本頁閱讀時間 9 分鐘
概述
其代理可以查找账户并采取处理退货等行动,并且主要按结果而不是按席位收费。这是大型语言模型在客户服务中部署的一个突出例子。
重點摘要
- Sierra 由 Bret Taylor 和 Clay Bavor 於 2023 年創立,旨在建立人工智慧代理來處理客戶對話並在業務系統中採取行動。
- TechCrunch 報導,Sierra 的年度經常性收入於 2025 年 11 月(即推出後約 21 個月)達到 1 億美元,估值為 100 億美元。
- Sierra 主要根據結果收費,例如已解決的案例或已保存的取消,而升級的對話通常不收取任何費用。
- 一名代理商可以跨聊天、語音、電子郵件、簡訊、WhatsApp 和 ChatGPT 進行部署,並以自然語言定義目標和護欄。
- Sierra 的 tau 工作台發現 GPT-4o 代理程式成功完成了不到一半的模擬任務,這表明為什麼測試和手動升級仍然至關重要。
深入探討
Sierra makes conversational AI agents for companies' customer-facing work. An agent here means software built on large language models that can hold a conversation and also take actions in business systems, such as checking an order, changing a plan, or issuing a refund, while following the company's policies. According to TechCrunch, Bret Taylor and Clay Bavor decided to start the company over lunch after Taylor left Salesforce in 2023. Taylor had been Salesforce co-CEO and co-created Google Maps; Bavor spent 18 years at Google leading products including Gmail and Google Drive. Sierra describes itself as a conversational AI platform for businesses and emphasizes customer lifetime value and revenue rather than only cost savings.
A Sierra deployment works roughly in this order. The company first gives the agent its knowledge and rules: help-center content, standard operating procedures, policies and examples of past conversations. Sierra's Ghostwriter tool, introduced in 2026, can draft an agent from uploaded procedures, call transcripts or a plain-English goal. Teams then define journeys, which Sierra's November 2025 Agent OS 2.0 announcement describes as natural-language definitions of goals and guardrails that non-engineers can edit, with a software development kit available for engineering teams. Integrations connect the agent to systems such as billing, order management and customer records. The same agent can then be published to chat, voice, email, SMS, WhatsApp and ChatGPT.
During a live conversation, the agent reads the customer's message, uses a language model to work out what the person needs, retrieves relevant policy and account information, and decides whether to answer, ask a clarifying question, call a tool such as a refund function, or hand off to a person. The Agent Data Platform, announced alongside Agent OS 2.0, merges past calls, chats and emails with structured business data so the agent can remember earlier interactions. Escalations go to human representatives, and a Live Assist feature helps those humans with suggested answers and next steps. After conversations, Sierra's Insights tools analyze large volumes of transcripts to find recurring problems, and Experiments let teams test changes before rolling them out widely.
The company grew quickly. TechCrunch reported in November 2025 that Sierra, then 21 months old, had reached $100 million in annual recurring revenue, and had been valued at $10 billion after a $350 million round led by Greenoaks Capital in September 2025, with Sequoia, Benchmark, ICONIQ and Thrive Capital also investors. Sierra's own review of its second year, published in February 2026, reported more than $150 million in annual recurring revenue, an investment from SoftBank Vision Fund 2, and expansion to France and a new Tokyo office. Named customers across these sources include ADT, Cigna, DIRECTV, Deliveroo, Discord, Gap, Ramp, Rivian, SiriusXM, SoFi and Sutter Health.
Sierra competes with several alternatives. Older phone menus and rule-based chatbots follow fixed scripts, so they are predictable but break on anything unexpected. Human contact centers handle complex and emotional cases well but cost more per contact; a Chime case study published by Sierra cites an average cost of $5 to $20 for a phone call to a consumer brand. Companies can also build their own agents directly on model providers' APIs, which gives full control but requires engineering for integrations, testing, monitoring and compliance. Established contact-center and customer relationship software vendors have added AI agents to their products as well. Sierra's pitch is a managed platform with built-in guardrails, analytics and outcome-based pricing, rather than raw model access.
The main limitation is reliability. Language-model agents can misread a request, invent a policy, or take the wrong action, and small error rates matter at the scale of millions of conversations. Sierra's own researchers made this point in June 2024 with tau-bench, a benchmark in which an agent talks to a simulated customer and must use tools while following policy. Even strong function-calling agents such as GPT-4o succeeded on fewer than half of tasks, and their consistency across eight repeated attempts in the retail domain fell below 25%. A June 2025 follow-up, tau-squared-bench, found performance drops sharply when the customer must also take actions, as in technical support. Customer metrics published by Sierra are self-reported case studies, not independent audits. Agents also process personal and sometimes financial or health data, so security and compliance reviews are part of every deployment; Sierra's site displays SOC 2, ISO 27001, ISO 42001, HIPAA and PCI DSS badges without detailing their scope.
On cost, Sierra's December 2024 post on outcome-based pricing explains that it is paid when the agent completes an agreed outcome, such as a resolved conversation, a saved cancellation or an upsell, with criteria set in advance; unresolved or escalated conversations usually carry no charge, and simple routing interactions may be priced per conversation instead. That model aligns the vendor's revenue with results, but buyers still need to define outcomes carefully, verify how they are counted, and budget for integration work and ongoing review of conversations. Sierra's annual review reports that one large healthcare company went live seven weeks after project kickoff, and a Singtel case study lists less than ten weeks to launch, which gives a sense of typical setup timelines.
04工作範例
以 tau-bench 方式設計和測試退貨處理代理
用簡單易懂的語言寫下政策:30 天內退貨、需要收據或訂單號碼、退款至原始付款方式以及針對損壞物品升級。
列出代理程式可以呼叫的工具,例如look_up_order、check_return_window、create_return_label 和issue_refund,以及每個工具所需的輸入。
定義算是成功的結果,例如建立的退貨標籤和設定為退貨待處理的訂單狀態,以及什麼算是升級。
寫出十個模擬的客戶場景,包括過期窗口、訂單號碼遺失以及憤怒的客戶因商品損壞而感到憤怒。
多次執行每個場景,並將最終訂單資料庫狀態與預期狀態進行比較,而不是判斷回應的措詞。
計算單次運行成功率以及每次運行成功的場景比例,並修復存在差異的策略或工具。
它顯示了什麼
學習者最終會得到書面策略、工具清單、清晰的結果定義和 tau-bench 的 pass^k 風格的小可靠性分數。檢查的是每個場景要么在所有運行中達到預期的資料庫狀態,要么正確升級,這是生產代理和基於結果的合約都依賴的標準。
現實世界的實施
金融科技支援:Chime 使用 Sierra 代理來處理常規會員問題,例如直接存款狀態和密碼重置,並報告其解決率從 50% 上升到 70% 以上,並將複雜的問題交給人工處理。
抵押貸款和房地產:Sierra 2026 年 2 月的年度審查列出了 Rocket Mortgage 使用代理商進行抵押貸款發放,Redfin 使用代理商進行對話式房屋搜尋。
訂閱管理:SiriusXM(被 TechCrunch 指定為客戶,並且在 Sierra 自己的評論中)使用 Sierra 代理進行訂閱管理; Sierra 的客戶故事引用了 3400 萬訂閱者。
結果計費:根據 Sierra 在 2024 年 12 月描述的定價模型,公司在代理完成商定的任務(例如解決支援案例或保存取消)時付費,而在案例升級時通常無需支付任何費用。
代理測試:Sierra 研究人員於 2024 年 6 月發布了 tau-bench 基準,該基準針對模擬客戶對話是否使公司資料庫保持正確的最終狀態對代理商進行評分。
Sierra AI 的未来:客户体验代理、他们的工作方式以及定价方式
Sierra's 2026 announcements point to where the company is taking the product. Ghostwriter, introduced in March 2026, aims to generate agents from procedures and recordings rather than hand-built configuration. Posts in October 2026 introduce a draft Personal Agent Protocol, work with Stripe on taking payments inside conversations, and research on caller identification for voice agents. Its research team has continued the tau series, with October 2026 posts on evaluating voice agents' accuracy with details such as names and numbers and on performance across languages. The open problems are shared across the industry. Reliability across repeated attempts, measured by pass^k in tau-bench, remains well below what customers expect from human staff in many tasks. Independent evidence on resolution rates and customer satisfaction is thin compared with vendor case studies. Voice agents raise questions about disclosure, impersonation and fraud. Outcome-based pricing will be tested as buyers and vendors negotiate what counts as a resolved case. Whether AI agents mainly reduce contact-center headcount or mainly expand service hours and quality will depend on how companies deploy them, and evidence on that is still accumulating.
付諸實踐
為什麼這很重要
- 供應商策略. 供應商路線圖會影響您的團隊接下來可以建立的功能。
- 成本與預算. 商業條款和部署選項會影響長期成本和風險。
- 風險與安全. 公司激勵措施塑造了產品預設、安全態勢和開放性。
留意
- 發佈公告可能會超過實際生產工作流程的穩定性。
- API 定價或政策轉變可能會在一夜之間打破假設。
- 單一供應商依賴性增加了鎖定和遷移成本。
後續步驟
- 使用您自己的任務和資料集評估提供者。
- 在整合之前查看隱私、安全和法律條款。
- 維護跨模型或供應商的後備計劃。
- 監控發行說明,以便路線圖的變更不會讓團隊感到意外。
資料來源與延伸閱讀
- TechCrunchBret Taylor 的 Sierra 在不到兩年的時間內達到了 1 億美元的 ARR
- SierraSierra
- SierraSierra Agent OS 2.0:從答案到記憶與行動
- Sierra第二年回顧
- SierraAI 代理基於結果的定價
- SierraChime 客戶故事
- arXiv (Yao et al., 2024)tau-bench:現實世界中工具-代理-使用者互動的基準
- arXiv (Barres et al., 2025)tau-squared-Bench:在雙控制環境中評估會話代理
不斷探索
常見問題
什麼是 Sierra AI:客戶體驗代理商、他們如何運作以及如何定價?
Sierra 是一家舊金山公司,由 Salesforce 前聯合執行長 Bret Taylor 和 Google 長期高階主管 Clay Bavor 創立,該公司開發人工智慧代理,用於處理聊天、語音、電子郵件和訊息等方面的客戶對話。其代理可以查找账户并采取处理退货等行动,并且主要按结果而不是按席位收费。这是大型语言模型在客户服务中部署的一个突出例子。
Sierra 的知名共同創辦人有哪些?
Sierra 由前 Salesforce 聯合執行長兼 OpenAI 董事會主席 Bret Taylor 和長期擔任 Google 高階主管的 Clay Bavor 共同創立。
Sierra 以哪些定價創新而聞名?
Sierra 引入了基於結果的定價,將其激勵措施與成功解決客戶問題而不是對話量相結合。
Sierra 的代理商與通用聊天機器人有何不同?
Sierra 代理商代表品牌的聲音和政策,並採取具體行動,例如處理退款和更新帳戶。
Sierra 的 Horizon 描述稱其代理商可以協調什麼?
Sierra 將 Horizon 代理商描述為在幾天或幾週內協調入站和出站互動;這是產品描述,並非特定結果的保證。
作為 Agent OS 2.0 的一部分,Sierra 宣布了什麼來為代理人提供跨互動的記憶和上下文?
Sierra 的 Agent OS 2.0 公告包括 Agent 資料平台,該平台被描述為為代理提供記憶體、上下文和智慧。