Sierra

Sierra is Bret Taylor's attempt to build conversational AI that enterprise brands will actually trust with their customers. Unlike chatbot builders where you wire up flows and pray, Sierra's pitch is autonomy with guardrails: the agent interprets intent, pulls from your knowledge base, and escalates when it hits uncertainty. I tested it against a mid-sized SaaS company's support queue and the difference from standard LLM wrappers is real. Sierra doesn't just pattern-match FAQs. It reasons through multi-turn conversations, remembers context across sessions, and handles edge cases without falling back to "let me transfer you" every third message. The voice capability is where it earns its keep. Traditional IVR systems are decision trees built in the 1990s. Sierra listens, clarifies, and acts more like a competent junior support rep than a phone menu. I watched it handle a billing dispute where the customer changed their mind mid-call about what they actually wanted. It adjusted, confirmed, and logged the outcome without a single "press 1 for..." prompt. Multilingual support works without separate training per language, which matters if you're scaling internationally and don't want to maintain five different bot personalities. Failure modes: Sierra is opaque when it makes mistakes. You get logs, but debugging why it chose path A over path B requires digging. The outcome-based pricing sounds clever until you realise you're negotiating what counts as a successful resolution. If your product has complex edge cases or requires deep technical knowledge, Sierra will escalate more than you'd like. It's built for volume CX work, not specialist troubleshooting. Compared to Intercom's Fin or Ada, Sierra feels more autonomous but less transparent. Fin gives you tighter control over responses; Sierra gives you fewer manual interventions. If you're a brand that values consistency and can tolerate occasional overreach, Sierra wins. If you need to audit every interaction or operate in a heavily regulated space, the black-box nature will frustrate you. One concrete workflow: a retail client used Sierra to handle post-purchase queries (order status, returns, size exchanges). It reduced ticket volume by 60% in the first month, but they had to manually tune the escalation thresholds twice because it was too eager to solve problems it didn't fully understand. Once calibrated, it worked well. The agent learned seasonal patterns (more return questions in January) without explicit retraining.