Skip to content
← Insights
Nov 14, 20264 min read

The invoice your AI should refuse to guess

Reading a field, checking its arithmetic, and authorising a record are separate decisions. An administrative workflow needs to preserve unresolved cases and measure correction work.

An invoice sheet under a magnifying glass beside a calculator, surrounded by a small document tray, check mark, and warning triangle.

An invoice reader should be able to leave a field unresolved. A plausible number is a poor substitute for a number the document supports.

The appeal of automating administrative work is easy to understand. Copying values between systems gives a small typing error a route into an order or an account record. An AI system can help read the material, but the workflow still needs a way to recognise when the proposed value is wrong.

Separate reading from permission to write

Consider an illustrative invoice line with a quantity of three and a unit price of 19.90. The line amount should be 59.70 before any separately specified adjustment. A workflow can calculate that relationship directly once it has the inputs.

The model's task might be to identify the printed fields. The arithmetic check has a different job: determine whether those proposed values agree with the line total. Combining both tasks into an instruction to produce a sensible-looking invoice makes it harder to identify which step failed.

The example also exposes the limit of checking arithmetic. If the system misreads several fields consistently, the values may still add up. A balanced total can pass a calculation while referring to the wrong quantity. Agreement with the source remains a separate test.

I would keep a reference to the part of the document supporting each extracted value. A reviewer investigating a questionable amount should be able to see the relevant image or text, without searching through an entire file. This is a proposed workflow design, not a claim that every extraction product supplies reliable source locations.

Use rules for relationships the business already knows

The JSON Schema validation documentation explains how a schema can require properties and constrain an object's structure. Such checks can reject a missing field or a value with the wrong type before another system accepts it.

A schema still needs to be paired with the business rules. A string shaped like an invoice number could belong to a different supplier. A date can be syntactically valid while referring to the wrong accounting period. Structural validity and agreement with the intended transaction are different checks.

For known calculations, use a defined numerical representation and an explicit rule. Currency amounts should be processed using the representation chosen for the accounting system, so a language model is not asked to invent a rounding convention. The workflow needs to document how that representation is applied at each boundary.

This division gives an unresolved case a useful destination. A missing supplier match can go to the person responsible for that account. An unreadable field can go back for a better document. Those outcomes are more specific than a general exception queue with no indication of what needs to be decided.

A confident response can still contain invented content

NIST's Generative AI Profile includes confabulation among the risks considered for generative systems. A workflow using generated text needs to account for that possibility when the output contains fields that will become records.

That concern changes what completion means. A response containing every expected field can be less useful than a response that marks an unreadable value as unresolved. The system should be evaluated against the original document, including its ability to stop when the evidence is insufficient.

Confidence also needs an interpretation. A high score is useful only if its relationship to actual errors has been assessed for the deployed task. A number printed beside an extraction does not automatically provide the probability that the field is correct.

Measure error after the whole transaction

The Generative AI at Work study found productivity improvements from an assistant used by customer-support staff. It studied a particular working arrangement. A result about issues resolved per hour cannot be copied into an estimate of invoice-field accuracy.

For an administrative pilot, I would count records that are accepted correctly after review. Include time spent correcting an extraction and any downstream work caused by a mistaken write. Otherwise a fast extraction stage can move effort into another team's queue while appearing to save time locally.

A comparison with the existing process needs the same acceptance standard. A human review may catch a model error before it reaches the account system. That is a successful review intervention, but it should still count as work required by the automated process.

The review should also look for repeated mistakes. A mapping error can affect every document from one supplier, so the individual cases may share a single cause. Retaining the mapping version makes it possible to identify which records need rechecking after that cause is found.

The useful automation is the one that leaves a transaction easier to verify. It reads the source, applies the rules that can be stated exactly, and sends genuinely unresolved decisions to an owner. Removing a typing step becomes an improvement when the accepted record is more reliable at a measured total cost.

AIAutomation