An invoice reader that extracts the invoice lines and checks them.

Built in this order

1. 1Read
2. 2Check
3. 3Hand on
4. 4Deploy in your cloud

The situation

• Supplier invoices arrive as PDFs and photos. Somebody reads each one and types the lines into the accounting system, or a document service does it per page at a price that adds up.
• The services that read documents charge per page. A language model charges per token. One invoice takes a few thousand tokens, most of them the fixed instructions, and costs about a tenth of a krone. The open question is whether the output can be trusted.
• Trust is the real question, so the reader checks its own output and keeps the source: every line item points back to the page it came from.

What it does, step by step

1. 1. Read

   Acts alone

   Comes in

   A PDF or a photo of an invoice, a receipt, a statement, a contract page.

   The reader

   Sends the document to the Gemini API with a fixed output format: supplier, invoice number, dates, currency, each line with description, quantity, unit price, VAT rate, totals. The raw file is kept unchanged.

   The people

   Nobody retypes a line.

   Sign-off

   "Acts alone". The reader only reads; it changes nothing.

   Traceable

   Every field records the page and the region it was read from.

2. 2. Check

   Asks first

   Comes in

   The structured line items.

   The reader

   Checks that the lines add up to the totals, that the VAT rates are valid, that the supplier is known, that the currency and dates make sense. A document that fails a check is flagged with the reason.

   The people

   The bookkeeper reviews the flagged documents, not the pile.

   Sign-off

   "Asks first" on a flag. A document that passes every check goes straight on.

   Traceable

   Each check records what it compared and what it found.

3. 3. Hand on

   The operator signs off

   Comes in

   Checked line items.

   The reader

   Writes them where you want them: the import format your accounting system reads, for approval there, or the warehouse, where they are stored with the rest of your data.

   The people

   The bookkeeper approves a batch inside the accounting system.

   Sign-off

   "The operator signs off" before anything is posted.

   Traceable

   The import file and the warehouse rows both point back to the document.

4. 4. Deploy in your cloud

   The operator signs off

   Comes in

   Your tech stack: Google Cloud, Azure, or both.

   The reader

   The builder agent is an AI agent that builds and maintains the data pipelines and knowledge workers for you. It deploys each step in your own project as a small function or job. The steps run as Cloud Run functions or Azure Functions and use that cloud's own scheduler, storage, and secret store.

   The people

   The operator is the person at your company who checks and signs off the knowledge workers' work. Every change to the cloud is reviewed by your IT, in-house or your IT partner, and the operator approves each one.

   Sign-off

   "The operator signs off" on every deployment.

   Traceable

   Every deployment has a plan that was shown first and an entry with an id.

The parts

The parts are small and independent, and each one deploys into the stack you already have. They already run in our own work; we adjust them and deploy them in your cloud.

The reader

Gemini API, a Cloud Run function or an Azure Function

The checks

the same function, your rules as configuration

The writer

your accounting system's import format, or BigQuery or Fabric

The runner

Cloud Scheduler and Workflows, or an Azure timer trigger

The storage

Cloud Storage or Azure Blob Storage, raw files kept unchanged

The MCP server

Cloud Run or Azure Container Apps, with sign-in, so people can ask questions of the data

The builder agent

Claude Code, in your repository, with the project's own guardrails

How we count the cost

Every cost figure on this page shows how we calculated it and the date we read it.

1. The reader ran on 491 supplier invoices in production in the week of 19 to 25 October 2026. The client whose invoices they were is not named.
2. Cost per invoice = the Gemini API spend for that week, read from the spend dashboard, divided by the invoices. That week also included our own test runs, so the figure is rounded up to about DKK 0.10; the true production cost is lower.
3. For comparison, Google's Document AI invoice parser costs USD 0.10 per invoice of up to ten pages, and Microsoft's Azure Document Intelligence prebuilt invoice model costs USD 10 per 1,000 pages, both read on 25 October 2026.
4. Prices change, so every price on this page shows the date it was read.

The tech choices, and why

The choices are ours and we would make them again. Each has one reason.

A language model with a fixed output format, not a per-page document service

The model returns the structure we need, and the checks decide whether to trust it.

Unchanged raw files

They are the evidence: a questioned line item points back to its page.

Checks as a separate step with named rules

A failed check can be explained, and you can change the rules yourselves.

Small functions instead of one application

Each function deploys, scales, and costs on its own, and runs in Google Cloud or Azure alike.

The builder agent and the repository stay with you

You can maintain it without us; you are not locked in.

What it would cost you per month

Move the slider to your invoice volume. The reader's price is the measured figure from this page; the others are the list prices, read on the dates shown.

• One page per invoice for the per-page services; a two-page invoice doubles their price and not the reader's.
• Prices change, so every price on this page shows the date it was read.
• Running cost only: the per-page services need no setup; the reader needs setting up in your cloud.

Before and after

• Every invoice line was typed by hand, or read by a service priced per page.

  The reader extracts the lines; the bookkeeper approves a batch and reviews the flags.

• When a line was questioned, someone had to find the PDF.

  Every line points to its page.

• Each new document type needed a new project.

  A new type gets a new output format and a new check, and both deploy as the same small functions.

• Before you could start, you had to build the pipelines.

  The parts already run; you pay for adjusting and deploying them.