AI-Assisted Salesforce Metadata Reconciliation Using TF-IDF Similarity: A Case Study on Multi-Org Enterprise CRM Consolidation

Abstract

Purpose: Large companies often maintain more than one Salesforce organization. Over time, acquisitions, regional changes, and separate release schedules cause those organizations to diverge. The same business concept may appear under different names, while similarly named components may no longer serve the same purpose. This study examines that reconciliation problem. Methods: We built a pipeline that retrieves metadata from two Salesforce organizations, converts the XML into comparable text records, and uses term frequency-inverse document frequency (TF-IDF) with cosine similarity to suggest matches. The pipeline also produces review queues and risk reports. The evaluation covered 4,152 comparison records, mainly from Account, Case, Contact, Lead, and Opportunity metadata, together with a smaller set of non-object artifacts. Results: Of the 4,152 records, 2,943 were classified as high-confidence matches (70.9%), 233 required review (5.6%), and 976 had no suitable match (23.5%). The process also found 240 possible renames and 34 possible drift cases. We manually reviewed a stratified sample of 300 records. On the binary subset, TF-IDF produced 94.5% precision, 95.9% recall, and a 95.2% F1 score, which was higher than the exact-name and Levenshtein baselines. Conclusion: The findings indicate that a relatively simple and explainable text-similarity method can be useful for Salesforce consolidation. It reduces the number of items that require detailed manual inspection and keeps the reasoning behind each candidate match visible to reviewers.