Power BI DAX: What LLM benchmark results show

Key insights

• Research notes report scores of 90% to 96% for six models in the DAX LLM Benchmark v2.
• Strong benchmark performance does not establish that a formula is correct for your semantic model.
• LLMs can help draft, explain, modify, and explore DAX expressions.
• People still need to define business meaning, check filter behavior, and assess model performance.
• The video description identifies a DAXBench v2 live discussion but provides no transcript or detailed results.

Power BI DAX formula writing has become a stronger use case for large language models (LLMs), but benchmark performance does not make generated measures safe to use without review. This matters to Power BI developers, analysts, and learners who use AI to draft or troubleshoot calculations. SQLBI is the source of the benchmark discussion referenced in its YouTube description, which identifies a DAXBench v2 live conversation with Marco Russo and Maxim Anatsko.

What's new

DAX, or Data Analysis Expressions, is the formula language used in Power BI semantic models, Analysis Services, and Power Pivot in Excel. It defines calculations such as measures, calculated columns, calculated tables, and queries. The research notes describe a marked improvement in LLM performance on DAX-writing tasks, with six models from three vendors scoring between 90% and 96% on the DAX LLM Benchmark v2.

Those results suggest that newer models can handle many common formula-writing tasks well, and may outperform people who are new to DAX. They do not show that LLMs are better than experienced developers across every task. In particular, a benchmark score is not a guarantee that a formula will return the intended result in a specific organization’s model.

The practical change is that an LLM can serve as a faster drafting and exploration partner. It can propose a measure from a written requirement, explain parts of an expression, or suggest an alternative implementation. The developer still needs to establish whether the requirement is clear and whether the result matches the business meaning.

Who it is for and what you need

This approach is relevant to Power BI analysts and developers who write or maintain DAX, as well as learners using an assistant to understand formulas. The research notes also mention AI assistants such as Microsoft Copilot, but the provided material does not specify a Copilot license, feature, setting, or availability. It does not establish that any particular assistant or model is available to every Power BI user.

To judge a proposed formula, you need the relevant model context: table and column names, relationships, measures, and the intended calculation. DAX results can depend on filter context—the active selections when a measure is evaluated—and row context, the current row used by calculated columns or iterator functions. A request that omits these details can leave an AI assistant guessing.

The material does not give a rollout stage, regional availability, or a prerequisite product version for the benchmark results. It also does not provide enough information to reproduce the benchmark or compare individual models. Treat the reported scores as research findings, not as a guarantee about the assistant available in your tenant.

How it works

Using an LLM for DAX is a drafting and checking process, not a replacement for model knowledge. SQLBI’s video description points to a live DAXBench v2 discussion, but there is no transcript supplied here. The following workflow reflects the DAX tasks and validation considerations described in the research notes.

1. Describe the business calculation and name the relevant tables, columns, existing measures, and expected behavior. Include how slicers or other filters should affect the result.
2. Ask the assistant to draft a DAX expression, or to explain or modify an existing one. Treat the response as a proposal rather than a confirmed calculation.
3. Review how the expression uses functions such as or iterator functions such as . Check whether it preserves, replaces, or changes the filters the requirement depends on.
4. Test the formula in the intended semantic model. Compare results with known totals and check relevant filter selections, aggregation levels, and edge cases before relying on it.
5. Assess performance in the actual model before adopting a suggested rewrite. A plausible alternative is not proof of a faster or more appropriate implementation.

This review matters because DAX behavior depends on evaluation context and model relationships. A formula can be syntactically valid yet analytically wrong if it assumes the wrong table, relationship, or filter behavior. Model design and business definitions also affect whether a technically sound expression answers the right question.

Limitations and open questions

The benchmark figures in the research notes are broad results; the supplied material does not include task-by-task scores, test prompts, or enough methodology to independently interpret the percentage. The video description itself asks whether LLMs are now better at DAX than humans, but the transcript is absent, so it cannot establish Marco Russo’s detailed answer or what was demonstrated during the live discussion.

The research notes describe limitations around complex model behavior, relationships, calculation groups, variables, newer syntax, and edge cases. They also warn that a formula that works in one visual or sample may fail under different slicers or levels of aggregation. These are reasons to test against the model’s expected business outcomes, rather than judging an expression by appearance or confidence.

For experienced developers, the work that remains includes clarifying ambiguous requirements, designing the model, interpreting business meaning, validating results, and making performance decisions. AI can reduce time spent drafting routine calculations, but the supplied material does not support treating it as an autonomous replacement for DAX expertise.

Sources

• SQLBI’s DAXBench v2 discussion
• SQLBI’s short on LLMs and DAX
• SQLBI website

Frequently asked questions

Are LLMs better than humans at writing DAX?

The research notes suggest advanced models can outperform many beginners on benchmarked DAX tasks. They do not establish that LLMs outperform experienced developers across all work or models.

Can I use an AI-generated DAX measure without testing it?

No. The notes warn that a valid-looking formula can still be wrong for a particular model or filter context. Validate results against expected business outcomes and relevant edge cases.

What do I need to give an AI assistant to get useful DAX?

Provide the calculation requirement and relevant model context, including table and column names, relationships, measures, and intended filter behavior. The supplied material does not specify a required Copilot license or setting.