Different formats from dozens of suppliers
Different fields, different names, different table structures.
Case study · Favex
From manual certificate checks to a measurable workflow
A steel-trading company used to retype hundreds of certificates into the ERP by hand. Today the system extracts the data, compares it against the standard, flags risks and the operator mainly handles exceptions.
The results in numbers
From hours a day to tens of seconds
The operator stopped retyping and started validating. AI extracts structured data, automation compares it against the standard and the original PDF and flags risks.
The workflow handles different suppliers, multiple certificates in one PDF and multilingual documents. Behind every extraction is an audit trail: what was in the source, what was in the standard, what ended up in the output and why.
4+ h
of manual work saved daily
30–45 s
average certificate processing time
95 %+
accuracy on key fields
16+
suppliers with automatic detection
Text PDFs · 95+ score · 0 HIGH warnings → candidate for auto-approve.
Client & starting point
Hundreds of certificates a month, each one different
Favex is a trading company in the steel segment. It receives hundreds of material certificates each month from dozens of different suppliers - each with a different format, structure and language. The operator retyped every certificate into the ERP by hand. Each day this cost 4+ hours of routine work.
Multilingual documentation
CZ, DE, EN, IT and their combinations in one PDF.
Multiple certificates in one PDF
The operator had to split the document into individual certificates by hand.
Repeated errors from routine
Hundreds of entries a day - mistakes from lapses in attention.
Dependence on a few people
The workflow rested on a handful of experienced operators.
No audit trail
For a given decision there was no way to trace back what had been in the source.
Application
The workflow as the operator sees it
The PDF next to the form, pre-filled fields, mismatches highlighted in colour. For certificates without issues, a single confirmation is enough.
Solution
Four phases of the workflow
We designed an end-to-end workflow in which the operator stopped retyping and started validating. Each phase has a clear role: AI extraction, rules for validation, a human at the decision, integration to hand over the data.
The PDF arrives, the system splits it and identifies the supplier
Web upload, automatic intake from email (IMAP), batch import of a folder. Multi-certificate PDFs are split automatically into individual certificates. The system detects the supplier and loads the right rule template.
Structured data from any layout
Header fields, chemical analysis, mechanical properties, heat treatment, technical tests and free-text notes in tables. CZ, DE, EN, IT - one flow, no per-language setup.
Limits, exceptions, QA scoring 0–100
Every value is compared against the limit values of the standard by steel grade and thickness. Conditional rules and exceptions are applied automatically. Each certificate gets a score of 0–100 and warnings at three levels of severity.
One-step confirmation, structured export
The operator sees the PDF, pre-filled data and mismatches highlighted in colour. For certificates without issues, a single confirmation is enough. The data goes straight into the ERP import format - no intermediate steps, no manual copying.
01 · Extraction and detection
Structured data from any format
AI handles different layouts, multi-certificate PDFs and multilingual documents. It keeps its own rules for each supplier - a new one is added as a data template, not a programming task.
Extraction of all certificate sections
Header fields, dimensions, weights, heat numbers, chemical analysis, mechanical properties, heat treatment, technical tests (Charpy, ultrasonic, hardness), markings and free-text notes in tables.
Detection of 16+ suppliers
Each supplier has a different format and different field names. The system handles it on its own. A new supplier is added as a data template, not as a programming task.
Multi-certificate PDFs & multilingualism
PDFs with multiple certificates are split automatically. Extraction handles CZ, DE, EN, IT and combinations - one standardised flow, no per-language setup.
02 · Control and validation
The standard as part of the workflow
The system has the technical standards loaded by steel grade and material thickness. Every chemical value and mechanical property is compared automatically against the limit values - the operator immediately sees what passes and what does not.
Validation against the standard
Every value is automatically compared against the limit values of the standard. The operator instantly sees what passes and what does not - with no manual searching through technical literature.
Exceptions from the notes
Standards contain conditional rules - element ratios, exceptions for certain grades, limits by thickness. The system applies these exceptions automatically and for each one states which note permitted it.
QA scoring 0–100
Each certificate gets a score and a list of warnings at three levels: HIGH (physically impossible), MEDIUM (suspicious), LOW (missing). Certificates above 95+ with no HIGH/MEDIUM are candidates for auto-approve.
03 · Operations and integration
The operator reviews instead of retyping
Next to the form they see the PDF, so they can verify every value against the source. The data is pre-filled, mismatches and exceptions highlighted in colour. For certificates without issues, a single confirmation is enough.
Intake from all channels
Web upload (drag & drop), automatic intake from email (IMAP with a subject filter), batch import of a whole folder, a built-in PDF splitter for manual splitting.
The operator only reviews
Pre-filled data, colour highlighting of mismatches, the PDF next to the form for verification against the source. Experienced colleagues step in only where it really needs attention.
Export to the ERP and API
Confirmed data goes straight into the ERP import format - no intermediate steps. A REST API is available for other systems. Connecting to reporting is a matter of configuration.
04 · What keeps quality over time
A workflow that improves in operation
The system runs efficiently without operator intervention and over time refines its extraction based on real corrections. A dashboard shows where it works reliably and where there is room for improvement.
Performance by complexity
Simple text PDFs go through a faster path, complex scans and multi-certificate documents through a more powerful one. The decision is automatic - operations run efficiently without operator intervention.
The system learns from corrections
Every correction from the operator is recorded. If a field is corrected repeatedly, the system remembers it and extracts it more accurately next time. Quality doesn’t drop over time - it improves.
Transparent operation
A dashboard shows processing status, extraction accuracy and the most frequently corrected fields. It’s clearly visible where the system works reliably and where there is room for improvement.
The same principle across processes
Certificates were the first live use case. We’re currently working with Favex on incoming invoice data extraction - a different domain, but the same principle: a document arrives in different formats from different suppliers, has to be extracted, validated and handed over to the ERP.
Process analysis
We start with how the work actually happens - not with whatever model we happen to have on hand.
Workflow design
What AI does, what a rule does, what an integration does, what a human does - and where the audit trail belongs.
Validation on real data
We assess accuracy, error rate, exceptions and impact - not the impression from a demo.
Production operation
We build for long-term maintenance, quality monitoring and expansion to further processes.