Context
A manufacturer sends product to trade fairs and sampling events throughout the year. An events coordinator needs stock at a stand on a particular date, and the request arrives as a message: a fair, its dates, some product codes, and some descriptions of product the coordinator knows by sight rather than by code.
Before this work the orders department retyped those requests from one spreadsheet into another, caught or missed typing errors in short product codes, lost items described in prose because they were not given as codes, left no trace in the company data of which event an order belonged to, and sent no status back.
Problem
The tempting response is a form. Give the coordinator a catalogue, a picker and required fields, and the ambiguity disappears at the point of entry. It also moves the whole burden onto somebody who is standing in a warehouse or travelling to a venue, and the request will continue to arrive as a message regardless, because that is how the work actually happens.
Accepting free text means accepting a matching problem. A phrase describing a wet dog food in a particular can size and flavour has to become a specific product code, and a model asked to do that will produce a code with complete confidence whether or not the product exists.
Approach
The message itself is the interface, so the coordinator writes what they need, the system extracts the event, the dates and the items, maps each item onto a product from the catalogue and comes back on the same channel with an interactive list asking for confirmation. Every proposed mapping carries a confidence score and a quoted fragment of the original message that produced it, so a reviewer sees what the model was reading rather than only what it concluded. Proposals below the confidence threshold go to the orders department for acceptance instead of back to the coordinator.
The AI proposes and does not decide, which in this domain has a precise meaning: no order line reaches the ERP without a human confirmation, and the confirmation is recorded against a person.
Reproducibility is handled the way it is elsewhere in the estate. The product catalogue is snapshotted at the moment of extraction, so a mapping can be re-examined later against the catalogue as it stood rather than as it is now, and the model, the prompt and the response are logged through the tracing layer. Message templates are versioned, because a change in wording changes what people send back.
The system is built as a separate bounded context alongside the planning platform rather than inside it, and the reasoning is worth stating. The domain language is different, the release cycle is different, and the security surface is materially different, because this application exposes a public webhook to a messaging provider while the planning platform does not. The catalogue and the order status are read through the planning platform's stable read interface, and this system never writes to that schema or to the ERP directly.
Privacy was treated as a design input rather than a review item. The coordinators are employees and their telephone numbers are personal data, so message contents are not held in logs in clear form and the tracing layer carries the appropriate classification.
Delivery was staged so that value arrived before the integration did. The first phase produces the spreadsheet the orders department already imports, which removes the retyping immediately, and the later phase writes the order to the ERP through an outbox with idempotency. Status from the ERP travels back to the coordinator on the channel they started on.
Outcome
A request that used to be retyped between two spreadsheets now arrives, is mapped, is confirmed by the people accountable for it and lands in the ERP under the event customer and series, with a trace connecting the order to the event it was for and a status that returns to the person who asked.
What it taught me
Meeting people on the channel they already use costs more engineering than a form and removes a failure mode a form cannot address, which is the request that never gets entered at all. The extraction problem is real, and it is a better problem than the compliance problem of asking busy people to use software they resent.
The second lesson is that a confidence score alone is not reviewable. A number attached to a proposal tells a reviewer how sure the model claims to be, and the quoted fragment of the source tells them whether that claim is worth anything. Shipping the second alongside the first is what makes a review faster than redoing the work.