Intelligent document processing
Structured data from PDFsand scans.
Intelligent document processing (IDP) is the engine that reads contracts, reports, forms or tender specifications, extracts the fields you need, classifies the documents and summarises long texts, in batches, with structured output you can review and import.
The data is in documents, not in systems
Important information — deadlines, amounts, parties, requirements — is locked in PDFs, scans and Word files. To get it into a spreadsheet or an ERP, someone reads and copies it, document by document.
- Slow manual transcription, where typing errors are hard to spot
- Different formats for the same type of document, from different suppliers or customers
- Entire archives left unclassified, impossible to analyse without reading everything
- Classic OCR tools read the text but do not understand which field it represents
Who it is for
- Teams that manually transcribe data from documents into spreadsheets or systems
- Companies with large archives that need to be classified, inventoried or migrated
- Technical teams that need an extraction component for their own application or workflow
- Departments that regularly analyse batches of documents: tenders, contracts, reports
What the document processing engine does
- 01
Field extraction
Captures values such as parties, dates, amounts, deadlines or requirements into a schema you define.
- 02
Classification
Sorts documents by type, category or risk level, according to agreed rules.
- 03
Summaries with references
Produces summaries of long documents, with the source page shown for each point.
- 04
Batch processing
Runs on large batches of documents, with a status report for each file.
- 05
Structured output
Delivers JSON, CSV or tables in a validated format, ready to import.
- 06
Confidence score
Flags uncertain or missing fields, so a person only checks what needs checking.
How it works
How a batch of documents goes through the engine
The steps marked in yellow are approved by a person.
- Input
- Action
- AI
- Human in the loop
- Result
- 01
Input
Batch of documents
Documents are uploaded or picked up from a folder or a system.
- 02
Action
Reading and OCR
The text and structure are read, including from scans.
- 03
AI
Extraction and classification
The model fills in the field schema, classifies the document and summarises it.
- 04
Action
Format validation
Fields are checked against rules: type, range, consistency between values.
- 05
Human in the loop
Exception review
A person checks only the fields flagged as uncertain or invalid.
- 06
Result
Structured data
The validated result is exported to a spreadsheet, a file or the target system.
What information extraction needs
- Representative examples of each document type, including difficult variants
- The schema of fields to extract, with the format and rules for each
- The classification categories and the criteria for assigning them
- A set of manually annotated documents, used as an evaluation benchmark
- Where the data goes: the export format or the system it is imported into
Possible integrations
- PDF, Word and scanned files
- Google Drive
- Microsoft SharePoint
- SFTP
- Amazon S3
- Odoo
- SAP Business One
- Microsoft Excel
- SQL databases
- PDF, Word and scanned files
- Google Drive
- Microsoft SharePoint
- SFTP
- Amazon S3
- Odoo
- SAP Business One
- Microsoft Excel
- SQL databases
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What people validate
- Fields flagged as uncertain, before import
- The field schema and classification categories, when defined and at every change
- A sample from each batch, including fields considered reliable
- Any decision made on the basis of the extracted data: payments, approvals, tender responses
Limitations and how we handle them
- Poor-quality scans, handwriting and complex tables reduce extraction quality
- The confidence score guides the review but does not replace it, which is why we also spot-check unflagged fields
- A new document type or a changed format may require adjusting the schema and retesting
- Summaries inevitably leave out details; for decisions, the original document is what counts
- The engine extracts and classifies; it does not run an end-to-end workflow with routing and approvals on its own
How we evaluate extraction quality
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Field-level accuracy
- For each field, the extracted values are compared with the set manually annotated by your team.
- Classification quality
- Documents wrongly classified in the test set, analysed by category.
- Share of exceptions
- How much of each batch goes to human review and for what reasons — from the run report.
- Confidence calibration
- Whether errors occur mainly in fields flagged as uncertain — checked by spot checks.
- Time and cost per batch
- Processing time and cost per document, tracked on every run.
How we implement data extraction
- 01
Sample and schema
We analyse real documents and define the fields, categories and output format together.
- 02
Annotated set
Your team manually marks the correct values on a sample, which becomes the testing benchmark.
- 03
Build and calibration
We configure the extraction, validations and confidence thresholds and test them on the annotated set.
- 04
Running in production
We start with small batches, check the exceptions and increase the volume gradually.
Use cases
These scenarios are conceptual examples showing what can be built. They are not delivered implementations or results achieved for clients.
- Construction company
Requirements from tender specifications
Tender documents are processed in batches, and the requirements, deadlines and criteria go into a spreadsheet that the team reviews.
Conceptual example - Leasing company
Inventory of existing contracts
A contract archive is classified by type, and the key data — parties, term, value — is extracted for migration to a new system.
Conceptual example - Energy audit firm
Summaries of technical reports
Long reports are summarised section by section, with the source page, so auditors can go straight to the relevant parts.
Conceptual example
Frequently asked questions
How is it different from the AI document processing agent?
The agent handles an entire workflow: it receives the document, extracts the data, checks it, passes it on and tracks approval. Intelligent processing is the extraction, classification and summarisation engine, used in batches or integrated into your application. An agent can use this engine.
What types of documents can be processed?
PDFs, Word documents, scans and images containing text. Quality depends on legibility and on how varied the formats are; we test it on your examples before any commitment.
What happens when the engine is not sure about a value?
The field is flagged as uncertain or left blank, with the reason, and goes to human review. Fields are not filled in with guessed values.
Where do the documents go and how are they protected?
Documents are processed in the environment agreed together: with model providers using settings that do not use the data for training, under their terms, or on privately hosted models for sensitive documents. Personal data and the retention period are handled in line with the GDPR.
Can we integrate the engine into our application?
Yes. It can be exposed as an API service that receives documents and returns structured data, or it can run on scheduled batches.
How is it different from classic OCR?
Classic OCR turns an image into text. Intelligent processing goes further: it understands what each value represents (supplier, amount, deadline), puts it in the right field of a schema and flags uncertain values for review.
Can’t find the answer? Write to us and we’ll reply directly.
Want the data from your documents straight in a spreadsheet?
Send us a few typical documents and the list of fields you need. We will show you what the structured output looks like and how we check its quality.
Related solutions
- AI agent for document processingTakes in incoming documents, extracts the data, checks it against your rules and systems and sends it on, with a person handling the exceptions.
- AI API integrationConnecting AI providers' APIs to your systems, with keys kept on the server, limits, logs, cost control and a fallback provider.
- AI evaluation and monitoringTest sets, checks at every change, human spot-check reviews and cost tracking, so you know how your AI performs over time.