Parseur, Docparser, Mailparser: a parser is not an answer

2 September 2026 · 本页面以英文提供。

Search for a way to get numbers out of the invoices in your inbox and you will find a well-built category of tools: Parseur, Mailparser and Docparser have been doing this for years. They work. The buyer who wanted his prices checked ends up disappointed by all of them for one reason: they answer a different question.

A parser returns fields; Price Double Check returns the finding. An invoice and a price list forwarded by email come back as a line-by-line comparison — this line was billed above the list, and by how much — rather than as JSON for someone else to act on. That difference is small on a slide and large in practice, and it is worth being precise about what sits inside it.

The gap itself is one sentence long. A parser turns a document into fields, and it does not tell you whether the price on line seven was the price you were promised.

What the category sells, and for how much

Note the bottom two rows. An enterprise accounts-payable product and a contingency auditor both exist because documents arrive as attachments and nobody checks them line by line. Their existence is the best available evidence that the problem is real and expensive. Their price is the reason a company with thirty suppliers has neither.

Where the work actually is

Assume extraction is solved and you have clean line items from every invoice. Here is what still stands between that and a sentence a buyer can act on.

• Which price list is in force. Suppliers do not send a new list every month; they send one when they feel like it. Comparing against the newest file you happen to hold is right most of the time and wrong exactly when it matters — in one real case the current year's list was never sent at all, and one line came in 17.75% above the old one.
• Which column is yours. A manufacturer's sheet often carries a recommended retail price, a retailer price and a distributor price. The right column is a fact about your relationship with that supplier, not something to be guessed from the file. And the columns are not proportional: across 177 rows in one real list the ratio between two of them took at least six distinct values, so there is no single contract discount to derive.
• How to join. Descriptions do not match between an invoice and a price list — 335 lines out of 442 in one set. The item code is the key, and codes come with undocumented suffixes, duplicates at different packaging, and lines that print no code at all.
• How to round. Four decimals in the list, two on the invoice, half-up, one cent of tolerance. Get this wrong and every invoice produces false alarms.
• What not to say. Freight has no reference price. A unit-of-measure collision looks exactly like a short delivery. An item that appears in no price list is an absence of basis, not an overcharge. A tool that reports all of these as problems gets ignored within a month.

None of that is extraction. All of it is judgement encoded once, and it is the whole distance between a table of numbers and an email that says: this line was billed at €13.23 against €11.235 in the list you were given, which is €47.76 on this invoice.

The build-it-yourself route

The arithmetic on the tools is tempting: about $39 a month for the parser, about $20 for the automation, and a spreadsheet you already own. What that leaves out is that the spreadsheet is the part with the rules in it, and the rules above are the ones that take the time. It also leaves out maintenance: a parser configured against one supplier's layout is a parser that breaks when that supplier changes its accounting package, and you find out through a silently empty field rather than an error.

There is a fair version of this route. If you have one supplier, a stable layout and someone in the building who enjoys this kind of thing, it works. It stops working at the point where you have thirty suppliers with thirty layouts and the person who built it changed jobs.

Where a chat with a model genuinely substitutes

It should be said plainly: for a single invoice, uploading it to a general AI assistant along with the price list and asking which lines differ works. It reads the PDF, it finds the columns, it does the subtraction, and for one document it will often be right. Anyone arguing otherwise has not tried it recently. As a way to answer one specific suspicion about one specific invoice, it is a good tool and it costs nothing extra.

What it does not give you is everything that makes the check a routine rather than an errand.

• It does not remember. Each conversation starts without the price lists you uploaded last quarter, so drift across months — a price that quietly became the new normal — is invisible by construction.
• It does not prove it read the document. There is no gate that adds up the extracted lines and refuses to speak until the sum equals the printed total, which is the check that catches misreadings a plausible-looking answer will not.
• It does not hold your rules. Which column is yours, the rounding convention, the categories that must not be judged — you would have to restate them every time, and the one time you forget is the time you get a confident wrong answer.
• It does not run without you. The invoices that go unchecked are not the ones you were suspicious about; they are the ordinary ones in a busy month, and those never get uploaded.
• It does not build a crosswalk. Lines with no item code were recovered here by using the supplier's own invoices from further up the chain, which only exists if something keeps every document you were ever sent.

What to look for, whatever you choose

Whether you end up with a parser and a spreadsheet, a chat window, or a service, the same three properties decide whether the check survives contact with a real month. It must prove that it read the document, by reconciling to the printed total. It must keep the history, because the most valuable findings are only visible across documents. And it must be honest about what it cannot judge, because a routine that cries wolf is a routine nobody reads.

Forward one invoice and the price list it should have been billed on to your Price Double Check address, and the comparison comes back by email. The prices, and an answer to read before deciding, are on the home page.