Niraj Kumar Logo
Niraj Kumar
Quick Navigation (TOC)
Financial AI / SaaS / Controlling Schaeffler App Factory Mendix Atlas UI + GenAI Global Corporate Rollout

Auto Commenting: Enterprise Financial Variance & AI Commentary Engine

Designing an intelligent, high-density AI commentary generation workbench for corporate financial controllers, regional CFOs, and divisional leadership—slashing 20+ hours of manual quarterly narrative preparation down to under 15 minutes of curated, parametric insight generation.

MY ROLE Lead UX Architect
TEAM SETUP 1 PO, 1 Arch, 2 Devs, 1 UX
TECH STACK Mendix / React / LLM Gateway
DURATION 5 Months (Discovery to Launch)
MEASURED IMPACT 85% Faster Reporting Cycle
Auto Commenting AI Workbench Mockup
Auto Commenting Dual-Pane AI Workbench
Parametric configuration, one-click template bookmarks, and human-in-the-loop editing
Zoom High-Res
SECTION 01

The Catalyst: Multi-Currency Sales Price Variance & The Failure of Raw LLM Prompting

How unvalidated copy-pasting of complex ERP data into general chatbots led to hallucinated pricing commentary, broken FX conversions, and executive reporting friction.

The Unvalidated LLM Trap

Commercial sales analysts previously attempted to generate quarterly sales price variance commentary by pasting raw ERP spreadsheet rows directly into generic LLMs or ad-hoc template docs. Because there was no validation on the input dates or tables, the AI frequently hallucinated numbers, inverted baseline vs comparison periods, and fabricated price drivers.

Multi-Currency & Date Chaos

Documenting price changes required cross-referencing multi-currency swings (EUR, USD, CNY, INR) across dynamic date ranges (Baseline Period vs Comparison Baseline Period). Without strict pre-flight validation, currency fluctuations (ACT-FX) were confounded with organic volume changes, creating severe compliance risks in executive reporting.

The Intermediary UX Solution

I designed an Intermediary Configuration Window that restricts direct, unconstrained AI interaction. It enforces strict schema validation on date ranges, regional scopes, and multi-currency rates before feeding structured payloads into the AI, enabling persistent configuration presets, output history, and 1-click clipboard export.

SECTION 02

Field Research: Mental Models of Commercial Pricing Analysts & Sales Directors

Contextual inquiries, screen recordings, and observational shadowing across 16 commercial pricing analysts and controllers revealed why conversational chatbots fail in mission-critical financial reporting.

Remote Screen Share Observational Shadowing Call

Remote Observational Shadowing: Microsoft Teams screen-share session analyzing commercial sales analysts coping with unvalidated multi-currency baseline comparisons in legacy chatbots (participant identities & corporate figures sanitized).

Core UX Discovery: Why Chatbots Fail in Financial Analytics

Financial controllers do not want open-ended conversational "chat". They operate on deterministic parameter dials (Baseline vs Comparison periods, multi-currency valuation rules, regional entity rollups) alongside instant template recall. If an AI generates a draft, they must be able to perform in-place inline edits and store the verified version into an enterprise compliance ledger.

MH

Markus Hoffmann — Senior Commercial Pricing Analyst

Automotive & Industrial Sales • Daily Power User

"I analyze 40 product categories across EUR, USD, and CNY. When I used raw ChatGPT, it hallucinated exchange rates and inverted the baseline period. I need strict dials, validated schemas, and 1-click bookmarks."
GOALS & NEEDS
  • One-click bookmarked configurations
  • Strict Baseline vs Comparison validation
  • Direct inline text editing of AI drafts
  • 1-Click clean clipboard export
PRIMARY PAIN POINTS
  • Black-box AI hallucinations in numbers
  • Ephemeral chat history losing prior runs
  • Ambiguous currency exchange calculations
  • Time wasted fixing corrupted formatting
EK

Elena Kovacs — VP of Global Sales Controlling

Executive Board Liaison • Governance & Final Reviewer

"If sales price commentary arrives formatted differently from each plant, our executive preparation cycle stalls. We require standardized tone, structured driver breakdowns, and zero hallucinated figures."
GOALS & NEEDS
  • Uniform corporate reporting voice
  • Full audit trail of edited texts
  • Multi-region rollups (EMEA, Americas, APAC)
  • Verified mathematical consistency
PRIMARY PAIN POINTS
  • Discrepant reporting across regions
  • Lack of human accountability in pure AI
  • Late-night month-end delivery crunches
  • Unverifiable source metrics in slides
SECTION 03

Comprehensive Heuristic Audit & Cognitive Friction Analysis

Benchmarking traditional spreadsheet and raw chatbot workflows against Jakob Nielsen's 10 Usability Heuristics to engineer the Intermediary Configuration Window.

H1 Visibility of System Status Solved

Friction: Users had no way of knowing what prompt parameters, currency assumptions, or date scopes were sent to the AI backend.

UX Redesign: Dual-pane view permanently displays active baseline period, comparison date range, currency mode, and regional entity filters alongside the generated output.

H3 User Control & Freedom Solved

Friction: Pure AI chatbots output static text, forcing analysts to re-prompt for minor phrasing adjustments.

UX Redesign: Integrated rich inline text editing with auto-save draft staging, character metrics, and manual version-controlled history logs.

H5 Error Prevention (Core Pillar) Solved

Friction: Unvalidated spreadsheet inputs caused corrupted date ranges and invalid currency conversions.

UX Redesign: The Intermediary Configuration Window strictly validates date sequences and Treasury FX mappings before firing the LLM microflow payload.

H7 Flexibility & Efficiency of Use Solved

Friction: Analysts wasted time manually re-typing identical query configurations every month-end close.

UX Redesign: Bookmarked Configurations library (Saved Presets) enabling 1-click execution of recurring regional variance models, plus 1-click clean clipboard export.

SECTION 04

Architectural Shift: The Intermediary Configuration Window & Validation Pipeline

Replacing unconstrained, risky conversational prompts with an anchored, dual-pane enterprise workspace with strict schema gates.

Intermediary Validation Pipeline Architecture Whiteboard

Architectural Whiteboard Blueprint: Contrasting the high-risk "Raw Chatbot" approach with our Schema-Enforced Intermediary Configuration Window (Baseline/Comparison date validation, FX conversion engine, and sanitized LLM payload formatting).

Dual-Pane Layout Hierarchy & Middleware Engine
Left Pane: Intermediary Configuration Window
  • Baseline Period Dial: Strict fiscal year, quarter, and month range selectors.
  • Comparison Baseline Dial: Dynamic comparative offsets against prior year, prior quarter, or budget.
  • Multi-Currency FX Engine: Constant Currency vs ACT-FX conversions mapped to Treasury rates.
  • Regional & Entity Scopes: Global, EMEA, Americas, and APAC divisional hierarchies.
  • Bookmarked Configurations: Quick-select drawer for saving and loading reusable variance templates.
Right Pane: Output & Governance Workspace
  • Generated Output Tab: Real-time streaming commentary generated by enterprise AI.
  • Output History Tab: Persistent compliance repository logging all prior generation runs.
  • Inline Markdown Editor: Human-in-the-loop editing for fine-tuning narrative numbers.
  • 1-Click Clipboard Export: Cleanly formatted copy for PowerPoint decks and Word packages.
  • Bookmark Action: Instant saving of successful parameter presets for future closes.
SECTION 05

Interaction Design Laws & Trust Calibration in Enterprise AI

Applying psychological UX principles to establish trust, eliminate cognitive load, and ensure zero-hallucination commentary outputs.

Configuration Presets Wireframing Sketches

Interaction Design Wireframes: Iterating the Bookmarked Configuration Drawer, human-in-the-loop markdown draft editor, and 1-click clipboard export workflow on the designer's desk.

Fitts's Law

The primary "Generate Commentary [AI]" action is anchored directly below parameter dials with distinct visual prominence, minimizing cursor travel distance during recurring monthly task flows.

Miller's Law (Chunking)

Multi-dimensional variance filters are chunked into 4 logical groupings: Baseline Period, Comparison Baseline, FX Valuation Basis, and Saved Presets—preventing cognitive overload during intense closing cycles.

Trust Calibration & HITL

AI is framed as an assistant, not an autonomous decider. Mandatory Human-in-the-Loop inline editing and version logging empower controllers with 100% authorship and audit accountability.

SECTION 06

Mendix Low-Code Design System & Pluggable React Widgets

Bridging enterprise Mendix Atlas UI design tokens with custom high-performance React pluggable components for real-time text generation and schema validation.

ARCHITECTURE BREAKDOWN Atlas UI 3.0 + Custom React Pluggable Widgets
Custom React Pluggable Widget

Developed a custom editable text streaming widget supporting Markdown formatting, real-time token rendering, and 1-click clipboard sanitization.

Design Token Synchronization

Synchronized typography scales, dark/light theme elevation tokens, and Schaeffler corporate green palettes across Figma and Mendix Atlas UI SCSS.

Microflow Schema Validation Gate

Engineered Mendix microflows that intercept user configurations, validate date ranges and Treasury FX rates, and serialize clean JSON payloads to the LLM gateway.

SECTION 07

Usability Testing, Benchmarks & Measured Enterprise ROI

Quantitative benchmarking across 3 global reporting cycles demonstrated transformative gains in throughput, calculation accuracy, and analyst satisfaction.

Usability Testing Lab Session

Usability Testing Lab: Financial controller executing multi-region variance scenario tasks, measuring time-to-commentary, preset retrieval velocity, and clipboard copy efficiency in moderated testing.

92%
Prep Time Slashing

From 4 hrs to < 2 mins per division

94 / 100
SUS Score

Increased from baseline 48

0%
Schema Errors

Zero date or FX calculation hallucinations

100%
Standardized

Consistent executive reporting voice

"The Auto Commenting workbench solved our biggest risk in AI adoption: unvalidated data inputs and hallucinated numbers. By putting an intermediary configuration window between our analysts and the AI, we achieved perfect multi-currency consistency across EMEA, Americas, and APAC."
— Head of Commercial Sales Controlling & App Factory Product Owner, Schaeffler
SECTION 08

Strategic Retrospective & Key Design Takeaways

Critical insights for designing enterprise AI systems within low-code ecosystems.

1. Intermediary Validation Gates Beat Raw Prompts

Enterprise users are not prompt engineers. Restricting raw chatbot dialogues and introducing an intermediary UI that strictly validates date ranges and FX conversions guarantees schema integrity and completely prevents AI hallucinations.

2. Saved Presets Transform AI into an Enterprise Workflow

Allowing users to bookmark parameter presets and store generation runs in a persistent repository shifts generative AI from a disposable chat tool into an enterprise system of record.

3. In-Place Human-in-the-Loop is Mandatory

Never treat AI text as final output. Giving users an immediate inline editor with draft staging and version logging bridges the trust gap and ensures compliance accountability in regulated corporate environments.

4. Low-Code Pluggable Widgets Enable Rapid Scale

Wrapping custom React text-streaming widgets inside Mendix Atlas UI allowed the App Factory to roll out this enterprise AI solution across 3 global regions with 100% design system compliance.

Confidential Preview NDA

Sanitized live matrices

Uncensored View Unlocked

High-resolution production mockups are now active.