Problem
PDF documents, invoices, receipts and payslips consumed administrative time and caused errors when renaming, classifying or extracting metadata.
Anonymised case
Document automation with Python and human review
Workflows to extract text, classify documents and generate structured indexes from administrative PDFs.
- Client type
- Administrative operations, anonymised
- Technologies
- PythonOCRYAMLCSVJSONLPDFLLM with review
Constraints
- Uneven quality of PDFs and OCR.
- Sensitive tax, banking and employment data.
- Need to preserve the original file and record exceptions.
- AI used only as support when fields are missing or confidence drops.
What was done
- Python tools for text extraction with OCR as a fallback.
- Rule-based classification, document routing and CSV/JSONL indexes.
- Extraction of metadata, amounts, dates, suppliers and centre codes.
- Dry-run options, text sidecars, auditing and controlled application.
Governance and security
- Local or controlled processing where privacy requires it.
- Dry-run before applying moves or renames.
- Preservation of the original document and traceability of rules.
What it demonstrates
- Practical automation of real administrative processes.
- Prudent use of AI on sensitive documents.
- Design of small tools that save work without imposing a full platform.
Result
Repetitive filing and extraction tasks moved to a workflow that is faster, reviewable and less dependent on manual memory.
Related services
Connected areas
Service
Fractional IT Director
A senior part-time IT function for organisations that are too small to maintain IT leadership of their own.
View service ServiceDocument and data automation
Processes PDFs, spreadsheets, surveys and repetitive exports with scripts, OCR, rules and human review.
View service ServiceIT governance, continuity and operations
Clarifies who controls the systems, where the data lives, how access is protected and what happens if something fails.
View serviceDoes this fit your situation?
Briefly describe the operational, data or AI problem. The first reply can help to guide the type of diagnostic.