# How to make your unstructured data AI-ready

Source: https://cplt.tech/blog/make-unstructured-data-ai-ready
Author: Stéphane Lepain
Published: 2026-11-18T21:52:54.739360+00:00
Tags: unstructured-data, metadata, small-business

---

To make unstructured data AI-ready, turn your documents, emails and scans into text an AI can read, label each file with its owner, version and who may see it, have a named person approve the sources for each task, and keep a record of what the AI used. Then test it with the questions your team really asks.

## What is unstructured data?

Unstructured data is information that does not sit in the rows and columns of a database: documents, emails, scans, slides, photos and chat messages.

In a small company, a lot of what people know lives in files like these. The price list is a PDF. The way you handle a complaint is in an email thread. The signed contract is a scan in a shared folder. These are the files an AI assistant has to work from.

## Why does AI give wrong answers from company documents?

Before blaming the model, look at the files. Three things go wrong.

**It cannot read the file.** A scanned contract with no text inside is invisible to an assistant that only searches text. The answer then comes from something else, and it sounds just as confident.

**It finds the wrong version.** With three copies of the same procedure in one folder, the assistant may quote last year's. I covered this in [how to train AI on your company documents](https://cplt.tech/blog/train-ai-on-company-documents).

**It knows nothing about who may see the file.** To an assistant, the salary sheet and the staff handbook are both just files. I wrote about this for one product in [can Copilot access all your company files?](https://cplt.tech/blog/can-copilot-access-company-files)

## How common is the problem?

[BARC](https://barc.com/research/harnessing-unstructured-data-for-ai-innovation/), a data and analytics analyst firm, surveyed 225 data, analytics and AI leaders in Europe and North America in February and March 2026. According to [BARC's summary](https://barc.com/news/unstructured-data-barc-study/):

- 29% fully know where their AI-relevant unstructured data resides.
- 70% say less than half of it can be found and used for AI.
- 79% are confident they can get value from it without breaking their governance controls.

The study is free thanks to two vendor sponsors, so keep that in mind. These respondents manage data for a living. I would not assume a ten-person company with no data team does better.

## How do you make unstructured data AI-ready?

I work in four steps, in this order. It is my own way of splitting the job, not an industry standard.

### 1. Read it: can the AI read every file?

Turn scans, photos and slides into text, page by page, so each answer can point to its page. Check the result: take ten documents, including your worst scan, and compare the text with the original. In the document system I built for my own work, a page that looks wrong goes to a person with the reason.

### 2. Label it: who owns the file, which version is it, who may see it?

This is metadata: information about a file rather than inside it. Four fields filled for every file beat twenty filled for some: owner, version, sensitivity and who may see it. A file called `final_v3` tells you very little.

### 3. Check it: has someone approved the sources for this task?

A named person decides which folders and versions the AI may use for a task, then reviews them when a document changes. My [acceptance checks for an internal knowledge assistant](https://cplt.tech/blog/internal-knowledge-assistant-acceptance-checks) show how to write this down.

### 4. Prove it: can you show what the AI used?

Keep a record of each request: who asked, which files the AI used and which version. When a customer or an auditor asks where an answer came from, you can show them. If those records must stay in house, so must the logs and backups, as I explain in [using AI on your documents without a third party](https://cplt.tech/blog/sovereign-ocr-document-ai).

## A short checklist to start

Start with one task. The rest of the drive can wait.

1. Write down the five questions your team asks most often.
2. Find the documents that answer them.
3. Archive the old copies so one current version remains.
4. Add an owner and a "last checked" date to each.
5. Mark the files only some people may see.
6. Ask the AI those five questions and check that each answer cites the current file.

If step 6 fails, the earlier steps tell you where to look. Want a second pair of eyes? Bring your five questions to a [free call](https://cplt.tech/contact) and I will tell you which step your documents need first.

## Frequently asked questions

### Can AI work with unstructured data?

Yes, once the files are in a form it can read. A typed document is usually fine. A scan or a photo of a page has to be turned into text first. Test with your worst scans and your oldest files before you rely on the answers.

### Are PDFs unstructured data?

Usually, yes. A PDF holds text laid out for people to read, not fields in a database. A PDF made from a scanner can be just an image of the page with no text inside, so check those first.

### What is metadata for AI?

Metadata is information about a file rather than inside it: who owns it, which version it is, when it was last checked and who may see it. An AI system can use it to pick the current version and leave out files a person should not see.
