AI for supply chain and logistics: the paperwork moves slower than the freight
A container can cross an ocean in three weeks, then sit for four days because a commercial invoice does not match a packing list. That is the quiet embarrassment of modern logistics. The physical network is fast and remarkably reliable. The administrative layer wrapped around it is not. Your team spends its day chasing status across carrier portals, answering "where is my order" emails, working exceptions out of shared inboxes, assembling customs documentation from attachments, and rekeying the same shipment data into a TMS, an ERP, and a customer's portal that were never joined up. None of this is the job your operations people were hired to do. It is the tax they pay because information moves through your business in emails, PDFs and portal screens rather than as structured data.
This is exactly the kind of work AI agents are now good at, and exactly the kind of work most software vendors have failed to fix, because it lives in the gaps between systems rather than inside any one of them. Below are the use cases we see hold up in production, with enough detail on how each one actually works that you can judge whether it fits your operation.
Answering "where is my shipment" without a human in the loop
The highest-volume question in any logistics operation is a status question, and answering it usually means a person opening two or three systems to assemble one sentence. An agent handles this end to end. It reads the inbound email or portal message, identifies the shipment from a reference number, a PO, or even a vague description matched against recent orders for that customer, then pulls current status from your TMS, the carrier's tracking feed, and any terminal or port data you subscribe to. It writes back a plain answer with the current location, the next milestone, and the revised ETA if one exists, and it logs the interaction against the shipment record so the history is in one place.
The handoff rule matters here. If the data sources disagree, if the shipment is flagged with an open exception, or if the customer's message contains anything beyond a status request, the agent drafts a reply and routes it to a person instead of sending. Routine questions get answered in minutes at any hour. Your team only sees the conversations that actually need judgement.
Working the exception queue instead of just reporting it
Most visibility platforms are good at telling you something went wrong and bad at doing anything about it. An agent sits on the exception feed and works it. When a container rolls, a pickup is missed, or an ETA slips past a committed delivery date, the agent reads the exception, checks which orders and customers are affected, and assesses the downstream impact: does this break a production schedule, a retail delivery window, a penalty clause? For routine cases it acts within rules you define, requesting a new pickup slot, rebooking on the next sailing the contract allows, or notifying the affected customer with a revised date, and it writes every action back to the shipment record.
For anything outside those rules, it prepares the decision rather than making it. A person gets a short brief: what happened, who is affected, the two or three realistic options with cost and time implications, and a recommendation. The human decides in two minutes instead of spending forty assembling the picture.
Claims and OS&D handled while the evidence is fresh
Freight claims are a paperwork race you usually lose. Deadlines are strict, evidence goes stale, and the file has to be assembled from the BOL, the POD, inspection photos, temperature logs, the carrier contract, and the original invoice value. Most operations under-claim simply because assembling the file costs more attention than the claim is worth.
An agent changes that arithmetic. When an overage, shortage or damage report comes in, it opens a claim file, pulls the relevant documents from your systems and email history, checks the notification deadline for that carrier and mode, and drafts the claim with the evidence attached and the value calculated from the commercial invoice. It tracks carrier responses and drafts follow-ups when they go quiet. A person reviews and submits every claim, and takes over entirely when liability is disputed or the value is high. The agent's job is to make sure no claim dies of administrative neglect, and that the strong ones reach a human with the file already built.
Customs documentation that gets checked before it causes a hold
Customs delays are mostly self-inflicted: a missing document, an inconsistent value, a description too vague to classify. An agent reviews the document set before it ships. It reads the commercial invoice, packing list, and certificates against each other, flags mismatched quantities, values and weights, checks that the goods descriptions are specific enough for classification, and confirms that the paperwork required for that origin, destination and commodity is actually present. It can propose HS codes based on the product description and your own classification history, and it drafts the broker instruction with the full document set attached.
The line the agent does not cross is the legal one. Classification decisions, valuation questions, and anything touching preferential origin claims go to your broker or compliance lead with the agent's analysis attached as a starting point. What changes is that your broker receives a checked, complete file instead of a pile of attachments, and the errors that cause holds get caught while the freight is still days from the border.
Freight invoice audit at line-item depth
Auditing carrier invoices against contracted rates is tedious enough that many teams sample instead of check. An agent audits every invoice. It reads the invoice, matches it to the shipment, and recalculates the expected charge from your rate cards, accessorial schedules and fuel surcharge tables. Matches within tolerance are approved and posted to your ERP with the right coding. Discrepancies generate a dispute draft with the calculation shown, which a person reviews before it goes to the carrier. Over time the dispute history becomes its own asset: a record of which lanes and carriers drift, which feeds your next rate negotiation.
Document intake that ends the rekeying
BOLs, PODs, packing lists, booking confirmations and arrival notices still arrive as PDFs and photographed paper. Someone reads them and types the contents into a system. An agent does the reading and the typing. It extracts the fields, validates them against the order and shipment it matches them to, and writes the data into your TMS or ERP through the same interfaces your team uses. When a document is illegible, does not match any known shipment, or contains a value outside expected ranges, it lands in a review queue with the agent's best guess highlighted, so the human is confirming rather than transcribing. This one is rarely the use case a leadership team gets excited about, and it is often the one that pays for the programme, because everything else depends on the data being in the system.
Carrier and supplier email triage
The operations inbox is where all of the above arrives mixed together. An agent triages it: classifying each message, extracting references, attaching it to the right shipment or order, and either handling it under one of the workflows above or routing it to the right person with context already assembled. Nothing gets deleted and nothing gets answered without either rules that permit it or a human approval. The inbox stops being a queue that a person drains by reading everything, and becomes a stream where humans only read what needs them.
Where to start
Resist the platform-shaped ambition. Pick one workflow that is high volume, painful, and measurable, with status responses and document intake as the two most common first choices, and get an agent into production on it in weeks. Run it in draft mode first, where every output is reviewed by the person who used to do the task. That builds trust, surfaces the edge cases no one wrote down, and gives you a clean before-and-after measure. Then widen the agent's autonomy rule by rule, based on evidence, and only then start the second use case. The organisations that succeed with this treat it as operations work with software attached, not an IT project with a go-live date.
Where AI fails here
Honesty is cheaper before deployment than after. Agents fail when the underlying data is wrong: if your TMS says the container is on the water and it is sitting in a yard, the agent will confidently tell your customer the wrong thing, faster than a human would. They struggle with genuinely novel situations, a port strike, a new trade restriction, a customer threatening to leave, where the right move depends on context no system holds. They should not be given final authority over customs declarations, high-value claims, or commitments that carry contractual penalties. And they will not fix a broken process; automating a workflow your team does not trust just produces mistrusted output at higher speed. The discipline that makes this work is boring: narrow scope, explicit handoff rules, humans reviewing early output, and logs you can actually audit. We recognise that this is less exciting than the demos suggest. It is also why the deployments that follow it are still running a year later.
The freight will keep moving whether or not you fix the paperwork. The question is how many of your best operations people spend their days rekeying, chasing and apologising, and what it costs you in delayed containers, missed claims and quietly frustrated customers while they do. If some part of your operation runs on inboxes and rekeying, describe the process to us at Adopt AI Partners and we will tell you honestly whether an agent can carry it.