Sierra didn’t build an Agent Data Platform because it wanted more product surface to sell. It built one because a $15.8B agent platform already running inside roughly 40% of the Fortune 50 still couldn’t act on a customer’s order or account from raw CRM and billing records alone. That gap is where Atlan enters the story, not as a Sierra competitor, but as the governed context layer underneath it and its closest rivals, Salesforce Agentforce, Intercom Fin, Decagon, Ada, and Zendesk AI. Sierra itself, founded in 2023 by former Salesforce co-CEO Bret Taylor and former Google VP Clay Bavor, charges per resolved conversation rather than per seat, a bet that only pays off if the agent can reach data good enough to resolve something real.
Sierra’s own product roadmap is the clearest evidence for that positioning. The company built an Agent Data Platform specifically because raw, scattered customer records across CRM, billing, and order systems were never usable by an agent on their own; Sierra’s own framing describes it as unifying “everything your company knows about a customer” into one layer. Atlan connects to that same MCP-native ecosystem through its own MCP server, carrying the lineage, ownership, and business-term definitions that a customer-facing agent needs but that no single vendor’s platform natively holds.
- Agent OS: the platform layer running every Sierra agent in production
- Agent Studio: the no-code layer for building and configuring agents
- Agent SDK: the developer layer for custom, code-first integrations
- Agent Data Platform: unifies a customer’s context across sessions, channels, and systems
- Enterprise Data Graph: Atlan’s lineage, ownership, and certification layer for what the Agent Data Platform doesn’t natively ingest
| Sierra: Quick Facts | |
|---|---|
| What it is | Conversational AI agent platform for enterprise customer service |
| Founders | Bret Taylor (ex-Salesforce co-CEO, OpenAI board chairman), Clay Bavor (ex-Google VP) |
| Funding/valuation | $950M Series E, May 2026, at a $15.8B post-money valuation; $1.5B+ raised to date |
| Pricing model | Outcome-based, per-resolution pricing |
| Key customers | WeightWatchers, SiriusXM, Sonos, ADT, Chime, Cigna, Nordstrom, Nubank, Ramp, Rivian |
| Core components | Agent OS, Agent Studio, Agent SDK, Agent Data Platform |
What is Sierra?
Permalink to “What is Sierra?”Sierra is a conversational AI agent platform founded in 2023 by former Salesforce co-CEO Bret Taylor and former Google VP Clay Bavor, built to let enterprises deploy customer-facing AI agents rather than scripted chatbots. Taylor and Bavor bring credentials that shaped Sierra’s founding thesis directly: Taylor as Salesforce’s former co-CEO and chairman of OpenAI’s board, Bavor as the former Google VP who oversaw VR and Labs (Forbes, Nov 2025). Taylor has described the company’s roots plainly: “To some degree, we probably wouldn’t have started Sierra if not for OpenAI. It’s just such an important organization, and we felt indebted to it.”
Sierra’s growth since has been fast even by AI-startup standards. The company crossed $100M in annual recurring revenue within roughly seven to eight quarters of launch, then reported around $200M ARR by mid-2026 (TechCrunch, Nov 2025), and raised a $950M Series E at a $15.8B post-money valuation in May 2026, up from a $10B valuation just months earlier (TechCrunch, May 2026). That trajectory is a genuine data point for how enterprises use AI agents at scale, not just a funding headline.
Sierra now serves roughly 40% of the Fortune 50, with named customers including WeightWatchers, SiriusXM, Sonos, ADT, Chime, Cigna, Nordstrom, Nubank, Ramp, and Rivian, a customer list that puts it well ahead of most other standalone conversational agent startups. Sierra’s own product surface is named plainly across four layers: Agent OS is the platform running every agent, Agent Studio is the no-code layer for building one, Agent SDK is the developer layer for custom integrations, and Agent Data Platform is the layer that unifies a customer’s context across sessions, channels, and systems. That last layer carries most of the weight in this page’s argument, and its full mechanism gets its own section next. Sierra sits in the same competitive lane as Salesforce’s Agentforce 360, Intercom Fin, Decagon, Ada, and Zendesk AI, each a vendor-native AI agent platform built to be the default place its customers deploy conversational agents.
How does Sierra’s AI agent platform work?
Permalink to “How does Sierra’s AI agent platform work?”Sierra’s architecture layers a “constellation of models” under a supervisor, with a separate Agent Data Platform responsible for unifying what the agent knows about a specific customer. Each layer below handles a distinct part of how an agent decides what to say and whether it’s allowed to say it.
Agent OS and the constellation of models
Permalink to “Agent OS and the constellation of models”Rather than routing every request through one monolithic LLM, Sierra coordinates a constellation of specialized models, each handling a different part of a conversation, the same layered pattern described generally in how AI agents work. Agent OS is the platform tying that constellation together in production, and Sierra’s own “Agent OS 2.0” framing describes the evolution as moving from answers to memory and action, not just single-turn responses.
The supervisor architecture
Permalink to “The supervisor architecture”Sierra’s supervisor is a real-time layer that checks a model’s output against scope and policy before an agent responds, described in the company’s own words as maintaining “confidence in every conversation”. That’s a genuinely credible piece of engineering, and it’s worth naming clearly here because the rest of this page returns to a distinction the supervisor doesn’t resolve on its own: catching a bad response is different from governing what customer data an agent is allowed to reach in the first place.
Agent Data Platform
Permalink to “Agent Data Platform”Sierra introduced the Agent Data Platform to unify “everything your company knows about a customer… across sessions, channels, and systems” into one layer an agent can query. Bavor has described what that unification gives an agent in concrete terms: “It’s starting a conversation on second or third base, as opposed to from scratch.” That’s a direct admission that raw, scattered enterprise data isn’t usable by an agent without a unification step first, the same gap context engineering exists to close one layer below any single agent vendor.
Agent Studio and Agent SDK (including the no-native-connectors reality)
Permalink to “Agent Studio and Agent SDK (including the no-native-connectors reality)”Agent Studio gives non-technical teams a no-code way to configure an agent’s behavior, while Agent SDK is the developer layer Sierra uses for every integration. Practitioner buyer-guide analysis reports that Sierra has no native marketplace connectors for Zendesk, Intercom, Freshdesk, or Salesforce; by that account, every connection into a customer’s own systems is bespoke, per-case engineering work. That single fact matters more than it first appears: it means the context an agent needs about a customer’s order or account still has to be piped in from enterprise data a company controls, not supplied natively by Sierra’s own platform.
The Agent Data Platform can be read as Sierra’s own acknowledgment that raw, scattered customer data isn’t usable by an agent without a unification step, though Sierra would likely frame it as forward-looking product design rather than a gap being closed. Either read points at the same underlying job: unifying context Atlan’s context layer also does one level below the agent itself, across systems Sierra was never built to natively hold.
| Aspect | Traditional scripted bot/IVR | Sierra’s agent architecture |
|---|---|---|
| Resolution approach | Fixed decision trees, keyword matching | Constellation of models reasoning over intent |
| Escalation logic | Hard-coded rules, frequent dead ends | Supervisor-checked handoff to a human when scope is exceeded |
| Context handling | Session-only, no cross-channel memory | Agent Data Platform unifies context across sessions and channels |
| Pricing basis | Per seat or per message | Per successful resolution |
| Update cadence | Manual script edits | Continuous model and policy updates via Agent Studio |
Sierra's Agent Data Platform already unifies session and channel context. What it doesn't resolve on its own, definitions like "resolved," cross-system lineage, and systems outside Sierra's native connectors, is where an external MCP server picks up.
Get the CIO's guide to context graphs
See the four-layer architecture data leaders use to give agents like Sierra's context beyond their native connectors, from metadata foundation to agent orchestration.
Get the CIO GuideWhat is Sierra’s outcome-based pricing model?
Permalink to “What is Sierra’s outcome-based pricing model?”Sierra charges per successful resolution rather than per seat or per message, a pricing model its founders position as proof the agent actually works. Sierra has not published an official rate card, and third-party business-model analysis reports figures that vary enough across sources that this page holds off on repeating specific per-resolution or contract-minimum numbers until Sierra confirms them directly. Independent analysts have called the model a genuine business-model innovation, but practitioner skepticism persists about whether “resolved” is defined consistently across contracts, and what happens when the same customer issue recurs days later.
A pricing model built entirely on “resolved” only holds up if every system involved, the agent, the billing platform, and the human team that handles escalations, agrees on what “resolved” actually means. That’s a governed-definition problem, not a pricing problem, and it’s the exact gap Active Ontology is built to close before an agent acts on a term whose meaning shifts depending on who you ask. Teams evaluating what a resolution-based model costs at scale should model it the same way they’d model the cost to run AI agents at scale on any platform, since resolution volume, not seat count, drives the bill.
Sierra’s enterprise use cases: where it delivers results
Permalink to “Sierra’s enterprise use cases: where it delivers results”Sierra’s clearest published results come from subscription and account-servicing workloads, though voice has now overtaken text as its primary interaction channel.
Subscription and cancel-flow handling (WeightWatchers)
Permalink to “Subscription and cancel-flow handling (WeightWatchers)”WeightWatchers’ Sierra agent contained roughly 70% of cases in its first week while maintaining a 4.5+ out of 5 CSAT score, the single most concrete customer proof point Sierra has published. That’s the kind of deflection enterprise-ready AI agents are meant to produce once a company trusts them with real subscription and cancellation conversations, not just demo scripts.
High-volume account servicing (ADT)
Permalink to “High-volume account servicing (ADT)”ADT’s Sierra agent handles roughly 2 million inquiries per month, run with tight guardrails and dense human-in-the-loop review (GetMacha, 2026), a scale that puts real pressure on how well the agent’s context about a specific account and service history stays consistent across every interaction.
Voice as the primary channel
Permalink to “Voice as the primary channel”Sierra acquired Receptive AI and launched native voice agents; by late 2025, voice had overtaken text as Sierra’s primary interaction channel by volume. Sierra has also expanded into sales lead qualification and account-management workflows in 2026, using the same AI agents for sales pattern many enterprises now apply beyond pure customer support, and the same scaling questions covered in how to scale AI agents from POC to production.
Why do Sierra deployments run into trouble?
Permalink to “Why do Sierra deployments run into trouble?”Independent coverage of Sierra consistently surfaces the same two friction points, a December 2025 guardrail failure and a persistent integration and data-dispersion complaint, and both point at the same underlying gap.
In December 2025, a coordinated bad-actor jailbreak attempt caused Gap Inc.'s Sierra-powered chatbot to respond to off-topic and inappropriate prompts, and Sierra CEO Bret Taylor publicly apologized for the incident (The Information, Dec 2025). To Sierra’s credit, the company was transparent about the cause: a “coordinated bad actor,” not a baseline guardrail failure, per its own account (CCB Journal, Dec 2025). That distinction matters and shouldn’t be overstated into proof the architecture is broken, but the incident is still a fair, dated example that supervisor-style guardrails catch some failure modes and not others.
Sierra’s own July 2026 product direction backs up that reading. The company shipped Agency, secure sandbox infrastructure for its agent-building tools, and Agent Studio Experiments, a framework for A/B testing agent behavior against resolution rate and churn. In the Agency launch post, Sierra’s engineering team wrote plainly that the hardest part of building great agents stopped being the model and became everything surrounding it, a genuine admission that a capable model and good guardrails were never going to be the whole story. Sandboxing and behavior A/B testing are real, useful infrastructure, but they solve a different problem than governing what an agent is allowed to know and say about one specific customer’s order or entitlement, a distinction covered more broadly in AI agent accuracy and in AI agent risks and guardrails.
The practitioner review pattern reinforces the same point. G2, Quiq, and eesel reviews consistently flag opaque, six-figure-plus pricing with no published rate card, a steep initial setup, and a data-dispersion problem where Sierra holds bot-conversation data while the existing contact-center tool holds human-agent conversations, with no unified inbox unless a company builds its own unification layer. The Forrester Wave named Sierra a Strong Performer among conversational AI platforms for customer service in Q2 2026, while flagging weaker legacy-system connectivity and live-agent escalation compared to incumbent contact-center vendors (Forrester, via CX Foundation). That specific critique is about integration and workflow completeness, not directly about context governance, and it’s worth keeping the two separate: “Sierra doesn’t connect to my old contact-center software” is a different problem from “Sierra lacks governed context,” even though both surface in the same practitioner-friction cluster.
The governed-context gap this section describes is exactly what an external context layer addresses, one level below any single agent vendor’s guardrails, the same argument covered in securing multi-agent systems in the enterprise and in AI agent governance broadly, and in AI agent access control for the specific case of an agent taking real actions like issuing a refund.
Check your context readiness before you scale
Most conversational agent rollouts stall on data readiness, not the reasoning engine. Run through the checklist enterprises use to spot gaps before scaling past a pilot.
Assess Your ReadinessHow does Sierra fit into an enterprise AI agent stack?
Permalink to “How does Sierra fit into an enterprise AI agent stack?”Evaluating Sierra alongside other conversational agent platforms means checking connector coverage and context governance before committing, not just which model sounds most fluent in a demo. Every platform in Sierra’s competitive set, Salesforce Agentforce, Intercom Fin, Decagon, Ada, and Zendesk AI, faces the same evaluation questions once an agent needs data the vendor’s own platform was never built to hold.
Before committing to a context strategy for any of them, ask whether it reaches data the Agent Data Platform doesn’t already hold, who resolves what “resolved” or “entitled to refund” means before an agent acts on it, and what happens when the platform needs a system it doesn’t natively connect to. Getting these answers right is what separates agents that hold up under how to build an AI agent tech stack planning from the ones that stall once they meet real customer data instead of a clean demo, the same pattern covered in how to make AI agents context-aware.
| Criterion | Why it matters | What to look for |
|---|---|---|
| Native connector coverage | Sierra has no native connectors for Zendesk, Intercom, Freshdesk, or Salesforce | A context source that reaches data across the estate, not just one platform’s Agent Data Platform |
| Context governance & definition consistency | Outcome-based pricing depends on a shared definition of “resolved” | A certified semantic layer resolving one definition per term |
| Pricing transparency | Per-resolution pricing has no published rate card | Clear, auditable pricing tied to a verifiable outcome |
| Guardrails vs. governed context | Supervisor guardrails check behavior, not data scope | Policy enforcement checked at the point the agent acts on a specific record |
| Cross-system reach beyond CRM/helpdesk | The Agent Data Platform unifies session and channel context, not every SaaS tool | An open, MCP-compatible context source alongside the Agent SDK |
Sierra’s rapid customer growth also raises the same operational question every fast-scaling platform faces: how to build an AI agent harness robust enough that adding volume doesn’t mean adding risk, and why more teams are treating why AI agents need an enterprise context layer as a prerequisite rather than an afterthought once an agent is handling thousands of real customer conversations a day. The enterprise context layer question isn’t specific to Sierra; it recurs across Gemini Enterprise and AWS Bedrock for enterprise agents too, and how to implement an enterprise context layer for AI walks through the same evaluation regardless of which vendor’s agent sits on top.
How Atlan approaches context for agents like Sierra
Permalink to “How Atlan approaches context for agents like Sierra”Atlan doesn’t compete with Sierra, and never enters the customer-service conversation itself. What Sierra’s own Agent Data Platform shows is that a conversational agent is only as trustworthy as the context it can reach about the specific customer, order, account, or policy in question. That context lives across the enterprise’s actual systems of record, CRM, order management, billing, entitlements, not inside any agent vendor’s own platform, Sierra’s included, the same systems-of-record dependency covered for AI agents generally.
Atlan’s MCP server gives any agent, including a Sierra agent wired up via its own Agent SDK, a governed way to query business context, definitions, lineage, ownership, access policy, rather than reasoning over raw, ungoverned exports from a CRM or data warehouse. What Is Atlan MCP? covers the mechanism directly, building on the broader Model Context Protocol standard and how MCP delivers business context to any compliant agent. The Enterprise Data Graph carries lineage and ownership across every connector in the stack, which matters directly for the “does this agent’s answer about a customer’s order actually match what’s true in the system of record” question practitioners already raise in reviews. Active Ontology resolves what a business term means, “resolved,” “active subscriber,” “entitled to refund,” before an agent acts on it, the same context graph discipline that separates a governed definition from a generic knowledge-graph lookup, directly relevant given Sierra’s own outcome-based pricing depends on a shared, auditable definition of “resolved.”
No internal or public source ties Atlan directly to a Sierra deployment, and none should be fabricated here; no Sierra-specific case study exists yet. The architecture claim stands on its own: any agent can query Atlan via MCP, the same real technical relationship regardless of which customer-facing agent vendor a company picks, whether that’s Sierra or an internal knowledge assistant built for employees rather than customers. Teams preparing for this kind of integration should also review how to prepare enterprise data for AI agents and how to connect enterprise data sources to LLMs securely, since Sierra’s own Agent SDK pattern mirrors the same custom-connection work either way.
Real stories from real customers: governed context for customer-facing agents
Permalink to “Real stories from real customers: governed context for customer-facing agents”The MCP-native pattern this page describes for Sierra isn’t specific to any one CX vendor. Atlan’s own MCP server and Enterprise Data Graph already run in production at enterprises solving the same shared-vocabulary and context problem outside any single agent platform.
"We're excited to build the future of AI governance with Atlan. All of the work that we did to get to a shared language at Workday can be leveraged by AI via Atlan's MCP server…as part of Atlan's AI Labs, we're co-building the semantic layer that AI needs with new constructs, like context products."
Joe DosSantos, VP of Enterprise Data & Analytics, Workday
"Atlan is much more than a catalog of catalogs. It's more of a context operating system…Atlan enabled us to easily activate metadata for everything from discovery in the marketplace to AI governance to data quality to an MCP server delivering context to AI models."
Sridher Arumugham, Chief Data & Analytics Officer, DigiKey
See **Context Agents** live
Watch how a governed context layer connects CRM, billing, and order systems, and everything outside them, to AI agents in real time, across systems and vendors.
Watch a Live DemoWhy Sierra’s own roadmap makes the case for an external context layer
Permalink to “Why Sierra’s own roadmap makes the case for an external context layer”Sierra is a real, credible platform, not a thin wrapper on a foundation model. The WeightWatchers containment rate is a genuine production result, and the supervisor architecture is a genuinely sophisticated piece of guardrail engineering. None of that is in dispute.
What’s worth noticing is what Sierra built alongside those wins: the Agent Data Platform, Agency, and Agent Studio Experiments all exist because a capable model and good behavioral guardrails were never going to be enough on their own. Sierra would likely frame all three as product maturity rather than an admitted gap, and either read still lands on the same fact: a well-funded, purpose-built agent platform still had to build infrastructure specifically to unify customer context, because the conversation layer and the context about a specific customer’s order or account are two different problems that don’t collapse into one vendor. The Gap.com incident is a fair, dated reminder that behavioral guardrails and governed context-and-scope control are related but distinct disciplines, not the same one wearing two names. As Sierra’s sandbox and behavior-testing tooling matures further, the open question shifts from whether the agent sounds confident to whether its answer matches what’s actually true in the systems of record it can’t see directly, the same test that applies to Salesforce Agentforce and every other vendor-native agent platform in this category, not just Sierra.
FAQs about Sierra
Permalink to “FAQs about Sierra”1. Who founded Sierra AI?
Permalink to “1. Who founded Sierra AI?”Sierra was founded in 2023 by Bret Taylor, former Salesforce co-CEO and chairman of OpenAI’s board, and Clay Bavor, a former Google VP who oversaw the company’s VR and Labs divisions. The two built Sierra as a platform for enterprises to deploy customer-facing AI agents rather than scripted chatbots.
2. What is Sierra AI used for?
Permalink to “2. What is Sierra AI used for?”Enterprises use Sierra to deploy conversational AI agents for customer service, most commonly subscription and account servicing, cancellation handling, and high-volume support inquiries. Sierra has also expanded into sales lead qualification and account-management workflows, and voice now carries more interactions than text.
3. Is Sierra AI the same company as Sierra Wireless or Sierra Nevada?
Permalink to “3. Is Sierra AI the same company as Sierra Wireless or Sierra Nevada?”No. Sierra AI, founded in 2023 by Bret Taylor and Clay Bavor, is a conversational AI agent platform for enterprise customer service. Sierra Wireless is an IoT connectivity company and Sierra Nevada is an aerospace and defense contractor; neither is affiliated with Sierra AI.
4. How much does Sierra AI cost?
Permalink to “4. How much does Sierra AI cost?”Sierra charges per successful resolution rather than per seat. The company has not published an official rate card, and third-party pricing estimates vary widely enough that they shouldn’t be treated as confirmed figures; enterprises evaluating Sierra should confirm exact per-resolution and contract pricing directly with Sierra.
5. What is the “supervisor” model in Sierra’s architecture?
Permalink to “5. What is the “supervisor” model in Sierra’s architecture?”The supervisor is Sierra’s real-time guardrail layer, a system that monitors a “constellation” of specialized models and checks their outputs against scope and policy before an agent responds. It is a behavioral quality-control layer, distinct from governing what customer data the agent is allowed to see.
6. Is Sierra AI safe after the Gap.com incident?
Permalink to “6. Is Sierra AI safe after the Gap.com incident?”In December 2025, a coordinated bad-actor jailbreak attempt caused Gap’s Sierra-powered chatbot to respond to off-topic prompts, and Sierra CEO Bret Taylor publicly apologized. Sierra’s supervisor guardrails are real and technically credible, but the incident shows behavioral guardrails alone do not guarantee scope control.
7. Does Sierra AI integrate with Salesforce, Zendesk, or Shopify?
Permalink to “7. Does Sierra AI integrate with Salesforce, Zendesk, or Shopify?”Not through pre-built marketplace connectors. Sierra has no native connectors for Salesforce, Zendesk, Intercom, or Freshdesk; every integration is custom, built through its Agent SDK and Integration Library on a per-case basis.
8. How is Sierra AI different from Agentforce or Decagon?
Permalink to “8. How is Sierra AI different from Agentforce or Decagon?”Sierra is a standalone, vertical customer-service agent platform with outcome-based pricing, while Agentforce is bundled into Salesforce’s CRM and priced on consumption credits. Decagon, Ada, and Zendesk AI compete in the same customer-service lane with their own pricing and connector models.
9. How does Atlan connect to an agent platform like Sierra?
Permalink to “9. How does Atlan connect to an agent platform like Sierra?”Atlan connects through its own MCP server, which a Sierra agent wired up via its Agent SDK can call as an external, governed context source. The Enterprise Data Graph and Active Ontology give that agent lineage, ownership, and consistent definitions the Agent Data Platform doesn’t natively hold.
Sources
Permalink to “Sources”- Inside OpenAI Chairman’s $10 Billion AI Customer Service Startup Sierra, Forbes (2025)
- Bret Taylor’s Sierra Reaches $100M ARR in Under Two Years, TechCrunch (2025)
- Sierra Raises $950M as the Race to Own Enterprise AI Gets Serious, TechCrunch (2026)
- Constellation of Models: The Architecture Powering Sierra’s Agents, Sierra Engineering Blog (2026)
- Sierra Agent OS 2.0: From Answers to Memory and Action, Sierra (2026)
- Confidence in Every Conversation, Sierra Engineering Blog (2026)
- Introducing Agent Data Platform, Sierra (2026)
- Best Sierra AI Alternatives That Integrate with Salesforce, Twig (2026)
- How Does Sierra AI Make Money: Outcome-Based Pricing at a $15.8B Valuation, Value Add VC (2026)
- How WeightWatchers Embraces AI to Engage Members with Empathy, at Scale, with Sierra, Sierra customer story (2026)
- Sierra AI’s $10B Valuation Marks a Turning Point for Conversational AI, CMSWire (2026)
- Gap.com Chatbot Was Targeted for Abuse by “Bad Actor,” AI Startup Sierra Says, The Information (2025)
- Gap Chatbot Slip Reveals Guardrail Gaps, CCB Journal (2025)
- Agency: Secure, Scalable Sandboxes for Agents, Sierra Engineering Blog (2026)
- The Forrester Wave: Conversational AI Platforms for Customer Service, Q2 2026, Forrester, summarized via CX Foundation (2026)
- Sierra Customers, Sierra (2026)
- Sierra AI: The Complete Guide, GetMacha (2026)
