Use a tool when someone needs an answer now
For the delivery enquiry, the adviser already has an order reference. The assistant needs to ask the order system for its current dispatch status. The Model Context Protocol (MCP) gives a compatible AI application a standard way to discover tools and call them with named inputs. The tool still needs a working connection to the order system behind it.
Scope a first tool to one question. Give it an order reference and return the dispatch status, the time of the last recorded change and a link the adviser can open. Check the adviser’s access before returning the record. If the system is unavailable, show that the lookup failed. A confident sentence about yesterday’s status would be particularly unhelpful to someone waiting for a van.
Before commissioning an MCP server, try the lookup in the software staff already use. If everyone works in the same order screen and the status is one click away, improving that screen may be cheaper. A tool becomes more attractive when advisers already handle enquiries in an assistant and have to interrupt each conversation to find the same information.
Use a webhook when the change starts the work
For the switch from collection to delivery, waiting for an adviser to ask a question leaves a gap. If the order system offers a suitable webhook, it can send a notification to your integration when the order changes. Check which changes actually generate that notification. A connector that only announces new orders will miss an amendment to an existing one.
Give the receiving service a simple job. Verify the notification, save it for processing and use the order reference to check what dispatch needs to do now. A fixed rule can put an order on the loading list. An AI assistant might later explain the change to the adviser, but the van should not depend on someone opening a chat.
Read the source system’s delivery rules. For example, Stripe documents that webhook events can arrive out of sequence. Its behaviour is a reason to ask your own supplier about retries, duplicates and ordering. Design the dispatch update so that receiving the same change twice leaves one loading instruction.
Use a scheduled check when a delay is acceptable
A scheduled job can compare the source with the loading list every few minutes, or produce a report before dispatch closes. Ask the manager how long an amended order can wait. A check every fifteen minutes can already be too late for an order changed five minutes before departure. Allow for the time taken to read and apply the changes as well.
Some services let you request only what has changed since a previous read. Microsoft Graph’s delta query is one example for supported resources. The application follows the returned page links until it receives a link for the next round of changes. Save that final link after applying the results successfully. Advancing it too early can leave a gap after a crash.
If your supplier offers only a full list, price the work of reading it and check its request limits. For a small depot, comparing today’s open orders with the loading list may be enough. Include a way to detect cancellations and orders removed from the source view. Otherwise the comparison can keep adding jobs without ever taking obsolete ones away.
Illustrative example — the van leaves at 4 pm
Suppose a depot closes its loading list at 3.55 pm. Its scheduled check starts on the quarter hour. At 3.51 pm a customer switches an order to delivery. The 3.45 pm check has already finished, and the 4 pm check begins five minutes after the list closes. Even flawless quarter-hourly checks cannot meet that particular deadline.
A webhook could flag the change in time if its delivery and processing delay stays within the four minutes available. The depot should measure that delay. A supervisor can run a final comparison before closing the list and decide what to do with changes that arrive during the comparison. Put a name against that decision so a late order has somewhere to go.
| Desk request | What starts it | What to check in a trial |
|---|---|---|
| Has this order left? | An adviser requests the current record. | The answer agrees with the source and shows when the status changed. |
| Add this amended order to the van. | The source reports an order change. | The change reaches dispatch before the cut-off, once. |
| Did we miss anything? | A scheduled or supervisor-led comparison. | A deliberately omitted update is found and assigned to someone. |
Ask for a demonstration with one real deadline
Take a recent anonymised order through the three requests with your supplier. Disconnect the test receiver briefly, amend the order, then reconnect it. Watch where the missing change appears, how the desk knows it has caught up and whether the same order appears twice. A successful live lookup settles only the first of those questions.
Choose the smallest connection that meets the deadline your team actually has. A daily discrepancy report may be sufficient for a back-office handover. A dispatch desk with a four-minute window needs evidence about late notifications and a named fallback. Ask the supplier to show that evidence before paying for a wider integration.