An order specifies a new delivery address. An AI agent connects email and order data, flags contradictions and prepares the order confirmation for your team.
Illustrative use case — not a report of a completed client project.
An existing customer places an order by email and requests delivery to a new site. The order system still holds the previous address, while the email thread also mentions a preferred delivery date. The order team needs to identify the current details and avoid confirming a draft based on outdated information.
The proposed workflow
The agent identifies the customer, items, quantities and requested change from the message. It links the email to an order and shows the relevant passages. If the customer or order cannot be matched reliably, the case goes to the order team.
The email details are compared with approved order data. The agent highlights differences in the street, postcode, city and recipient. A formally complete address is not proof that delivery is possible or that the customer has authorised it.
For a new delivery address, the agent drafts a question showing the proposed change. It is sent through the agreed approval process. Only a confirmed change may be written to the order; a passing reference in an old email is not sufficient.
The agent drafts the confirmation using checked order data. Delivery dates and prices come from the designated systems and are not invented. The order team reviews outstanding points and approves the confirmation. A unique case reference prevents duplicate processing.
The pilot needs access to the order mailbox, readable order data and a reliable way to identify customers. The team first defines which data may be read and which may be changed after approval. Tests include multiple addresses in a thread, missing building numbers, duplicate messages and unknown senders.
A delivery-address change can have commercial consequences. Confirmation therefore stays with an authorised person or an agreed customer verification process. If deliverability checks are required, a suitable data source must be connected. A language model alone cannot verify a real-world address.
Evaluate value in the pilot
The potential benefit is a prepared case instead of an unstructured email thread. Measures include time to an approvable draft, the number of corrections and the share of address discrepancies correctly detected. Incorrect changes are tracked explicitly so speed does not come at the expense of quality.
That is not the goal. The first step is to assess the read and write interfaces of your existing system. If a secure integration is unavailable, the scope is adjusted or the workflow initially stops at a reviewable draft.
