Paper in, records out.
Every business has a queue of inbound documents that somebody reads and retypes — invoices, delivery notes, orders, remittances, certificates, forms. The reading is the easy part to automate. The part that decides whether it is usable is what happens to the ones it gets wrong.
Any inbound document·Validated against records·Exceptions reach a person
The queue as it exists today.
This is the work that is invisible until somebody is off, and then it is the only thing anybody talks about.
- 01
A shared mailbox as a workflow
Documents arrive, get read, get actioned, and get filed — with the state of the queue existing only as which emails are still unread.
- 02
Retyping into the system
A person reading a PDF on one screen and typing the same numbers into a form on the other, several hundred times a week.
- 03
Formats defeat templates
The template-based scanning tool works until a supplier changes their layout, and then quietly does not.
- 04
Errors surface downstream
A transposed figure found at reconciliation, weeks later, with no easy way back to the document that caused it.
- 05
No queue, no ageing
Nobody can say how many documents are waiting, how long they have waited, or which ones are stuck.
What runs instead.
Reading the document is the first step and the least interesting one. What matters is that it is validated against records you already hold, and that failures are visible.
Read without a template
Layout is inferred rather than mapped, so a supplier changing their invoice format does not break the process — the common failure of template-based capture.
Resolved against your masters
Supplier, customer, item, and account matched to the records you actually hold, including their own part numbers and references.
Validated, not just extracted
Totals checked against lines, tax against the applicable template, and quantities against the open order — arithmetic before acceptance.
Into a real document
A draft purchase invoice, delivery note, order, or payment entry — a record with a workflow, not a row in a staging table.
Confidence-based routing
High confidence and validated goes to a draft for release; anything else routes to a person with the source document beside the extraction.
A visible queue
How many are waiting, how long they have waited, and where each one is stuck — with ageing, because a queue without ageing hides its own problems.
What the agent may and may not do.
The honest position on document reading is that it will be wrong sometimes. The design question is what happens then, and these limits are the answer.
- May read
- Inbound documents, and the master and transaction records needed to validate them
- May do
- Create draft documents, resolve parties and items, and flag for review
- May not do
- Submit a document, release a payment, or overwrite a master record
- Thresholds
- A confidence floor per document type, below which a person always sees it
- Validation
- Arithmetic and tax checks must pass before a draft is offered at all
- Provenance
- The source file attached to every record it produced, permanently
- Correction
- A person's correction is recorded against the document, not silently applied
The modules underneath this.
Document processing is not a module — it feeds them. These are the ones inbound documents most often become.
How to tell whether it worked.
Straight-through rate on its own is a misleading figure. Read it beside the error rate, or a system that accepts everything will look excellent.
- Straight-through rate
- Documents becoming a validated draft with no human touch, by document type and by sender.
- Error rate on accepted documents
- The important one. A high straight-through rate with an unmeasured error rate is not a result.
- Queue age
- How long documents wait, and whether the oldest are old because they are hard or because they are forgotten.
- Exception reasons, ranked
- What actually fails — unknown supplier, missing order, tax mismatch — because most of it is fixable upstream.
- Time to first touch
- From arrival to appearing as a record, which is what determines whether early-payment terms are reachable.
- Rework after acceptance
- Documents corrected after they were posted, which is the true cost of accepting a bad extraction.
The programmes this normally arrives inside.
Next step