A useful exercise, before any AI roadmap meeting: ask everyone in the room which of the leading foundation models they're planning to use. Note the confidence with which the answer comes. Then ask whether the data those models will run on is in a fit state — schemas defined, quality known, ownership clear, access controls aligned to who can see what. Note the silence.
That gap between confident answers about models and shrugs about data is the gap that explains why so much enterprise AI spend produces nothing.
The published numbers on data readiness in 2026 are bracing. Cloudera and Harvard Business Review Analytic Services, surveying senior leaders in October 2025, found that just 7% of organisations said their data was completely ready for AI adoption, while 27% reported their data as not very or not at all ready. Dun & Bradstreet's 2026 AI Momentum survey produced a similar picture from a different angle: 97% of organisations report active AI initiatives, but only 5% say their data is ready to support them.
These are not edge cases or outliers. They are the median. And they explain a lot.
The model is not the moat
There is a comforting story about AI strategy that goes: pick the best model, write good prompts, ship. It is comforting because it puts the decision somewhere visible and somewhere external. Vendors compete publicly. Benchmarks update monthly. A team can pick a model with confidence in an afternoon.
The discomforting truth is that the model is, increasingly, not where competitive advantage lives. The frontier model you can use is also the one your competitor can use. Three new ones will be released between now and the end of the quarter. None of them know anything specific about your business, your customers, your contracts, your processes, your history. None of them will be better than the next one in six months.
What they will all do, equally well, is operate on the data you give them. Which means the differentiator — the part nobody else has — is the data layer underneath. Your taxonomy. Your historical decisions. Your customer records. Your support transcripts. Your contracts. Your operational telemetry. These are unique. These compound over time. These are difficult to copy. These are the moat.
The trouble is that the moat, in most companies, is also a swamp.
What "AI-ready" actually means
The term "AI-ready data" gets used loosely. Gartner has done the most rigorous job of defining it, and the definition is more demanding than most teams realise. Their working description is that AI-ready data is data aligned to specific use cases, actively governed at the asset level, supported by automated pipelines with quality gates, and continuously quality-assured.
Note what is and is not in that definition.
What is in it: alignment to use cases (not just "clean data" — clean for what?), governance at the level of individual assets (not annual audit cycles), automated pipelines (not manual exports to a notebook), and continuous quality assurance (not point-in-time checks).
What is not in it: volume. AI-ready is not "we have lots of it." Volume without alignment is noise.
Gartner makes another point that is easy to miss but important: high-quality data by traditional standards is not the same as AI-ready data. Traditional data quality wants outliers removed, anomalies cleansed, edge cases tidied away. AI training and AI evaluation often want exactly the opposite — representative data including the errors, the outliers, and the unexpected-but-valid cases, because that is what the model needs to learn from and be tested against. A dataset that passes traditional DQ checks can be a poor fit for AI work. A dataset that fails them might be exactly right.
This is the kind of nuance that gets lost when "we should clean up our data for AI" becomes a programme of work. Clean for what? Aligned to which use case? Governed by whom, at what cadence? Without those answers, the data-cleanup project produces tidier data that is still not AI-ready.
The four conditions, in plain language
If you strip the analyst vocabulary away, the conditions data needs to meet to be useful for an AI feature are roughly these.
Representative. It contains the patterns the model needs to see — including the awkward ones. A customer-service knowledge base that documents the easy cases and omits the gnarly ones produces a system that handles the easy cases beautifully and falls over on the gnarly ones.
Governed. Someone owns it. Access is controlled at a level fine-grained enough that you can answer "should the AI feature have been allowed to see this?" before the regulator asks. The Cloudera/HBR research found that 73% of respondents say their organisation should prioritise AI data quality more than it currently does, with siloed data (56%) and lack of a clear data strategy (44%) as the top obstacles.
Pipeline-fed, not exported. The data your AI feature operates on updates automatically. If someone is running a weekly SQL export and dropping a CSV into a vector database, you do not have a system, you have a person whose holiday will break things. As Futurum's 2025 survey put it, the rise of AI has transformed poor data quality from a persistent annoyance into an existential threat to the top business priority.
Continuously checked. The pipeline includes quality gates that fail loudly when the upstream data changes shape. Schema drift, missing fields, format changes, duplicate explosions — these will happen, and the question is whether you find out from a quality gate or from a customer.
A data set that meets all four for one specific use case is AI-ready for that use case. A data set that meets none of them is the median enterprise data set in 2026. There is a lot of room between those two points, and most of the work happens in that room.
The honest cost of skipping this
The temptation, especially early in an AI project, is to skip the data work. The model is exciting. The demo is achievable. The data prep is unglamorous and uncertain and hard to scope. Teams skip ahead.
The numbers on what happens next are now well-documented enough to be quoted. Gartner predicted in 2025 that through 2026, organisations will abandon 60% of AI projects unsupported by AI-ready data. S&P Global Market Intelligence reported in 2025 that 42% of companies were abandoning the majority of their AI initiatives before reaching production, up from 17% the prior year — a more-than-doubling in a single year. MIT's NANDA initiative, in its July 2025 State of AI in Business report, found that 95% of organisations were seeing zero measurable return from generative AI investment.
These numbers can be argued with at the edges. Methodology matters; sampling matters; the definitions of "abandoned" and "zero return" vary across the studies. The MIT figure has been pushed back on by reasonable people. But three independent studies pointing in the same direction is not a coincidence. The structural pattern is real: most enterprise AI projects do not produce value, and the most-cited reason is not the model.
What does produce value, in the same studies, is doing the data work. The MIT report's most-cited differentiator between success and stalled cases was not which model was used or which framework. It was integration: connecting the AI system to actual workflows, fed by data the system could actually use.
The work that nobody puts on slides
What does "doing the data work" actually look like, in a team that is shipping AI features in 2026? It is not a single project. It is a discipline that lives next to the development team.
It looks like data engineers and AI engineers in the same standup, working on the same backlog. It looks like data contracts between source systems and downstream consumers, with breaking changes flagged before deploy. It looks like a knowledge base where the entries have lineage — who wrote it, when, against what version of the policy. It looks like automated evaluations that run on a sample of production data nightly, with alerts when retrieval quality drops below threshold. It looks like access controls applied before retrieval, not after generation. It looks, in short, like normal engineering hygiene applied to a layer that most organisations have historically treated as someone else's problem.
This is unglamorous work. It does not photograph well. It does not demo well. It is also the work that turns the 95% who get zero return into the 5% who get serious value.
The teams we see succeed treat this layer as a first-class part of the AI stack, not an afterthought. They invest in it before they invest in the application layer. They name an owner. They build pipelines instead of running exports. They write down what AI-ready means for their specific use cases, and they hold themselves to those definitions. They abandon projects where the underlying data cannot be made ready in a sensible timeframe, rather than building on top of unreliable foundations and discovering the cost later.
What this means for buyers and builders
For teams currently planning or scoping AI work, two practical things follow from this.
First, before you pick a model, audit your data against the use case. Be specific. "Can our existing customer support transcripts, in their current state, support an AI-driven first-response feature with the quality bar we'd want?" is a much more useful question than "should we use Model A or Model B?" The honest answer is often "not yet, and here's the work in between." That answer saves quarters.
Second, treat the data layer as infrastructure, not as a project. Projects end. The data layer does not. It needs an owner who is responsible for its health continuously, not a kick-off and a Gantt chart. The teams that get this right end up with a capability that compounds across every AI feature they build afterwards. The teams that don't end up rebuilding the same brittle pipeline for every new initiative.
There is a third implication, and it is uncomfortable to say cleanly without it sounding like marketing: most companies should not be building this infrastructure from scratch in 2026. The patterns are now mature enough, the failure modes are now well-enough documented, the engineering effort to do it well is now expensive enough — that for most use cases, the right answer is to use a layer that has been built for this purpose and is operated by people whose job it is. We've written about that build-versus-buy question separately. The point here is more upstream: even if you intend to build it yourself, you cannot escape doing the work. The data has to be ready, however you get there.
The companies pulling away from the pack on AI are not the ones with the best model access. They are the ones who realised, earlier than their competitors, that the model was the easy part. The data was the moat. And the moat was theirs to dig.
References
- Cloudera and Harvard Business Review Analytic Services — Taming the Complexity of AI Data Readiness (October 2025). https://www.cloudera.com/about/news-and-blogs/press-releases/2026-03-05-only-7-percent-of-enterprises-say-their-data-is-completely-ready-for-ai-according-to-new-report-from-cloudera-and-harvard-business-review-analytic-services-reveals.html
- Dun & Bradstreet — AI Momentum Survey 2026, summarised in CIO. https://www.cio.com/article/4170978/nearly-every-enterprise-is-investing-in-ai-but-only-5-say-their-data-is-ready.html
- Gartner — Lack of AI-Ready Data Puts AI Projects at Risk (February 2025). https://www.gartner.com/en/newsroom/press-releases/2025-02-26-lack-of-ai-ready-data-puts-ai-projects-at-risk
- Gartner — AI-Ready Data Essentials to Capture AI Value. https://www.gartner.com/en/articles/ai-ready-data
- Gartner — Hype Cycle for Artificial Intelligence 2025. https://www.gartner.com/en/articles/hype-cycle-for-artificial-intelligence
- Alation — How AI-Ready Data Drives AI Success (analysis of Gartner's "Quick Answer: What Makes Data AI-Ready?"). https://www.alation.com/blog/gartner-report-ai-ready-data-drives-success/
- Futurum Group — 1H 2025 Data Intelligence, Analytics, and Infrastructure Decision Maker Survey Report. https://futurumgroup.com/press-release/new-futurum-survey-shows-enterprise-demand-for-an-ai-ready-data/
- MIT NANDA Initiative — The GenAI Divide: State of AI in Business 2025 (July 2025), summarised in Fortune. https://fortune.com/2025/08/18/mit-report-95-percent-generative-ai-pilots-at-companies-failing-cfo/
- S&P Global Market Intelligence — AI in Business: Path to Maturity 2025, summarised in QuickLaunch Analytics. https://quicklaunchanalytics.com/bi-blog/data-ai-readiness-enterprise/
- Informatica — 2025 CDO Insights Survey, summarised in QuickLaunch Analytics. https://quicklaunchanalytics.com/bi-blog/why-80-of-ai-projects-fail-before-they-start-its-your-data-foundation/
Written by JP, Sixees Labs. Last reviewed May 2026.