TL;DR: Data modeling for AI isn’t optional. Organizations struggling with inconsistent AI outputs almost always trace the problem upstream to ungoverned data definitions. Semantic inconsistencies that once produced wrong dashboards now produce wrong decisions at machine speed and enterprise scale. Data modeling establishes the shared semantic layer that data governance and AI governance both depend on. Solve the data modeling problem first, and your AI strategy has something solid to stand on.
Here is a question worth sitting with: if your organization launched an AI initiative in the last 18 months, how confident are you that every system consuming your data is working from the same definition of “customer”? Or “revenue”? Or “churn”?
If the answer is anything other than completely confident, you are not alone. And this is not a new problem that AI created. It is an old problem that AI has made catastrophically more expensive.
“Data modeling is the foundation of data governance, and data governance is table stakes for AI governance and succeeding with it.”
– Danny Sandwell, Industry Strategist, Quest Software
I explored this challenge recently in a webinar with Danny Sandwell, a Quest industry strategist and one of the most seasoned voices in the data modeling discipline, and Robert Lutton, VP at Sandhill Consultants, who works inside enterprise modeling teams day in and day out. What came through clearly in that conversation is something that practitioners at every level are now running into: the foundational layer most organizations skipped or deprioritized is the exact layer their AI strategy depends on.
That framing landed as the central organizing principle of the entire conversation. As Danny Sandwell put it, “Data modeling is the foundation of data governance, and data governance is table stakes for AI governance and succeeding with it.” It is not marketing language. It is the operational reality data teams are running into every day. The semantic inconsistencies that used to produce wrong dashboards now produce wrong AI decisions, at machine speed and enterprise scale, with no human analyst in the loop to catch them.
Danny Sandwell, Quest industry strategist, explains why data modeling is the non-negotiable foundation beneath both data governance and AI governance.
Why AI elevated the stakes
For most of data modeling’s history, the cost of getting it wrong was a bad dashboard or a delayed report. A human analyst would catch the inconsistency, escalate it, and the team would fix it. Painful, but contained. AI removes that containment entirely. When a machine learning model trains on data where “revenue” is defined three different ways across three systems, it does not surface the conflict. It learns all three definitions and delivers three kinds of wrong output, at scale, with confidence. The downstream cost is not a wrong spreadsheet. It is a flawed recommendation engine, a biased risk model, or a compliance failure embedded in thousands of automated decisions before anyone notices.
“AI does not fix bad data or bad foundations. It amplifies them.”
– Robert Lutton, VP, Sandhill Consultants
Robert Lutton, VP at Sandhill Consultants, on why AI doesn’t fix a broken data foundation: it amplifies it.
The hierarchy your AI strategy is built on
According to Gartner, at least 30% of generative AI projects will be abandoned after proof of concept due to poor data quality, inadequate risk controls, or unclear business value.1 Notice that the top reason is not a technology problem. It is a data foundation problem.
The same research points at poor data quality as the leading culprit, and poor data quality is almost always downstream of unclear, inconsistent, or ungoverned data definitions.
Gartner also predicts that by 2027, 60% of AI projects will miss their value targets due to fragmented, reactive governance structures that do not align with business objectives.2
The pattern here is not subtle: fragmented stacks produce fragmented meaning, and fragmented meaning produces AI systems that cannot be trusted. You can invest in the best model infrastructure available, but if the data feeding that infrastructure lacks consistent semantic definitions, the output cannot be trusted at scale.
The hierarchy Danny described is worth unpacking in practical terms:
- Data modeling for AI provides the shared definitions, naming standards, and structural blueprints.
- Data governance operationalizes those definitions across the organization.
- AI governance depends on data governance to enforce consistency, traceability, and trust at the point where models are trained and deployed.
You cannot shortcut this hierarchy. You cannot govern AI well if you have not governed your data. And you cannot govern your data well if you have not modeled it.
That is what Robert Lutton was getting at when he said that “AI does not fix bad data or bad foundations. It amplifies them. You have to be very careful in making sure that your AI is as good as the platform that it’s enabled on.” The logical data model, the layer that defines shared business meaning rather than just physical structure, is the platform everything else depends on. And it has never had higher stakes attached to it than it does right now.
Where the friction actually lives
Most organizations understand this in theory. The friction is in practice.
Legacy data modeling workflows were built for a different operating environment: desktop-bound tools, expert-only access, and sequential design cycles that could take weeks or months to complete. Meanwhile, your data engineers are working in Snowflake, Databricks, and Microsoft Fabric on two-week sprint cycles. Your business analysts are making real-time decisions without access to the governed definitions that would make those decisions reliable. Your AI teams are waiting on data products that are blocked upstream by modeling backlogs.
The demand for data modeling has outpaced the traditional delivery process, and teams are turning to AI-assisted approaches not because the discipline has become less important, but because it has become more important and more urgent than the old process can support.
Three problem areas show up consistently across organizations navigating this challenge:
- Slow time to value, compounded by fragmentation. The volume of modeling work has grown faster than the number of qualified modelers available to do it. Backlogs accumulate, timelines stretch, and downstream teams work around the bottleneck by making their own assumptions about data meaning. Meanwhile, most enterprises now operate across multiple cloud platforms simultaneously, and keeping meaning consistent across that fragmentation is not a problem that physical schema precision solves. Both pressures feed the same outcome: semantic drift starts before anyone is watching for it.
- Semantic drift. This is the quietly dangerous one. It does not announce itself. It accumulates gradually, as different teams settle on slightly different definitions for the same concepts, and those differences compound across every downstream system that consumes the data. By the time an organization recognizes it has a semantic drift problem, the drift has usually been embedded in production systems for months.
- The AI readiness gap. AI systems are only as reliable as the definitions they are trained on. Organizations investing in AI that have not invested in data modeling for AI are building on an unstable foundation, because the semantic layer is what AI governance requires. The AI does not know what it does not know, and it will not surface its own inconsistencies. That responsibility belongs to the data modeling discipline.
“The organizations that get ahead of this are the ones that recognize data modeling not as a prior-generation discipline that AI is replacing, but as the foundational layer that AI governance depends on.”
The job of data modeling tools is no longer just to draw the picture. Ryan Crochet explains why protecting shared meaning is now the whole point.
What data modeling for AI actually requires
Solving these problems is not primarily a tooling decision. It is an organizational commitment to treating data modeling as a living practice rather than a static documentation exercise.
That commitment requires a few things:
- Broader participation. Business analysts, data engineers, and business stakeholders all depend on the data model, but they are rarely contributors to it. Closing the gap between business intent and technical implementation requires modeling environments that cross-functional teams can actually work in, without demanding specialist expertise as the price of admission.
- Real-time collaboration. The iteration cycles that legacy modeling workflows produce, where a data architect takes requirements offline and returns two weeks later with a design, are too slow for how modern data teams operate. The people who consume the model need to be able to participate in it, in the same environment, at the same time.
- Enterprise governance infrastructure. Speed and collaboration are only valuable if they are accompanied by the governance guardrails that prevent a fast, collaborative process from becoming a fast, collaborative path to a new set of inconsistencies. Version control, naming standard enforcement, check-in/check-out controls, and a central repository are not optional features for mature organizations. They are the difference between a modeling practice and a modeling free-for-all.
- Connection to the modern data stack. Modeling environments that require teams to step outside their existing workflows will not get adopted at the pace the organization needs. Integration with the tools your data engineers already use, dbt, Git, and the cloud warehouses where your data actually lives, is the practical requirement for making modeling a continuous practice rather than a periodic project.
The organizations that get ahead of this are the ones that recognize data modeling not as a prior-generation discipline that AI is replacing, but as the foundational layer that AI governance depends on. They are building semantic consistency before drift sets in, expanding modeling participation before backlogs become blockers, and governing their data foundations before their AI systems scale the problem.
“The AI governance problem is a data modeling problem in disguise. Solve the right one first.”
The shift you need to make
If you are a few months into an AI initiative and you are starting to feel the friction, the most productive question to ask is not “what is wrong with my AI model?” It is “how consistent are the definitions feeding it?”
Trace the problem upstream. The inconsistency in your AI output almost always originates in an inconsistency in your data foundation, which almost always traces back to a data modeling process that was not designed to keep pace with the way your organization is now working.
The fix is not to abandon data modeling. The fix is to evolve to data modeling for AI: broader participation, real-time collaboration, enterprise governance infrastructure, and continuous integration with the modern data stack.
The AI governance problem is a data modeling problem in disguise. Solve the right one first.
Sources:
- Gartner, “Gartner Predicts 30% of Generative AI Projects Will Be Abandoned After Proof of Concept By End of 2025,” July 2024
- Gartner, “Gartner for IT LinkedIn post,” September 2025
