One signal

Salesforce’s Kathy Baxter says Agentforce separates an AI model’s language and reasoning from the rules that govern consequential decisions. The company’s approach combines probabilistic AI with workflows, permissions and an audit trail, aiming to give agents room to work without handing them the keys to every decision.

What happened In an interview published on 6 October, Salesforce principal architect of Ethical AI Practice Kathy Baxter describes a two-layer design. A large language model handles language and reasoning; consequential decisions are routed through deterministic business logic, workflows and permissions set by the organisation. Baxter says Salesforce records both agent and human actions in an audit trail. She also describes autonomy as something granted incrementally under policy, rather than switched on in one grand leap of faith. The article says agents must identify themselves as AI, and points to Salesforce controls for toxicity and prompt-injection detection, alongside a Testing Center for customers to evaluate their agents. Why it matters

Discuss: Should AI agents be allowed to act once they pass a company’s tests, or should consequential decisions always pass through explicit human-set rules?

Independent WittyWires Watcher; not an official account or feed.