What "AI-Ready" Data Actually Means

The Gap Between Confidence and Reality

A Komprise survey of 500 technical decision-makers, released in October 2026, put a number on something many organisations have quietly suspected for a while. Between 93% and 96% of respondents said most of their data is not ready for AI use. Only 7% reported that 70% or more of their data had been cleaned and classified. We think that second figure is the more useful one. It is not a measure of confidence or scepticism about AI. It is a direct measurement of how much structured, governed data most organisations actually hold before they start layering AI systems on top of it.

Most organisations are not short of AI ambition. They are short of the groundwork that makes AI outputs trustworthy once deployed. Komprise frames this as a gap between perceived AI maturity and actual AI results, and that framing matters because it points at the right failure mode. A model given disorganised, unclassified, badly permissioned data will still produce answers. The answers are just less likely to be correct, and far harder to audit once something goes wrong. Reporting coverage of the same survey made a similar point: the shortfall is not a lack of enthusiasm for AI, it is a lack of the unglamorous data work that AI quietly depends on.

What "AI-Ready" Actually Means

"Clean" data is the easiest bar to clear, and the one most reporting tools measure for. It covers deduplication, consistent formatting, and removing obvious junk files. It says nothing about what the data actually is. A file share holding customer contracts, HR records and old product specs can be perfectly clean and still be unusable for AI, because nobody has worked out which files contain personal data, which are commercially sensitive, which are safe for a general-purpose assistant to read, and which should never leave a narrower, access-controlled system.

Classification is the slower step most organisations skip. It means going through unstructured data (file shares, email archives, years of scanned documents) and tagging what each thing is and how sensitive it is. Permissioning follows from that: deciding, category by category, which systems, which AI tools and which staff roles get to see it. Skip that work and an AI deployment either sees too little to be useful, or too much to be safe. Both are failure modes, and both tend to surface only after the system is already live.

This is also why "most of our data is clean" and "most of our data is AI-ready" are different claims, and why Komprise's respondents were right to separate them. A spreadsheet of sales figures can be clean and fully classified in an afternoon. A twelve-year archive of contracts, scanned invoices and departmental notes is a different scale of problem, and it is usually where an organisation's actual risk and actual value both sit.

Why the Estimate and the Reality Diverge

The gap between 93 to 96% of decision-makers admitting most of their data is not ready, and the general confidence many organisations project about their AI maturity, is not really a contradiction. It is what happens when readiness gets judged informally. A team can genuinely believe its data is in reasonable shape if it is going by what is visible: the databases it queries daily, the systems everyone uses. Unstructured data built up over years, spread across file shares, retired collaboration tools and departmental drives nobody fully owns, does not get the same scrutiny. It is also where most of the volume sits, and where most of the sensitive material hides.

There is an incentive problem too. Classification and permissioning work does not show up on a roadmap. It produces no visible output until an audit, a data incident, or a stalled AI pilot forces the question. So it gets deferred, year after year, while the data pile grows faster than the governance underneath it. By the time an organisation notices, the backlog is not a quarter's work, it is several years of accumulated file shares, inherited systems from acquisitions, and departmental archives nobody currently owns.

What the First Steps Actually Look Like

The practical starting point is smaller than it sounds. It does not require solving the whole estate at once. It starts with an inventory: what data exists, where it lives, who created it, and who currently has access, often the first time anyone has asked that question in years. From there, classification can be prioritised by risk and value rather than attempted everywhere at once. Start with the data that would cause the most damage if exposed, or add the most value if an AI system could use it safely. Permissioning follows the classification, not the other way round. Deciding who gets access before you know what you are granting access to is how organisations end up with AI tools that are either too restricted to be useful, or too open to be safe.

This is the specific problem our AI Readiness assessment is built around. Before any AI deployment, we help an organisation classify its historic data (what it is, how sensitive it is, who should be able to see it) and put staged authorisation around it, rather than assuming a general-purpose assistant should simply be pointed at everything on the file server. It is unglamorous work, closer to an audit than a product launch, and that is roughly the point. The Komprise numbers suggest most organisations are trying to run AI on top of data that has never been through that process, which is a reasonable explanation for why so many AI pilots underdeliver against the expectations set for them.

None of this makes the 7% figure a surprise. Cleaning, classifying and permissioning a decade of unstructured data is slow, unglamorous work with no product launch at the end of it. But it is the work that determines whether an AI system is actually safe to point at an organisation's real data, or just fast at producing answers nobody can fully stand behind.

Sources

• GlobeNewswire, 'Komprise Survey of Unstructured Data and AI Reveals Wide Gap Between Perceived AI Maturity and AI Results', globenewswire.com (15 October 2026).
• Blocks and Files, 'Komprise: AI needs AI-ready data, and there isn't enough of it', blocksandfiles.com (19 October 2026).