Every open order, held against every supplier email.

1An order changes in the inbox first.

A purchase order is a promise: a quantity, a price and a date. Suppliers answer it in email. An acknowledgement confirms it or changes it, a ship notice says what is actually on the truck, a delay message moves the date, and a question waits for an answer before anything is made. None of these changes the ERP by itself. A buyer reads each one, works out which order line it is about, and updates the ERP by hand.

Both steps fail in familiar ways. People copying values from one document into another by hand get a share of them wrong: 6.57% in a meta-analysis of 93 studies.¹ The larger risk is the email that is never acted on. In a study of more than 2 million email users, the share of each day’s email people replied to fell from about a quarter on light days to under one in twenty at around 100 messages a day.² That count includes newsletters, so it measures overload, not missed orders. A delay announced in an email nobody answered reaches the plan when the material fails to arrive.

2Match every message to its order line.

Each morning Neil reads the purchasing inbox: acknowledgements, ship notices, delay messages, invoices and questions, in the body, in attachments and in the replies quoted below. Each message is matched to the order line it concerns, by order number, part, quantity, supplier and thread, and what it says is read into fields: a date, a quantity, a price, a question. A message Neil can’t match for certain is shown as unmatched, never guessed.

Reading is where an AI model helps, and where it still slips. In a 2025 preprint, the best model read 92.71% of fields correctly on scanned invoices and 96.50% on clean ones.³ So every value Neil reads keeps a link to the line of the email or the page of the attachment it came from, and an uncertain value is read twice.

3Compare in code with the ERP.

Then code compares each message with the ERP: the supplier’s date against the ERP’s, the quantity shipped against the quantity ordered, the price acknowledged against the price on the order, the days an order has gone unacknowledged and the days a supplier’s question has gone unanswered. Your buyers set how long is too long, for each supplier or material.

4Show what each change reaches.

A changed date matters by what depends on it. Neil follows each change to the materials, work orders and customer orders it reaches, from the same morning’s run as material demand planning and production schedule alignment. A three-week slip on a material with months of cover is a note. The same slip on a material that is already short is the first thing a buyer sees.

5Five messages at Acme Medical.

Acme Medical is a fictional ISO 13485 maker of tubing and catheter components with nine regular suppliers and two buyers. On Thursday 5 November 2026, Neil’s 06:00 run raised five orders from the purchasing inbox.

The first matters most. PO 4500-1182 is the PEBA 72D meant to cover Acme Vascular’s ramp, and the ERP still shows 20 November. At the new date the material runs short, and the work order that needs it can’t run as planned.

6The buyer sends the follow-up.

Neil drafts; the buyer decides. A draft opens beside the message it answers, the order line and the ERP’s figures. The buyer edits it, sends it from their own mailbox and updates the ERP. Neil never sends an email, and never places, changes or cancels an order.

Agents left to work through email alone are not there yet: in a 2026 preprint, the best agent completed 33.5% of realistic email tasks, though 99.7% of its tool calls ran without error.⁴ Neil’s part is the reading, the comparison and the draft. The judgement and the send stay with the buyer.

Each follow-up is kept with the message it answers and what the buyer did. FDA does not routinely inspect component makers;⁵ it relies on device makers to control what they buy, and those device makers ask their suppliers for the record.

7Validate it on three months of mail.

Take three months of one buyer’s supplier email and the open-order exports from the start of that period. Neil reads the mail as if each day were live. Compare what it matched and raised with what the buyer did and what the ERP showed at the end: the dates that moved, the partial shipments, the prices that changed and the questions that waited.

Count what Neil raised in time, what it missed and what it raised that didn’t matter, and have the buyer read the drafts as they would a colleague’s.

8Start with one buyer’s inbox.

Bring three months of one buyer’s supplier email, exported, with the open-order and receipt exports from the same period. No live access to the mailbox or the ERP is needed to prove it.