Decision Firewall: what building an AI governance framework taught me

A refund example was used to illustrate the decision-making process when building an AI governance framework. The scenario involved a duplicate charge claim, which required verification of a second payment record and cancellation eligibility based on timing and usage. A refund policy allowed for automatic refunds for verified duplicates and unused subscriptions cancelled within seven days, with eligible amounts up to ₹1,000 processed automatically and larger amounts requiring review.

A daily automated limit of ₹10,000 was set to handle excess demand, with simulated payments used for demonstration purposes.

The article describes the development of Decision Firewall, an open-source reference for AI action authorization and audit. The author started with a refund-recovery example to explain the framework's purpose and used simulated payments to test its functionality. The refund recovery process involved interpreting the request, authorizing the exact action, dispatching and losing the response, and recovering without a second refund.

The framework made provisions for idempotency keys and execution state durability to ensure accurate refunds.

As the project progressed, the author incorporated more features, such as history ingestion, dataset snapshots, named Python rules, and optional context retrieval and telemetry. These features aimed to make the framework more versatile and useful for developers. However, the author later realized that the framework's benefits needed to be demonstrated compared to other methods of implementing AI governance. This realization prompted a reevaluation of the design and implementation process.