Most AI initiatives don't fail because the models are wrong, they fail because the underlying data lacks the metadata and governance needed for AI to reason over it reliably. This post argues that AI can only work with context that's been made explicit, unlike experienced analysts who carry undocumented institutional knowledge, and that missing context doesn't stop AI, it just makes AI guess confidently. Using Databricks Genie Spaces as a working example, it explains why Unity Catalog is the governance layer that turns tacit knowledge into something AI can actually use safely.
Do your AI Ambitions Have a Context Problem?
According to Gartner, 63% of organisations either do not have or are unsure if they have the right data management practices for AI. That’s significant because despite the attention given to models, prompts, and copilots, AI success is increasingly determined by something far less glamorous: the quality, governance, and context of your data — alongside the organisational readiness needed to act on it.
Organisations across every sector and size are investing in Generative AI. They are implementing productivity tools like ChatGPT, deploying conversational analytics tools like Databricks Genie Spaces, and experimenting with agents that interface with data across their business and execute operational processes. Many of those initiatives will underdeliver. Not because the models are wrong or because the technology is immature, but because the data underneath lacks the metadata, documentation, and governance needed for AI to reason over it reliably.
In general, the quality of AI output is only as good as the data and context it is given. That has always been true of human analysts. AI is no different.
Context Is Everything
AI is impressive, but it cannot draw on the knowledge and experience that people carry – only on what has been made explicit, in documents, metadata, and semantic layers. A model doesn’t know that “Customer” and “Client” refer to the same business concept unless it’s been told or has been able to infer it based on its training data or related context.
With the right access and instrumentation, models can increasingly use signals like usage patterns to infer which version of a table is authoritative or which pipeline has been superseded – but even with that, it is just an educated guess, and confident assumptions can often the hardest mistakes to detect.
For years, organisations compensated for this with experienced people: analysts knew which datasets to trust, which reports had gone stale, which definitions had evolved. That knowledge was rarely documented, but it existed. AI has no equivalent to fall back on – it can only reason over the context and training data it’s given.
Consider two scenarios:
- In the first, an AI is pointed at tables with names like cust_trx_v3, tbl_rev_final2, and staging_export_OLD. No descriptions, no column-level annotations, no documented relationships. AI does its best, but it is effectively guessing, and it will guess with confidence.
- In the second scenario, the same tables have clear descriptions, column-level context explaining what each field represents, classifications indicating sensitivity, and documented relationships to related datasets. The AI now has what it needs to reason more accurately. The outputs are more trustworthy, and the business will be more confident to act on them.
That difference (between contextualised data and raw tables with opaque names) is not a documentation problem. It is an AI-readiness problem. Generative AI simply makes it impossible to ignore.
Data has always been more valuable when it is accompanied by rich metadata. The problem is that governance and metadata have historically struggled to excite delivery teams, too often dismissed as a documentation exercise rather than recognised as a business enabler. Until now, organisations could rely on the knowledge of experienced people to fill those gaps. AI can’t.
How This Works in Practice: Genie on Databricks
Let’s take the use of Genie Spaces as an example. Genie Spaces is Databricks’ natural language analytics capability. It allows business users to ask questions of their data in natural language and receive explainable answers without writing SQL or depending on a data analyst to build a report.
The quality of Genie’s responses is directly dependent on the metadata and business semantics available to it. In practice, Databricks increasingly recommends expressing these business semantics through Unity Catalog Metric Views, which provide a governed semantic layer containing business metrics, relationships, synonyms, and descriptions. Genie uses this metadata, alongside table and column descriptions, governed tags and documented relationships, to translate business questions into accurate queries.
A Genie Space with access to rich, well-maintained metadata produces responses that are more precise and trustworthy. An environment with sparse or absent metadata leads to responses that are approximate at best and misleading at worst.
This is not a limitation of Genie – it reflects the environment it operates in. Databricks explicitly recommends rich metadata, business definitions, and Metric Views, because they provide the semantic context Genie needs to accurately interpret natural language and generate trustworthy SQL.
Responsible AI Requires Governed Access
Metadata does more than improve output quality. It enables organisations to control what AI systems have access to, and under what conditions. Unity Catalog uses governed tags applied to data assets to drive access policies automatically – a column classified as containing personal data can be masked for AI workloads without a legitimate purpose, without manual intervention at every step.
This matters because AI interacts with data at a scale and speed that makes human oversight at every point impractical. A well-governed metadata model is what makes it possible to give AI access to production data responsibly, by enabling automated access rules to restrict data rather than relying on broad permissions that expose more data than an AI workload needs.
If Your AI Projects Are Not Delivering, Context May Be Why
For organisations running Databricks, Unity Catalog is the governance layer that makes all this possible – rich metadata, granular access control, automated lineage, and the architectural foundation for Genie Spaces, Metric Views, and the platform’s most advanced AI capabilities.
If your organisation is investing in AI on Databricks and not seeing the return you expected, it is worth asking whether your platform is leveraging Unity Catalog and whether the data those models are working with has enough context to be useful. Poorly described, inconsistently governed data produces AI that under-delivers.
Getting Unity Catalog right, and metadata more generally, is not trivial, and it is worth doing properly. It’s something our team have a great deal of experience in and have written an article on to share their top six steps that determine Databricks Unity Catalog implementation success Six Steps That Determine Databricks Unity Catalog Implementation Success – worth a read!
At Analytics8, we help organisations build the governed data foundations their AI ambitions require. If that sounds relevant to where you are, we would be glad to talk. Book a free, no-pressure strategy session with one of our senior consultants. We’ll talk through where you are, what’s blocking progress, and what a practical path forward could look like.
