Cognous Keeps Your AI in Check

Tone-deaf emails sent by AI agents have become a significant risk for organizations. Automated agents can generate emails without any human oversight, leading to potential compliance issues. To address this problem, Cognous has introduced the Open Control Stack, an open-source infrastructure designed to govern AI agents effectively. The stack comprises four key components: Declare, Control, Replay, and Evidence.

Declare, the Agent Action Manifest, is a written document that outlines specific actions an AI agent can perform before it executes them. For instance, it allows reading customer data but blocks reading contracts and requires human sign-off before sending an email. This gives organizations greater control over their agents' behavior, preventing them from performing unauthorized tasks.

Control, the Agent Control Plane, operates alongside the AI agent. It checks each proposed action against the rules defined in the Declare component in real-time. Depending on the outcome, it either allows, blocks, or escalates the action for human approval. Each of these outcomes is recorded, providing a comprehensive log that can be used for auditing purposes. This proactive approach ensures that no action is performed without proper authorization.

Replay, the Agent Replay Bundle, is an export of all recorded actions from a run. Instead of relying on application logs that are only available after the fact, Replay provides a consolidated and portable format of every action taken by the agent. This includes reads, drafts, blocks, escalations, and outcomes. By having a complete record of every step, organizations can easily trace the agent's actions and understand why a particular decision was made.

Evidence, the Agent Governance Evidence Pack, is assembled after multiple runs. It compiles a record of the agent's behavior, demonstrating compliance with the established policies. Instead of simply stating that there are controls in place, Evidence provides tangible proof. For example, it might show that 40 emails required human sign-off in a given month, along with the details of each case. This shift from vague assurances to concrete evidence ensures that organizations can prove their compliance and accountability.

In conclusion, Cognous's Open Control Stack offers a comprehensive solution to the problem of uncontrolled AI agents. By providing a framework for authorization, enforcement, and proof, organizations can now govern their AI agents more effectively. The stack promotes transparency, accountability, and proactive risk management, preventing automated agents from sending tone-deaf or unauthorized emails. To learn more about Cognous and explore the Open Control Stack, visit their website at cogno.us.