A Universal Translator: Agentic AI as Supply Chain Babel Fish
Agentic workflows offer a way to finally turn disparate partner data into a shared language.
Having lived in the UK during the 1980s, I was lucky to enjoy a weekly TV ritual watching The Hitchhiker’s Guide to the Galaxy every Friday night. Among the many crazy ideas this comedic sci-fi series floated was the Babel Fish, which, when inserted in one’s ear, allowed the listener to understand any language perfectly.
Updating the concept 40 years later, I’m happy to report on the prospect of a modern-day Babel Fish for supply chain: Agentic AI systems for data translation between trading partners.
By automating the often clumsy task of ingesting supplier data, running it against LLMs alongside classical algorithms for interpretation and standardization, and then enriching it with metadata for downstream analytics, agentic systems could act as universal translators between supply chain partners.
You Say Tomayto, I Say Tomahto
Intercompany data transfer is a notoriously messy problem. Different naming conventions, units of measure, abbreviations, synonyms, part numbering schemes, and many other quirks pollute the process of assembling clean, complete data for analyses like cost reductions, lead-time promising, and materials planning. One supplier’s “4.000 kg of aluminium” may be exactly the same as another’s “8,818.49 lbs of aluminum,” but the data is at least 50% different on a character-by-character basis.
Legacy systems choke on this.
Humans use common sense, especially when it’s their job to manage inbound materials, to correct for these kinds of data glitches. Data integration between trading partner systems has traditionally relied on such human intervention for translation. To quote former Amazon Consumer CEO and Auger founder Dave Clark: “For decades, enterprise software operated on a quiet assumption: humans would always be the coordination layer.”
The result is a slow, error-prone process that impedes value chain automation.
Semantic Resolution with Agentic AI
Agentic systems may be a good solution for this problem because the task is repetitive and the underlying logic for harmonizing data lends itself to machine learning technology. Here’s a quick cheat-sheet on common AI/ML techniques for data cleansing:
- Entity Resolution
• Algorithms: Jaro-Winkler distance, Levenshtein distance, Soundex
• ML Models: Siamese networks for learning similarity
• Graph-based: Connected components for transitive relationships- Duplicate Detection
• Blocking strategies: Reduce comparison space (eg, compare only within same category)
• Similarity metrics: Cosine similarity, Jaccard index
• Ensemble approaches: Combine multiple signals (text, numeric, categorical)- Data Validation
• Rule-based systems: Domain-specific constraints
• Anomaly detection: Isolation forests, autoencoders
• Statistical validation: Z-scores, IQR methods- Natural Language Processing
• Named Entity Recognition: Extract product attributes
• Text classification: Category mapping
• Semantic similarity: BERT, Sentence-BERT for understanding meaning- Active Learning
• Human-in-the-loop for ambiguous cases
• Model improves from corrections
• Confidence thresholds determine when to escalate
Early Use Cases Are Promising
Use cases are emerging at companies like Kraft Heinz, which used TealBook’s AI-powered supplier database to automate entity resolution, removing all duplicates and enabling further data enrichment. Another comes from a chemical company that has automated supplier contract data translation to standardize terms, resulting in over $100M saved on contract leakage. A third example shows packaging supplier Veritiv using Pactum AI to manage master data collection and cleansing while autonomously negotiating with long-tail suppliers.
Among Zero100 members, less public uses of agentic AI systems for managing inbound supplier data are beginning to show promise. One CPG company, for example, is training AI agents to understand different invoice formats and data sources, while another is using genAI to identify tariff/supplier-price clauses in sourcing contracts. Meanwhile, an industrial company is using AI to normalize supplier quotes across assorted data formats; another uses an agent to standardize and consolidate supplier data to help surface savings opportunities. And finally, a pharma company is using a system of agents to cleanse inbound supplier data as part of a continuous IBP process.
It’s Still Early, But Not Too Early
A common thread across these use cases, both public and NDA, is the huge amount of human work still needed to start. Human-in-the-loop is critical to training, refining, and using these systems for real work. Plus, any value created by automating data reconciliation depends on people being curious and capable enough to ask smarter questions with their improved data and extra time.
After all, universal translation is only useful when there is a conversation worth having.