Salesforce Combines AI Probabilistic and Deterministic Systems for Safer Deployment

Salesforce is introducing a hybrid approach to AI agents that combines probabilistic and deterministic systems, ensuring both innovation and safety. The company's Agentforce framework splits the AI agent's brain into two layers: a generative layer for language and reasoning, and a deterministic layer for consequential decisions. This dual approach allows businesses to deploy AI agents with greater trust and accountability, as all actions are subjected to preset business logic, workflows, and permissions.

Kathy Baxter, Salesforce’s Principal Architect of Ethical AI Practice, explains that the deterministic layer ensures agents follow prescribed policies and procedures, while an audit trail records every action for review. This structure prevents AI agents from making biased or unethical decisions by grounding them in the factual realities of enterprise workflows. The Atlas Reasoning Engine further enhances this by forcing agents to explain each of their actions, providing a level of accountability.

Salesforce emphasizes the importance of human oversight with its "human at the helm" philosophy, ensuring humans remain in control. Autonomy is granted incrementally based on policies, allowing businesses to monitor and adjust AI behavior. The company also mandates that AI agents identify themselves as non-human and includes a Trust Layer for toxicity and prompt injection detection.

To address bias, Salesforce offers a Testing Center where customers can evaluate agent responses and set customizable controls to manage the data used in decision-making. Baxter highlights the critical need for continuous monitoring of AI agents, as probabilistic systems can produce varying results over time. This ongoing evaluation ensures agents remain safe and effective in providing the best customer experience.