"Where is my order?" is the delivery question that keeps arriving long after the parcel has left your warehouse. In online retail it is common enough to have its own abbreviation, WISMO. Most people asking are not upset. They want one reassuring answer and a clear next step before they start wondering whether the order was lost.
An AI agent can take a large share of this load, but only if you are precise about which delivery questions it may answer on its own. This article sorts delivery questions into three groups, shows what each group needs behind it and describes how to test the result before customers see it.
Why delivery questions need their own plan
Delivery questions look alike on the surface and are very different underneath. "How long does shipping take?" can be answered from a page you wrote once. "Where is my parcel?" depends on a live status that sits with the carrier. "My parcel arrived damaged" is a case that needs a person with the authority to decide on a replacement or a refund.
If an AI agent treats all three the same way, it will either hand too much to your team or, worse, answer a question about a specific order without having the facts. The second failure is the expensive one. A confident but wrong answer about a delivery date creates a new complaint on top of the old question.
The three groups of delivery questions
| Group | Typical question | What the answer depends on | Who handles it |
|---|---|---|---|
| 1. General information | How long does delivery take? What does shipping cost? Do you ship to Norway? | Approved shipping and returns information | The AI agent, from knowledge you have written |
| 2. Tracking guidance | How do I track my parcel? Where do I find the tracking number? | Your process and the carrier's tracking page | The AI agent can explain the steps and point to the carrier |
| 3. Order-specific cases | Where is order 4821? My parcel is late. It arrived damaged. | The actual order, the carrier status and your policy | A person, or an AI agent connected to your order system and working under clear rules |
Live order status requires a connection to your shop and your carrier. What is possible depends on your platform, your carrier and the setup, so ask what is supported in your case before you promise customers an automatic answer. Until that connection exists, group 3 belongs with your team, and the AI agent's job is to collect the details and pass them on.
What an AI agent can answer from approved information
Groups 1 and 2 are the safest place to start, because the answers do not change from customer to customer. Write them down once and keep them current in the knowledge base.
- Delivery times per shipping option and per country, including the cut-off time for dispatch on the same day.
- Shipping costs and any free-shipping threshold.
- Which carriers you use and what each option means: pickup point, home delivery or parcel locker.
- How and when the customer receives a tracking number, and where to use it.
- What happens if nobody is home or a pickup point is full.
- Your return window, how to start a return and who pays for return shipping.
Write these as short, plain statements. If your shipping page is a long document with exceptions in footnotes, the AI agent inherits the confusion. The advice in knowledge base best practices applies directly: one fact per statement, clear ownership and no contradictions between pages.
Order-specific questions: connect or hand over
When a customer asks about one particular order, you have two honest options. The first is a connection to your order system and carrier data, so the AI agent can read the status and answer from it. The second is a clean handover to a person. Both are better than letting the AI agent guess.
Either way, decide in advance what the AI agent should collect so your team does not have to ask again:
- The order number and the email address or phone number used at checkout.
- The carrier and the tracking number, if the customer has them.
- What the customer sees on the tracking page and what they expected to see.
- What they want: an update, a delivery change, a replacement or a refund.
- Whether the matter is urgent, for example a gift with a deadline.
The human handoff checklist goes through what a good handover contains, and human handoff explains how escalation works in Aveexia. For rules on when the AI agent should stop and pass a case on, read when the AI should escalate to a human.
What the AI agent should never do
- Guess a delivery date. If it does not have the data, it should say so and explain the next step.
- Promise a refund, a replacement or compensation unless you have written a rule that allows it.
- Describe a tracking status it cannot see.
- Blame the carrier on your behalf. Keep the tone factual and helpful.
- Change a delivery address without an approved process for it.
Put these in the instructions as explicit boundaries. Customers forgive "I can't see that, so I'm passing it to the team with your details." They do not forgive a confident answer that turns out to be wrong.
Prepare for peak periods
Delivery questions rise when volumes rise: Black Friday, the weeks before Christmas, campaign launches and carrier disruptions. Treat these as a recurring checklist rather than a surprise.
- Update delivery times and cut-off dates before the period begins.
- Add a short, accurate note about known delays, and remove it as soon as it no longer applies.
- Check that the handover route is staffed, or that customers are told when they will hear back.
- Run your test set again, because a changed delivery time can quietly break older answers.
Maintaining an AI agent's knowledge base after launch describes a simple ownership routine for exactly this.
Test delivery questions before launch
Build the test set from your real inbox. Take the most recent fifty delivery messages, sort them with the three-group model above and write down the expected behaviour for each. Include the awkward ones: a customer who gives the wrong order number, a message that mixes two languages, an angry message about a late gift and a question about a country you do not ship to.
Then run them the way a customer would and compare the answers with what you wrote. How to test an AI agent before you launch it has a full test plan you can reuse. Fix the source or the rule behind a poor answer rather than adding more wording on top of it.
Measure what matters
Three simple measures show whether delivery handling is improving:
- The share of delivery questions answered without a person, split by group.
- The share of handovers that arrive with complete details, so nobody has to ask again.
- Repeat contact: customers who write again about the same order.
Avoid counting a conversation as resolved just because the customer stopped replying. KPIs for AI chat shows how to define resolved in a way you can trust.
A practical way to start
- List the last fifty delivery questions and sort them into the three groups.
- Write approved answers for groups 1 and 2 and add them to the knowledge base.
- Decide what happens in group 3: a connection to your order system, a handover, or both.
- Write the boundaries and the details to collect for each handover.
- Test with real messages, fix the sources and then open the web chat to a small share of your traffic.
If you run an online store, Aveexia for e-commerce shows how this fits into the wider customer flow, and integrations explains how connections are set up. Availability depends on your platform and plan, so confirm your case before you build a process around it. Pricing describes what a pilot involves.
