--- title: "An AI Agent in Logistics: Shipment Tracking and Exceptions" description: "Ninety percent of shipments arrive without trouble and need no attention. All the work is in the rest — and that is where an agent earns its place." canonical: "https://interfase.sk/en/news/an-ai-agent-in-logistics-shipment-tracking-and-exceptions" lang: "en" last_updated: "2026-10-16" --- # An AI Agent in Logistics: Shipment Tracking and Exceptions > Ninety percent of shipments arrive without trouble and need no attention. All the work is in the rest — and that is where an agent earns its place. Published: 2026-10-16 - The value is not in tracking shipments that go to plan, but in recognising the ones that have drifted off it. - The agent should catch an exception before the customer calls — a carrier's silence is often a stronger signal than an error status. - Carrier statuses are not comparable; without mapping them onto your own scale you cannot evaluate anything reliably. - Customer communication needs a limit — informing yes, promising new delivery dates no. A company sending hundreds of shipments a day does not need to know about each one. Most arrive on time and interest nobody. The problem is that among them are fifteen stuck somewhere — and the company usually finds out when the customer calls. ## The work is in the exceptions This is the whole idea behind the deployment. Tracking shipments is something a spreadsheet can do. Recognising which one needs intervention, and preparing that intervention, is the work a person currently does by clicking through carrier portals. An [AI agent](/riesenia/ai) suits this because the judgement "this is fine / this is not" depends on context rather than a single rule. A shipment with no movement for two days is ordinary over a weekend and suspicious on a Wednesday. Normal for an international parcel, not for one crossing town. ## Unify carrier statuses The first practical obstacle: every carrier reports something different. One says "in transit", another "on the way", a third has five intermediate states that do not line up with anyone else's. Some send webhooks, others have to be polled. So it pays to define **your own status scale** and map carrier statuses onto it. Without that you cannot simply say how many shipments are late, because each carrier reports lateness differently. How these links get built is covered in the piece on [automating dispatch and carrier integration](/blog/automatizacia-expedicie-a-dorucenia-prepojenie-s-prepravcami). ## Silence is a signal too This gets underestimated in practice. An error status from a carrier is easy — it arrives and you know what is happening. The harder case is when **nothing arrives at all**. A shipment sitting on the same status for three days has no error, but it almost certainly has a problem. So the agent should react not only to events but to their absence: compare the actual progress against what is expected for that route and carrier, and speak up when they diverge. ## What the agent should do with an exception By type: - **Stuck shipment** — prepare the claim for the carrier, with all the details and the status history. - **Failed delivery** — contact the customer with the option to change address or date. - **Damage** — start the claims process and reserve a replacement from stock if one is available. - **Wrong address** — compare against the order data and propose a correction. In every case the agent **prepares**; it does not "handle". A claim against a carrier is a relationship between two companies and should pass through human review. ## Customer communication has a limit Telling a customer their shipment is late is right and reduces phone calls. **Promising a new delivery date is not** — the agent has no way to guarantee it, and a broken promise does more damage than silence. The formulation that works is to describe the situation and offer options rather than predict. And always leave a route to a person. ## Connecting to the warehouse Delivery exceptions affect stock: a returned parcel has to reappear in the warehouse, a damaged item must not be sold again. If the warehouse does not know, you get exactly the gap described in the piece on [automating stock movements and inventory](/blog/automatizacia-inventury-a-skladovych-pohybov). ## What to measure - **The share of shipments flagged as exceptions** — and how many really were problems. - **Time from a problem arising to it being caught** — the headline figure. - **The share of cases where the company knew before the customer did.** - **Number of support calls about shipment status.** The second and third say whether the deployment works. The first is for tuning sensitivity: too many exceptions means nobody can work through them, and the system starts being ignored. ## Where to start One carrier and one type of exception — typically the stuck shipment, because it is the most common and the easiest to measure. Once that has run for a month, add a second carrier. If you are trying to get a single view of shipments across several carriers, [get in touch](/kontakt). We build integrations like this routinely — examples in our [case studies](/referencie). --- Web version: https://interfase.sk/en/news/an-ai-agent-in-logistics-shipment-tracking-and-exceptions llms.txt: https://interfase.sk/llms.txt