The order arrives before the information does
A hospital pharmacy takes delivery of biologics at six in the morning. The purchase order calls for two units, the delivery note for three, and the logger records an excursion of 1.4 degrees lasting eleven minutes. Nobody at the dock can say whether that matters, because the stability data sits in a dossier they cannot see.
The order was not wrong because somebody was careless. It was wrong because a commitment expressed in a message had no structured form in which to be checked. What made it expensive were this industry's constraints: finite shelf life, and quality held as a record rather than an appearance.
Why a small error becomes a total loss here
In many businesses a mis-keyed quantity is a rework cost. Here it is closer to a write-off. A consignment with less remaining shelf life than the customer will accept cannot be returned, held against demand, or used. The head of supply chain owns the commercial consequence; the qualified person owns the disposition, and is not at the dock.
| Where it breaks | What gets lost | Who owns the repair |
|---|---|---|
| Order capture | Lot, expiry, remaining shelf life, storage | Commercial owner of the account |
| Acknowledgement | Confirmation the supplier can meet the terms | Procurement and supplier manager |
| Booking and handoff | The cold chain requirement on lane and carrier | Logistics lead |
| Receipt and release | Temperature history, certificates, lot records | Quality lead |
The information breaks in four predictable places
- Capture, where a required attribute has no defined home. The minimum shelf life a customer will accept lives in a purchase order clause and one buyer's memory.
- Acknowledgement, where the supplier replies in free text and the reply is reconciled by eye. A supplier that can refuse a condition in seconds finds it slower to refuse than to ship.
- Re-keying, where the same order is entered into the order system, the quality record and the carrier booking.
- Exception, where the routine path is handled and the unusual case by whoever notices it. A short-dated consignment found at goods-in has no owner.
Structured capture is a precondition rather than a feature
Before any system is chosen, the business has to say what a line of a life sciences order is: which fields exist, which system is authoritative for each, and what happens when a field is missing or contradicted. It also means deciding the rules. A cold chain line cannot be booked onto standard service, and a line with too little shelf life cannot be confirmed.
Only then does the technology question become sharp: read order lines, acknowledgements, purchase orders and logger files, extract the defined fields, check them against the rules, present them for confirmation, and retain the source as evidence. Extraction from prose and scanned documents is where a language model earns its place, because supplier variation defeats a fixed parser.
What should not be delegated to a system
The consequential judgements here have regulatory consequence, and they should stay with a named qualified person. Whether a temperature excursion is acceptable for a given product. Whether quarantined material may be released. Whether a short-dated consignment should be held or returned. These are not questions about what a document says, but about what the business will stand behind.
Traceability is a records problem before it is a data problem
The question a quality auditor asks is rarely about one order. It is about a family: for this lot, what was ordered, what was received, how it was transported, at what temperature and who released it. Answering it means joining events across systems by identifiers that mean the same thing everywhere, as described in how a business ontology works.
Exception handling is where the loss concentrates
The routine path is not where the money goes; it is the well-understood case that arrives in good order. The loss concentrates in four exceptions: a consignment outside its temperature range, one with unusable shelf life, one without the required certificate, and one in the wrong quantity against a confirmed order.
For each, four things need naming, and the absence of any one turns an exception into a complaint. Name the decision. Name the owner. Name the time limit, because a qualified person is not on site at dawn. Name the evidence needed to close it. An approval path that survives an audit is covered in designing AI workflows with approval and exception handling.
What to start with on Monday
- Choose one order type where the stakes are clearest, and write down every attribute that should have been captured.
- Count the correction rate on that order type over a fixed period, and record it as a baseline.
- Define the fields, the authoritative source for each, and the rule for a missing or contradictory field.
- Specify the acknowledgement you want back, and the conditions a supplier must confirm or refuse.
- Name the exception types, the owner of each, the clock on each, and the evidence to close one.
- Agree the measures for review at ninety days, before building anything.
The measures that matter are unglamorous: order lines still needing manual correction, time from order to acknowledgement, consignments rejected on temperature or documentation grounds, and the hours needed to assemble the history of a named lot.
A structured, traceable order record is the only way to know quickly whether what arrived was what was meant, and the only way to answer a quality question months later. The engineering discipline that keeps such a record maintainable is in enterprise agent engineering, and the neighbouring discipline of running exceptions across separate organisations is in managing exceptions before they become expensive failures.