What 87 Real Invoices Actually Contain: A Field-by-Field Study

Published 22 September 2026. Where other products are mentioned we describe the kind of product and its pricing model, not figures, which change.

Most advice about invoice spreadsheets assumes every invoice carries the same fields. It does not. We took 87 publicly published invoices, transcribed all 91 invoices and 221 line items inside them field by field against the original page, and counted what is actually printed. The short version: build a column for every field you expect and roughly half of them will be empty.

What the 87 documents are

Two different things, and the difference matters for every number below. 45 are real supplier invoices released by UK public authorities under freedom-of-information requests — actual paperwork from Screwfix, Deloitte, Jurys Inn, Interserve and others. The other 42 are sample invoices published by Microsoft and Google to demonstrate their own document-reading tools.

A sample invoice is tidy by construction, and it shows. 57% of the vendor samples print a due date against 24% of the real invoices — more than twice as often. Where the two populations disagree we say so below, because a blended figure describes a population that does not exist. Where they agree, as with the tax rate (55% real, 52% samples), the blended number is the honest one.

The headline numbers

Across all 87 documents (135 pages) containing 91 invoices and 221 line items, this is how often each invoice-level field was actually printed on the page. Read the due-date row with the split above in mind: among the 45 real invoices alone it is 24%, not 40%.

Three findings that will break a naive spreadsheet

1. Due dates mostly do not exist. Only 36 of 91 invoices printed one. A payment-tracking sheet keyed on due date will be 60% empty, and the usual workaround — assume 30 days — is an invention, not data.

2. A VAT amount is far more common than a VAT rate. 78 invoices printed the tax amount but only 49 printed the rate. If you need the rate, derive it from tax ÷ subtotal and expect rounding noise, rather than expecting a column to read.

3. One in ten invoices has no unambiguous date. Five printed no invoice date at all, and on four more the date was genuinely ambiguous — several dates on the page with nothing marking which was the invoice date. Sorting a spend report by date silently misplaces those rows.

The line-item trap nobody warns you about

Line items are less consistent than invoice headers, and one difference matters more than the rest — whether a row total already includes tax:

Of the 221 rows, 173 had totals excluding tax, 26 had totals including tax, and on 22 it could not be determined from the page. So roughly one row in eight is VAT-inclusive while most are not. Sum a mixed column expecting net figures and the total is overstated — and nothing in the spreadsheet tells you it happened. This is why a line-item table needs a basis column, not just an amount.

Documents are not invoices

Four of the 87 documents contained more than one invoice, so a one-file-one-row import would have lost data. Twenty-two were multi-page. And two were not invoices at all — a gas-usage spreadsheet and a blank bilingual template — despite sitting in folders of invoices. Any bulk process needs to handle "this file is not an invoice" as a normal outcome rather than an error.

How we did it, and what this does not prove

The documents are invoices published openly: open-source document-processing test sets and invoices disclosed in UK public-sector freedom-of-information responses. Each was transcribed field by field from the rendered original and then reviewed in a second field-by-field pass, recording for every field whether it was printed, absent or ambiguous. Line-item rows were annotated the same way, including whether the row total was inclusive or exclusive of tax.

The limits are real and worth stating. Ninety-one invoices is a small sample. It is not a random sample of UK invoices — it is skewed towards public-sector suppliers and towards documents that sample sets chose to include. Transcription and review were done by AI agents against the original page rather than by independent human coders, so the field counts should be read as careful transcription, not audited ground truth. The counts above are recomputed from the annotation files by an automated test, so they cannot drift out of date with the data behind them.

What to do with this

• Treat due date and VAT rate as optional columns, not required ones.
• Store a tax basis alongside every line-item total, or you will sum inclusive and exclusive rows together.
• Let a row exist with a missing date rather than guessing one, and flag it for review.
• Expect a small share of files to contain several invoices, or none.

7 days · 50 documents · No card needed

Try it on one invoice, no account needed · The full 14-tool comparison table · More guides · What Invoices to Spreadsheets is · Invoice to Excel