Hello fam 👋

Something different this week!

This is our first in a series of deep dives on the most ambitious (and often misunderstood) AI startups. My goal is to make their tech, business models, and long-term bets as clear as possible, so we can all see the future they’re aiming for.

Thoughts? Just reply and let me know what you liked, what you didn’t, and what you'd want to see more of. Let’s go.

Sierra is building an outcome-priced agent OS for customer service, betting that reliability engineering and supervision can make probabilistic systems dependable enough for Fortune-500 scale.

That’s our kind of bet.

Last week my Wi-Fi died. I went through the usual rituals to fix it. ChatGPT, YouTube tutorials, Reddit. Then I committed the unspeakable. I called customer support.

Dave picked up. He listened, or maybe just waited for his turn in the script. Then came the line: “Have you tried turning it off and on again?”

We marched through troubleshooting steps that had nothing to do with my actual issue. I stopped feeling like a person and started feeling like an input. Not a paying customer, just a ticket in a queue, a variable in someone’s formula for lowering average handle time.

This experience is cliché because it’s everywhere. But the problem is not Dave.

The problem is the system that produced Dave. A system obsessed with metrics, so allergic to inefficiency, that it has scrubbed humanity out of one of the most important contact points between a company and the people funding it.

I read an essay that captured the feeling from the other side of the phone. A delivery driver described his job like this:

“They see it as automation. They see you as just a system. You’re essentially working as a vending machine.”

Human vending machines. We press buttons, and if we’re lucky, the right product drops.

And it’s expensive. Poor customer experiences cost $3.7 trillion in global sales every year. So we turn to an automated chatbots. But anyone who has typed “I WANT TO SPEAK TO A REP” knows how badly that often goes…

To build something better, we need to look back at how the old nervous system was pieced together, one kludgy patch at a time.

Customer support has never stood still. Each phase solved the pain of the last, only to introduce a new limitation.

1. Voice

Initially, support was purely human. Call centers in the 1960s, like Birmingham Press and Mail, pioneered the use of automatic call distribution and brought empathy. You could hear a sigh, a pause, a smile.

The limiting factor was presence. But you could only help as many people as you had reps and chairs.

2. Scripts

To solve for scale, businesses turned to scripts. IVR menus (“Press 1 for billing”) and rule-based bots like Clippy helped handle volume. But they couldn’t think. The second a query broke the tree, they broke too.

3. Fluency

Transformers changed the game. ChatGPT made it cheap to sound smart. Bots could now talk like people. But still: no hands. They could explain the return policy, but not start the return.

4. Agency

Now comes the shift to Human-to-Agent. Here, agents not only answer, but can take action on the customer’s behalf. Like initiating the return and troubleshooting my Wi-Fi router.

Presence gave support a voice. Scripts gave it scale. Language gave it fluency. Agency finally gives it hands. This is the frontier Sierra is betting on.

Before we zoom in on Sierra, it’s worth asking why this shift to AI agents feels more like a foregone conclusion. The short answer is money and momentum.

Support is expensive. There are 15 million people working customer service jobs worldwide. Each one costs a business about $110,000 a year, once you stack salary, benefits, equipment, and the cost of replacing them when they burn out. Every phone call costs $3 to $6.5 per minute.

AI changes that math overnight. An AI-powered interaction costs pennies, about $0.03 to $0.25 per minute. That’s a 95 percent cost collapse. Some nuance: the human fee includes labor, facilities, and churn, while the AI figure covers model inference and platform fees only.

According to KPMG, companies that make the switch to use AI for customer support see a return of $3.50 on every dollar spent, with 5% of companies reporting a return of $8.

We extrapolated this, and our ROI progression curves show businesses reaching 3.5-8x returns within 12-18 months, with break-even occurring at 3-6 months.

The first-order story is savings. The second-order story is what companies do with that reclaimed capital. The dollars pulled out of repetitive support loops can be reinvested into growth, rewiring a company’s unit economics in ways that slower competitors can’t keep up with.

So it’s not surprising that venture dollars are clustering around the companies positioned to be category winners. Sierra raised $175 million at a $4.5 billion valuation in October 2024, just over a year after its launch. That’s quite WOW.

Technology, economics, and customer expectations are all pointing in the same direction.

After ChatGPT, every company’s boardroom conversation will include some version of, “What’s our AI strategy?” Even the most cautious industries are running pilots, if only to avoid falling behind.

On the demand side, expectations continue to ratchet higher. We stream movies instantly, get packages in hours, and expect a chat reply in seconds. Being told to hold is now commercial malpractice. 83% of customers expect to resolve complex problems in a single interaction, without having to bounce around like human pinballs.

The status quo is no longer tenable. Something has to fill the gap between customer impatience and corporate cost constraints.

Okay, I’m not saying that AI agents are a magic bullet. Hallucinations, trust, and governance are real challenges. But the total addressable market is swelling from $12 billion in 2024 to nearly $48 billion by 2030. That scale, paired with the acute pain of existing systems, means the shift is going to happen.

The “why” has already been answered. The real conversation has moved on to the “how.”

Most attempts to build agents fall into two traps.

The first is wiring an LLM directly into a workflow. It works for demos, but real-world inputs are messy. Accuracy drops fast once you add noise or multi-step logic. The Air Canada case, where a chatbot gave false refund policies and the company was forced to pay damages, shows how expensive a single failure can be.

The second is stitching together frameworks and no-code tools into a “DIY agent.” This gives you control, but at a heavy cost. Engineering teams become the bottleneck, CX teams get locked out, and the maintenance overhead balloons. What looked simple turns into endless QA, audits, incident playbooks, and rollback drills.

The demo is always seductive, but it’s only the visible tip trap of the iceberg.

Clay (Sierra co-founder), in his recent interview, mentioned:

"We've had many of them come back nine months later and it's like hey... it was deeper and darker than we expected under there".

The real challenge is not picking the right LLM. It is managing the operational friction of a probabilistic system at enterprise scale. Without discipline, teams end up in firefighting mode, spending more time maintaining agents than improving their product.

Sierra’s answer is a vertically integrated platform. Instead of bolting parts together, they built Agent OS as a full stack where dialogue, business logic, and integrations live in one place. The focus is on speed-to-value for the business user, without sacrificing reliability.

Every path to building an agent carries a trade-off between speed and control.

Sierra’s thesis is that this trade-off is artificial. They argue you shouldn’t have to pick between speed and enterprise-grade rigor. Their answer is a full-stack platform (Agent OS, SDK, and Experience Manager, we’ll get into these soon) designed around one specific job: customer service.

The focus matters. Because I see three distinct markets in AI today:

This last layer is where I’m convinced most of the opportunity is today. This moment in AI feels like the Age of Exploration (15th-17th centuries), the glory days of Europe, when explorers like Christopher Columbus charted unknown lands and established connections between different parts of the world.

Back then, ships were marvels. Built at great cost, captained by daring men, they carried empires’ hopes into the unknown. A few owned the ships. Others drew the maps, the tools to navigate, but rarely correct.

But the real wealth didn’t come from the ships. It came from the settlements, which became economic engines. They generated trade, taxes, and territorial control.

Models are the ships. Tools are the maps. Agents are the settlements where people actually live and pay rent.

Settlements win because:

Sierra is built this way. They charge per issue resolved. If the agent fixes the problem, they earn; if not, they don’t.

That builds a feedback cycle:

More resolutions → better data on edge cases → higher accuracy → stronger margins → expansion.

(h/t Acquired podcast for these notes, was a great listen)

Sierra started over grilled octopus and sparkling water at a Mediterranean restaurant in Palo Alto. In early 2023, Bret Taylor received a call from Clay Bavor, a Google executive he'd known for nearly two decades. What Taylor assumed was a casual check-in turned into a multi-hour deep dive into what ChatGPT had just made possible, which had coincidentally launched just as Taylor was leaving Salesforce.

By the time the check came, Taylor’s sabbatical was over. Sierra had its first heartbeat.

This wasn’t some random tech bros following a trend. Their entire careers were a two-part solution to a problem that hadn't been solvable until that exact moment.

They met at Google in 2005. Taylor helped build Google Maps, then founded FriendFeed (acquired by Facebook) and Quip (sold to Salesforce for $750M) before becoming co-CEO of Salesforce itself. There, he led the Customer 360 initiative, where he had a front-row seat to the Gordian knot of enterprise customer data.

Bavor stayed at Google for 18 years, running AR/VR and Google Labs. He was in the room with early LLMs like LaMDA and PaLM. He watched a model explain the 2008 financial crisis as “a lot like the movie Inception, except instead of dreams within dreams, it was debt within debt.”

That was the moment he knew: the phase shift had happened. Models were actually reasoning.

While the idea was born over lunch, the actual plan came from Taylor’s paranoia.

FriendFeed had taught him a brutal lesson. Beautiful tech means nothing without distribution. This time, he picked up the phone. For weeks, he grilled CIOs and enterprise execs with one question:

“If you had a magic AI wand, what fire would you put out first?”

The breakthrough came from a late-night Zoom with Anthony Tan, the CEO of Grab. Tan didn’t hesitate: Customer service.

It was the perfect wedge: a universally despised, money-hemorrhaging, and most obvious mess in every single enterprise.

Taylor and Bavor had stumbled into a rare alignment, rare synthesis of capabilities. Taylor was a walking index of enterprise pain. Bavor had watched language models think in metaphor while the rest of the world was still impressed they could fetch the weather.

But their vision was far bigger. Customer service wasn't their end goal. It was the point of insertion. When you fix the most broken thing in the room, you do more than sell a product. You earn their trust.

And once you’re the trusted intelligence solving a company’s hardest problems, you’ve earned the right to plug into everything else, their entire customer lifecycle.

Sierra raised $110M in early 2024 while still in stealth. By October, just months after launch, it secured another $175M at a $4.5B valuation (more than many long-standing public SaaS firms).

Investors see Bret and Clay as the kind of entrepreneurs who could build “the Salesforce of AI support”.

The company crossed $20M in annualized revenue within its first year.

This means Sierra has a 225x revenue multiple, way above established players with rational (lower) valuations. However, AI-first companies typically have 25-40% higher valuations than non-AI peers, with some cases exceeding 100x revenue multiples.

This valuation seems less a reflection of its financial metrics and more a measure of the belief in the founding team's ability to create an entirely new category of AI-native records for customer relationships.

For one, they’re not selling a chatbot maker. They’re offering an AI operating system (Agent OS) for enterprises to run smarter, more reliable conversational agents at scale.

The core of that system is what Sierra calls a Constellation of Models.

Rather than relying on one dominant LLM, Sierra blends frontier models like GPT‑5 or Claude with specialized open-source or proprietary models. One model might parse intent, another craft the response, and another supervise for hallucinations and policy alignment.

Since new models appear almost weekly, the ground is always shifting. Sierra absorbs that volatility. If a stronger model arrives, they plug it in. If one breaks, they route around it. Customers don’t care which model is running; they just want their problems solved. Sierra makes sure the tech changes in the background without disrupting the business.

And Agent OS isn’t bound to chat windows. It’s omnichannel. You design your agent once, and it works via voice, chat, email, SMS, and more, to provide a consistent customer experience across all channels.

Within the Agent OS, they have one product for devs (Agent SDK) and one for non-devs (Agent Studio).

This is the platform-as-a-service for developers. It uses a declarative programming language. This means you define what needs to happen, not every step.

Set the goal, like “process a return request.” Add guardrails, like “no returns past 30 days.” The SDK then orchestrates a multi-step workflow: check order status, validate product damage images from the customer, and initiate a refund. If the rules allow, the agent can adapt when the logic needs flexibility.

The Agent SDK uses composable skills that can be combined into custom workflows. Once built, these agents plug directly into standard development practices: version control, CI/CD pipelines, release gates, and rollback.

They also integrate cleanly with enterprise systems such as CRMs, order databases, or APIs. That lets the agent move beyond conversation to actual execution, like updating a subscription or escalating to a human.

Each agent release is captured as an immutable snapshot. Teams can roll back instantly if a new version misbehaves.

For non-technical teams, Sierra built a no-code interface. CX and operations leaders can lay out customer Journeys, like “Subscription Cancellation”, in plain English.

That’s how the apparel brand Chubbies created their agent, Duncan Smuthers. The ability for non-devs to build and adjust agents directly gives companies speed that would otherwise be bottlenecked inside engineering.

Inside the studio, agents can be hooked into live systems like order databases, CRMs, APIs. So they don’t just talk, they act. The platform also includes tools for testing: simulated conversations, automated tagging to surface common topics, and AI-based grading to evaluate performance before an agent ever goes live.

The studio has testing features like simulations to run conversations, automated conversation tagging (to discover common topics), and AI evaluations that grade the agent’s performance on sample dialogues.

The biggest risk with customer-facing AI agents is not traditional software bugs, but probabilistic errors. An agent that is “mostly right” still creates brand-level disasters in the moments when it is wrong.

Sierra tackles this with supervisory models. These are additional layers of AI and rules that monitor the main agent’s output in real time. They catch hallucinations, off-policy replies, or sensitive content before it reaches the customer.

Clay Bavor summed it up well:

“Large language models are remarkably better at detecting errors in their own output than at avoiding those errors in the first place.”

A supervisor that flags 90% of errors on top of an agent that is 90% accurate

= near 99% reliability.

Supervisors can take many forms: classifiers that answer questions like “does this violate policy?”, explicit filters for banned words, or workflows that require certain actions to be approved by a human.

The control stack in practice

Just like a little Jiminy Cricket looking over the agent’s shoulder.

If any test fails, the system blocks, rewrites, or escalates to a human. Every extra check adds latency, but Sierra keeps the drag minimal. Small classifiers handle obvious bans; bigger models kick in for edge cases. Supervisors run in parallel, not in sequence, and cache results for repeated calls. The entire mesh adds just 80 to 120 milliseconds, which is effectively invisible in chat and voice.

Other providers often rely on single-model “constitutional” prompts or static filters. They use one LLM with a baked-in constitution or a filter like Llama Guard. That catches offensive content, but it doesn’t verify tool calls or track business-state transitions. Sierra’s mesh does all three, in one pass, with audit logs for compliance teams to trace every decision.

Supervisors that share the same training data as the primary model can inherit its blind spots. Sierra mitigates this by mixing providers and retraining small specialty models on freshly discovered failure modes. The team is experimenting with cross-model voting (letting two unrelated LLMs critique each other) plus cryptographic “watermarks” on approved replies to detect tampering in transit.

Sierra also builds in escalation. The system can hand a conversation to a human at exactly the right moment.

This matters because the dream of 100% automation is misguided. As Balaji (ex-Coinbase CTO) notes:

“The optimal amount of AI is not 100%. After all: 0% AI is slow, but 100% AI is slop. So the optimal amount of AI is actually between 0-100%.”

This is where Sierra completely torches the old business model.

For decades, customer service software has been priced on capacity, not outcomes. Vendors like Salesforce or Zendesk sell seats: licenses for human agents. Companies pay whether call volumes are high or low, whether problems are solved or not. The vendor wants to sell more seats, the company wants to push through tickets, and no one is directly incentivized to deliver better results.

Sierra breaks that loop. Sierra is following an outcome-based pricing model. Customers pay only when the AI agent achieves a specific business outcome, like a resolved case, a saved cancellation, or a cross-sell. If the conversation has to be escalated to a human agent, in most cases, there's no charge.

This is not unheard of in the BPO world, but rare among software providers. And it’s a sign of their confidence that the product is delivering actual value.

This gets even more profound because it turns customer support from a cost center into a value generator.

Let's look at the unit economics.

A US-based customer service rep makes about $21/hour. The fully loaded cost (including benefits, training, and overhead) is approximately $30/hour.

The average call lasts approximately 6 minutes, allowing one agent to handle 10 calls per hour. This puts the fully-loaded cost per human interaction at roughly $3

Sierra’s agents typically handle about 70% of cases autonomously. The remaining 30% still go to humans at $3 each. This brings the effective cost for that portion to $0.90 per interaction ($3 * 30%).

Of course, Sierra itself isn’t free. Similar AI platforms usually charge between $0.50 and $1.25 per fully resolved case. Using a midpoint estimate of $0.85, the blended cost comes out to approximately $1.48 per interaction.

That’s still roughly half the cost of handling everything with humans, and in high-volume cases with lower AI pricing, savings could reach 60–70%.

For the last 20 years, enterprise software has been about selling better shovels, but you still had to do the digging. Now, the software is digging the hole for you.

The only thing that matters for businesses is whether the job gets done. Did you save money? Did you increase sales? Did you retain more customers? If not, the vendor doesn't get paid. This is a fundamental reordering of the value chain.

In enterprise, security and compliance are not optional.

When Sierra’s agents act within backend systems, such as issuing a refund or updating an account, those steps are executed through controlled API calls with strict templates. Even if the conversational layer goes off course, the agent cannot touch money or data outside its permissions. Every action is deterministic, logged, and auditable.

Customer data is also protected by default. Sierra automatically encrypts and masks personally identifiable information (PII). If a customer shares a phone number or medical detail in chat, it is redacted in the logs. This safeguard matters in an era where hackers have already tricked general-purpose AI systems into leaking sensitive data.

On top of this, Sierra carries the expected certifications: SOC 2 Type II, HIPAA, and GDPR. It has also secured ISO 42001:2023, the new international standard for AI Management Systems.

Sierra treats voice as a first-class channel. Since launching in late 2024, voice has been its fastest-growing segment.

The company invested directly in custom voice models through its acquisition of Receptive AI. These models integrate with existing telephony and call-center systems, which means a Sierra voice agent can transfer a call to a human seamlessly.

Only a few competitors, such as PolyAI, match this level of voice capability. Most others lean on generic third-party voice services that are less natural in conversation. By building its own stack, Sierra can tune voice for customer service: matching brand tone, handling interruptions gracefully, and even customizing pronunciation for product or company-specific terms.

The hardest problem in the agent world is not raw capability but dependability. Enterprises need to know what “good” looks like, and most benchmarks don’t measure it.

Standard tests like WebArena are too shallow. They assume a single-turn interaction where all information is handed to the agent upfront. Real customer interactions are nothing like that. Users are vague, messy, and often contradictory.

(I know this well because I am a messy customer myself)

Sierra built τ-bench to close that gap. Instead of one-off tasks, it measures whether an agent can perform the same complex job repeatedly, even when the user is being difficult. The results were telling: vanilla GPT-4 agents succeeded <50% of the time, and only 25% when asked to repeat tasks reliably. No enterprise can tolerate an agent that fails 3 out of 4 times on the same request.

This pushed Sierra to invest heavily in reliability engineering, a core differentiator when selling to large customers. The effort also shaped the industry: Anthropic has adopted τ-bench internally to stress-test its models.

Sierra then introduced τ²-bench, which measures collaboration between human and agent. It asks whether the AI can guide and empower a person through a process, not just complete tasks alone. LLM providers like Kimi are now adopting τ²-bench alongside other evaluation suites.

For competitors, τ-bench and τ²-bench are research milestones. For investors, they signal that Sierra is steering the field rather than chasing it. And for customers, they are proof that the product has already been through the gauntlet.

Less than two years after launch, Sierra is already embedded inside some of the world’s largest enterprises. Nearly half its customers generate more than $1 billion in annual revenue, and close to 15% exceed $10 billion.

The results tell the story.

At Casper, Sierra resolves more than half of all customer issues, driving a 20% lift in satisfaction. The agent also surfaces a new kind of customer intelligence: people open up about their sleep habits more freely to an AI than to a human rep.

At WeightWatchers, Sierra contained nearly 70% of support volume within its first week. The tone was so empathetic that members started sending the AI heart emojis. The pilot worked so well that WeightWatchers quickly scaled Sierra across more use cases and touchpoints.

At Sonos (they make great speakers, I have a set in my room), Sierra powers onboarding during the crucial first 30 days, or what they call “time-to-music.” The agent even troubleshoots beyond Sonos hardware. Like diagnosing home Wi-Fi issues, delivering the kind of surprise-and-delight moment that builds loyalty.

Sierra agents embody the brand. Their personas can be customised. Chubbies gave theirs a playful identity named Duncan Smuthers, cracking jokes in line with their irreverent style. OluKai (footware) designed theirs with the warmth of Hawaiian “Aloha.”

Sierra is operating in one of the busiest corners of enterprise AI, but the real competitors fall into three groups, each shaped by the constraints of their business models.

The giants of customer service, like Salesforce and Zendesk, rely on seat-based pricing. Their AI strategy has been one of containment: bolt-on features like Einstein or Zendesk AI that improve agent productivity rather than replace agents entirely. The model protects their license revenue, which ranges from $2,000 to $5,000 per agent each year.

This reinforces their main business because they have their client's operational dependency and terabytes of proprietary data to create a sticky ecosystem where the AI's value is intrinsically tied to the platform, making switching to a new vendor extremely difficult.

Even Intercom, which has experimented more boldly, layers outcome-based pricing on top of its platform by charging $0.99 per resolved ticket with its Fin AI product. The gesture is forward-looking, but still tethered to a system that depends on seats. The reluctance to fully embrace outcome-based economics leaves an opening for new entrants.

This defensive posture creates an opening. As Bret Taylor points out,

"Changing a business model is inherently difficult for established companies which leads to missed opportunities."

Specialized AI platforms

PolyAI, Moveworks, and Ada dig deep into narrow problem sets. PolyAI has built some of the best natural-sounding voice agents in the market. Moveworks focuses on IT and HR helpdesk automation, with high resolution rates in that vertical. Ada is strong in orchestrated flows for chat. Their advantage is depth. The drawback is fragmentation: customers end up stitching together multiple tools, with conversations breaking when they cross domains. Outcome-based pricing is rare, and most rely on conventional SaaS economics.

Emerging Agent Platforms

Decagon, Cognigy, and Kore.ai are younger entrants chasing full autonomy from the start. Decagon AI, for example, has reported helping Chime reduce contact center costs by 60 percent while doubling NPS. Cognigy and Kore.ai market broad orchestration layers that cover chat and voice. These platforms show strong momentum, but most are still packaging LLMs without the full enterprise architecture to make them dependable at scale.

This is where Sierra stands apart. It offers a complete stack: Agent OS, SDK, Experience Manager, Supervisor AI for safety, and τ-bench as a reliability benchmark. This focus on delivering a secure and verifiable solution is what has attracted Fortune 500 enterprise clients and a $4.5 billion valuation.

Sierra does not own the base models. It relies on providers like OpenAI and Anthropic, which means the raw intelligence is not proprietary. In theory, another startup could wire the same models into its own system and claim similar capabilities.

On paper, that makes Sierra look differentiated more by go-to-market moves (outcome-based pricing, credibility of its founders) than by hard technology.

In practice, the moat lies in execution. Running agents reliably at enterprise scale is not trivial. Sierra has built up the scars and know-how to make a probabilistic system behave predictably. That depth of operational engineering is difficult to copy.

The approach does create challenges. Deployments are hands-on and require significant customization, which makes scaling labor-intensive. And there are customer segments that may resist automation altogether. Luxury and premium brands often prize human interaction as part of their identity and may view AI-driven service as a downgrade.

Even with those caveats, Sierra’s traction shows it has solved problems that competitors have only circled.

Every era has a dominant interface. Microsoft owned the command line. Apple defined the smartphone. Google captured search. Natural language is next, and the platforms that matter will be the operating systems running AI agents.

Sierra’s wedge is to combine autonomy, safety, and enterprise credibility under an outcome-based model. That combination explains why it has been able to win billion-dollar logos so quickly, despite entering a market already crowded with “AI-powered” solutions.

What they’re really building is an intent clearinghouse. Support stops being a ticket queue. It becomes a real-time exchange, matching user intent to backend action.

And it works. Enterprise software isn’t supposed to be lovable. But Sierra’s agents get thank-you notes and heart emojis. That says everything.

Next time I’m stuck on hold, listening to the lie that my call is important, I’ll wonder if a Sierra agent is spinning up to prove it.

If so, I’ll press 1.

Thanks for reading,

Teng Yan & Ravi

Want More? Follow Teng Yan & Chain of Thought on X

Subscribe for timely research and insights delivered to your inbox.

Check out our YouTube channel if you like to watch more than read.

If you found this essay interesting, please share it with a friend who you think will appreciate it. We pour many hours into our research, and we want as many people as possible to read it. More deep dives on AI agent companies coming soon!