Enterprise AI Exposes Data Modeling Gaps That Require More Than Validated Numbers

An enterprise dataset can be validated and still produce a misleading answer if an AI system lacks the business context needed to interpret it. The same activity may appear in order, payment, and accounting systems, each recording a different part of the process. As AI tools consume analytical data directly, organizations need clear definitions of what datasets represent, which systems are responsible for them, and how they should be used.

Silja Märdla is the Head of Analytics Engineering at Bolt, with extensive experience in data engineering, business intelligence, and analytics infrastructure. Previously, she led data engineering and business intelligence teams at Omniva, overseeing data warehouse development and improving reporting capabilities. At Bolt, she leads analytics engineering teams focused on building reliable data models and establishing consistent practices across multiple business lines.

"If you're relying on an automated system, the rules need to be written down. Otherwise, it isn't going to work well," Märdla says.

Growing business complexity calls for more structured data models

Medallion architecture provides a foundation for organizing enterprise data, with gold-layer datasets intended for validated consumption. However, managing analytical models across multiple products and teams introduces complications that the architecture alone can't resolve. Medallion is still valid, but it's not enough. Märdla believes additional structure and rules are necessary to tell people what has to happen inside every layer of the medallion. Large tables combining different business processes may initially support reporting needs, but they become difficult to maintain as new requirements emerge.

Märdla places particular emphasis on the intermediate modeling layer, where teams can divide analytical structures into smaller, manageable components. "You can theoretically build something once, but the business will move on," she explains. "It will introduce new products and processes that you still need to measure."

At Bolt, her team addresses this by identifying recurring business entities, including users, drivers, cars, cities, and product categories, alongside the processes and events involving them. These concepts provide a consistent basis for organizing analytical models as operational requirements evolve. "The more clearly you can define those concepts, the easier it becomes to map them into data structures and then apply a system or framework to maintain each of those structures over time," she says.

Reliable analytics depend on choosing the right data sources

While consistent models make analytical data easier to maintain, teams still need to determine which measures are appropriate for particular business decisions. At Bolt, product teams calculate revenue using recorded orders and their expected value, while finance teams rely on accounting or payment systems to measure money moving into and out of the business. Both figures describe revenue, although they serve different purposes. "Those are legitimately different things, and different teams need both," Märdla says. "We should define them separately and expose them as distinct datasets, with explanations of when to use each one."

Choosing the appropriate measure also requires identifying which system is responsible for the underlying business process. "The invoicing system is responsible for invoicing. If I want invoicing metrics, I start with invoicing-system data. If I want order metrics, I start with the order system."

Once teams identify those authoritative sources, they can enrich the foundational datasets with information from other systems while preserving the distinctions between them. For AI systems consuming the resulting data products, those relationships need to be explicit enough to support accurate interpretation.

Consistent data models support more reliable AI analytics and automation

Communicating the distinctions mentioned above becomes easier when analytical datasets follow shared modeling conventions. Märdla sees value in establishing recognizable patterns for representing business entities, processes, and metrics. "If you can generalize those conventions and explain them once to your AI or human consumers, there's a better chance they'll use the data correctly," she says.

Legacy datasets complicate this work because their structures may have developed without common conventions. Teams can compensate with extensive documentation that explains how individual datasets should be interpreted, although Märdla sees greater potential in building consistent structures from the beginning. "You can't always build from beginning," she notes. "It's also okay to unify later, but I do support unification to achieve the consistency. It takes some extra time, but also buys that back for you later."

Documented conventions can also support automation within data engineering itself. When teams establish clear rules for constructing analytical models, they create a basis for automating parts of the development process. "Unless you understand how to do something yourself, it's very difficult to hand it over to anyone else," Märdla says.

As AI takes on more analytical work, trust in data becomes central

Successful automation depends on a clear understanding of analytical workflows and the standards used to validate their outputs. Märdla emphasizes the importance of making validation status visible to consumers, distinguishing certified datasets from provisional information whose limitations may still be acceptable depending on the business context. "If you're reporting to investors, I'd double-check it," she says. "If you're using it internally to make a quick decision, some data is usually better than no data, unless there's a known, serious issue with it."

As AI systems assume more responsibility for delivering analytical results, confidence in the underlying data becomes increasingly important. Clear definitions, traceable origins, and transparent validation help users assess the reliability of the information they receive. For Märdla, maintaining that confidence remains a fundamental responsibility of data teams. "Users losing trust in data is the worst outcome of conversational AI systems. They will not only stop trusting the LLM, but stop trusting data in general," she concludes.