The companion article — Why Most AI in Finance Fails — explains why data models designed for dashboards break when AI consumes them. This one answers the next question: what does a controller’s data model actually need to look like?
Six things. They apply regardless of BI tool, cloud platform, or ERP. They define what “AI-ready” means for controlling infrastructure.
1) Single Computation Path per Metric
Every metric — revenue, gross margin , EBITDA, working capital — needs exactly one computation path. Not one formula in the ERP, another in the dashboard, and a third in last quarter’s Excel.
Human-only model: Gross margin exists in three versions. ERP uses standard costs. Dashboard uses actual costs with manual adjustments. Controller’s Excel applies a third methodology. Humans know which to use when. The numbers “work” because a person navigates between them.
Dual-consumer model: One formula in the semantic layer . Allocation methodology defined once. Dashboard, AI, any future consumer — same result. Contribution margin after allocation means one thing, everywhere.
2) Governed Hierarchies with Explicit Roll-Up Rules
Cost centre → department → business unit → group. Defined in the model, not in dashboard drill-down configurations. The hierarchy includes rules: which entities consolidate, which eliminate against each other, at which level management reporting aggregates.
Human-only model: Roll-up logic lives inside Power BI. The Excel budget uses a different grouping. Controller applies grouping rules from memory. AI queries flat tables with no hierarchy awareness.
Dual-consumer model: One hierarchy definition. Every consumer rolls up and drills down identically. AI asked “What is EBITDA for the manufacturing division?” follows the same path a controller would use.
For multi-entity groups, this intersects directly with consolidation. Five subsidiaries, three countries — the model defines which entities consolidate, which intercompany relationships exist, at which level the group reports. Connects to Principle 4.
3) Explicit Variance Decomposition
Price/volume/mix decomposition encoded as model logic. Not calculated ad hoc in a spreadsheet each cycle.
Human-only model: Controller calculates price/volume/mix manually in Excel each month. Dashboard shows total variance only — actual vs. plan, no decomposition. AI asked “Why did margin drop?” can’t decompose. Produces either “margin decreased by X%” or a hallucinated explanation from raw correlations.
Dual-consumer model: Variance components (price, volume, mix, FX, cost) are pre-defined dimensions. Any consumer queries any component, any period, any entity, any product line.
This most directly addresses the question controllers actually ask: “What drove the change?” With decomposition logic in the model, AI answers with the same rigour a controller applies manually — but across every product, every region, every period simultaneously.
4) Intercompany Elimination Rules as Model Logic
Intercompany transactions and their elimination at consolidation — defined in the model. Not in a spreadsheet the group controller maintains during close.
Human-only model: Eliminations happen in an Excel workbook. Controller identifies matching transactions, applies entries, produces a consolidated view. AI asked “What is group revenue?” sums all entities without eliminating intercompany sales. Overstates by the full value of internal transactions.
Dual-consumer model: Entity-pair relationships and elimination rules in the model. Consolidated and unconsolidated views as model dimensions. AI asked “What is group EBITDA?” applies eliminations automatically.
For a group with subsidiaries across Slovakia, Czech Republic, and Poland — each on a different ERP, with intercompany sales, management fees, shared service charges — the elimination logic is the group controller’s expertise. Making it model logic rather than spreadsheet logic is what enables AI-powered consolidation.
5) Temporal Consistency — Close Cycle States
Every data point carries temporal metadata: open, provisional, or final. Per entity, per period.
Human-only model: Controller knows March is closed but April is provisional — three subsidiaries closed, two haven’t submitted. This knowledge lives in the controller’s head or a status email. Dashboard shows both months identically. AI treats April’s provisional numbers as final.
Dual-consumer model: Status flag in the model. AI asked “What was April revenue?” responds: “April is provisional — 3 entities closed, 2 still open. Revenue from closed entities is €X.”
Prevents the most common AI failure mode in management reporting : presenting incomplete data as complete.
6) Semantic Definitions as Model Metadata
What “revenue” means, how “EBITDA” is calculated, what “FTE” includes — attached to the model as metadata. Not in a data dictionary PDF. Not on a Confluence page. Not in a governance policy approved two years ago.
Human-only model: Revenue definition lives in a finance manual. Different people interpret it slightly differently. New controller learns through trial and error. AI has no access to any definition and guesses from column names.
Dual-consumer model: Every metric carries its definition, formula, source lineage, and owner as metadata. AI asked “How is gross margin calculated?” answers from the model itself.
Closes the loop. Principles 1–5 ensure the model computes correctly. Principle 6 ensures every consumer can understand what the model computes and why.
What This Changes
These principles don’t just enable AI. They improve how humans work with the same data.
Board reporting becomes deterministic. Same number in the dashboard, management pack, and AI summary. No reconciliation.
Close cycles gain clarity. Everyone sees the same close status. No status emails.
Variance analysis scales. Price/volume/mix for every product line, every region — work that previously took days of spreadsheets per cycle.
AI produces trusted answers. Not smarter AI. Better foundation. The shadow process disappears.
Roland Berger: 25% reduction in monthly close time from AI — where the data foundation supported it natively. GoodData: 50–80% reduction in semantic complexity through governed semantic layers. Not better AI. Better infrastructure.
Where to Start
Not a year-long project. Design decisions applied incrementally.
Start with 1 — audit which metrics have multiple computation paths. If gross margin is calculated differently in the ERP, the BI tool, and a spreadsheet, that’s the first fix.
Then 6 — attach definitions to the model. Even basic metadata (formula, owner, last validated date) transforms how both humans and AI interact with the data.
2–5 follow as complexity requires. Multi-entity groups need hierarchy governance and intercompany elimination early. Single-entity companies may not need them at all.
The test: can the organisation produce consistent answers without heroic human context? If yes, the model is dual-consumer. If not, these six tell you where the gaps are.
Sources: AtScale — Why AI Redefined the Semantic Layer , GoodData — How to Modernize Your BI for the AI Era , KPMG — Rebuilding Data Governance in the Age of AI , Roland Berger — Mastering AI in the Finance Function