Data Readiness - How to Prepare Your Company's Data for AI Before You Spend a Dollar on Implementation

A manufacturing company in southern Poland signed a contract in 2025 for an internal AI assistant meant to answer employee questions about procedures, product specs, and historical design decisions - a classic RAG system built on company knowledge. The model worked flawlessly on a technical level. The problem showed up a month after launch: the assistant regularly cited outdated safety procedures from three years earlier, confused the specs of two similar products because both sets of documentation sat in the same folder with no distinguishing detail in the file name, and - worst of all - leaked HR folder data to someone outside the HR department in one response, because the permissions from the network drive had never been carried over into the index the model searched. The language model wasn't at fault. The data nobody prepared before anyone wrote a single line of integration code was.

Gartner projects that by the end of 2026, organizations will abandon 60% of AI projects specifically because they lacked AI-ready data - not because the model was bad. MIT NANDA's July 2025 report went further: with $30-40 billion invested by companies in generative AI, 95% of pilots delivered no measurable return. The common denominator in both cases isn't technological - it's the data mess nobody cleaned up before signing the vendor contract. Here's how to audit your data sources, assess their quality, and clean them up BEFORE the rollout, not while putting out fires afterward.

This post is the missing first link in a chain I've been building from the middle outward: where to start with AI implementation assumes you already know what you want to build; building an internal RAG knowledge base and advanced RAG - chunking and hybrid search both assume the data you're loading into the system is already in decent shape. In practice, that assumption fails most often - and it fails at stage zero, before anyone touches the model.

Why rollouts fail on data, not on models

This isn't a marginal issue - it's the leading cause of failure, not one of several. Gartner's February 2025 forecast is explicit: by the end of 2026, organizations will abandon 60% of AI projects that lacked AI-ready data. This isn't a distant-future problem - it's already happening. MIT NANDA's "The GenAI Divide: State of AI in Business 2025" report, released in July 2025 and based on a review of more than 300 publicly disclosed AI deployments plus interviews with 52 organizations, went further: with $30-40 billion invested by companies in generative AI, 95% of pilots delivered no measurable impact on the bottom line. Only 5% of rollouts extracted real value.

An interesting, practical finding from the same report: buying a ready-made tool from a specialized vendor and building a partnership around it succeeds roughly 67% of the time, while building an internal solution from scratch succeeds only about a third as often. That doesn't mean building your own systems is a mistake - it means internal teams systematically underestimate how much work data preparation takes before a model performs correctly, while external vendors more often force that step as a condition of the rollout.

Auditing your sources - where your company's data actually lives

Before you organize anything, you need a map. Most companies I talk to can't clearly answer "where exactly does our data live" - and without that map, there's no way to assess what's ready and what needs work.

I walk through the full process of automatically scanning and categorizing documents from one of these sources in practice in AI integration with Google Drive - a good starting point for the audit itself, before you decide what to do with the data next.

A data quality matrix - how to assess what's ready and what isn't

Once your sources are mapped, you need a way to score each one against the same criteria, instead of deciding by gut feel:

1. 1.Freshness. When was the document last verified (not: when was it created) - a crucial distinction, since a file can carry a year-old modification date and still be accurate, or have been edited yesterday and already be substantively wrong.
2. 2.Uniqueness. Does more than one version of the same information exist, and is one of them marked as "the authoritative one" - conflicting sources are the fastest path to hallucination in a RAG-generated answer.
3. 3.Ownership. Who is substantively responsible for this document - if nobody can answer that in 30 seconds, that's a signal the document is probably abandoned.
4. 4.Permissions. Who currently has access to this document in the source system - and can that restriction be carried over into the AI layer, or does it need to be rebuilt by hand.
5. 5.Format and structure. Can the document even be meaningfully split into chunks - a scan with no text layer, a table pasted in as an image, or a PDF with broken column layout is a different level of difficulty than clean text.

A document that fails criterion 1 or 2 shouldn't enter the system in the first place - no RAG technique, however advanced, will fix conflicting or stale source data. This is literally the "garbage in, garbage out" principle transplanted onto semantic search: the model doesn't judge truthfulness, only similarity - if the index holds two conflicting versions of a fact, both look equally credible to it.

Preparing data for RAG - formats, versioning, owners, permissions

Once you know which sources are ready and which need work, cleanup splits into four parallel tracks:

• Format standardization. Turn scans into searchable text (OCR), extract tables out of screenshots, standardize heading structure - the foundation without which advanced chunking and hybrid search techniques have nothing to work with.
• Versioning and marking the "source of truth." Every document with multiple versions needs a clear marker for which one is current - ideally enforced structurally (one location, the rest archived), not just described in a note.
• Assigning owners. Every document or document category needs a person responsible for keeping it current - without that, data cleaned up today will be a mess again in six months.
• Permissions as a technical layer, not a policy statement. This is the most commonly skipped step, and the riskiest one to skip: if a document isn't visible to a given user in SharePoint, Confluence, or the source system, it must be equally invisible to the AI search mechanism - even if it's a perfect semantic match for the query. Access control has to be enforced BEFORE a result is returned, not after. Your choice of vector database affects how easy this is to implement - I break down the options for RAG and AI memory in vector databases: pgvector, Pinecone, Qdrant, or Weaviate.

Structured vs. unstructured data - different prep paths

These two categories of data require completely different work, so treating them with the same cleanup process is a common mistake:

Structured data is mostly solved at the integration and deduplication level - classic engineering work. Unstructured data additionally requires editorial judgment calls: who decides a document is current and worth keeping. That second task is harder to automate and more often skipped - which is why it accounts for most of the failures in RAG-based rollouts.

Quick wins before the rollout - what to do in week one, before paying anyone

You don't need to wait for a full audit to start. These steps cost your team's time, not tooling budget, and meaningfully reduce the risk of the rollout failing:

1. 1.Pick one narrow area for the pilot - one category of documents (say, one department's procedures), not "all company knowledge" at once. A narrower scope means fewer sources to check and faster feedback on whether the approach works.
2. 2.List the documents you're certain are outdated, and remove or archive them - the cheapest action in the whole process, and it eliminates the largest source of conflicting answers.
3. 3.Assign an owner for every document category in the pilot's scope - one person, one name, not "department X" as an abstract owner.
4. 4.Check whether permissions in the source system can even be read programmatically - if the AI vendor you're talking to can't explain how they plan to respect them, that's a red flag before you sign anything.
5. 5.Count how many documents actually fall within the pilot's scope - if the number runs into the hundreds of pages on the first pass, the scope is too wide to start with.

Security and permissions are part of data readiness, not a separate topic

Data readiness and data security are, in practice, the same project seen from two angles - cleaning up data for AI without simultaneously thinking through who should have access to it after the rollout produces a system that's technically ready and legally risky at the same time. The scenario at the start of this post - an HR data leak from improperly carried-over permissions - isn't an outlier; it's the most common mistake in first-time RAG rollouts on internal documents. I go deeper into not handing over your company's secrets during an AI rollout in AI data security - worth reading alongside your data cleanup, not after it.

Data readiness checklist - before you sign a vendor contract

1. 1.You have a map of your data sources - you know exactly where every category of data you want to use physically lives.
2. 2.You've scored every source against the same criteria - freshness, uniqueness, ownership, permissions, format.
3. 3.You've removed or archived documents you're certain are outdated - the cheapest step with the biggest impact on answer quality.
4. 4.Every document category has an assigned owner - a specific person, not a department.
5. 5.You know how permissions from the source system will translate into the AI layer - and you have that confirmed by the vendor, not just promised.
6. 6.The pilot's scope is narrow and quantified - one document category, not all company knowledge at once.
7. 7.You have a plan to keep things organized after the rollout - data owners, a review cadence, not a one-time cleanup.

---

I help companies audit their data and prepare it for an AI rollout before they pay for an integration that won't work on top of a mess - a source map, a quality assessment, a cleanup plan, and a permissions architecture, all in one engagement. I do this through AI consulting and AI automation engagements. Get in touch - I'll start with a quick audit of your main data sources and point out where the biggest risk sits.

Worth reading next:

• Where to Start with AI Implementation in Your Company
• How to Build a Company Internal Knowledge Base with AI (RAG in Practice)
• Advanced RAG - Chunking Strategy, Hybrid Search and Reranking
• AI Integration with Google Drive - Automatic Document Scanner
• AI Data Security - How Not to Hand Over Your Company's Secrets