About the Client
The client is a Texas-based private investment and property management firm with a diversified portfolio spanning the United States, Canada, and Europe. Focused on delivering consistent, long-term value for investors, the organization manages financial operations at scale across geographically distributed properties, accounting systems, and business entities.
Centralized financial oversight plays a critical role in maintaining reporting accuracy, supporting audits, and ensuring consistency across the portfolio.
Background
Every month, the client’s finance team processes large volumes of financial data, including thousands of General Ledger (GL) entries generated across multiple properties.
These records form the foundation of financial reporting, audits, compliance, and management decision-making. As the portfolio expanded, however, reviewing GL activity became increasingly resource-intensive.
Monthly reviews relied heavily on manual validation. Finance professionals had to examine thousands of entries, identify potential miscoding, compare posting patterns, and determine whether transactions had been assigned to the appropriate accounts.
While essential for financial control, the process was repetitive, time-consuming, and susceptible to oversight—particularly when entries contained subtle contextual differences across properties, accounting periods, and transaction descriptions.
The client needed a more intelligent approach that could accelerate GL validation without compromising governance or removing financial decision-making from the finance team.
The Challenge
Thousands of GL entries flow into the client’s financial environment each month from systems such as Yardi. Some transactions can be assigned to incorrect accounts or categories due to several factors.
Ambiguous Descriptions: Transaction descriptions may lack enough detail to immediately determine the appropriate account classification.
Inconsistent Data: Similar transactions may be described differently across properties, teams, or accounting periods.
Human Oversight: Manual data entry and classification can introduce miscoding, particularly when processing transactions at scale.
Limited Historical Context: Reviewers may not always have immediate access to historical posting patterns that could help validate current entries.
Manually identifying and correcting these issues consumed significant finance team capacity. At the same time, increasing the volume of manual review introduced the possibility of additional human error.
The client needed a scalable, AI-powered review mechanism that could analyze GL entries using historical context, surface potential anomalies, and provide actionable recommendations—while ensuring that final decisions remained with finance professionals.
The Solution
The client partnered with Supply Medium to design and implement an LLM-driven AI-powered General Ledger Review System combining intelligent automation, explainable recommendations, continuous feedback, and human oversight.
The solution was designed to complement existing financial controls rather than replace them.
Secure Financial Data Ingestion
Finance teams upload monthly GL files through a Streamlit-based web application protected by Microsoft Entra ID authentication.
Before processing begins, uploaded files are validated for structure and completeness to help prevent incomplete or malformed datasets from entering the analysis workflow.
Validated data is ingested into PostgreSQL, enriched with vector embeddings for contextual analysis, and securely backed up in Amazon S3 to support traceability, resilience, and historical reference.
LLM-Powered GL Analysis
Before AI analysis, GL entries are normalized and deduplicated to improve consistency and reduce unnecessary processing.
Vector embeddings preserve contextual relationships across transaction descriptions, historical records, properties, and accounting periods.
Using AWS Bedrock, the AI engine analyzes current GL activity against historical posting patterns with large language models such as Cohere, Claude, and LLaMA 2.
For potentially miscoded entries, the system can:
- Flag transactions requiring additional review
- Recommend more appropriate GL accounts or categories
- Assign confidence scores to recommendations
- Provide contextual explanations supporting each suggestion
Rather than presenting finance teams with unexplained AI outputs, the system provides review-ready recommendations that help users understand why an entry has been flagged.
Human-in-the-Loop Review & Feedback
Following analysis, finance teams receive downloadable Excel or CSV reports containing flagged transactions, recommended classifications, confidence scores, and supporting explanations.
Reviewers can evaluate each recommendation and mark it as correct or incorrect based on their financial expertise and business context.
The annotated file can then be uploaded back into the platform, feeding validated outcomes into the AI workflow and helping refine future recommendations based on real-world finance team feedback.
This creates a continuous improvement cycle in which AI performance evolves alongside the organization’s actual accounting practices.
Finance Retains Final Decision Authority
A key design principle was maintaining clear human accountability.
The AI system does not automatically modify entries within Yardi or any other source accounting platform.
All final approval, correction, and posting decisions remain with the client’s finance team. AI serves as an intelligent review layer—supporting financial professionals with faster analysis while preserving existing governance, compliance, and control requirements.
The Outcome
The AI-powered GL Review System transformed a highly manual validation process into a more efficient, scalable, and intelligence-driven workflow.
Faster Monthly Reviews: AI-assisted analysis significantly reduced the manual effort required to review large volumes of GL transactions.
Earlier Error Detection: Potentially misclassified entries could be identified earlier in the reporting cycle, reducing the need for downstream corrections.
More Consistent Validation: Historical patterns, contextual analysis, and standardized recommendations provided finance teams with additional intelligence when reviewing complex or ambiguous transactions.
Scalable Architecture: The cloud-native solution provides the flexibility to support increasing transaction volumes as the client’s multi-property portfolio continues to grow.
Continuous Improvement: Finance team feedback feeds into the review workflow, allowing recommendations to become increasingly aligned with real-world accounting patterns and organizational requirements.
Human-Controlled Decision-Making: Finance professionals retain complete authority over final classifications and posting decisions, ensuring AI enhances rather than replaces financial judgment.