In 2023, a major financial services firm deployed a generative AI model to flag suspicious transactions, only to have it flag the same legitimate high-value customer repeatedly because three different source systems defined “customer tier” inconsistently—one by revenue, one by tenure, one by product count. The model had no way to know which definition was correct, so it applied all three simultaneously, generating 40% false positives and burning through human review capacity within weeks. The root cause wasn’t the model. It was data.
This isn’t an outlier. According to a 2024 Gartner survey, 68% of enterprises report that poor data quality directly impacts their AI initiatives, yet only 23% have formal data governance frameworks that extend to AI workloads. The disconnect is stark: traditional data governance was built for humans reading dashboards and reports. AI systems read the entire dataset, at scale, at machine speed. A decimal place in the wrong column doesn’t matter to a spreadsheet user. To a language model fine-tuned on financial data, it’s signal.
The infrastructure shift happening now isn’t about better tools. It’s about treating data governance as an operational layer that AI agents depend on, the same way cloud infrastructure depends on networking. Without it, trust collapses.
Why Traditional Governance Models Fail at AI Scale
Legacy data governance typically operates on a compliance cadence: annual audits, quarterly reviews, documentation tucked into a wiki nobody reads. The assumption is that errors get caught in a validation layer before they reach critical systems. That model breaks completely when AI is the consumer.
A traditional BI pipeline might surface 50 metrics to 200 business users. Each user applies context: they know that Q4 numbers include one-time adjustments, they understand that “active user” has a different definition in mobile versus web, they notice when a number looks wrong and investigate. An LLM trained on the same data doesn’t apply that context. It sees patterns. If “revenue” is defined three ways across your warehouse—gross, net, and blended—the model doesn’t pick the right one. It averages them into something meaningless.
Databricks conducted an analysis of 200 companies using their Unity Catalog and found that 34% had duplicate or conflicting definitions for core business concepts across their data platforms. Snowflake’s 2024 State of Data Governance report showed that enterprises without formal semantic layer definitions saw 2.3x higher error rates in AI model outputs compared to those with centralized business glossaries.
The problem compounds when you add multiple AI agents working in parallel. If Agent A queries your revenue table and Agent B queries your customer table, and neither has a guaranteed lineage back to source truth, they’re operating on different assumptions about what the data means. That’s not a data quality issue. That’s a trust architecture problem.
Metadata and Lineage: From Audit Trail to Operational Input
For decades, data lineage was treated as a compliance artifact. “Show me where this number came from for the auditors.” Tools like Apache Atlas and Collibra tracked lineage retroactively, mostly for documentation.
That’s changing. Lineage is now operational infrastructure for AI.
When an LLM is asked, “What’s our total revenue this quarter?”, the system needs to:
- Identify which table contains revenue
- Verify that table’s upstream sources and transformation logic
- Confirm that the transformation was validated after the last schema change
- Check whether the user querying is authorized to see that specific grain of data
- Return not just an answer, but a confidence score based on lineage integrity
This is why dbt has become essential infrastructure in AI-first data organizations. dbt’s model documentation and test results create machine-readable lineage that automation systems can consume. When dbt runs tests and they pass, downstream AI systems can trust that specific transformation. When tests fail, AI query engines should refuse to serve that data, not just log a warning.
Companies using dbt with formal metadata enrichment—documented owners, SLOs, freshness guarantees—report 50% faster time-to-resolution when AI models produce unexpected outputs because the metadata pinpoints which transformation or source system changed.
Semantic Consistency as a Precondition for AI
Semantic layer tools like Looker and dbt Semantic Layer aren’t optional anymore. They’re foundational.
Consider “active user.” In your mobile product, it means “logged in within the last 30 days.” In your SaaS platform, it means “has a current subscription and at least one API call in the last 90 days.” In legacy reports, it meant something else entirely. A human analyst reading two reports knows these are different. An AI system doesn’t. It just sees the metric name.
When Doordash scaled their internal AI agents, they discovered that 17 different business definitions of “order” existed across their data ecosystem—some excluded refunded orders, some didn’t; some counted test orders, some filtered them out. They built a centralized semantic layer that defined “order” once, with variants explicitly documented, and connected it to all downstream systems. Adoption took three months. It prevented six different AI-driven projects from shipping broken models based on conflicting definitions.
A semantic layer forces the hardest question first: “What does this word actually mean?” It moves the answer from tribal knowledge to code. That explicitness is what makes AI systems trustworthy.
Access Control and Identity for AI Agents
Most data governance frameworks apply access controls to humans. You’re in the finance group, so you can see revenue data. AI agents need the same logic, plus additional constraints.
An AI agent shouldn’t have blanket access to all tables just because it’s running in production. It should have the same row-level and column-level security as the human it’s representing, plus time-based restrictions and query pattern monitoring. If an agent suddenly tries to query 100x its normal volume, that’s a signal to audit the request.
Systems like Snowflake’s Dynamic Data Masking and Databricks’ row and column access control are now being deployed specifically for AI workloads. The policy enforcement is identical to what governs human queries—but it’s evaluated in milliseconds, at scale, before the query returns data to the model.
Data Governance and AI Governance Are Converging
You can no longer separate these disciplines. Data governance policies now determine what an AI system can learn. AI governance policies now determine how data must be stored.
This means your data governance team needs to include someone who understands transformer architectures and fine-tuning. Your AI governance team needs someone who can explain schema design. That integration is happening in leading organizations through cross-functional data councils that meet monthly to align policies before they’re implemented.
Building AI-Ready Governance Without the Bureaucracy
The goal is automation, not process overhead. This means:
Implement single-source-of-truth metadata that’s consumed programmatically. Use dbt’s built-in testing and documentation, not a separate governance tool that duplicates that information. Store the semantic layer in code, version it, and make it part of CI/CD. Deploy access controls through policy-as-code frameworks that integrate with your data platform’s native enforcement engines.
Start with your three most critical datasets. Define their lineage in dbt, document business logic in your semantic layer, and establish access policies. Don’t try to govern your entire warehouse at once.
If you’re building similar solutions in your organization, consider sharing your approach on LinkedIn Daily’s write-for-us program—the community benefits from real implementations, not abstractions.
Your Next Action
This week, audit one critical dataset that feeds AI systems in your organization. Map its lineage in whatever tool you use (dbt, Atlas, or manually). Identify every transformation step and every point where its definition could diverge from business intent. Document which humans currently know the answers to those questions. That’s your governance gap. Close it by moving that knowledge into code.
Write for us
We publish contributed reporting on cryptocurrency. If you have something specific and checkable, see what we are looking for.