Skip to content

Solutions · Data extraction · Custom project Capture the details you need from your documents.

Invoices, delivery notes, orders by email: using your examples, we assess which fields can be read automatically and where a manual check is still needed.

Common hurdles

The details are in the document. They still reach your system by hand.

Open the document, find the detail, type it in, double-check. Document after document, day after day.

  1. The challenge: Typing from PDFs and scans

    Invoice number, amount, order number: every detail is looked up and keyed into an entry screen. That time is then missing for the cases that really need attention.

    With deepsight: Extract instead of re-keying

    The application reads the agreed fields from incoming documents and stores them as records.

  2. The challenge: Every sender, a different layout

    The same detail is called order number in one place and order no. or PO number in another, and it sits somewhere different every time. Reading by fixed position means a separate template for every sender.

    With deepsight: Describe fields by meaning

    We describe which detail is meant – not where it sits. Using your examples, we check whether that holds across your senders’ layouts.

    See the methodology
  3. The challenge: Errors surface later

    A transposed digit in an amount or a wrong order number often only shows up during reconciliation, at payment or when someone follows up.

    With deepsight: Check before handing over

    Values that are hard to read or break an agreed rule are flagged by the application for your team to review.

    See how it runs in operation
  4. The challenge: Sensitive documents, strict requirements

    Contracts or HR records don’t belong in just any online tool. Your IT often requires processing to stay in your own environment.

    With deepsight: Operation on your terms

    Where the application runs is defined in the project – with local language models in your infrastructure if you prefer. We set up, maintain and operate the models; the sample shows whether a local model is good enough for your documents.

    Go to custom development
Example: incoming invoice

One document, three fields, one review view

Sample Office Supplies GmbHIndustriestraße 12 · 49074 Osnabrück, Germany

Invoice no. RE-2026-0412

Invoice date: 08/14/2026 · Customer no.: 20931

Printer paper A4, 40 boxes ........ €612.00

Toner, 6 units ........ €437.20

plus 19% VAT ........ €199.35

Total €1,248.55

Review view

Select a field to see where it appears in the document.

Starting point

What we clarify at the start

Document type

Which documents arrive regularly, in which format and in what quality?

Required fields

Which details do you need, and which rules apply when checking them?

Target system

Where does the data go: a spreadsheet, a business application or an ERP system? The connection is defined in the project.

Approach

From sample documents to operation

  1. Review samples

    A selection of typical documents, including difficult cases.

  2. Define fields

    Which details are extracted and when a field goes to manual review.

  3. Measure a sample

    We measure how often fields are read correctly using your documents.

  4. Connect

    Export or connection to your target system, depending on the project.

Use cases

What your documents should answer

  • Which invoice number, which amount, which due date – and which order number does the invoice state?
  • Which pickup address, which quantities and which delivery date are in an order sent by email?
  • Which items and quantities does the delivery note list?
  • Who are the contracting parties, how long does the contract run, and what notice period applies?
  • Has every mandatory field in an application been filled in?
  • Which measurements and designations appear in test or lab reports?
Documents and sources

What we work with

Documents that arrive regularly and whose details you need again and again.

Digital PDFsInvoices, order confirmations or reports created on a computer. The text is already in the document.
Scans and photosScanned delivery notes or photographed receipts.
EmailsOrders or inquiries whose details are in the body of the email or in its attachments.
Forms and applicationsRecurring forms with fixed fields, even when several versions are in circulation.
Contracts and reportsLonger documents in which the details you need sit in running text rather than in a field of their own.
Tables in the documentLine items with quantity and price, for example from invoices or bills of materials. Whether each row becomes a record of its own is decided when we define the fields.
In operation

The path of a document

  1. Intake

    Documents arrive through the agreed channel, for example a mailbox, a folder or your business application.

  2. Extract

    The application reads the defined fields and puts dates, amounts and numbers into the format your target system expects.

  3. Review

    Your team sees flagged fields next to the document and confirms or corrects them.

    See the review view
  4. Hand over

    Reviewed records go to your target system – as a file or via its interface.

  5. Adjust

    New senders or revised forms? We then check the affected fields again using examples and adjust the rules, as agreed.

    Project phases

The application reads. Your team decides.

Doubtful cases are resolved by the people who know your processes.

Results

What you receive in the project

Requirements descriptionDocument types, fields, review rules and target system in writing, with an effort estimate – as the basis for your decision.
Prototype to try outA first version using some of your documents that your team can test itself.
Measurement resultThe result of the sample – on request for each field and with the documents in which a value was wrong.
Application in operationThe finished application, connected to your target system, with hosting and support as agreed.
Quality assurance

What matters in extraction

Assess each field on its own

Misread bank details matter more than a wrong delivery date. That’s why we discuss for each field whether, after measurement, it is captured automatically or always goes to review.

Rules from your expertise

Mandatory fields, expected formats such as the structure of your order numbers, or cross-checks such as line items against the total – we define these together. This also flags values that are readable but implausible.

Traceable for audits

On request, we also trace automatically captured values back to their place in the document and build logging and review steps for your internal audit into the project.

Good to know: Need a custom-trained model? As part of a custom project, we assess whether this approach suits your use case.

FAQ

Questions about data extraction