The problem, precisely
Orders arrive as free text, PDF attachments, spreadsheets, and occasionally a photograph of a handwritten list. Someone reads each one and types it into the order system, twice on a bad day.
The typing is the visible cost. The invisible one is the delay: an order arriving at 4pm is processed the next morning, which is where order-to-delivery time actually goes.
What we build
- A mailbox watcher that picks up orders the moment they arrive, including attachments
- Extraction that pulls products, quantities, dates and delivery details from free text and documents
- Catalogue matching, because customers write “2 boxes of the usual 40mm” and mean a specific SKU
- A review queue for anything below a confidence threshold, with the original email beside it
- A write into your order system, with the source email attached to the record
Customer shorthand is the interesting part
Every regular customer has their own vocabulary. The system learns it: once a human confirms that this customer's “usual 40mm” means that product code, it is never asked again.
That learning is what takes the review rate from uncomfortable to negligible over the first couple of months.
Where the human stays
Low-confidence matches, new customers whose shorthand is unknown, quantities that look wrong against history, and anything above a value threshold you set.
Everything else goes straight through. In the wholesale build we delivered, that combination removed the morning keying entirely and contributed to order errors falling by 93%.
What it costs and how long
Typically £10,000–£20,000 and 6–10 weeks, driven mostly by whether your order system has a usable API. Ask your vendor in writing before anyone quotes — it changes the price more than the extraction does.