Headers

Case study

A day of retyping replaced by a check

Figures from supplier PDFs were retyped into spreadsheets by hand, taking one person a full working day. The software now pulls them out, flags what it is unsure about, and a person confirms the result before it is used.

Industry
Chemical distribution
Team size
under fifteen people
Regulatory regimes
industrial, pharmaceutical, food and cosmetics
Document languages
EN, CN, PL
What we delivered
a web application for data extraction and human review
Where the data sits
In the EU, with a provider in Germany

A full day of retyping, a quote in days

The client works in three regulatory regimes at once: industrial chemicals, pharmaceutical ingredients, and components for food and cosmetics. The team is under fifteen people.

To issue a quote they had to retype figures from supplier certificates of analysis (COA) and specifications (SPEC) into their own spreadsheet templates. Every manufacturer uses a different layout and a different language: English, Chinese, Polish.

One person spent a full working day on it. A quote for thirty items stretched to several days. Manual retyping also produced slips in figures that could reach the customer's quote.

The model pulls the data, a person confirms it

We built a web application that pulls the required fields out of a PDF with a language model. It handles different layouts and languages without a separate template per supplier.

The application works out whether it is a COA or a SPEC, pulls the data into one schema and gives every field a confidence score. A person then checks and confirms. The decision stays with them, not with the model.

  • Extraction by a language model

    The model reads the document and writes it into one schema. It handles English, Chinese and Polish without a template per supplier.

  • Document type detection

    The application decides whether it is a certificate of analysis or a specification and processes the data accordingly.

  • A confidence score per field

    Every extracted field carries a confidence score. Low-confidence fields are highlighted on the review screen, so a person knows where to look.

  • Human review

    The review screen shows the PDF beside the extracted data. A person corrects the unclear fields, approves the result and the system records the changes.

  • Export into their own template

    Approved data exports straight into the client's spreadsheet template. No further retyping.

  • Concurrent processing

    The application handles several documents at once. On failure it retries or falls back.

What the acceptance tests showed

100 %
document type detected correctly in 860 documents
over 97 %
mandatory fields extracted correctly
~40 s
p95 per document

Specification table rows came out around 97 percent at confidence 0.8 and above. The figures describe acceptance tests on a sample of the client's documents, not long-run operation. The application replaced a full-time retyping job and cut multi-day quote preparation to minutes.

All work

Is there a process where somebody retypes data all day?

Describe it to us. We will show you whether software can take it over and where a person should stay.

Talk about your problem