Skip to content
An ochre rendered wall with faded panels
Guide

Forty to eighty invoice formats

Parser coverage is the moat and the treadmill. This is the arithmetic behind that sentence and what it asks you to build.

Type
Guide
Updated
10 September 2026
Reading time
6 min

Count the carriers a group across fifteen countries actually pays. Then count the products: fixed data, mobile, voice, hosting, and usually a legacy line nobody will admit to. Then count the layouts, because a single carrier commonly issues a different document for each product line, and a different one again after a billing-system migration.

Our estimate from that count, not a survey: 40–80 invoice formats across 15 countries for a group of this shape. Carriers change layouts on their own schedule, and the first sign is usually a column that has quietly moved.

So uncertainty is declared rather than smoothed over, and some of it lands in a queue on purpose.

A person signing documents at a dark desk

What opens a review item

Anything below the confidence floor for that specific carrier opens an item. One global threshold would be wrong for both a clean structured XML and a scanned bilingual PDF, so the floor is set per carrier and moves as the format is learned.

  • A field extracted below the carrier's own confidence floor
  • A charge that could not be classified, which is kept and routed rather than dropped
  • A document whose lines do not sum to its total in minor units, or whose tax lines do not sum to its declared tax
  • A date parsed from a Hijri calendar, which always requires confirmation
  • An identifier that did not match any service in the inventory

Making month two cheaper than month one

Every item is also a chance to learn a format once. A mapping is saved against the fingerprint of the header row, together with the carrier and the entity. Two layouts from the same carrier each keep their own, so a layout that reappears next month costs nothing.

Bilingual labels are stored as synonyms rather than translated at parse time. Digit sets, decimal and thousands separators, date formats and currency handling are transforms attached to the mapping. A mapping found wanting in week six is undone in one move back to the pre-batch state.

That machinery is what makes onboarding survivable, and it only pays off if the last few per cent are taken as seriously as the first ninety.

A stack of receipts and invoices

Why the last five per cent matters

Extraction at 95% accuracy sounds excellent until you put it in front of a group CFO. It is wrong one time in twenty, in a document whose purpose is to be arithmetic. One wrong figure in a board pack costs more credibility than forty right ones earn.

The consequence is structural rather than motivational. Human review belongs in the product as a first-class surface with its own queue, its own permissions and its own audit trail. Review designed as the place the work happens will be good; review designed as the place failures go will be thin.

The review surface

An item shows the page it came from and the exact location of the claim beside the field being questioned. The reviewer chooses between ranked candidates rather than retyping a value from a second screen.

Every correction supersedes rather than overwrites. The earlier claim stays, with its source and its confidence, which keeps a conflict between the CMDB and the invoice visible instead of settled by whoever wrote last. It is also what makes a correction reversible.

The queue, the arithmetic checks and the currency handling behind all of this live on the bills page.

Bill processing

People crossing a plaza in low sun, seen from above

Send three invoices. A person replies with what they show.

Request a pilot