Step by step: which parts of delivery tracking are plain automation, where AI reads carrier messages and drafts updates, and which decisions about late or lost shipments stay with a person.
Mostly yes, but the biggest part is plain automation rather than AI. Carrier APIs and webhooks already report where a parcel is, and simple rules can match each shipment to its order and flag the ones running late. AI earns its place on the messy edges: reading free-text carrier exceptions and customer emails, and drafting a clear status update. A person still decides what to do when a delivery is late, damaged or lost. Below you will find the split per step, what you need in place first, and a small pilot you can run on one carrier.
What each step needs
01● Standard automation
Pull tracking events from carriers
Every major carrier offers tracking through an API or webhooks, and many shops use an aggregator that normalises several carriers into one feed. This is a mechanical job: a scheduled call or an incoming webhook stores each event with the tracking number, the timestamp and the status code. AI adds nothing here and can only add errors if it is asked to guess a status. Keep the raw event next to your own cleaned-up status, so that when a customer disputes a delivery you can show exactly what the carrier reported and when.
⛨Works when you have tracking numbers for every shipment and API access or an aggregator for each carrier. Smaller local carriers may only offer a web page, which needs a different approach.
02● Standard automation
Match each shipment to its order and promised date
The tracking number has to be linked to the order, the customer, the items and the delivery date you promised. That is a lookup in your order system or warehouse system, and it should be exact. Split shipments, partial deliveries and returns are the usual trouble spots, so the rules need to handle one order with several parcels. If the link is wrong, every later step is wrong, so this is worth testing on a batch of real orders before anything is automated downstream.
⛨Needs the tracking number written back to the order at dispatch. Without that link, someone has to match shipments by hand and the rest of the flow stalls.
03● Standard automation
Flag late, stuck and at-risk shipments
A shipment is late when the expected date passes without a delivered status, and it is stuck when no new scan arrives for a set number of days. Both are simple date comparisons, and fixed thresholds per carrier and service level work well. Start with a short list: no first scan within a day of dispatch, no movement for three days, delivery attempt failed, and returned to sender. Rules give the same answer every time and are easy to audit, which matters when you later measure how many late parcels you caught early.
⛨Thresholds must be set per carrier and service, since an international parcel and a next-day parcel behave differently. Review them every quarter against real transit times.
04● AI candidate
Read carrier exception messages and customer emails
This is where AI helps. Carrier exceptions arrive as short, inconsistent text such as address issues, weather delays, access problems or a customs hold, and customers write in their own words about a missing parcel. A language model can classify each message, extract the tracking number and the cause, and route it to the right queue. Treat its output as a label with a confidence, not a decision. Anything it cannot place goes to a person, and a sample of its labels should be checked every week to catch drift or a new carrier phrasing.
⛨Works best with a fixed list of exception types and a few real examples of each. Keep the model away from anything that implies a refund or compensation.
05● AI candidate
Draft proactive status updates for customers and colleagues
Telling customers about a delay before they ask reduces repeat contacts. AI can turn the structured facts into a short message in the customer's language and tone: what happened, the new expected date if the carrier gave one, and what happens next. The facts come from the data, not from the model, so it should only phrase them. For routine notices a template with filled-in fields is often enough. Use the model for the cases that need a more careful message, and have a person approve anything that mentions fault or compensation.
⛨Needs a short style guide and approved wording for apologies. Start with internal summaries before sending anything to customers automatically.
06● Human review
Resolve lost, damaged and repeatedly late deliveries
Deciding whether to reship, refund, file a carrier claim or switch carrier involves money, customer relationships and contracts. A person weighs the order value, the customer history and the carrier's claim terms, and takes the conversation with the customer or the carrier when it is sensitive. AI can prepare a case file with the event history and the order details, which saves time, but it should not decide compensation. Over time, patterns in these cases also tell you which carrier or lane to review at the next contract round.
⛨Always, for lost or damaged parcels and high-value orders. Set a clear rule for when a support agent may act alone and when it goes to a manager.
A sensible first experiment
Pick one carrier and one week of outbound orders. Run the rules for matching and late flags next to your current manual process, and let AI only classify exceptions and draft messages for internal review. Measure three numbers: how many late or stuck shipments you caught before the customer contacted you, how many 'where is my order' contacts came in compared with the previous week, and how many AI labels a person had to correct. Keep the sample small enough that someone reads every exception. If the correction rate is high, tighten the exception list before you add a second carrier. Do not turn on automatic customer messages until the labels have been right for several weeks in a row.
The trap to avoid
The usual mistake is to buy or build an AI layer before the order-to-tracking link is reliable. If tracking numbers are missing, duplicated or attached to the wrong order, no model can fix the output, and customers get confident messages about the wrong parcel. A second trap is letting a model infer a delivery date the carrier never gave, which turns a vague delay into a promise you cannot keep. Use the carrier's own date, say clearly when none exists, and let a person own any commitment.
Questions teams ask
Do I need AI for delivery tracking at all?
Not for the core. Pulling tracking events, matching them to orders and flagging late shipments is rule-based automation, and many shipping and e-commerce tools already do it. AI becomes useful when you have a lot of free-text exceptions and customer emails that need sorting, or when you want clearer proactive messages. If your volume is small, rules and templates may be all you need.
Can AI predict when a delivery will arrive?
It can estimate from historical transit times for a carrier and lane, and some carriers and tracking platforms publish their own estimates. Treat any estimate as a guide, not a promise, and always show the carrier's date when there is one. Avoid giving customers a date that comes only from a model, because you will be held to it.
Is it safe to let AI message customers about delays?
It is safe for routine, factual notices where the data comes from the carrier and the wording is approved. Keep a person in the loop for anything that admits fault, offers a refund or involves a high-value order. Start by having the model draft messages for review, and only automate the categories where a sample has been correct for weeks.
Which carrier data should I store?
Store each tracking event with its timestamp and status code, the tracking number, the order reference and the carrier. Keep the raw event unchanged next to your own cleaned status. This gives you evidence for claims and disputes, and it lets you measure real transit times per carrier over time.
Illustrative workflow guidance by Arcgent. Each business needs its own assessment. No integration or savings claim has been verified for your systems.