For the past two years, the conversation around generative AI (GenAI) has been dominated by models – their size, their speed, their novelty. As always, early “era” experience forces the conversation to change.
As large language models (LLMs) rapidly commoditise, the source of competitive advantage is shifting decisively. It is no longer about which model you use. It is about what you feed it: the representativeness, context and applied governance of your data.
This shift is already exposing a hard truth. Through 2027, Gartner predicts that 30% of GenAI projects will be abandoned after proof of concept, not because the technology fails, but because organisations are unprepared to operationalise it. Poor understanding of the alignment of data, weak risk controls, rising costs and unclear business value are quietly derailing scale.
The implication for IT and D&A leaders is clear: GenAI does not fail at the model layer. It fails to differentiate at the data layer.
The end of “model-first” AI
Early GenAI adoption followed a predictable pattern: experiment with a model, test a use case, and demonstrate potential. But scaling those experiments into production has proven far more complex.
This is because GenAI introduces a level of opacity that traditional machine learning did not. Organisations are now working with systems where training data, inference pathways and outputs are not fully transparent. This contributes to what Gartner defines as a trust barrier – uncertainty around relevance, reliability and risk.
In this context, simply adding more data is not the answer. In many cases, it amplifies noise rather than improving outcomes.
What matters is representative data: data that is not aligned to specific business outcomes but is annotated and tagged based on its uses. This is enriched with metadata and context tagging to maximise efficacy for its usage, adding valuable context every time the data is used, reused, monitored and audited. This helps determine how the data should respond to any given governance model that is authorized to use it. Only through this approach is it possible to reduce risk, and organisations that fail to make this distinction will continue to see GenAI stall at pilot stage.
From data management to data readiness
Many organisations assume they are “data ready” because they have invested in data platforms, pipelines or governance frameworks. GenAI raises the bar significantly.
Data readiness for GenAI is not a static capability. It is a continuous, iterative discipline that evolves alongside models and business needs.
Gartner’s 2025 State of AI-Ready Data Survey showed that organisations using automated data readiness assessments, such as continuous profiling and regression testing, are 2.3 times more likely to achieve high effectiveness in data engineering for AI.
This signals a broader shift: from managing data as an asset, to engineering data as a dynamic system. Data must be continuously aligned to business problems, actively governed to mitigate risk, and constantly refined through feedback and evaluation.
This is not a one-off transformation. It is an operating model.
Why metadata has emerged as the new differentiator
If models are commoditised, then context becomes the differentiator, and context is delivered through metadata—and assures representativeness.
Metadata transforms raw data into something meaningful by explaining how it has been interpreted over-and-over, and where different interpretations either reinforce or challenge each other. Without it, even high-quality datasets can produce inconsistent or irrelevant GenAI outputs.
Gartner finds that organisations implementing metadata management effectively are 4.3 times more likely to achieve high effectiveness in AI data engineering.
This is because metadata continually identifies conditions in which one specific use-case may see noise, while another identifies signal. In probabilistic systems like GenAI, that signal is what anchors output to real business value.
Trust is engineered, not assumed
One of the most persistent barriers to scaling generative AI is trust.
Executives are right to question outputs that may hallucinate, expose sensitive information or produce inconsistent results. But limiting adoption is not a viable long-term response. Trust must be engineered into the system itself.
Leading organisations are embedding observability into their systems that interact with control mechanisms that make GenAI outputs more reliable, auditable and usable. They are introducing secure data gateways to filter sensitive information based upon conditional analysis before it reaches external models. They are identifying when structured outputs ensure consistency and support downstream integration. They are also developing domain-specific evaluation datasets so that performance is measured against real organisational requirements, rather than generic benchmarks.
Organisations with comprehensive AI security policies are significantly more likely to achieve both governance effectiveness and measurable business impact.
The rise of the data flywheel
Perhaps the most important shift in GenAI is how value is sustained over time.
Traditional systems tend to degrade without intervention. Generative systems, when designed correctly, can improve with use. This is the principle of the data flywheel: a continuous cycle where usage generates data, data improves models, and improved models drive further usage.
However, this cycle must be intentionally designed. It depends on human feedback, automated evaluation and ongoing refinement of both data and prompts. Without these mechanisms, GenAI deployments remain static and fail to deliver compounding value.
Organisations that operationalise this approach move beyond experimentation. They create adaptive systems that learn continuously and scale effectively.
From pilot to production
To move from isolated pilots to production-scale GenAI, organisations must take a more disciplined approach to data.
This begins with leadership. Clear accountability within a governance model ensures that GenAI initiatives remain aligned to business priorities rather than fragmented experimentation. It extends to enriching data with metadata to provide context and relevance. It requires implementing governance controls that actively identify risk and protect sensitive information where required before it reaches models.
Equally, organisations must address the variability of GenAI outputs by enforcing process standards and ensuring those processes exhibit consistency and usability across systems. They must ground performance in real-world expectations through expert-led evaluation data “testing” to evaluate for confirmation or challenger situations. They must also establish continuous feedback loops that refine both models and data management over time.
Finally, leading organisations are applying AI techniques to data preparation itself, improving efficiency, reducing cost and accelerating readiness across the entire lifecycle.
Taken together, these shifts redefine how organisations approach AI. They move GenAI from promise to performance.
The next phase of GenAI: data as strategy
The narrative around generative AI is evolving. What began as a technology story is now a data story.
Organisations that continue to focus on models will find themselves competing on commoditised capabilities. Those that invest in data, its qualification, its governance and its contextual relevance, will define the next phase of advantage.
This is the data reckoning. GenAI success will be determined by who feeds it best, not who adopts it fast.
Gartner analysts are exploring GenAI-ready data, AI engineering foundations and strategies to scale AI from pilot to production at this week’s Gartner Data & Analytics Summit 2026 in London.
Mark Beyer
Mark Beyer is Distinguished VP Analyst at Gartner‘s Data and Analytics practice, specialising in data architecture, governance and AI-ready data strategies. His research focuses on the intersection of data management and generative AI, including active metadata, data fabric, data mesh and distributed governance. With experience spanning application development, data architecture and project delivery, Mark is known for combining strategic insight with practical, real-world implementation expertise.
