Applied AI May 2026 · 5 min read

70% of enterprise data is unstructured. Classical automation can't touch it.

TL;DR

Up to 70% of enterprise data is unstructured – emails, PDFs, voice recordings, handwritten notes – and classical automation can't touch any of it.
This isn't a technical problem. It's a business problem: decisions get made on a fraction of the available information.
AI that reads free text turns invisible data into usable input – without changing the processes that already work.

Most automation running in companies today does one thing well: it reads fields from forms. A structured invoice with the right number in the right place – the system processes it. Fast, reliably, cheaply.

The problem is that this describes a minority of enterprise data.

Where business knowledge actually lives

An email from an enforcement agent with non-standard phrasing. An accountant's note on a receivable written three years ago. A workshop transcript capturing requirements for a new ERP system. A medication leaflet with dosage instructions broken out across six age groups.

This is unstructured data. Text where the important information exists – just not in the field the system is looking for.

According to long-standing estimates from IDC and Gartner, unstructured content accounts for 70–90% of all new enterprise data. The remainder – structured records, databases, tables – is the smaller share.

Classical automation, RPA, rules-based systems: designed for that smaller share. For the 70%, they have no answer.

What this means in practice

This isn't an abstract data architecture problem. It's something nearly every back-office team runs into every day.

At Dôvera, one of Slovakia's three health insurers, staff were manually processing hundreds of documents from enforcement agents. Each document phrased differently, each agent with a slightly different style. The system couldn't read them. People could – but at the cost of hours of work per day.

That's not an exception. It's the rule in any sector where documents arrive from regulators, courts, partners, or customers. Each slightly different. No single format. And behind each one sits a decision someone has to make.

Why “put it in a table” doesn't work

The instinctive response is to standardize inputs. Introduce forms, templates, mandatory fields. Force people and partners to send data in one format.

That works where you control the sender. With your own employees. On internal processes.

It doesn't work where the data sources are outside your control – enforcement agents, courts, suppliers, customers. Or where the value lies precisely in the nuance of the text: the comment on a receivable, the reasoning behind a decision, the tone of a customer review.

Structuring the input doesn't eliminate the problem. It just moves where the pain shows up.

What AI does differently

Language models don't wait for structure. They read free text – and pull out what you need.

Not universally. Not for free. But if the task is clear – “from every document, I need to know X, Y, and Z” – a solution can be built that handles it reliably, at scale.

At Dôvera, we deployed a pipeline that reads incoming enforcement documents, extracts the relevant data, and writes it into internal systems. The result: 60% back-office time saved, 100% of correctly formatted documents processed automatically.

For Vestberry, we automated KPI extraction from portfolio reports – each company sends them in a different format, in a different language. Over 100 KPIs tracked across any format, any language. Hours of manual work reduced to minutes.

The logic is the same in both cases: free text has structure – just not a visible one. AI reads it.

Where the limits are

Not every document makes a good input for AI. If the text is too short, too inconsistent, or if the value lies in something other than the text content itself – a solution may not make sense.

The goal isn't to deploy an LLM on everything. It's to identify the processes where:

The input is text or documents whose format you don't control

The output is a repeating decision or extraction of specific data

The volume is large enough that manual work is becoming a real constraint

Where that's true, classical automation covers part of the problem. AI that reads unstructured text covers the rest.

What to do about it

The right first step isn't technology. It's a process map – where in your organization does unstructured data enter? Where does it cause manual transcription, decisions, escalations?

Those are the candidates. Not every one of them warrants a solution – but without the map, you can't see them.

At INSYNAPS, we start every engagement with an AI Workshop, and that map is exactly what it produces. Not a demo, not a pitch. A clear-eyed look at your processes and an honest answer to where AI will help – and where it won't.

If your team spends hours on documents the system can't read, it's worth a conversation.

Applied AI

Frontier or open-weight? Companies don’t choose a model. They choose the level of risk.

August 2026 · 15 min

Applied AI

From Pilot to Production: What Real AI Implementation Looks Like

May 2026 · 6 min

AI Implementation

AI won't fix a broken process. The problem runs deeper than that.

May 2026 · 5 min

A similar challenge in your organisation?

Tell us about your workflow. We'll tell you whether AI fits – honestly, even if the answer is no.