In January 2026, Microsoft repositioned Copilot at the centre of Power BI, introducing conversational analytics as the primary way business leaders are expected to interact with data. By December 2026, Microsoft will retire the legacy Q&A visual entirely. The message is unambiguous: the future of business intelligence is AI-powered, and organisations that have not prepared will lose natural language query capabilities without a replacement.

But a significant number of organisations are rushing to enable Copilot on data that was never designed to support it. The results are predictable. Inaccurate responses. Inconsistent answers to the same question. Executives who distrust the output and revert to manually built dashboards. Adoption collapses within weeks.

The problem is not Copilot. The problem is that Copilot is being treated as a feature to switch on, rather than an outcome that requires a governed data foundation underneath it.

What Copilot Actually Does With Your Data

Power BI Copilot does not query raw data. It interacts with semantic models. It reads field names, table descriptions, measure definitions, relationships, and synonyms to interpret what a user is asking and generate a response. When those elements are well-structured and clearly defined, Copilot performs reliably. When they are absent, ambiguous, or inconsistent, Copilot guesses.

The core risk

Enabling Copilot without optimising the underlying semantic model is the equivalent of hiring a highly capable analyst and giving them access to a filing cabinet with no labels, no structure, and no context. The analyst will produce answers. They will simply often be wrong.

The specific failure modes are well documented. A column named simply 'Amount' without context gives Copilot no way to determine which table's Amount field to use. Measures without descriptions produce misleading narratives. Missing relationships cause incorrect aggregations. And when multiple datasets are available, Copilot has no way to know which one represents the single source of truth for a given metric.

Microsoft introduced the 'Approved for Copilot' control precisely because of this. Administrators can now mark specific datasets as trusted for AI use and restrict Copilot to querying only those vetted sources. But that control only works if the datasets being marked as trusted are actually trustworthy in the first place.

The Governance Foundation Copilot Requires

A governed semantic layer

The semantic model sitting underneath Power BI reports needs to be deliberately designed for AI consumption. This means a lean schema that exposes only business-relevant tables and measures, removes technical artefacts like staging tables and surrogate keys, and uses naming conventions that reflect how the business actually talks about its data. Every measure and column should carry a concise business-oriented description. Every hierarchy should be explicitly defined.

A single source of truth for key metrics

One of the most common governance failures in large organisations is metric sprawl. Revenue is defined differently across Finance, Sales, and Operations. Customer count uses different deduplication logic depending on which team built the report. When Copilot queries across these inconsistent definitions, it will return different answers to the same question depending on which dataset it draws from. Consolidating into enterprise-wide models with standardised metric definitions is a prerequisite, not an optimisation.

Sensitivity classification and data access controls

Copilot operates across whatever data it has access to. In regulated industries or organisations handling commercially sensitive data, this creates a real risk. Sensitivity labels applied through Microsoft Purview ensure that classified datasets are either excluded from Copilot access or handled under appropriate compliance controls. Without this layer, Copilot can surface data that should be restricted, and there is no audit trail to demonstrate what was queried and by whom.

Data quality at the gold layer

Copilot should only be enabled on data that has been through quality remediation. In a Medallion architecture, this means the gold layer: standardised, deduplicated, validated, and ready for business consumption. Pointing Copilot at bronze or silver layer data exposes all the quality issues that those layers are specifically designed to contain. The result is AI-generated insights built on incomplete or inconsistent records, which erodes confidence faster than any governance failure.

What Happens Without This Foundation

The pattern is consistent. An organisation enables Copilot. Early users ask straightforward questions and get plausible-sounding responses. Then someone notices that the number Copilot returns for total revenue does not match the number in the finance pack. Investigation reveals that Copilot used a different measure definition. Trust collapses. Adoption stops. The Copilot licence is underutilised and the investment is written off as a technology disappointment rather than a governance failure.

Research published in 2025 found that among organisations that experienced AI implementation issues, data quality was the top concern, cited by 41% of respondents. Copilot does not create this problem. It makes the problem visible, and it does so in a very public way when business leaders ask a question and receive a wrong answer.

The Right Sequence

The organisations that get the most from Power BI Copilot follow a consistent sequence. They do not enable Copilot and then try to fix the data. They fix the data first and then enable Copilot as the interface to a foundation that is already trustworthy.

The Broader Shift

Microsoft's decision to retire the legacy Q&A visual by December 2026 is a forcing function. Organisations that have been deferring semantic model governance no longer have the option to stay on the old approach. The deadline is fixed. The only question is whether organisations arrive at it with a governed data foundation or without one.

Power BI Copilot is not a technology risk. It is a governance readiness test. The organisations that pass it will have AI-assisted analytics that is trusted, auditable, and scalable. Those that do not will have a Copilot that produces answers that nobody believes.

How VE3 Can Help

VE3 works with organisations to design and implement the governed semantic layer that makes Power BI Copilot reliable. This includes data quality assessment, semantic model optimisation, Microsoft Purview sensitivity classification, and user enablement programmes. If your organisation is preparing for Copilot adoption or reviewing an existing deployment, speak with our data and analytics team.