Why AI Success Depends More On Data Foundations Than Models

Enterprise AI conversations often focus on the model. Which one performs best? Which one is fastest? Which one offers the largest context window? Should the organisation use a commercial model, an open-source alternative, or several different models routed through the same platform?

These are reasonable questions. Models shape what an AI system can do. But they're also becoming easier to access. Most enterprises aren't building foundation models from scratch. They're using models provided through cloud platforms, software vendors, and managed AI services. Their data is a different story.

Enterprise data carries years of business history, operational knowledge, customer relationships, internal rules, and business context. A powerful model can't compensate for data it can't find, definitions it doesn't understand, or policies it doesn't know it must follow.

That gap is becoming increasingly visible. According to a report from Cloudera and Harvard Business Review Analytic Services, only 7 per cent of respondents believed their organisation's data was completely ready for AI.

AI success increasingly depends on whether organisations can provide trusted information, clear business meaning, and the right controls when systems need them. That may prove much harder to replicate than any model.

Why Models Are Becoming Easier To Access

Foundation models are large AI models trained on broad collections of information. They provide the base capabilities behind many generative AI tools, including the ability to interpret language, generate content, write code, and answer questions. 

Building one requires enormous amounts of computing power, specialist talent, time, and money. Most organisations have no reason to do it. Instead, enterprises can choose from models offered by AI companies, cloud providers, software vendors, and open-source communities. 

They can also use different models for different workloads. A smaller model may handle a repetitive internal task, while a more capable model supports complex analysis or customer-facing applications. This flexibility is changing where competitive advantage comes from.

Two companies can use the same model. They may even access it through the same cloud provider and build with similar tools. Yet one can produce useful, dependable results while the other struggles to move beyond demonstrations. The difference often lies in what surrounds the model.

• Can it retrieve the right information? 
• Does that information reflect current business conditions? 
• Are customer, revenue, risk, and product definitions consistent? 
• Can the system tell which sources are approved? Does it know what a user is allowed to see?

Without those foundations, model comparisons can become a distraction. Teams spend weeks debating small performance differences while the data feeding every option remains fragmented or unclear. The more useful question is no longer only, “Which model should we use?”

It’s also, “Can our AI data foundations help any suitable model produce an answer the business can use?”

Why AI Changes What A Good Data Foundation Looks Like

Data quality isn’t a new concern. Organisations have spent decades trying to improve reporting, analytics, forecasting, and decision-making by cleaning records and bringing information together. But traditional analytics environments were mainly designed for people.

A human analyst can notice that two reports use different revenue totals. They can ask the finance team which definition applies. They may recognise that a blank field doesn’t mean “no”, or that a sudden change probably came from a system migration rather than customer behaviour. AI systems don’t arrive with that organisational awareness.

They can find patterns and generate convincing responses, but they don’t automatically understand why one dataset is trusted and another isn’t. They won’t know that a business unit uses a different definition of an active customer unless that distinction has been recorded somewhere the system can access.

This becomes more serious as organisations introduce AI agents. Unlike a chatbot that waits for a question, an agent can complete several steps, use tools, retrieve data, and trigger actions with less direct human involvement.

That means business context can’t remain trapped in policy documents, informal processes, or the knowledge of individual employees. It needs to become machine-readable. In practical terms, an AI-ready data foundation must tell a system:

• What the data represents
• Where it came from
• How recently it was updated
• Who owns it
• How it has changed
• Which definitions apply
• Who can access it
• What it can be used for

The challenge isn’t simply collecting more enterprise data. It’s giving that data enough structure and context for an intelligent system to interpret it consistently.

The Five Data Capabilities Becoming AI Differentiators

No single tool can make an organisation’s data ready for every AI use case. Readiness comes from several connected capabilities, each answering a different question about whether information can be trusted and used appropriately.

Data quality

AI data quality describes whether information is suitable for the task an AI system is being asked to complete. Accuracy is part of that, but it isn’t the whole picture. Data also needs to be complete enough, consistent across relevant sources, current enough for the decision, and representative of the situation being analysed.

A customer record may be factually correct but still be unsuitable for predicting current purchasing behaviour if it hasn’t been updated in two years. A risk dataset may contain no obvious errors but still create distorted results if it excludes important groups or operating conditions.

This is why quality needs to be assessed against the use case. The standard required for summarising internal documents won’t be the same as the standard required for approving credit, detecting fraud, or changing a customer’s account. The model can only reason from the version of reality it receives.

Metadata and business context

Data rarely explains itself. A column may contain a value called “customer status”, but that label doesn’t tell an AI system how the status was calculated, who owns the definition, or whether it can be used outside the sales platform. 

Metadata management adds this information. Metadata is data about data, including descriptions, classifications, owners, formats, update histories, and usage rules. For AI, metadata is becoming more than a catalogue people consult when they need help finding a dataset. 

It can provide active context that helps systems distinguish approved information from an old copy, recognise sensitive content, and understand how business concepts relate to one another. 

That changes the role of the data catalogue. Instead of acting only as an inventory, it starts becoming part of the operating environment around AI.

Data lineage and traceability

When an AI output looks wrong, someone needs to work backwards. 

• Which source supplied the information? 
• What transformations were applied? 
• Which version reached the model? 
• What other systems used the same data? 
• Did the output influence a report, recommendation, or automated action?

Data lineage records that path. Traditional lineage often followed data from its original source into a warehouse, dashboard, or report. AI extends the chain. Organisations increasingly need visibility from source data through the model or agent and into the decision or action that followed.

This supports auditing and regulatory evidence, but it also helps with everyday operations. When a flawed dataset is discovered, teams can identify which AI systems used it and which outputs may need to be reviewed. Without traceability, one data problem can turn into a long search across systems nobody sees together.

Semantic consistency

Many organisations have several correct definitions of the same thing. Finance may calculate revenue one way. Sales may use another. Regional teams may apply local rules, while an executive dashboard uses a group-wide definition. People can often work around these differences because they understand the purpose behind each report.

An AI system may simply choose one. A semantic layer creates shared definitions and relationships that sit between raw data and the tools using it. It can define what terms such as “active customer”, “net revenue”, or “high-risk account” mean, including which calculations and conditions should be applied.

This gives AI a more reliable understanding of the business. It also reduces the chance of an answer being technically valid but commercially wrong. The growing investment in semantic technologies shows how important this problem is becoming. 

Snowflake’s Horizon Context, for example, is designed to make governed business definitions and relationships available across AI agents, applications, and analytics tools. Its own guidance notes that large language models can generate plausible queries while still misunderstanding joins, metrics, or the structure of enterprise data.

Governance and policy enforcement

Governance has often depended on written policies, approval meetings, and scheduled reviews. Those controls may work when a person requests access to a dataset and waits for permission. They’re less suited to an agent retrieving information and completing a workflow in seconds.

Data governance for AI needs to operate closer to the point of use. Access restrictions, privacy requirements, retention rules, and approved purposes must travel with the data and be applied when an AI system tries to use it. This doesn’t remove human accountability. It gives people a practical way to enforce the decisions they’ve made.

Databricks’ 2026 State of AI Agents report found that organisations using AI governance tools moved more than 12 times as many projects into production. This is an observed correlation rather than proof that governance alone caused the difference. 

Even so, it suggests that controls and oversight are helping mature organisations deploy AI rather than holding them back.

Why Data Platforms Are Shifting Towards Context And Control

Technology vendors tend to build around the problems their customers keep bringing to them. Right now, many of the largest data platforms are investing in metadata, semantic context, lineage, governance, observability, and policy enforcement. 

Storage and computing capacity still count, but they’re no longer enough to support the systems enterprises are trying to build. Snowflake is expanding its catalogue into a context layer for AI. Databricks is connecting data, models, agents, evaluation, and governance within the same environment. 

Collibra is extending lineage and oversight across AI assets, decisions, and actions. Alation is bringing models and agents into its governance framework. Informatica is embedding AI across integration, quality, metadata, and policy workflows.

These companies take different approaches, and their claims should be judged against each organisation’s architecture and requirements. Still, the direction is remarkably consistent. The market is moving from storing and processing data towards helping AI understand which information to use, what it means, and what it’s allowed to do with it.

Collibra’s May 2026 AI Command Center reflects this shift. It brings models, agents, use cases, ownership, behaviour, decisions, and risk into a shared control environment. The underlying goal is continuous oversight rather than a governance review performed before deployment and forgotten afterwards.

This doesn’t mean every enterprise needs to replace its data stack. It does mean leaders need to look beyond whether their platforms can connect to an AI model. The harder question is whether those platforms can provide enough context and control for the model to operate safely once it’s connected.

Questions Leaders Should Be Asking About Their Data Foundations

AI readiness assessments often begin with infrastructure, skills, use cases, and model availability. Those areas still belong in the conversation, but they can create false confidence when the supporting data can’t carry the workload. A more useful assessment begins with the decisions the AI system will support and works backwards.

Leaders should ask:

• Can our AI systems identify which data sources are approved for a specific use?
• Are important business terms defined consistently across systems and teams?
• Does lineage continue from the original data into models, agents, outputs, and actions?
• Can access and usage policies be applied automatically when data is retrieved?
• Could we identify which decisions were affected if a dataset was later found to be wrong?
• Is our metadata current enough for an automated system to rely on?
• Does every high-value dataset have a clear owner who can make decisions about its quality and use?

A “no” doesn’t mean the organisation has to stop every AI initiative until its data estate is perfect. No enterprise has perfect data, and waiting for it is an excellent way to remain in planning meetings forever.

It does show where controls, ownership, or context need to improve before a use case becomes more autonomous or reaches more of the business. That’s the purpose of an AI readiness assessment. It isn’t to produce a comforting score. It’s to identify which weaknesses could change the outcome of a real system.

Final Thoughts: AI Advantage Is Becoming A Data Foundation Problem

Models will continue to improve. They’ll become faster, cheaper, and easier to use. Organisations will gain access to capabilities that would’ve required specialist research teams only a few years ago. But access doesn’t equal advantage. The information an organisation has built over years of operating is far more difficult to reproduce. 

So are the definitions, relationships, controls, and institutional knowledge that make that information useful. The next phase of enterprise AI may be decided less by who reaches the strongest model first and more by who can give a suitable model the clearest understanding of the business.

That requires data quality suited to the decision, metadata that provides context, lineage that preserves traceability, semantics that create shared meaning, and governance that works while systems are running.

For technology and data leaders, the work is no longer limited to adopting AI. It’s building the conditions that allow AI to act with enough context, confidence, and accountability to earn a lasting place in the organisation.

EM360Tech continues to follow how enterprises are strengthening those conditions across data strategy, governance, architecture, and AI operations, because the systems organisations build next will only be as dependable as the foundations they inherit.