Service · AI and ML engineering

Document intelligence

Turn invoices, forms, registers and reports into clean data you can use.

Relevant to Manufacturing, Service businesses, Education, Growing businesses

The problem

Your most important data arrives as paper and PDF: supplier invoices, lab reports, admission forms, delivery challans, handwritten registers. Someone types it in, slowly, with errors.

Reading documents is something current models do well. The engineering is in validating what they read, handling the exceptions, and putting the result where the business uses it.

What it costs today

Typical: a day or more of entry time per thousand documents, a 2 to 5 percent error rate on retyped fields, and history that cannot be analysed because it was never digitised.

This is for you if
  • Hundreds of documents a week, or an archive of thousands you want to use.
  • The documents follow a handful of layouts, even if the layouts vary by supplier or by year.
  • You know where the data should land: ERP, accounting, LMS, a database.
What we build
  • Extraction of the fields you need from scanned and digital documents, including handwriting where it is legible.
  • Validation against what you already know: supplier masters, purchase orders, rate lists, student records.
  • An exception queue for the few that do not pass, with the page and the field highlighted.
  • Digitisation of archives: years of registers or reports turned into tables you can analyse.
How we build itIn this order
  1. 01

    Collect a real sample

    A few hundred documents across suppliers, years and scan quality. Bad scans are part of the sample because they are part of the work.

  2. 02

    Label the truth

    We build a set of correct values with your team. It measures the system before launch and after every change.

  3. 03

    Extract, then validate

    Models read; rules check. A total that does not match its lines, a vendor not in the master or a date in the future goes to the queue, not the ERP.

  4. 04

    Route the exceptions

    The queue shows the document and the field. A person corrects it in seconds, and the correction becomes training data.

  5. 05

    Land the data where it works

    Into the ERP, accounting system or database you already have, with a record of which page each value came from.

Data you already haveThis is usually enough to start
  • The documents themselves, paper or digital
  • The masters they should match: suppliers, items, students, assets
  • Past manual entries, which become the labelled set
  • The system where the data should land
In practice

Illustrative example

Reading supplier test certificates in minutes, not by hand, for a pharma plant

Every incoming raw-material lot comes with a supplier certificate of analysis, as a PDF or a scan. QC staff read each one, check the values against the specification and type them into the lab system. It is slow and it is exactly the kind of work where tired people miss things.

Read the case study
Typical engagement
Duration2-week assessment, then 6 to 8 week pilot on one document type
TeamML engineer, data engineer, your process owner
You receiveExtraction live for one document type, measured field accuracy, exception queue in use
PriceFixed price, quoted after the readiness assessment
What you own at the end
  • Extraction pipeline, validators and queue, running on your server or account.
  • The labelled set and the accuracy report, per field.
  • Digitised archives as clean tables, with provenance back to the page.

Everything runs in your environment and is handed over with documentation. Nothing stops working if we part ways.

Questions we are usually asked

Our documents are handwritten.

Legible handwriting in a fixed form works reasonably well; free handwriting on a register is harder. The assessment measures it on your pages and gives you the number before you commit.

How accurate is it?

Reported per field, on your documents, against your labelled set. Totals and dates are usually very accurate; free-text descriptions less so. Anything below your threshold goes to the queue.

Does our data go to a cloud model?

Only with your written agreement. Where documents are sensitive, we run open models on your own server.

Can you digitise ten years of registers?

Yes. Archive digitisation is scoped by page count and quality, and it often pays for itself through the analysis it makes possible.

Does this replace our data-entry team?

It replaces the typing. The team moves to the exception queue and to the checks nobody had time for. Volumes usually grow into the capacity freed.

Often combined with

Further reading

Not sure if your data is ready?

Take the free 20-minute readiness check. You get a short scorecard and a plain answer about what, if anything, is worth doing next.

Start the readiness checkTalk to us