Sierra Co-founder Clay Bavor on Open Models, Enterprise AI and Building Teams

Clay Bavor says he and Bret Taylor met through Google’s associate product management program and stayed in touch for years. After several earlier opportunities to work together, they decided to start Sierra when language models were improving and both the partnership and timing felt right.

Bavor recalls enjoying his 18 years at Google, where managers gave him responsibility and opportunities to learn. He had wanted to start another company since childhood, but waited until he saw a suitable partner and a technology shift that could create an opening for a smaller team.

One lesson from Google was to invest as deeply in the technology stack as needed to deliver a product. Bavor says Sierra built its own agent frameworks and architecture early, before agents became a common topic. This is his account of the company’s engineering approach.

Sierra briefly considered pretraining a foundation model but decided against the continuing capital expense. Bavor says the company instead uses investments made by model labs and cloud providers, while fine-tuning open-weight models for its own needs. He distinguishes that work from training a frontier model from scratch.

Stebbings asks whether cheaper open models customized for companies undermine frontier providers. Bavor expects organizations to use both: routine tasks may not require the most capable system, while complex work may justify higher cost. His view of future demand is a forecast, not a demonstrated universal outcome.

Bavor suggests model capabilities that were once expensive can become cheaper and move into open-weight systems. The right choice depends on a task’s difficulty and the cost of errors. His cost comparison is illustrative and does not establish identical quality across models or applications.

The host asks what share of enterprise work open models can perform. Bavor says relatively little work is completely automated and identifies several possible constraints: models, applications and organizational adoption. He does not provide a measurement that separates those factors.

Bavor argues that demand for high capability in coding, discovery and other complex tasks could be very large. He calls it effectively unbounded, but that is an expression of his expectation rather than evidence of literally unlimited demand.

On token economics, he distinguishes the price of an individual token from total spending on reasoning and agent workflows. A model that uses more compute at inference time can consume more tokens even if hardware becomes more efficient. The relationship between additional compute and quality depends on the model and task.

Bavor identifies hardware efficiency, open-weight models, energy and available compute as influences on price. If demand grows faster than supply, he says, infrastructure constraints can limit how quickly costs fall. This is an economic scenario, not a guaranteed price floor.

Stebbings asks whether running models on devices can ease the pressure on data centers. Bavor expects local hardware to improve some consumer applications, but believes demanding workloads will still need substantial data-center compute. His view about future hardware and demand remains a prediction.

The conversation turns to open-weight models from the US and China. Bavor speculates that differences may partly reflect large-scale distillation of frontier models; he offers no model-specific evidence. That suggestion is his impression and should not be treated as proof of any company’s training practices.

Stebbings asks whether AI enables dramatically smaller teams. Bavor reports that some Sierra engineers estimate they ship features several times faster with coding tools. These are employees’ estimates of a specific activity, not a controlled study of total productivity or a prediction for every organization.

Sierra serves large and sometimes regulated companies, Bavor says, and successful deployment requires understanding their systems, goals and relationships. He reports working with 40% of the Fortune 50 and describes customer revenue distributions; those are company figures not independently verified in the interview.

Bavor introduces an internal Sierra agent called Pinecone and a shared MCP gateway connecting company systems. He says employees use their own permissions through that gateway rather than gaining access to everyone’s private documents. The interview describes intended controls, not an independent security audit.

The internal agent uses a purpose-built harness for engineering and shared skills that employees can develop. Bavor says he uses one to help review hiring packets, while retaining his own approval role. He describes the tool as increasingly useful, not yet indispensable.

He also describes a strategy assistant grounded in company documents, board letters and operating reviews. The purpose is to help him reason about priorities with context, not to hand organizational decisions entirely to a model. Its quality and safeguards are not independently assessed in the interview.

Asked whether to limit employees’ token spending, Bavor says high usage has sometimes signaled experimentation. He has heard of or observed some engineers spending at an annualized rate above $100,000, but says Sierra’s own use is more modest. The figure is not an industry-wide budget or requirement.

He predicts future compensation planning may include both pay and an AI usage budget. For now, he says, Sierra prefers learning quickly to imposing tight limits. He expects the engineering bottleneck to move from writing code toward reviewing and deciding what should be built.

Stebbings cites reported AI spending at another company and asks whether token costs will reach a larger fraction of developer pay. Bavor predicts a figure closer to 20% than the smaller example offered by the host. This is a speculative estimate, not a forecast established by financial data.

Asked how an enterprise vendor stays close to customers, Bavor says he and Taylor still build and use agents themselves. He acknowledges that founders cannot recreate every complex customer environment in-house and must also understand the end user’s experience.

The host describes a crowded market of startups and incumbents. Bavor agrees competition is intense but claims Sierra is larger and growing faster than comparable young rivals. He offers no independently verified comparison, and his expectations about market concentration are his own.

Bavor says Sierra learned a forward-deployed approach by working closely with early design partners. Engineers were embedded in customer environments to understand business processes and integrate the first version of the platform. He describes this as a way to learn as well as deliver.

He says the product can be used without a forward-deployed team, but that expert help can accelerate early implementation. Bavor cites examples of deployments taking weeks rather than months; these are company-reported timelines, not independent measures of implementation quality or outcomes.

He believes demand for customer-facing agents is strong and says Sierra is expanding internationally, including through an acquisition in Japan. His description of market leadership and demand reflects the company’s perspective and is not a verified industry ranking.

Bavor describes applications beyond support, including assisting customers with discovery, refinancing and product recommendations. He presents these as examples of a broader customer lifecycle strategy. The interview does not independently assess conversion, accuracy or other customer outcomes.

The interviewer compares Sierra with Palantir. Bavor says the company borrowed elements of a forward-deployed model but aims to serve more customers by accumulating reusable expertise in specific industries. His comparison is a strategic ambition, not an established measure of either company’s scale.

The company looks for common features that can strengthen its platform, he says, while sometimes building a unique extension for a large customer. Bavor believes coding agents can make such work more feasible and that a solution for one client may later help another. Those are management expectations.

Sierra holds board meetings approximately every six weeks, alternating longer and shorter sessions. Bavor says that cadence helps the company reassess its assumptions when model capabilities and customer needs change rapidly.

Instead of slide decks, the founders send written board memos in advance. Bavor believes writing forces clearer thought and gives directors time to prepare. The meetings then focus on substantial questions rather than simply presenting results.

He says the memos also identify areas where the company could improve. One early example was failing to hire recruiters quickly enough despite signs of demand. Bavor presents the episode as a lesson in acting on information already available to the team.

Asked how Sierra thinks about financing, Bavor says the founders focus on the capital needed to reach the next milestone and consider dilution without trying to maximize valuation at all costs. He says they accepted lower prices than they might have obtained; the interview provides no independent financing comparison.

The company’s stated values include craftsmanship, intensity, family, trust and customer focus. Bavor says excellence consists of many details across people, processes and product, and that the care shown to customers helps establish trust in how the company will serve their users.

As an example, he recalls senior staff personally monitoring agent conversations during an early major retail period. The anecdote illustrates the attention the team says it paid; it does not establish the agents’ accuracy or customer outcomes.

Bavor describes intensity as maintaining pace and high standards in a competitive market. He says founders must set an example, but should choose carefully where their direct involvement changes the outcome instead of entering every detail.

He says ambitious goals can help a team work backward from what must be true. His example is expanding in Japan and considering an acquisition to build local presence. He also warns that deadlines can cause work to expand to the time available.

Bavor and Taylor developed values by first thinking independently and then comparing their notes. Family reflects their belief that demanding work can coexist with important commitments outside the office. He presents this as the culture they want to build, not a guarantee of every employee’s experience.

Bavor says having children encourages him to use working hours efficiently. He describes maintaining connectivity during a long commute and setting aside time for family. This is his personal arrangement rather than a prescription for others.

He favors in-person work at a young company because he believes it helps build shared norms, relationships and informal apprenticeship. He acknowledges the challenge of creating the same connection remotely. Other organizations and employees may make different tradeoffs.

Bavor recommends learning from skilled colleagues and cites Richard Hamming’s talk on research. He compares early learning and sustained effort with compounding: skills acquired early can support later work. The advice reflects his own experience.

Asked what young people should do in an uncertain labor market, he encourages them to become comfortable with AI tools and seek opportunities to learn. He says some young Sierra employees are particularly effective with those tools. His examples do not settle the broader question of entry-level employment.

Sierra changed its engineering interview to let candidates build an application with their chosen AI tools and a token budget. Bavor says the company still examines architecture, systems design, product thinking and cultural fit, rather than measuring only tool usage.

On cybersecurity, Bavor expects the importance of defensive work to grow as offensive capabilities improve. He says he is not an expert and is uncertain whether general AI tools or specialized products will provide the best defenses. This is a forecast, not a product assessment.

Bavor recalls a disagreement with Taylor over whether to improve a process or change leadership to move faster. He says they eventually concluded that both mattered. Their approach is to challenge each other’s assumptions and seek a workable answer.

The co-founders divide responsibilities into overlapping majors and minors rather than separate silos. Bavor says Taylor spends more time on sales and engineering, while he focuses more on running the company, though both remain involved in product and important decisions.

Asked about Sundar Pichai, Bavor praises his ability to move between long-term strategy and product detail while staying attentive to people. His account is a personal lesson drawn from working together.

He recommends David McCullough’s The Wright Brothers as a history of invention, persistence and building on earlier technology. Bavor sees the story as a useful portrait of entrepreneurship and the repeated experiments behind an eventual achievement.

As a parent, Bavor says he creates rituals such as family dinners and making things with his children. He tries to learn about their individual interests so he can participate and support them. His approach reflects his family’s circumstances.

He describes marriage as a partnership with shared goals while allowing each person to pursue distinct interests. For him, helping the other person thrive is part of that shared project. This is a personal perspective rather than a universal rule.

Bavor thanks his parents for supporting his interest in computers even though technology was not their profession. He recalls their willingness to encourage a fascination they did not fully understand and the difference that made to his early learning.

He remembers family trips to buy a computer and visit Macworld, including his father encouraging him to get more memory. Bavor sees that early encouragement as an important contribution to his later career at Google and to starting Sierra.