Outcome-priced AI agents for customer experience
HQ: San Francisco, CA, United States | Founded: 2023 | Employees: 700-850 (estimates vary; Tracxn lists 855, PitchBook 488) | Stage: Series E; $950M at a $15.8B post-money valuation | Website: https://sierra.ai
Sierra sells AI agents that do customer-facing work, and the shape of that work has changed twice since launch. The company was incorporated in March 2023 after Bret Taylor resigned as co-chief executive of Salesforce in January and met Clay Bavor for lunch in Palo Alto. Six paid design partners ran from late 2023 into February 2024 and every one converted to a paying customer. Sierra launched publicly in February 2024 with $110 million at a reported $1 billion valuation, thirty employees, and customers including WeightWatchers, SiriusXM, Sonos and OluKai.
The first product was the Agent SDK: a developer defines a goal and the system composes multi-step workflows of skills such as triage, respond and confirm. The second was Agent Studio, a visual build surface with knowledge retrieval over help centres and documents, guardrails, simulations and an Expert Answers feature that turns contact-centre know-how into grounded articles. By November 2025, Agent OS 2.0 added an Agent Data Platform, Insights analytics, and deployment into ChatGPT alongside voice, chat, email, SMS and WhatsApp — fifty-nine languages in total.
Voice became the centre of gravity. General availability arrived in October 2024, and by October 2025 Sierra was handling more voice calls than chats. The engineering behind it is specific: a custom voice-activity-detection model, a concurrent rather than sequential reasoning graph, streaming text-to-speech, IVR-tree navigation, and Level 1 PCI-compliant card collection that avoids an interactive-voice-response handoff entirely. Voice Personas, added in August 2026, let a brand control tone and pronunciation across languages.
The second act is Horizon, launched in July 2026 from the acquisition of Takeoff, a three-person company that reached roughly eight figures of revenue in fourteen months. Horizon agents pursue goals across days or weeks rather than a single conversation — loan origination, prior authorisation, claims, subscription upgrades — supported by a Context Engine that stitches customer history into next-best-action decisions. The economics are deliberately inverted: Sierra states that customers pay for outcomes delivered and that it bears the token cost. Ghostwriter, released in March 2026, is the agent that builds the agents — you describe the behaviour in prompts, feed it standard operating procedures or audio interviews with subject-matter experts, and it generates and simulates the agent, diagnoses failures and fixes them.
Around all of this runs a research programme that is unusually public for a private company. Sierra open-sourced τ-bench in June 2024 and found every one of twelve tested frontier models below a 50% pass rate, introducing the pass^k reliability metric to measure consistency rather than average performance. It has since released τ²-bench, τ³-bench, τ-voice and, in September 2026, hyper-τ-bench, which evaluates agents that build agents. The work is authored with Princeton Language and Intelligence.
Sierra has also bought its way into new geographies and capabilities: Receptive AI for voice in March 2025, Opera Tech in Tokyo in March 2026, Fragment in Paris in April, and Takeoff in July. It runs roughly 700 to 800 people across twelve-plus countries, and 51 published customer stories.
Sierra has done the hardest thing in enterprise software: invented a pricing model that a customer can accept without a procurement fight. Charging per resolved outcome, and not charging at all when the conversation ends in a human handoff, aligns the vendor with the buyer in a way that per-seat or per-conversation pricing never did. That, more than model quality, is why the company reached $200 million of annual recurring revenue in nine quarters while the biggest helpdesk vendors were still charging by conversation.
The second thing Sierra did was treat the frontier labs as suppliers rather than rivals. Its agents are assembled from fifteen or more frontier, open-weight and proprietary models with automatic failover when a provider degrades, and founder Bret Taylor — who chairs OpenAI's board — has said he does not view the model providers moving down-stack as a threat. The bet is that the durable value sits in context, integrations, outcome measurement and brand control rather than in the model.
The third is a genuine expansion of scope. Horizon, launched in July 2026 from the Takeoff acquisition, takes the product from resolving a ticket to pursuing a goal over weeks — loan origination, healthcare prior authorisation, insurance claims — with Sierra explicitly bearing the token cost. Figure's home-equity deployment reportedly lifted funded-loan conversion 143%.
What has not been resolved is whether the accuracy can hold at scale. Sierra's own τ-bench found every frontier model it tested scoring under 50% on real customer-service tasks, the company is not yet profitable, and a $15.8 billion valuation is roughly 79 times revenue. The regulatory exposure is unusually specific: the FTC's July 2026 policy statement names deceptive claims about AI replacing human representatives as an existing enforcement theory, which is precisely the pitch Sierra's customers buy.
Browse: Home | Blog | About | All Startups A-Z | View full profile on StartupWiki