“Where is my order?” is one of the few customer questions that already has a system of record. On Shopify, fulfillment status, tracking, and the shipping window live on the order. A chatbot is useful here only if it reads that order and answers from it. A chatbot that improvises a delivery date is worse than the contact form you already have, because it sounds sure.
This note is an illustration of the workflow, not a claim that Nexen shipped a named store or measured a conversion. Shopify is a product a lot of small shops already pay for. The design below is what we would ask for before calling the bot done. It does not require a custom model. It requires a rule and a lookup.
What the bot is allowed to say
Limit the first version to status. The customer gives an order number and the email on the order, or they are already logged in. The system reads the order. If the fulfillment is unfulfilled, it says so and repeats the handling time you publish on the product or the shipping page. If a tracking number exists, it returns the carrier and the link. If the order is cancelled or refunded, it says that and stops. It does not invent a new date. It does not promise a refund. It does not offer a discount to make the person go away.
- Read the order. Do not answer status from memory or from a generic policy alone.
- Repeat only the window the shop already publishes.
- If the order cannot be matched, ask for the email and the order number again, then hand off.
- Refunds, address changes, and damaged items go to a person with the order attached.
A walk-through, not a case study
Picture a shop that sells a physical product and gets the same question every morning. The customer types an order number. The bot looks up the order in Shopify. The fulfillment says “in transit” and the tracking link is present. The reply is the status plus the link plus the sentence already on the shipping page. That reply can be automatic if you have decided the status text is safe. The same customer then says the box arrived crushed. That message is a different intent. The bot should stop, keep the thread, and show a person the order, the status it already read, and the new complaint. The person decides the replacement or the refund. The bot does not.
If the order number does not match any order for that email, the bot should say it cannot find the order. It should not guess that the package is “on the way.” A wrong confident answer creates a second ticket. The failure message is part of the design. So is the log: which order was read, at what time, and which fields were used in the reply. When the customer argues, the person needs that log, not a paraphrase.
Where it should sit
Put the status answer where the question already happens. That might be the shop’s chat, the order-status page, or a WhatsApp thread if that is how your customers write. A new widget that cannot see Shopify is a dead end. Shopify’s admin and the customer account are the records. The bot is a reader with a narrow pen. If the shop also uses a help desk, the handoff should open a ticket there with the order id, not paste the customer into a blank form.
Authentication matters more than the model. Anyone who knows an order number should not be able to read another person’s address. Match the order to the email, or require the customer to be logged in, before you reveal tracking or a shipping city. Status without the street address is usually enough. Do not echo the full address back into a chat log you also email to a vendor.
What not to automate in version one
Do not let the bot edit the order. Do not let it refund, cancel, or change the ship-to address. Do not let it apply a discount code it found in an old campaign. Do not let it answer product-advice questions that your staff would refuse. Those can be later steps, each with its own approval. Version one is done when status replies match the admin screen on a pile of real orders you already shipped, including delayed ones and cancelled ones. If you only test the happy tracking link, you have not tested the bot.
Also write the hours. If a person is not watching the handoff, the bot should say that a person will see the damaged-item thread on the next working day, using whatever reply-time sentence you are willing to keep. Do not let the model invent “within two hours” because that sounds helpful. If you have no published reply time, say you do not promise one. Nexen does not publish a reply-time promise on its own contact page for the same reason.
How to judge it
Judge it on mismatches, not on a vanity deflection rate. Take thirty real order questions from the last month. Run them against the admin screen. Count how often the bot’s status matches the order, how often it refuses when the order is missing, and how often a person had to correct a sent reply. A single wrong “delivered” is more important than a week of correct “in transit” messages. Keep those corrections. They are the test report. They are not a marketing statistic, and they should not be published as one unless you are willing to show the method.
The cost of this pilot is mostly the connection, the rules, and the handoff, not a mysterious model fee. The next note on this site talks about ranges without pretending there is one price. If order status is the queue you want mapped, the sample plan can sketch the pilot. A consultation is separate, and neither is a quote.
Written by Nexen AI Solution. Published 14 October 2026.

