An AI architect combines technical judgment with decisions about an agent's voice, behavior, and business goals.
2
The best AI strategies start with a narrow customer problem, accept some risk, and reorganize teams around improving the agent.
3
Building an agent involves a much larger system than choosing a model, including testing, upgrades, voice handling, and tools for nontechnical teams.
Summary
Clay Bavor describes the AI architect as the person responsible for a company's customer-facing agent, much as a webmaster once shaped a company's early web presence. The role combines an understanding of agent capabilities with decisions about personality, brand, customer experience, and business outcomes. Bavor says successful teams start with a specific problem, such as processing a product return, then improve from real usage. They also create new teams that review conversations and coach the agent. He is skeptical of the idea that companies can quickly build their own agents by assembling a model, framework, and database. Much of the work sits below the surface, including regression tests, model changes, interruptions, and voice complexity. Sierra uses simulated conversations before launch and feedback from live interactions afterward. Bavor expects agents to move beyond chat and voice, with glasses and other wearables providing an always-present personal interface.
The AI architect is the webmaster of the agent era
Bavor compares the AI architect with the webmaster role from the early internet. A webmaster shaped a company's digital storefront while making choices about its technology and presentation. The AI architect does something similar for an agent, but the work also includes coaching and improving a probabilistic system. Bavor says the role has appeared inside Sierra customers as teams took responsibility for building, managing, and refining their agents. The person needs a working feel for what agents can do, even without training a model or using a framework directly. The role also covers the agent's voice, values, tone, persona, and connection with customers, along with the business outcomes the company wants.
The role combines technology, experience, and business judgment
Bavor describes three parts of the job. The first is understanding agent technology and its capabilities. The second is experience and aesthetics, including how the agent sounds and what personality it presents. He gives Chubbies as an example: its agent, Duncan Smothers, makes irreverent jokes and uses a funny tone that fits the brand. The third part is business. A company needs to decide what outcomes the agent should drive while speaking with customers. This makes the role broader than software architecture. It requires taste about the customer experience and a clear view of what the business is trying to achieve.
The strongest AI architects begin with a small real problem
Bavor says successful teams do not wait for perfect technology. They accept that agents are probabilistic software and take enough risk to try them in practice. Their work begins with a concrete customer or business problem rather than a general instruction to apply AI. One customer started with processing a single return. The team celebrated when the agent could answer the phone, arrange a replacement pair of shoes, and provide a shipping label. Bavor says the first use case does not need to be complex. The team can start somewhere narrow, learn from it, and expand.
Bavor says companies should not simply carry old team structures into work with agents. Some successful customers have reorganized customer experience and service teams around helping the agent improve. At one customer, people review a couple hundred conversations each day. They coach the agent on what to say, which decisions to make, how to show empathy, and how to use better judgment. Bavor says this team did not exist in the company's earlier way of working. He is most interested in people who have been close to customers through support, service, care, or retail, then become hands-on enough with the technology to shape the agent.
The visible parts of an agent hide a large engineering system
Bavor calls the gap between the apparent simplicity of an agent and the work underneath it the agent iceberg. A technical team may think it only needs to pick a language model, choose LangGraph or LangChain, select an embeddings model and vector database, then connect tools. Under the surface are regression and unit testing, model migration and upgrades, and many voice problems. These include separating primary and secondary speakers and handling interruptions. Sierra's Agent OS combines tools for technical builders with no-code tools for nontechnical people who refine, coach, edit, and update the agent. Bavor says companies that chose to build everything themselves have sometimes returned after nine months because the work was deeper than expected.
Sierra treats agent work as a new software development life cycle because the software is nondeterministic. Its testing harness creates many user personas, simulated accounts, and even simulated devices with different states, such as an amber light being on or off. Teams first map the important customer journeys and model them in code. The agent can handle unexpected turns in a conversation, while remaining deterministic where rules require it, such as compliance language. Before an agent goes live, Sierra runs tens or hundreds of thousands of simulated conversations. The results expose missing knowledge and corner cases that need more work.
Model progress should shape what teams build today
Bavor recommends staying hands-on with new models and tools, reading research papers, and watching adjacent areas such as video models. He says knowing where capabilities are going matters more than knowing only where they are now. At Sierra, he keeps a Google Doc with problems that were too hard for GPT-4, then checks whether newer models can solve them. This gives the team a way to track the slope of improvement, alongside factors such as latency. He says Sierra began with the expectation that token costs would fall and model capabilities would grow, so the company tried to build toward the near future rather than only the current model limits.
Agents will use many interfaces, with glasses as a personal access point
Bavor expects agents to move beyond the current chat box and voice call. He describes them as shape shifters that can summon text, voice, video, images, user interfaces, and other ways of interacting. He also expects glasses and wearables to become a strong vehicle for a trusted personal AI. Such a system could perceive what a person sees and hears, then whisper or provide a visual prompt. Bavor argues that an assistant meant to help throughout the day should not require repeatedly retrieving a phone from a pocket or bag. He sees glasses as part of a future in which the assistant stays present as people move through the world.
"I think the AI architect I would say is kind of the AI era and AI agent version of the webmaster in a way."02:37
Who should watch
You are responsible for defining an AI agent's customer experience and need a practical description of the role.
Your team is deciding whether to build an agent internally and needs to account for testing, upgrades, voice behavior, and ongoing coaching.
You want a way to evaluate changing model capabilities against product decisions and future interface ideas.
