AI can find and compare conflicting definitions across your reporting, SQL, and documentation. Deciding which one the enterprise standardizes is a business decision, and the real risk is a model that makes that decision without anyone noticing.
If you read nothing else
AI can compare definitions across reports, SQL and documentation, but it should not quietly select an enterprise standard when none has been approved. Retain distinct measures, trace each to its evidence, and let the accountable business owner decide whether a shared definition is needed.
Imagine asking an AI system to build a Customer 360 data model.
You give it the available documentation, reporting definitions, database schemas, and transformation logic. It identifies the customer entities, suggests relationships, and produces a candidate logical model. Everything looks reasonable.
Then someone asks a simple question: What makes a customer active?
In this hypothetical, the revenue team counts a customer as active if they've generated revenue in the last 90 days. The customer master system uses an ACTIVE status. The product team counts anyone with an active product.
Each definition may serve a legitimate purpose. They also describe different kinds of information: one is calculated from transactions over a time window, one comes from a maintained status field, and one is derived from product records. The revenue measure needs an as-of date; the stored status and product activity may have their own effective dates. The question is whether the organization needs separate definitions or an approved shared meaning.
An AI system can find all three and explain how they differ. Choosing one as the enterprise standard is a different kind of decision. If the system makes that choice quietly, the model can look perfectly reasonable while representing the business incorrectly, and nobody reviewing the diagram will know a choice was made.
The situation itself is common in large organizations. Business rules accumulate. Teams define terms for their own reasons, applications change, companies acquire other companies, and reporting logic evolves without every definition being reconciled. Eventually someone is asked to bring it all together. AI can shorten that investigation considerably. Accountability for the answer stays where it was.
For technology leaders, this reaches beyond a diagram. An unresolved definition can flow into management reports, operational dashboards, and AI applications. Teams may believe they are using the same metric while making decisions from different populations. Finding the discrepancy during design is usually easier than explaining inconsistent numbers after deployment, especially when nobody has been assigned to own the definition.
What the existing evidence can and can't tell you
Experienced data modelers already know how to structure entities, define relationships, and work through normalization or dimensional design. In an existing enterprise, the difficulty is understanding what the organization has actually built.
Take a customer domain. A company may have a CRM customer, a billing account, a legal entity, a master customer record, and several product-specific identifiers. Some refer to the same organization; others represent different parts of the relationship. The database doesn't always say which. Documentation helps, but some pages are out of date. A Power BI report may use a definition that differs from the customer master. A transformation may contain logic added years ago to handle an acquisition that nobody has revisited.
An experienced modeler can work through this. The problem is volume: how much evidence has to be read before reaching a conclusion. AI is useful here because it can read the supplied documentation, compare reporting calculations, examine SQL transformations, propose candidate relationships, and flag inconsistencies for the modeler.
The mistake is treating that first pass as a finished model. Finding a relationship is not the same as proving it is correct. Two tables may share a column name without sharing a business meaning, and a customer ID in one system may be a billing account in another.
It also helps to start from one business concept rather than asking AI to design the ideal domain model. A broad request invites the model to fill gaps with general industry patterns, which is useful when designing something new and risky when the task is to reconstruct what an existing enterprise already means. With Active Customer, find where the term appears, how each report calculates it, and whether different functions use different logic, then follow those calculations back to the source systems. In a Microsoft environment the evidence may sit across Power BI, Fabric, Purview, SQL, notebooks, and Confluence. Other organizations will have different tools, but the investigation is much the same.
Lineage helps establish available dependencies between data assets, but it is not a substitute for reading the actual measure expression, SQL, or notebook transformation. Nor does a matching field name prove shared business meaning. Microsoft documents specific limits for Microsoft Purview's scanned Fabric lineage: for non-Power BI items, cross-workspace lineage and notebook-to-pipeline lineage are not supported. These are limitations of the Purview integration, not a blanket description of every Fabric lineage view. Teams should check current product documentation and their own environment before assuming a trace is complete.
Sometimes the documentation says one thing while the SQL does another. That difference shouldn't be corrected quietly to produce a cleaner model. It may reveal an outdated rule, an intentional exception, or a genuine disagreement between business functions, and the modeler needs to know which.
Narrow questions also work better than broad ones, because they are easier to check. Give the model a specific question and the relevant evidence:
"These three reports use the term Active Customer. Compare their definitions, identify the underlying source fields, and explain where the logic differs. Do not select an authoritative definition."
That gives the AI a useful analytical job without asking it to make a decision it isn't authorized to make. Asking it to cite the report, measure, or SQL fragment behind each finding lets a reviewer check the claim against the source instead of rediscovering it, and keeping those findings and the decisions that follow outside the chat means they outlast the session.
One label. Three valid meanings.
A shared label does not establish a shared business rule. These examples are hypothetical, not POC results.
Keep the definitions separately named. Raise a governance issue only if they are being used interchangeably or a single approved enterprise definition is required.
A relationship can look right and still produce the wrong numbers
Grain describes what one record represents. A customer table may hold one row per legal entity, an account table several accounts per customer, and a transaction table thousands of records per account. If the relationships are misunderstood, combining these tables multiplies records and distorts the results.
An AI system may suggest a join based on matching column names and available metadata. The join may even run. That doesn't mean the result is correct. Consider a customer with several contracts and several products. Joining both to the customer in one query returns every combination of contract and product, rows that never existed as business events, and a revenue measure can jump because transactions were duplicated, not because the customer paid more. History produces the same trap more quietly. If a customer dimension keeps a new row each time an attribute changes, a join on the business key alone returns every historical version, and a relationship that looks one-to-one in the current data turns out to be one-to-many.
The SQL runs. The dashboard loads. The numbers look plausible. Finding the problem takes someone who checks the actual grain, keys, and cardinality.
One customer, six joined rows
Two independent one-to-many relationships can multiply rows when joined together without the right grain.
Why it matters: A measure from one branch may be repeated across combinations from the other. Aggregating after that join can inflate totals. Validate the grain and aggregate at the correct level before combining branches.
Identity is a related but separate problem. Legacy applications and acquisitions leave a company with several identifiers for the same customer, and matching them is rarely as simple as selecting columns with similar names. Where permissions allow, profile the proposed relationships: how many records match, how many don't, whether keys are duplicated, and whether a supposedly unique relationship returns multiple matches. Read the results as evidence, not proof. A high match rate can come from two systems whose unrelated numeric IDs happen to overlap, and a low one can be legitimate when many customers simply have no account. Where profiling isn't permitted, record that limitation rather than substituting an estimate.
This is where experienced modelers add the most. Their understanding of grain, history, entity boundaries, and business meaning helps them see structural problems that a convincing output can hide.
Make the model show what it doesn't know
A finished diagram is easier to present than one that comes with unresolved definitions and open questions, and AI strengthens that pull, because it can generate a detailed structure even when the evidence is thin.
A better convention is to give each finding a status: supported by current technical evidence, inferred, in conflict with other evidence, or unknown. A customer-to-account relationship might be directly supported. Another might be inferred only from naming conventions. Two reports might contradict each other on a business rule, and a historical identifier might stay unexplained because its source is unavailable. These shouldn't carry the same weight when the model is reviewed or handed to another team. Without the distinction, an early assumption can harden into an accepted design decision simply because it appears in the documentation.
It also changes how AI is used. Rather than asking whether the model is correct, ask the model to attack it: what contradicts this relationship, what was assumed about customer identity, could this join duplicate records, and was this attribute found in an existing source or added because similar systems usually have it. The most useful question is the plainest. What would make this conclusion wrong?
Not every conflict needs a new business decision. If the organization has already approved an enterprise definition, in a ratified glossary entry or a governance decision, the AI's job is to apply it and flag the reports that deviate. A decision is needed when no approved definition exists, or when the approved one doesn't match what the reports compute.
In the Active Customer case, the working deliverable should preserve the three distinctly named definitions, each with its expression, source, evaluation date or time semantics, and consuming reports. The definitions themselves are not necessarily conflicting: revenue activity, master-record status, and product activity can be three legitimate measures. Record a separate unresolved governance issue only when teams use the same label interchangeably or need one enterprise standard with no approved rule. If a ratified definition already exists, apply it and flag deviations. Otherwise, ask the owner a focused question: "Should these remain distinct named measures, or should one become the approved enterprise-wide definition?" The decision, evidence, owner, and date should be saved outside the AI conversation.
What the team should actually record
| Named definition | Evidence to retain | Status |
|---|---|---|
| revenue_active_customer | 90-day logic, as-of date, report and expression reference | Evidenced* |
| master_active_customer | Status field, effective date and source definition | Evidenced* |
| product_active_customer | Product-activity rule, source and related reports | Evidenced* |
Is an enterprise-wide “Active Customer” definition needed? If so, who can approve it? Record the decision, owner, date and downstream consumers.
*Evidenced here means the calculation exists in the hypothetical source; it does not imply business-wide approval or empirical validation.
Decide what the AI is allowed to see
None of this works until the data-handling question is settled. Enterprise data modeling involves information that can be confidential even when it contains no customer records. Schemas expose internal application structures, and SQL transformations can reveal pricing logic, financial calculations, and segmentation rules. Giving that material to an external AI service without understanding the controls isn't a sensible starting point.
Teams need to settle which AI environment is approved, what it may process, who can access the source material, what gets logged, how long information is retained, and whether the provider may use it for other purposes. No single deployment model fits every organization. Some teams will use approved enterprise AI services with appropriate contractual and technical protections; others will prefer private or self-hosted environments for certain workloads. Self-hosting gives more control over deployment but doesn't remove the need for access management, monitoring, and security review.
The same restraint applies to access. An AI system shouldn't be allowed to run arbitrary queries against production databases. For many modeling tasks, a controlled set of metadata, approved documentation, and scoped aggregate queries is enough. Give the system the evidence it needs and not everything around it.
The output still needs review. OWASP describes indirect prompt injection as malicious instructions embedded in external material such as files and websites. An everyday modeling risk is quieter: a stale wiki page or misleading column comment that the model treats as authoritative. Both are reasons to review retrieved material, separate evidence from instructions, and maintain clear source references.
How to tell whether AI is actually helping
It's easy to show that AI can produce a data model quickly. That isn't the same as showing it reduced the effort required to produce an accepted one. Suppose AI generates a logical structure in an hour, but the modeler spends several days correcting relationships, tracing unsupported attributes, and resolving assumptions. Has the approach saved time? Possibly, but it takes a baseline to say so. Preparing evidence, reviewing outputs, correcting errors, and documenting decisions all count as effort.
Quality counts as much. Each important attribute should trace to a source, grain and cardinality should be validated, conflicting definitions should be visible, and open questions should be specific enough for the right owner to answer. It also matters whether the work duplicates what the organization already has. If a lineage system maintains source dependencies, asking AI to recreate that documentation adds nothing; the useful task is interpreting what exists, finding the gaps, and connecting the technical evidence to the business model.
Some situations don't favor AI at all. If the relevant evidence is small and clear, a modeler may resolve it faster without building a workflow around it. If the source material is unreliable or inaccessible, AI may produce more uncertain conclusions than useful findings. If what remains is a policy decision, more analysis won't help until the right owner answers.
Centizen has recently begun a proof of concept to evaluate AI-assisted enterprise data modeling. No technical outcomes or effort improvements have been established yet. The evaluation is intended to test how AI can support analysis of existing evidence while experienced modelers retain responsibility for validating conclusions. Any assessment of value will need to cover the full process, not just the speed of a first draft.
A model you can defend
Return to Active Customer. A useful system finds each definition, traces its underlying logic, and explains why the populations differ. If an approved enterprise rule exists, it can apply that rule and identify deviations. If no rule exists, it should preserve the alternatives and route the unresolved choice to the accountable owner.
For engineering and data leaders, a sensible way to begin is to name who has authority to settle definition conflicts before the work starts, since a conflict with no owner has nowhere to go. Then pick one meaningful business concept, see how well the technology extracts and interprets the evidence, note where modelers have to step in, and measure the total effort an accepted result takes. Then decide whether to expand.
The test of the final model is whether someone accountable can explain why each definition in it is there and what evidence it rests on.
Frequently Asked Questions
Can AI create an enterprise data model automatically?
AI can generate candidate entities, attributes, and relationships from existing schemas and documentation. The proposed model still needs validation against source data, implemented business rules, and approved definitions before it can be accepted.
Why can AI-generated data models be inaccurate?
AI may infer relationships from similar column names, misunderstand data grain, overlook historical identifiers, or assume that conflicting business definitions mean the same thing. The result can be a convincing structure with the wrong business meaning.
Can AI resolve conflicting business definitions?
It can compare implementations and apply an existing approved rule, subject to verification. If there is no authoritative rule and the enterprise needs one, an authorized business owner must decide; distinct legitimate measures need not be treated as conflicts.
Does data lineage tell you what a business term means?
Not by itself. Lineage shows available dependencies, while the meaning of a term also depends on calculation logic, documentation, and approved business definitions. Coverage varies by tool and asset type.
Is it safe to use enterprise data with an LLM?
That depends on the deployment environment, the information involved, access controls, retention policies, and the organization's security requirements. Teams should use approved environments and limit access to the evidence the task requires.
How should companies measure the value of AI-assisted data modeling?
Compare total effort and accepted quality against a defined baseline. Include evidence preparation, human review, corrections, validation, and unresolved issues, not just the time to generate the first draft.
For organizations reviewing enterprise data architecture and AI readiness, explore Centizen’s engineering and data services.
Continue the Centizen AI Engineering series
For the broader engineering workflow, read Rethinking the Software Development Lifecycle in the Age of AI and How to Build a Reliable AI-Assisted Development Workflow.
Our services:
- Staffing: Contract, contract-to-hire, direct hire, remote global hiring, SOW projects, and managed services.
- Remote hiring: Hire full-time IT professionals from our India-based talent network.
- Custom software development: Web/Mobile Development, UI/UX Design, QA & Automation, API Integration, DevOps, and Product Development.
Our products:
- ZenBasket: A customizable ecommerce platform.
- Zenyo payroll: Automated payroll processing for India.
- Zenyo workforce: Streamlined HR and productivity tools.