Quotes, invoices, unit price schedules (BPU), priced bills of quantities (DPGF), framework agreements, price lists. Pulling a unit price, a discount, a quantity or a due date out of a document that was never designed to be read by a machine. This step is the silent foundation of procurement, finance and accounting control. When it is rushed, everything that follows (matching, anomaly detection, the decision itself) becomes uncertain.
Most organisations invest in sophisticated analysis and reporting tools, yet neglect the link that feeds them. A badly extracted figure does not produce a degraded analysis: it produces a wrong one, with exactly the same appearance of reliability as a correct one. That is precisely what makes data extraction both a technical and a strategic matter.
Data extraction: what are we really talking about?
Data extraction means turning an unstructured document (a PDF, a spreadsheet, a scan, a summary e-mail) into structured, usable and comparable information. Concretely, it means identifying the decisive elements in a document (nature of the service, quantities, unit prices, discounts, payment terms, due dates, contract references) and returning them in a logical form, ready to be cross-checked against other sources.
Extraction has to be distinguished from its neighbouring steps. Capture means retrieving the document. Classification means identifying its type (is it a quote, an invoice, a unit price schedule?). Matching means comparing the extracted data with the reference data. Extraction is the hinge between the raw document and intelligence: it decides the quality of everything that comes next. A perfect match on badly extracted data is still a wrong match.
Why extraction is a harder problem than it looks
The first intuition is that reading a document is trivial. A human does it in a few seconds. That is exactly what hides the real difficulty of the problem at scale.
Heterogeneous formats
No legal standard dictates the structure of a quote or a supplier invoice. Every issuer applies its own layout, its own wording, its own item codes. The same product can appear under three different names from three suppliers. A quote can be a PDF generated by an ERP, an Excel table laid out by hand or a plain e-mail. Before two documents can be compared, they have to be translated into a common language, something the human eye does without thinking but that a system has to learn to reproduce reliably and repeatably.
Density and variability of documents
A unit price schedule (BPU) can hold several hundred lines, each with its code, unit of measure and price. A priced bill of quantities (DPGF) follows its own logic in every contract. Extracting four hundred service lines cleanly, without missing a single one and without confusing a quantity with an amount, quickly exceeds what careful manual processing can do, especially under time pressure.
Unstructured data
Most of the information that travels between a supplier and a procurement department is not stored in neat boxes. It is buried in free text, borderless tables, annotations and special conditions written in natural language. A discount can sit in a clause rather than in a column. Ancillary charges can appear under a label that matches nothing in the original quote. Extracting is therefore not just copying numbers: the context has to be understood to know which number counts.
Not all documents are equal: the scale of complexity
Understanding why extraction resists automation means accepting that documents are not all at the same level of difficulty. The more variable a document is in its form and the richer in its content, the more demanding extraction becomes. This reality can be pictured as a scale of increasing complexity. The wider context does not work in companies’ favour: according to IDC, close to 90% of enterprise data is unstructured, an order of magnitude confirmed by other players in the field such as Databricks, which put it between 80% and 90%.
At the first level sit simple, structured documents, such as a standardised purchase order or a normalised spreadsheet. Their form is predictable and their extraction relatively straightforward. At the second level, semi-structured documents (invoices, forms) offer a few structural cues, but every issuer has its own presentation. At the third level, unstructured documents (quotes in a free-form format, price lists, financial statements) mix text and figures with no common logic from one supplier to the next. At the top, highly complex documents (framework agreements, BPU, DPGF, technical specifications) combine hundreds of lines, clauses in natural language and indexed conditions.
The decisive point for procurement, finance and accounting functions is this: most of the documents they handle every day are concentrated at the top of this scale. The easy cases do not dominate; the complex ones do. That is precisely why an extraction designed for simple documents collapses on real ones.
From OCR to agentic AI: the long trajectory of document extraction
Automated data extraction was not born with recent artificial intelligence. It follows a long trajectory, each stage of which has pushed back the rigidity of the template a little further.
As early as the 1950s, the first optical readers (OCR) converted typewritten documents into machine data. The 1970s brought omni-font OCR, able to recognise characters whatever the typeface. From the 1980s, computer vision algorithms made it possible to handle more varied layouts. In the 2010s, a first generation of document processing tools emerged, layering friendly interfaces over OCR and starting to draw on machine learning and natural language processing (NLP) to gain flexibility.
The thread running through this history is clear: every generation has tried to break free from the constraint of the predefined model, the fragility that makes a simple change of layout break the extraction. The current break, driven by the combination of NLP, vision models (VLM) and agentic AI, extends that trajectory by crossing a threshold: reading a document without a template, by understanding its content rather than tracing it.
The classic approaches and their glass ceiling
Faced with this challenge, organisations have historically relied on three remedies, each hitting the same limit.
Manual entry
It is still the norm in many departments. It is flexible but costly and fallible. Between 40% and 70% of a buyer’s time can be absorbed by document handling, a significant share of it re-keying. Every entry is an opportunity to introduce an inaccuracy (a miscoded reference, a rounded discount, a misplaced decimal) that will then contaminate every downstream analysis.
Classic OCR
Optical character recognition digitises a document and extracts text from it. It is a useful first step, but it stops at the surface. OCR built on rigid templates works as long as the document matches the expected template exactly. The slightest change of layout breaks the extraction. Above all, OCR reads without understanding: it does not know that the price per kilo it has just read must be checked against a price list negotiated six months earlier, nor that a different label refers to the same service. It moves the problem into a spreadsheet; it does not solve it.
The spreadsheet
Excel remains the universal refuge. It compares cells, not contexts. It does not detect that a “handling included” line corresponds to “free shipping” on another document, and a simple formula error propagates silently through every calculation that depends on it. At low volume, it helps. As soon as the number of documents and the complexity of the conditions grow, it reaches its structural limits.
What these three approaches have in common: they treat extraction as a mechanical operation, when it is an operation of interpretation.
The business stakes: what reliable extraction really delivers
Focusing attention on extraction is not a technical refinement. It is where a concrete share of the margin is decided.
Time given back to the teams
Automating extraction frees the hours currently spent reading, re-keying and normalising documents. That time is not strategic time: it is mechanical verification time paid at an expert’s rate. By recovering it, procurement and finance teams can redirect it to negotiation, supplier relationships and analysis, in other words to what really creates value.
Reliable decisions
Reliable extraction produces a comparable database. Without it, structuring purchasing decisions (renewing a contract, consolidating volumes, renegotiating a price grid) are taken on partial data or data reconstructed by hand. With it, every decision rests on what the documents actually contain, not on what they are assumed to contain.
Downstream detection
Extraction mechanically conditions the quality of discrepancy detection. A discount that was not extracted is a discount that will never be checked. A misread line of charges is a discrepancy that will slip under the radar. Error margins sit between 3% and 5%. It is precisely because extraction is made reliable upstream that the detection of document discrepancies can reach high accuracy downstream. Today, one supplier document in two still escapes control, and failed extraction is often the first cause.
Conclusion
Data extraction is the most discreet building block of procurement and finance control, and yet the one that decides everything else. It does not show up in dashboards, it is not the subject of dedicated strategic projects, and that is exactly why it is a costly blind spot. As long as data enters the system through fallible manual entry or blind scanning, no downstream analysis tool can make up for that flaw in the foundation.
Making extraction reliable means refusing to build decisions on uncertain data. It is also the first condition for recovering time, securing margin and giving teams the visibility they have never had over what their documents really contain.
Discover Zylio on your own documents.
