Use case
Document processing
From incoming document to posting in the target system: fields extracted, checked against order and master data, clean cases posted. Exceptions go to a person with a stated reason.
01The problem
What the work looks like today
Invoices, order confirmations, certificates, delivery notes, contracts. Each arrives in a slightly different format, gets read by a person, checked against something in a system, and typed or posted. Accuracy is non-negotiable and the volume never drops.
02Failure modes
Where time and quality are lost
Three patterns we see in almost every operation of this kind.
- 01
Documents are checked twice because nobody trusts the first check
- 02
Data is retyped between systems
- 03
Exceptions follow rules that live in one person's head
03The agent
What the agent does, step by step
Reads each document regardless of layout, extracts the relevant fields, validates them against the order, contract or master data, posts clean cases to the target system and routes exceptions with a stated reason: quantity mismatch, unknown supplier, missing certificate.
04Human checkpoint
Where a person decides
Exceptions are decided by a person. Clean cases within defined tolerances are posted automatically once the shadow phase has proven the error rate. Financially critical postings keep a four-eyes check.
05Control set
What is enforced and logged
Field-level confidence with thresholds per document type. Full audit trail from document to posting. Tolerances and auto-post rules are owned by finance, not by IT.
06Measurement
What the monthly report shows
Four numbers, against the baseline from the assessment.
- 01
Straight-through rate
- 02
Extraction accuracy per field
- 03
Exception handling time
- 04
Cost per document