Unstructured Data

Unstructured data is any information that does not fit the rigid row-and-column schema of a relational database. We are talking about PDF invoices, emails, contracts, photos, audio, video, social posts and WhatsApp messages: content rich in meaning that a traditional SQL query cannot read directly. According to IDC (2023), it accounts for 90% of the data organizations generate, and it is almost always the part left out of the dashboards.

The flip side is structured data (CRM records, journal entries, an inventory spreadsheet), which already comes in tabular form, ready to filter and sum. In between sits semi-structured data (a JSON file, an XML file, an email with headers): it has some tags, but no fixed schema.

The business challenge is not storing it but turning it into actionable information. In the documents and finance domain, reading a scanned invoice or a delivery note and loading its fields into a system is exactly what Arconte automates by combining OCR and AI, without anyone typing in the data field by field.

Why it matters so much in a company

For decades, business systems were built around what fit in a table: customers, products, amounts, dates. But most of a business's real knowledge lives outside those tables. The contract that defines a penalty clause, the email where the customer clarifies a payment term, the photo of the shelf that proves compliance with a trade agreement, the audio of a complaint to the call center: all of that is unstructured data. Ignoring it means making decisions with only a fraction of the information available.

The historical problem was technical: a computer could not "understand" a PDF or a recording as easily as it could add up a sales column. That is why, for years, that content stayed in folders, inboxes and shared drives, with no one able to query it at scale. The arrival of generative AI and LLMs changed that picture: today a model can read a text, extract the key fields, classify it and answer questions about its content.

How it becomes usable

The typical path for turning unstructured data into useful information combines several techniques:

• OCR to convert a scanned document or a photo of text into machine-readable characters.
• Natural language processing and language models to interpret the content, not just read it.
• Embeddings that translate text into numeric vectors, so you can search by meaning and not just by exact word.
• Vector databases that store those vectors for fast retrieval.
• RAG so an assistant answers using those documents as its source, instead of making things up.

When this pipeline works, a stack of PDFs stops being a dead archive and becomes searchable: you can ask "how much did we invoice this supplier last quarter?" and get the answer by reading the invoices themselves.

Structured vs unstructured

A concrete example in Argentina

Picture a wholesale distributor that receives hundreds of supplier invoices by email every month, each in a different format. The back-office team prints them, reviews them by hand and enters the data into the accounting system. It is slow, prone to typing errors and no one has real-time visibility of how much is owed. By applying OCR and AI to those PDFs, the relevant fields (tax ID or CUIT, number, date, amount, VAT rate) are extracted automatically and matched against the purchase order to catch discrepancies before paying. The same data that used to be trapped in an attachment now feeds reconciliation and the collections dashboards.

Common mistakes

A first mistake is underestimating the volume: many companies believe their "data" is only what sits in the ERP, when most of it is in emails and documents. The second is digitizing without extracting: scanning a paper to PDF does not make it structured; it is still an image until OCR and interpretation are applied. The third is blindly trusting AI without human validation on critical fields such as amounts or ID numbers, where a misread has a real cost. That is why serious workflows keep a human-in-the-loop step for low-confidence cases.

The good news is that the ceiling has moved: what for years was inaccessible noise is now a source of information as searchable as any table, as long as it is processed with the right tools.