Agent Concepts & Structure
An Agent is a specialized machine learning model designed to automate spend classification. It functions by learning the relationship between raw transaction strings (descriptions, vendors, GL codes) and your organizational categories.
🏗️ The Anatomy of an Agent
To function, an Agent requires two foundational dependencies at the moment of creation. These define what the Agent "knows" and how it "labels" data.
- Training Dataset: A segment of clean, vendor-normalized spend data used to "teach" the Agent. The model analyzes these verified examples to establish its internal classification rules.
- Target Taxonomy: The master classification hierarchy (e.g.,
Indirect->IT->Software) that serves as the "answer key." The Agent will only predict categories that exist within this linked taxonomy.
📊 Classification Output & Accuracy
Once an Agent is trained and reaches the Ready to Classify status, it provides three layers of validation:
| Metric | Definition | Impact |
|---|---|---|
| Overall Confidence | The aggregate accuracy of the Agent across a dataset (e.g., 99.83%). | Indicates the general reliability of the model. |
| Transaction Confidence | A specific Low, Medium, or High score assigned to every individual line. | Prioritizes which records require manual human audit. |
| Status Tag | A system indicator (e.g., Training, Ready, Classifying). | Communicates the current state of the AI's processing lifecycle. |
📈 Strategic Business Value
1. Speed and Scale
Agents reduce the time required to categorize millions of transactions from weeks to minutes. This allows procurement teams to move immediately from data ingestion to Strategic Sourcing without manual intervention.
2. Continuous Learning Loop
The Agent system utilizes an iterative feedback loop. When an auditor corrects a Low confidence prediction, that correction is automatically ingested by the model.
3. Data Governance
Because an Agent is strictly tethered to a specific Taxonomy, it acts as a digital enforcer of your data standards. It ensures that "Cloud Storage" is classified identically across every business unit, regardless of the source system.
Key Definitions
- Ready to Classify: The status confirming the model has finished its learning phase and is ready for production.
- Auditing: The process of a human reviewer verifying Agent predictions to improve future accuracy.