable within days. Run with a person approving every outgoing message for the first weeks, then track one number: how often the person edits before sending. High means the rules or the data are wrong, which you have now learned for the price of a small build. Low and stable means you can drop approval on routine cases and keep it on exceptions.
On price: a single agent covering one process end to end costs 3,000 to 5,000 USD over 2 to 3 weeks. A coordinated fleet spanning quoting, tracking, and documentation costs 15,000 to 50,000 USD over 6 to 12 weeks. What pushes a project up within its tier is almost always carrier data that has to be scraped or collected rather than fetched. The full breakdown is in our cost guide, and comparable deployments are described in our production cases.
──
If you want to know which process in your operation is worth automating first, that is a 30 minute discovery call. Your systems, our questions, and an honest answer including the case where the right first step is fixing a data feed rather than building an agent.