FactCraft

Study · November 8, 2026

What would an AI have to guess?

We ran FactCraft's audit on 53 public Power BI semantic models from GitHub: 1,404 tables, 6,374 measures, 1,424 relationships, read as metadata. What a model declares is what an agent can read about a measure: its name, its description, its DAX. Most of these models say very little.

48 of 53

models came back Not ready for AI: Copilot has no AI instructions to follow.

44 of 49

models with visible measures leave some of them undescribed.

13

models carry measures with identical DAX. Some of those are wrong numbers.

The verdicts

A good score does not make a model ready.

The score is an average across the rules, weighted by severity, so the median stays high. The verdict is stricter: a model is not ready as soon as one blocking rule affects more than half of what it examines. 48 of the 48 models not ready have no AI instructions: that alone is enough. The feature is recent and still in preview, so read this as how new it is more than as neglect.
  • 48 Not ready for AI

  • 1 Fix before plugging in AI

  • 4 Ready for AI

  • 81Median score 81 / 100

Rule by rule

What the audit found.

Each rule reads declarations only. Rules are sorted by how many models they affect, and a rule that found nothing is left out.
Models affected by each audit rule that found something, out of 53 models
RuleModels affectedFindings
Visible technical columnFC-LIS-05 · minor50 of 532,148 findings
AI instructions missingFC-IA-04 · blocking48 of 5348 findings
Visible measure without a descriptionFC-LIS-01 · important44 of 494,572 findings
Missing or automatic date tableFC-STR-02 · important37 of 53147 findings
Same generic column name in several tablesFC-COH-04 · important34 of 53506 findings
Visible measure without a formatFC-LIS-06 · minor28 of 49496 findings
Visible table with no relationshipFC-STR-05 · minor28 of 53127 findings
Measure and column with the same nameFC-COH-03 · important24 of 49162 findings
Bidirectional relationshipFC-STR-01 · important23 of 53121 findings
Implicit measures allowedFC-IA-01 · blocking22 of 52247 findings
Cryptic name on a visible objectFC-LIS-03 · important20 of 53437 findings
Two measures with the same DAXFC-COH-01 · important13 of 4852 findings
Several filter paths between two tablesFC-STR-03 · blocking13 of 5353 findings
Description that says nothingFC-LIS-02 · important4 of 1984 findings
No figure definedFC-IA-06 · blocking4 of 534 findings
Visible copy or obsolete objectFC-LIS-04 · important3 of 539 findings
Calculated column that aggregatesFC-STR-04 · important3 of 127 findings

Every bar uses the same scale: the full width is all 53 models. The solid bar shows the models a rule affects, the faint bar the models it applies to.

What stands out

Four patterns behind the numbers.

  1. 01

    Almost no model tells Copilot what to do.

    48 of 53 models carry no AI instructions: no business terms, no measure behind each key figure, no default date. They are the piece that moved Copilot's answers when we tested it.

  2. 02

    The meaning is missing.

    4,572 of 6,055 visible measures have no description (76%). 30 of the 49 models with visible measures describe none of them. The 5 that describe them all are demos and portfolio projects. Copilot does not use these descriptions to answer, but an agent reading the model has little else.

  3. 03

    A description is not always an answer.

    Of the 19 models with descriptions to read, 4 carry 84 descriptions that say next to nothing: they repeat the name of the measure, or stop at three words or fewer.

  4. 04

    Duplicates can hide wrong numbers.

    13 models have measures with identical DAX (52 findings). We read about ten cases by hand. Most are two names for one total. Some are wrong numbers: a two-year comparison that returns last year's gap, a margin that returns the profit amount. The formula runs without error, so Power BI has nothing to report.

How to read this

What these numbers can and cannot say.

  • This is not a random sample.

    GitHub code search found 999 models in repositories under a permissive license, forks left out. We picked 53 by hand for variety: sectors, languages, sizes. Some we picked because they are well described, and 4 because they ship AI instructions, which makes both look more common here than they may be. Many are demos, samples or portfolio projects. Company models may look better or worse: these figures describe these 53 models, not the market.

  • What we tested on Copilot.

    These 53 models never went through Copilot: this page shows what they declare, not what Copilot answered. Separately, we asked Copilot 15 questions on one fictional model in four versions, once each, on a paid F2 capacity (November 2026). Right figures out of 15: 11 with no preparation, 10 with descriptions only, 13 with AI instructions, 13 with descriptions, instructions and an AI schema. One model, one pass, and instructions written knowing the questions: a direction, not a rate. It matches Microsoft's documentation (July 2026): Copilot's answers about a model depend on the AI schema, verified answers and AI instructions; descriptions only count for search and DAX queries.

  • The rules are FactCraft's own.

    So are the score and the verdict, and FactCraft sells what comes after the audit. Keep that in mind, and run the audit on your own model to judge the rules for yourself.

  • Metadata only.

    The audit reads declarations: names, types, descriptions, DAX, relationships. It never reads data. It looks for AI instructions in the model's linguistic schema, where Power BI Desktop saves them, and in the project's Copilot folder: instructions kept elsewhere are not seen.

  • No model is named.

    The figures on this page describe the set as a whole. The models were under the MIT or Apache 2.0 license, read as text and never run. Audited on November 8, 2026.

Run the same audit on your model.

It is free and runs in your browser: your model never leaves your machine.