Every organization has people whose job includes retyping documents: supplier invoices into the accounting system, contract terms into a spreadsheet, application forms into a customer or employee record. The work is slow and error-prone, and it is rarely anyone’s idea of a good use of their time.
Document intelligence changes the economics. Models trained on large numbers of documents read a page much as a person does, recognizing the layout, the tables and the labels, and return each field with a confidence score. That is the part that demonstrates well. Whether the investment pays back is decided by what happens after the reading.
Read, extract, check, route
A working document process has four steps, and only the first two depend on artificial intelligence:
- Read. Documents arrive by email, portal or scanner, are converted to text and are classified: invoice, credit note, contract, application form.
- Extract. The model returns the fields that matter, such as supplier, invoice number, date, purchase order, lines and total, each with its confidence.
- Check. Business rules compare the fields with what the organization already knows: the supplier master, the purchase order and the goods receipt, contract prices, tax rules, totals that must add up, and invoice numbers already paid.
- Route. Documents that pass every check are posted without a person. The rest go to a queue, with the doubtful fields highlighted next to the original page.
Where it pays back first
The best first candidates combine high volume, recurring layouts and a system of record to check against. Supplier invoices matched to purchase orders are the classic case for intelligent automation, followed by remittance advices, proofs of delivery, claims and onboarding forms for suppliers, customers or employees. Rank the candidates by the hours spent on them today and by the cost of their errors: late-payment penalties, missed early-payment discounts and duplicate payments.
Contracts are different. Volumes are lower and each document matters more, so the value lies not in speed but in coverage: renewal dates, notice periods, price escalators and liability limits extracted into a register that someone actually monitors. A single missed renewal or unapplied price increase can outweigh the cost of the whole project. Keep a contract manager as the reviewer of every extracted term until the register has earned that trust.
Forms deserve a question before any model is built: if the organization designs the form, can it be made digital instead? A well-designed online form captures clean data at the source, and reading paper becomes the exception rather than the process.
Rules do more of the work than the model
A field read with high confidence can still be wrong for the business. The invoice number is read perfectly, but the invoice has already been paid. The total is correct, but the price is above the contract. The purchase order exists, but it is closed. None of these is a reading error, and no better model will catch them. Rules that compare each document with the systems of record will.
Model confidence and business rules together decide what is processed without a person. Set thresholds by field, not by document. A supplier name read with moderate confidence can be confirmed against the supplier master, while bank details should never be changed on the strength of a document alone, however confident the model is.
Duplicate detection deserves rules of its own. The same invoice often arrives twice, by email and by post, or comes back with a new number after a payment reminder, and a model that reads both copies perfectly will happily post both.
Design the exception queue before the model
The exception queue is where the economics are decided. Every document that fails a check should arrive with the reason, the doubtful fields highlighted, the original page beside them and, where possible, a suggested correction. The queue is sorted by what matters, such as payment due date, early-payment discount or value, and every item has an owner.
Corrections made in the queue are the best training data available, so they should flow back to improve extraction. The reasons for exceptions are just as valuable. When the same supplier never quotes a purchase order number, the fix is a conversation with that supplier, not a better model.
Who works the queue matters as much as how it is built. Accounts payable clerks stop being typists and become reviewers of exceptions, and their knowledge of suppliers turns into the rules that keep the next invoice out of the queue.
Measure touchless processing
The headline measure is touchless processing: the share of documents that go from receipt to posting without anyone opening them. Field-level accuracy is a useful technical measure, but it says nothing about how many invoices a person still had to handle. Track the touchless rate next to the time from receipt to posting, the exception rate by reason and by supplier, and the errors found after posting, which should be close to none.
Expect the rate to climb in steps rather than overnight. Each rule that is tuned, each supplier that changes how it invoices and each batch of corrections fed back to the model moves more documents out of the queue. Report it monthly by document type and by supplier, and the next improvement is usually obvious. Watch for the opposite failure as well: a touchless rate that rises together with errors found after posting means the thresholds are too loose.
Start with one document type
Start with one high-volume document type, one system of record to check against and an exception queue designed with the people who will work it. Prove the touchless rate there, then add the next document type. Log every check, its result and the confidence of every field, so that when auditors ask how a document was posted without a person, the answer is already on record. The reading keeps getting better; the checks and the queue are what make it pay.