Semantic models · Power BI · AI
Your agents and Copilot are only as right as the numbers you feed them. FactCraft audits your Power BI models, has your key measures validated by the business, and tells you the moment a validated calculation changes. Metadata only, never your data.
Or audit your model now, free and in your browser: nothing leaves your machine.
Before
FactCraft Audit
After · fixes validated by the business
Validated
Validated
Validated
Validated
Validated
Renamed, described
Removed
The problem
3 versions
Finance, sales and controlling each have their own revenue. The agent picks one, and nobody knows which.
CD_STAT_X
Cryptic columns, measures with no description, technical columns left visible. A human guesses. An AI makes things up.
Column or measure?
A calculated column where a measure belonged, a fact table at the wrong grain: inflated totals, and not a single error message.
Before and after
Off
The agent reads the names and descriptions of the measures, picks one or writes its own calculation, and answers with confidence.
On
The audit pins its findings on the measures that misled it: a description that says nothing, a copy, a stale validation.
Fixed
The business validates the fixes. The agent reads the model again, and the number changes.
Question to the agent
What was our revenue this year?
What the agent reads in the model
Picked
After FactCraft
€4.87M off by 13%
€4.31M
[Revenue] now says what it is, the copy of [Net Sales] is gone: the agent picks [Net Sales], validated by the Finance department.
Example amounts, fictional. After FactCraft, the agent answers €4.31M, instead of €4.87M.
How it works
01 · The model
4 tables, 7 measures, 3 relationships. Copilot and agents choose the measure that answers a question from these declarations: names, descriptions, DAX, relationships.
02 · Validated
On 5 Oct 2026, the Finance department rereads Net Sales, Gross Margin and Margin %. Each validation is written into the model, with a fingerprint of the calculation.
03 · A change
Someone adds shipping costs to Gross Margin. Its fingerprint no longer matches, and Margin %, which calls it, falls with it. Net Sales still holds.
04 · The audit
15 findings, a score of 77 out of 100. Not ready for AI. Hover or select a flagged object to read what is wrong and the proposed fix.
Swipe the diagram sideways to see the whole model.
[Turnover] in Sales
blockingFC-LIS-01Visible measure without a description
Proposed fix, to validate with the business: Add a /// description above the measure: what it computes, in which unit, under which rule.
importantFC-COH-01Two measures with the same DAX
Proposed fix, to validate with the business: First check with the business that the two names mean the same figure. If they do, keep a single calculation: replace the DAX of “Turnover” with [Net Sales], or delete “Turnover” if nothing uses it; hiding it is not enough, the measures that reference it still read it. If they do not, one of the two calculations is wrong: correct it.
Validated facts
The owner reads the definition: what the measure computes, in which unit, under which rule. The validation is written into the model itself: who, when, and a fingerprint of the calculation.
It covers the measure's name, description, format and DAX, and the calculation of every measure it calls. Reformatting the DAX leaves it intact; changing the calculation does not.
The audit reports the stale validation, and a check in your CI pipeline stops before an AI quotes a number nobody has reread.
In the model, not beside it
annotation FactCraft_Validation = {"by":"Finance department","on":"2026-10-05","fingerprint":"sha256:…"}
A TMDL annotation on the measure: it travels with your .pbip project, reads in a git diff, and Power BI keeps it without interpreting it.
/// Gross margin in euros: net sales minus the purchase cost of the products sold.measure 'Gross Margin' = [Net Sales] - SUM ( Sales[Purchase Cost] ) formatString: #,0 € annotation FactCraft_Validation = {"by":"Finance department","on":"2026-10-05",…}Validated fact
by the Finance department, 5 Oct 2026 · fingerprint sha256:da9b…3c53
The session
Measure 1 of 6
Is this what your team means by “Gross Margin”?
“Gross margin in euros: net sales minus the purchase cost of the products sold.”
Its definition or its calculation changed after Finance department validated it on 5 October 2026: reread it before renewing.
It uses
Net Salesvalidated
The measure
Gross Marginchanged since validated
Skip for now
[Net Sales] - SUM(Sales[Purchase Cost]) - SUM(Sales[Shipping Cost])
What we do
Free, in your browser
FactCraft reads your semantic model (.pbip, TMDL) and scores it out of a hundred: readability, consistency, structure, AI readiness. Every finding comes with a proposed fix.
Subscription · coming soon
Your key measures, validated by the business and fingerprinted in the model. The moment a validated calculation changes, the audit and your CI pipeline flag it.
Join the waitlistNext
An MCP server on top of the semantic model: the agent answers through your validated measures, never by inventing its own calculation.
Engagement
Guided modeling, validation workshops with the business, and the whole chain made usable by AI, from Databricks to the semantic model, inside your environment.
Where it sits
Model authors
Power BI Desktop, Tabular Editor
The business
Validates the key measures
Your semantic model
TMDL files, in git
FactCraft check
Audit and validated facts, in your CI
Coming soon
Published model
Power BI service, Fabric
Copilot
In Power BI
Data agents
In Fabric
Your agents
Querying the model
The audit
Sales.SemanticModel
4 tables · 7 measures · 3 relationships
77 / 100
Not ready for AI
Readability81
Fail · Visible measure without a description, 3 of 7
Consistency95
Fail · Two measures with the same DAX, 1 of 7
Structure86
Fail · Bidirectional relationship, 1 of 3
AI readiness56
Fail · Implicit measures allowed, 2 of 8
Validated facts
Validated facts: 1 of 7 visible measures
First fix · FC-LIS-01 · to validate
[Turnover] in Sales: Add a /// description above the measure: what it computes, in which unit, under which rule.
Where to calculate?
| Business need | Where | Why |
|---|---|---|
| Margin per customer, filterable | Measure | It has to respond to the report's filters. |
| Age band for segmentation | Power Query | A fixed value per row, compressed better than a DAX column. |
| Conversion at the daily exchange rate | Source | A rule shared by every tool, not just Power BI. |
| Customer status shown in a slicer | Column | Rare: only when it depends on another table in the model. |
Security
Table, column and measure names, descriptions, DAX and relationships. Never a single row of data.
The free audit is computed in the page, which sends nothing. Load it, cut your network: it still works.
On engagements, we work with your access, in your Fabric or Databricks, or on a copy of the model under NDA, deleted afterwards.
What the audit opens
Its TMDL files and, if the project has them, its AI instructions for Copilot: nothing else. The data cache that Power BI Desktop keeps beside them is never opened by the audit, in your browser or on the command line.
Who's behind FactCraft
6 years
of engagements at a large European energy group, across several departments.
~50M rows
in SSAS Tabular semantic models, with advanced DAX.
.pbip
Creator of ChartCraft, which generates Power BI projects.
Source: the founder's own engagements.
From facts to decisions
FactCraft
AI, built on facts.
Upstream: data, semantic models, validated measures, agents.
ChartCraft
Decisions, by design.
Downstream: Power BI reports designed around decisions, delivered as .pbip.
Discover ChartCraftRun the free audit in your browser, then send us its JSON report: object names, never your data. We review it for free and come back with the priority fixes and a plan to validate your key measures.
A full audit, run on your model's definition shared under NDA, hands back an audit dossier like this sample, built on a fictional model.
contact@factcraft.io