Version control (AI)

The systematic tracking and management of changes to AI models, training data, code, configurations, and documentation throughout the model lifecycle, enabling reproducibility, auditability, and rollback capability.

Source: AIEOG AI Lexicon (Feb 2026), adapted from NIST AI 100-1 and SR 11-7

Version control for AI is the practice of keeping a detailed record of every change made to an AI system — including the model itself, the data it was trained on, the code that runs it, and the configuration settings that shape its behavior. It allows teams to see exactly what changed, when, by whom, and why, and to revert to a previous version if something goes wrong.

In traditional software development, version control (tools like Git) tracks changes to code. AI version control extends this concept to encompass everything that makes a model work: datasets, model weights, hyperparameters, feature engineering pipelines, evaluation results, and deployment configurations.

This matters because an AI model’s behavior is defined not just by its code but by the specific combination of data, training process, and configuration that produced it. Changing any one of these elements can alter model behavior in significant and sometimes unexpected ways.

Version control is a regulatory expectation and an operational necessity in financial services:

• Auditability and examination readiness. Examiners expect institutions to demonstrate exactly which version of a model is in production, what changes have been made since the last validation, and whether those changes were properly approved. SR 11-7 and OCC 2011-12 require documentation of material model changes and their rationale.
• Reproducibility. Regulators and internal validators need to reproduce model results. Without version control, reproducing the exact conditions that produced a particular output — which version of the data, which model weights, which configuration — becomes nearly impossible.
• Rollback capability. When a model update introduces unexpected behavior — biased outcomes, degraded performance, compliance violations — institutions need the ability to quickly revert to the previous known-good version. Without version control, rollback is manual, error-prone, and slow.
• Change management compliance. Most financial institutions operate under change management policies that require formal approval for production changes. Version control provides the infrastructure to enforce and evidence these approvals.
• Incident investigation. When model outputs are questioned — by customers, regulators, or internal stakeholders — version control enables the institution to determine exactly which model version produced the output and under what conditions.

• Documentation — the broader practice of recording AI system information
• Model integrity — ensuring models remain accurate and uncompromised
• AI lifecycle — the end-to-end process that version control supports
• Performance monitoring — detecting when model versions need updating
• Training data — a key artifact that must be versioned alongside models