Skip to main content
Labs // AI sandbox

Watch an invoice turn into structured data.

A safe, synthetic walkthrough of the review step in the document extraction pipelines we build. No uploads, no real data, no real model.

Local synthetic preview // No file upload or OCR

Explore an extraction review

Run an invented sample through a local preview endpoint to explore the review flow. It does not accept documents, run OCR, or send data to an external service.

Document preview

Synthetic sample

Sample invoice

Run the preview to see sample details

Illustrative text only. No real business or person is represented.

Illustrative extracted fields

Preview not run
FieldSample valueSignal

    Run the preview to load illustrative fields.

    The signal values are invented endpoint fixtures, not model measurements. A real workflow needs an OCR service, validation rules, and a human review policy.

    Synthetic samples only

    The sandbox uses two invented documents and fixed example values. It accepts no uploads and runs no OCR or model. Production pipelines run in your environment with your retention rules.

    What the preview shows

    Choose a sample invoice or purchase order and run the preview. The sandbox returns the fields a pipeline would extract, such as supplier, document number, dates, totals, and line items, each with a signal value. Fields below 75 percent are flagged for review, which is how a real pipeline decides what a person must check.

    Some fields are deliberately problematic: a purchase order match that needs a person, or a supplier confirmation that isn't present. Mark the sample as reviewed to see how the status changes once someone has checked it.

    The signal values are fixed example numbers chosen to illustrate the flow. They are not measurements of any model.

    What a production pipeline adds

    StageIn production
    IntakeEmail inboxes, uploads, scanners, or API, with file checks
    ExtractionOCR and models returning fields in a strict schema
    ValidationBusiness rules such as totals, supplier records, and purchase order matches
    ReviewA queue where people confirm or correct uncertain fields
    Hand-offApproved data created as draft records in your ERP

    Why review is built in

    Even very good extraction makes mistakes, and the costly ones are in totals, tax, and bank details. A review step routes exactly those cases to a person, while confident, rule-passing documents move on. Over time, measured accuracy decides how much can be approved automatically.

    What changes with your own documents

    Real document pipelines meet variety the sandbox doesn't: dozens of supplier layouts, scans at odd angles, multiple languages, and documents that turn out to be something else entirely. A pilot on your documents measures accuracy per field and per document type, so you can see where automation is safe and where review should stay.

    Most pilots start with one high-volume document type and a review screen for the team who checks the results today. The numbers from that pilot decide what happens next.

    What a pilot measures

    • Accuracy for each field and document type
    • How often documents need a person, and why
    • Time saved per document compared with manual entry
    • Which business rules catch errors the model misses

    Privacy in a real pipeline

    Production pipelines handle real invoices, contracts, and forms, so data protection is designed in: documents are processed in environments you control or with providers configured not to train on your data, access is limited to the people who review them, and every change is logged. The sandbox uses only synthetic examples, so nothing real is ever involved here.

    Frequently asked questions

    Is this the production system?
    It's a demonstration of the workflow. Production systems are configured and evaluated on your documents.
    Can I upload my own documents?
    Not here. The sandbox only runs invented samples, so nothing you own is ever sent. Real documents are evaluated in a private pilot.
    Can it read our document types?
    An evaluation on a sample of your real documents answers that.
    How would a pilot on our documents work?
    We agree the document types and fields, run a representative sample through a private pipeline, and measure accuracy per field with your reviewers before anything is automated.
    Why are the confidence values fixed?
    They are example numbers chosen to show how review routing works. Real confidence comes from the models and is measured on your documents.
    Which document types work best?
    High-volume documents with a fairly stable structure, such as supplier invoices, delivery notes, and order forms. Free-form letters are possible, but usually need more review.