Sierra

Sierra is a San Francisco company that builds AI agents that businesses deploy to handle customer interactions over phone, chat and messaging. It was co-founded by OpenAI board chair Bret Taylor and former Google executive Clay Bavor, and it builds on foundation models from OpenAI and Anthropic rather than training its own frontier models (Source: cnbc.com). It raised a $950 million Series E at a $15.8 billion post-money valuation in May 2026, with annual recurring revenue having crossed $150 million within eight quarters of launch. On July 23, 2026 it announced the acquisition of Takeoff, a 14-month-old, three-person startup building long-horizon agents that had reached nearly eight figures of annual recurring revenue (Source: sierra.ai; theinformation.com).

Snapshot

Valuation

Revenue

Business and products

Sierra sells AI customer-service agents to large enterprises. Taylor told CNBC in May 2026 that the company uses a "constellation of models" with its own fine-tuned proprietary layers, and that agents had "digitized the last remaining analog channel, which is the telephone line". He estimated that $400 billion is spent each year on customer service, most of which he expects to move to AI agents (Source: cnbc.com). Shahidi, Rusak, Manning, Fradkin and Horton cite Sierra, alongside Decagon and Harvey, as an example of an agent provider that customizes other developers' foundation models for particular uses, in contrast to model builders such as Anthropic and OpenAI. They note that the two business types have different cost structures: training carries high fixed costs, while operating agents carries mostly variable costs (The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025)).

Sierra's products include:

• Agent OS, its conversational agent platform. Sierra states that it powers phone calls and chats for hundreds of companies, including Santander, Rocket Mortgage and Cigna, and almost half of the Fortune 50 (Source: sierra.ai).
• Ghostwriter, announced March 25, 2026, which Sierra describes as an agent that creates and optimizes other agents (Source: sierra.ai).
• Horizon, announced July 16, 2026, a platform for agents that pursue goals across many interactions over days, weeks or months, such as originating a loan or obtaining prior authorization for a healthcare procedure. Sierra describes it as adding a "context engine" and long-horizon planning to Agent OS, and calls it the largest expansion of the company since its 2024 launch. A US healthcare provider scheduling specialist referrals is named as a design partner (Source: sierra.ai).

Sierra charges for results rather than usage. Its Horizon announcement states that customers "don't pay for tokens" but for "business outcomes delivered", with Sierra bearing the cost of managing token spend. It presents this as a response to customer concerns about controlling token costs and about "ceding value" to model providers, citing remarks by Microsoft chief executive Satya Nadella (Source: sierra.ai). Taylor has discussed outcome-based pricing at length, including the difficulty of defining outcomes for agents whose work, unlike a customer-service case, has no clear resolved-or-not state (Source: cheekypint.substack.com).

Customers and deployments

Sierra's customers are mostly large enterprises, including Prudential, Cigna, Blue Cross Blue Shield and Rocket Mortgage, as well as one in three of the world's largest banks. Taylor said in May 2026 that Sierra serves more than 40% of the Fortune 50 (Source: cnbc.com). In its funding announcement Sierra said its agents handle insurance first notice of loss, claims and sales; mortgage origination, refinancing and servicing; bank sales offers; revenue-cycle management between healthcare providers and payers; telecommunications subscriptions; and retail product discovery. It reported that Nordstrom launched its voice agent in five weeks, that Singtel launched in 10 weeks with resolution rates above 70%, and that Cigna reached production in eight weeks and cut patient authentication time by 80% (Source: sierra.ai).

Taylor has described deployments in which a payer's agents and a provider's agents call each other in English over the public telephone network. He has also cited a SoFi case study in which net promoter score rose by 33 points, and a retailer whose customer-conversation volume grew two to three times after deployment while its costs did not fall significantly (Source: cheekypint.substack.com).

Funding and valuation

Sierra announced its $950 million Series E on May 4, 2026, led by Tiger Global and Google's GV, at a $15.8 billion post-money valuation, up from $10 billion in the fall of 2025. Sierra said the round gave it more than $1 billion to invest (Source: cnbc.com; sierra.ai). The Information reported that the new valuation was 50% above a September 2025 round led by Greenoaks (Source: theinformation.com). Benchmark general partner Peter Fenton, one of the first investors, called Sierra the leader in the customer-experience category by revenue and customer base. Taylor said an IPO is "definitely in our future" but that he views staying private as an advantage during rapid scaling. He also said he expects an AI market correction within two years, with a "culling effect" in which capital dries up for all but market leaders (Source: cnbc.com).

Acquisitions

On July 23, 2026, Sierra announced that it is acquiring Takeoff, which describes itself as a long-horizon agent runtime and whose agents worked in five industries from lending to healthcare. In a post on Sierra's blog, Takeoff founder Aakash Thumaty wrote that Takeoff grew from $0 to nearly eight figures of annual recurring revenue during 2026 with a three-person team, winning several seven-figure contracts on outcome-based terms, and that the two teams are building Horizon together (Source: sierra.ai). The Information described the acquisition as a move to diversify Sierra's business (Source: theinformation.com). Horizon itself had been announced by Sierra a week before the acquisition (Source: sierra.ai).

Research and benchmarks

Sierra's research team published τ-bench in June 2024, a benchmark in which a language agent with domain-specific API tools and policy guidelines holds conversations with a user simulated by a language model. Success is scored by comparing the final database state with an annotated goal state. The paper reported that even state-of-the-art function-calling agents such as GPT-4o succeeded on fewer than 50% of tasks and were inconsistent, with pass^8 below 25% in the retail domain (Source: τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains (Yao et al., 2024); sierra.ai). Successor versions, maintained as τ²-bench, extend the benchmark to further enterprise domains and to voice (Source: taubench.com).

Frontier developers report results on these benchmarks. Anthropic's system cards describe τ²-bench as "an evaluation from Sierra" that measures an agent's ability to interact with simulated users and programmatic APIs while following domain-specific policies, split into retail, airline and telecom sections. The Claude Opus 4.5 card reported policy loopholes in the airline section, built a corrected version, and submitted the fixes to the benchmark's authors (System Card: Claude Opus 4.5 (Anthropic, November 2025)). The Claude Opus 4.6 and Claude Sonnet 4.6 cards omit airline results because those fixes had not yet been incorporated upstream, and report 99.25% (telecom) and 91.89% (retail) for Opus 4.6 averaged over 5 trials, and 97.9% (telecom) and 91.7% (retail) for Sonnet 4.6 averaged over 10 trials (Claude Opus 4.6 System Card; Claude Sonnet 4.6 System Card). Google reported that Gemini 3.8 Live Extended Thinking, released September 15, 2026, scored 35.1% on Sierra's τ-Voice-banking benchmark (Source: blog.google; see Gemini 3 / Gemini 3 Pro). See AI Benchmarks and Evaluation.

People

• Bret Taylor, co-founder and chief executive, chairs OpenAI's board (Source: cheekypint.substack.com). He is a former co-CEO of Salesforce and former chief technology officer of Facebook, and was Twitter's chairman when Elon Musk bought the company. He is credited with helping create Google Maps (Source: cnbc.com).
• Clay Bavor, co-founder, led virtual-reality efforts and Google Labs at Google (Source: cnbc.com).
• Former Sierra research scientist Noah Shinn went on to lead Instinct, a personal-assistant agent operated by Spear Street Technology (Source: techcrunch.com).

Relationships

• depends-on: OpenAI, Anthropic — foundation models underlying Sierra's agents
• deploys-in: Financial Services — AI Deployment, Insurance — AI Deployment, Healthcare — AI Deployment, Retail — AI Deployment
• instance-of: Agentic AI — long-horizon agents as an acquisition target category
• related: The Coasean Singularity? Demand, Supply, and Market Design with AI Agents — Shahidi, Rusak, Manning, Fradkin, Horton (2025) — cites Sierra as an agent provider on others' foundation models
• related: AI Benchmarks and Evaluation — τ-bench and τ²-bench
• related: System Card: Claude Opus 4.5 (Anthropic, November 2025), Claude Opus 4.6 System Card, Claude Sonnet 4.6 System Card — report τ²-bench results
• related: Harvey, Salesforce, Inc., Spear Street Technology