Prioritizing Enterprise Service Stacks Over Frontier Model Development

The true competitive advantage in the AI era is not owning the model, but mastering the service-as-software stack. Clay Bavor, co-founder of Sierra, explains that while the industry obsesses over frontier model capabilities, the real value lies in the capability overhang: the gap between what models can do and what enterprises actually deploy. By embedding forward-deployed engineers directly into client workflows, Sierra bypasses traditional vendor-client friction and creates a feedback loop that improves product quality over time. This conversation matters for leaders who need to distinguish between theoretical AI productivity and the operational reality of building durable, agentic systems that deliver board-level ROI.

The hidden cost of off-the-shelf strategy

Most startups feel pressured to build their own foundation models to gain credibility. Bavor argues this is a mistake for all but a few companies. The capital expenditure required to maintain a perishable bag of floating point numbers creates a structural drag that distracts from the actual problem: solving enterprise-specific workflows.

By using the massive capital investments of hyperscalers and fine-tuning open-weights models, Sierra avoids the mega-cluster trap. This creates a lasting advantage: while competitors burn cash on training, Sierra invests in the application layer, the agentic intelligence that interacts with customers.

"If you can't build frontier models yourself, okay maybe the next best approach is to distill them and offer them up. I think that's probably the main driver of the difference."

-- Clay Bavor

The forward-deployed moat

Conventional wisdom suggests that enterprise software should be sold as a finished product and thrown over the wall to the customer. Bavor rejects this. He notes that because no one has deployed AI agents at this scale, the implementation is the product. By embedding engineers directly into client organizations, to the point where they are effectively part of the customer team, Sierra gains a deep understanding of the snowflake complexities inherent in large enterprises.

This approach is uncomfortable and expensive in the short term, but it creates a massive moat. It accelerates time-to-value, allowing them to go live in weeks rather than months, and it provides a stream of proprietary data on how to handle specific industry edge cases that competitors cannot access.

Why token budgeting is the new OPEX

As AI moves from chat interfaces to autonomous agents, token consumption is no longer just a variable cost; it is a proxy for organizational leverage. Bavor observes that top-tier engineers are already hitting $100,000 in annual token spend.

The system-level shift is clear: CFOs will soon treat token budgets as a fundamental line item, equivalent to headcount. When developers spend more on tokens than their own base salary, they are not just using AI; they are augmenting their output by orders of magnitude. The companies that successfully integrate this into their financial planning, rather than treating it as an experimental line item, will outpace those still gatekeeping token usage.

"I think for CFOs in the future like capital allocation will look more like how do we allocate OPEX and then headcount, and headcount will be both headcount for salaries and SBC and also tokens associated with headcount."

-- Clay Bavor

The engineering interview as an AI-native event

Traditional coding interviews are becoming obsolete. If the goal is to assess how an engineer builds in an AI-first world, asking them to solve algorithmic puzzles on a whiteboard is a lagging indicator. Sierra’s shift to AI-native interviews, where candidates are given a budget and tasked with building an application using their own agentic stack, reveals a fundamental truth: the skill of the future is not writing syntax, but orchestrating intelligence to build at pace.

Key action items

• Implement "Think Apart, Think Together" (Immediate): To avoid groupthink and capture independent insights, have key decision-makers draft their positions separately before converging. This forces clarity and prevents the loudest voice from dominating.
• Establish a 6-Week Board Cadence (Next Quarter): If your industry is moving at AI speed, quarterly updates are too slow. Shift to a tighter cadence to update priors and course-correct before small issues compound into systemic failures.
• Shift to "Board Memos" over Decks (Immediate): Replace presentations with 6-10 page memos. Writing is thinking; this forces leadership to confront the uncomfortable truths they might otherwise gloss over in a slide deck.
• Audit Your Token Spend (Next Quarter): Treat token usage as a proxy for AI adoption. If your top engineers are not hitting significant token spend, they are not leaning into the leverage available to them.
• Pilot a Forward-Deployed Team (12-18 Months): For enterprise-grade products, stop selling software and start selling outcomes. Embed engineers into your most complex customers' environments to identify the unsolved problems that will define your next product iteration.
• Normalize Token Budgeting (12-18 Months): Begin the transition to treat token spend as a standard component of total compensation or headcount cost, rather than a discretionary engineering expense.