Customer orders arrived as PDFs and spreadsheets attached to emails, in roughly thirty different layouts. Four people spent most of their day reading them and typing the contents into the ERP. Errors were rare but expensive, and the backlog grew every quarter end.
Why the obvious answer was wrong
The first instinct is to ask customers to use a portal. That had been tried. The largest accounts, who sent the most orders, were the least willing to change how they bought. Any solution had to accept the formats already in use.
What we built
Extraction with a confidence score
Each document is parsed into structured line items with a confidence value per field. Layout templates handle the recurring senders; a model handles the rest. The important design decision was that nothing is auto-accepted on a low score.
An exception queue, not an approval queue
High-confidence orders post to the ERP directly. Anything uncertain, a new sender, an unrecognised product code, a quantity outside the usual range, goes to a review screen showing the document and the extraction side by side. The team now looks at the hard cases instead of all cases.
Validation against the ERP, not just the document
Product codes, pricing and credit status are checked before posting. Several classes of error that used to surface at dispatch are now caught at entry.
Rollout
We ran extraction silently alongside the manual process for five weeks, comparing results without acting on them. That gave a measured accuracy figure before anyone was asked to trust it, and it caught two product families the parser handled badly.
Where it landed
About seven in ten orders now post without a human touching them. The order desk did not shrink; it moved onto customer queries and expediting, which had previously been squeezed.
This sits across our automation and process and integration work. If re-keying is eating a team's day, tell us about it.