An AI-Drafted Variance Narrative Misattributes the Driver Behind a Margin Decline
Executive Summary
Illustrative Scenario
This case study is a composite, educational scenario built from patterns commonly observed in AI-assisted FP&A reporting workflows. It does not describe a specific, identifiable company or team, and any resemblance to a particular organisation is coincidental.
Background¶
An FP&A team at a mid-sized manufacturing company had recently adopted a generative AI tool to draft first-pass variance commentary for its monthly management report, aiming to reduce the time spent on routine narrative drafting each reporting cycle.
The Problem¶
For a given month, the team's summary income statement showed a gross margin decline versus budget. The generative AI tool, prompted with the summary income statement but not the underlying product-mix detail, drafted commentary attributing the decline to a pricing change referenced elsewhere in the company's recent communications. The commentary read fluently and plausibly, and was published in the monthly board report without being tied back to the underlying general ledger detail.
Findings¶
A board member questioned the pricing change referenced in the report, since it did not match their own understanding of recent pricing decisions. This prompted the FP&A team to trace the commentary back to source, at which point the actual underlying driver, a shift in sales mix toward a lower-margin product line that month, became clear from the general ledger detail the AI had not been given or had not incorporated into its drafted explanation.
Root Cause¶
The AI-drafted commentary was accepted and published without being checked against the actual underlying numbers, the number tie-out step addressed in AI for Financial Analysts and Generative AI in Financial Modelling. The tool produced a plausible-sounding explanation consistent with general patterns rather than the specific driver actually visible in that month's detailed data, a hallucination-adjacent failure mode addressed in full in AI Hallucination Risk.
Risk¶
The published management report, already circulated to the board, attributed a real margin decline to an incorrect cause. Had the error gone uncorrected, it risked prompting a pricing-focused management response to a problem that was actually driven by product mix, a materially different, and potentially ineffective, corrective action.
Resolution¶
The FP&A team corrected the board report with an addendum identifying the actual product-mix driver, and introduced a mandatory number tie-out step for all AI-drafted variance commentary before publication, requiring the underlying general ledger detail supporting any stated driver to be confirmed before the commentary is finalised.
Lessons Learned¶
- AI-drafted variance commentary should always be tied back to the actual underlying numbers before publication, since fluent, plausible-sounding narrative carries no inherent guarantee of accuracy.
- Supplying a generative AI tool with summary-level data alone, without the underlying detail needed to identify the actual driver, increases the risk of a plausible-sounding but incorrect explanation.
- The number tie-out discipline described in AI for Financial Analysts exists specifically to catch this class of error before it reaches a board-level report.
- The AI-Assisted Modelling & Analysis Checklist's number tie-out check exists specifically to surface this class of misattribution before publication.
Continue Reading¶
Related Pillars¶
Related Technical Guides¶
Related Checklists¶
How OXXON tests thisRun a free structural check with FMAE
Frequently Asked Questions
Is this a real company's reporting incident?
No. This is an illustrative, composite scenario built from patterns commonly observed in AI-assisted FP&A reporting workflows. It does not describe a specific, identifiable company or team.
What did the AI-drafted variance commentary get wrong?
It attributed a month's gross margin decline to a stated pricing change, when the actual underlying driver, confirmed against the general ledger, was a shift in product mix toward a lower-margin product line that month, an entirely different cause.
Why did the AI draft the wrong driver?
The generative AI tool drafting the commentary was prompted with the summary income statement but not the underlying product-mix detail, and produced a plausible-sounding explanation, a pricing change, consistent with a general pattern in its training data rather than the specific cause visible in this month's underlying data.
How was the error identified?
A board member questioned the pricing change referenced in the report, prompting the FP&A team to trace the commentary back to source, at which point the actual product-mix driver became clear from the underlying general ledger detail.
What was the effect of the error before it was caught?
The management report, already circulated to the board, attributed a real margin decline to an incorrect cause, risking a pricing-focused response to a problem that was actually driven by product mix, a materially different corrective action.
What should the FP&A team have done differently?
Tied the AI-drafted commentary back to the actual underlying general ledger detail before publication, following the number tie-out discipline set out for AI-drafted commentary across this domain, rather than publishing the draft on the basis that it read plausibly.
Related Articles
Generative AI in Financial Modelling
Generative AI, large language models applied to drafting and language tasks, has a specific and bounded role in financial modelling: accelerating structure, formatting, and narrative drafting, not producing verified numerical output. This guide sets out that role in detail, the specific failure modes generative AI introduces into a modelling workflow, hallucinated figures, plausible-but-incorrect formula logic, and unverifiable citations, and the concrete review practices that contain each failure mode.
AI Hallucination Risk
Hallucination, a generative AI model producing plausible-sounding but fabricated content, is the single most consequential risk in applying generative AI to finance. This guide explains why hallucination occurs as a structural property of how language models generate text, the specific finance contexts where it carries the most consequence, citations, figures, and factual claims feeding a material decision, and the layered controls, source grounding, verification checkpoints, and ongoing output monitoring, that manage the risk in practice.
AI for FP&A
FP&A is one of the finance functions where AI adoption has progressed furthest, spanning budgeting, forecasting, variance analysis, and management reporting. This guide maps AI application across the FP&A cycle, distinguishes tasks where machine learning-driven prediction is well matched from tasks better suited to generative AI drafting, and sets the frame for the more specific budgeting, forecasting, and analysis guides that follow it in this domain.
AI-Assisted Modelling & Analysis Checklist
This checklist covers the verification checks specific to AI-assisted financial modelling and analysis, on top of the general financial model audit baseline. It focuses on confirming the AI technique used was matched to the task, that AI-drafted formulas, figures, and citations have been independently verified, and that AI-generated analysis is being presented as a decision input rather than the decision itself. It is intended for financial modellers, FP&A teams, and reviewers checking AI-assisted work before it supports a material decision.