A controller asks the AI copilot: “What drove the gross margin drop in Q2?”
The AI answers. The number doesn’t match the dashboard. Doesn’t match the Excel analysis from last week either. Twenty minutes of reconciliation later, the controller uses the Excel number — the one a person built and checked manually.
The AI didn’t fail. It queried the data model and computed an answer. The data model was never designed for a machine to consume. It was designed for humans who read dashboards. Humans fill gaps that machines cannot.
Most AI in finance underdelivers for this reason. Not wrong tools. Wrong foundation.
The Two-Consumer Problem
Until recently, every reporting consumer was human. Dashboards, management packs, board presentations — all built for people who bring context and institutional knowledge to every number.
Humans navigate inconsistent data well. The ERP shows revenue as €10.2M. The BI dashboard shows €10.4M. A controller knows why: one includes returns processing, the other doesn’t. Mental adjustment, footnote, move on. The human does interpretation work that no system captures.
AI agents cannot do this. An AI copilot asked “What was Q1 revenue?” doesn’t know which definition to apply. It queries whatever tables it finds. Sometimes €10.2M. Sometimes €10.4M. Sometimes €9.8M from a different calculation path entirely. Technically correct. Operationally useless.
This is the dual-consumer data model problem. Most data models were designed for one consumer. Now they serve two. The model hasn’t changed. The number of consumers has.
What Happens When AI Meets a Single-Consumer Model
The failure doesn’t look like a crash. It looks like erosion.
First time: someone checks and explains the discrepancy. Second time: “we should double-check that.” Fifth time: the team has quietly built a shadow process — manual validation running alongside the AI — that defeats the entire investment.
The numbers confirm it. Forrester: only 15% of AI decision-makers reported an EBITDA lift. Fewer than one-third could tie AI value to P&L changes. MIT research (via AtScale): roughly 95% of generative AI pilots show no measurable P&L impact. Gartner and Deloitte: more than 40% of agentic AI projects will be abandoned by 2027.
Same pattern everywhere. The technology works. The infrastructure doesn’t support it.
Where the Model Breaks: A Controlling Example
Gross margin in a mid-market company with two ERPs and a BI tool. Not a simple metric — it involves cost allocation, returns processing, promotional discounts, shipping adjustments. Three versions exist:
The ERP calculates it using standard costs. Consistent within the system. Doesn’t reflect actual costs for the period.
The BI dashboard uses actual costs with manual adjustments. A controller built the logic, applied corrections for promotions, validated against management accounts. Accurate — for anyone who understands the adjustments.
The AI agent queries whatever tables it can find. Doesn’t know about manual adjustments. Doesn’t know which cost allocation to apply. Finds revenue and cost columns, divides, produces a number. Defensible given the data. Useless for a decision.
Three versions. A controller reconciles them because the institutional knowledge lives in that controller’s head. An AI agent cannot access knowledge that isn’t in the data.
Governance as Policy vs. Governance as Infrastructure
Most organisations have data governance policies. Somewhere in a shared drive sits a document defining what “revenue” means, how “EBITDA” is calculated, which entities consolidate at which level. AI cannot read that document. It cannot open a PDF, interpret a definition, and apply the correct formula.
Governance as documentation works for humans. AI needs governance as infrastructure: metric definitions encoded as computation logic, hierarchy rules as model relationships, allocation methods as transformation steps. Policy — humans enforce it. Infrastructure — the system enforces it.
KPMG: 62% of organisations cite insufficient governance as the top barrier to scaling AI. Not technology. Not budget. Governance. Building better AI on top of policy-based governance is adding speed to a car with no steering.
The Semantic Layer Gap
Between raw data and consumers sits — or doesn’t sit — a semantic layer . It defines business metrics, hierarchies, and relationships once. “Revenue” means one thing. “Gross margin” applies one cost allocation. “Cost centre hierarchy” rolls up one way. Dashboard, spreadsheet, AI agent — same definitions, same numbers.
Without it, every consumer interprets raw data independently. The BI developer builds logic into the dashboard. The analyst rebuilds it in Excel. The AI agent guesses from column names. Three consumers, three interpretations, three numbers.
GoodData: “Traditional BI platforms trap business definitions inside dashboards designed for human consumption, not for AI agents.” The logic is locked inside tools only humans can access.
Gartner now classifies semantic layers as “critical infrastructure.” The Open Semantic Interchange (OSI) specification was finalised in January 2026. No longer a convenience. A requirement.
The Mid-Market Compounding Factor
Enterprise companies throw data engineering teams at this. Mid-market companies — €5–60M in revenue — don’t have that option.
Most didn’t design their data architecture. They inherited it. First ERP chosen at 15 people. Second added with an acquisition. Third runs the warehouse. The chart of accounts mostly lines up. Cost centres look consistent until you roll them up across entities.
This is where AI should help most — and where it fails hardest. The inconsistencies a controller navigates daily through expertise and spreadsheets are invisible walls for AI. It doesn’t know entity A uses different cost centre numbering than entity B. It doesn’t know Q4 revenue in the Czech subsidiary is provisional. It computes answers from whatever it finds and presents them with false confidence.
Same dual-consumer problem as enterprise. Fewer resources to solve it.
What Doesn’t Fix It
A better AI tool. Smarter model, same broken foundation, more confidently wrong answers.
A data quality layer on top. Validation catches dirty records. It doesn’t fix three different gross margin calculations in three different tools. Missing semantic definitions, not dirty data.
Connecting AI to the same dashboards. Helps — until the AI queries the data differently than the dashboard’s pre-built model. The definitions live inside the BI tool, not in a layer the AI can read.
Cleaning the data first, then adding AI. Necessary but insufficient. Clean data with inconsistent definitions still produces inconsistent answers. The problem is semantic quality, not data quality.
What Does Fix It
A governed semantic layer between raw data and all consumers. Defines business logic once.
One computation path per metric. One roll-up rule per hierarchy. One KPI definition encoded as model logic, not documented in a PDF. Dashboard reads it. AI reads it. Spreadsheet reads it. Same numbers. Every time.
The test for AI-readiness is simple: can the organisation produce consistent answers without heroic human context? If a controller needs to reconcile, validate, or interpret before anyone trusts the number — the foundation isn’t ready.
Roland Berger: 74% of CFOs plan to invest in AI for controlling. Same research: 25% reduction in monthly close time from AI — but only where the data foundation supported AI consumption natively. Intent is clear. Infrastructure is mostly absent.
The Path Forward
If AI answers require manual validation, the data model serves one consumer and is being asked to serve two.
The fix is embedding business logic in the data model layer — not in dashboards, spreadsheets, or governance documents — and making it available to every consumer identically.
Not a technology problem. A design problem. The companion article — Designing a Data Model That Works for Both Humans and AI — covers the six things your data model needs to serve both.
Sources: Forrester Predictions 2026 , KPMG — Rebuilding Data Governance in the Age of AI , Roland Berger — Mastering AI in the Finance Function , Deloitte — CFO Guide to Tech Trends 2026 , AtScale — Why AI Redefined the Semantic Layer , GoodData — How to Modernize Your BI for the AI Era
- Glossary: Dual-Consumer Data Model | Semantic Layer | Data Governance | Data Consistency
- Data Governance & AI Readiness | The Financial Data Governance Framework | Chart of Accounts Architecture
- Designing a Data Model That Works for Both Humans and AI — companion article with the six design principles