How Can Customer-Facing AI Agents Delight Users?

TL;DR

Customer-facing AI agents can give people a simpler way to shop, obtain support, manage policies, and complete other business interactions through natural conversation. Sierra helps companies build branded agents that represent their voice, values, policies, and processes, while additional AI-based checks can help detect errors that language models may fail to prevent during initial generation.

Transcript

one of the more interesting learnings from the past you know year and a half of working on this stuff is that the solution to many problems with AI is more Ai and it's somewhat unintuitive but one of the remarkable properties of large language models is that they're better at detecting errors in their own output than in not making those errors in t... Read More

Key Insights

• Large language models are capable of reasoning through metaphors and analogies. An early Google model summarized the 2008 financial crisis through an Inception comparison, connecting nested dreams with nested debt and demonstrating an ability to combine concepts rather than merely repeat a memorized description.
• Generative image systems can combine concepts that have never appeared together in a specific image. Early DALL-E examples created avocado-shaped chairs pixel by pixel, showing Bavor that AI models could understand qualities associated with separate concepts and synthesize them into something new.
• Sierra is a platform for creating branded, customer-facing AI agents. Companies can use these agents for interactions ranging from customer service to commerce, giving customers a conversational interface that represents the business rather than requiring them to navigate conventional menus and organizational structures.
• Conversational AI changes how customers access products and services. A shopper can describe the desired type of running shoe and refer to an earlier purchase instead of navigating a hierarchy of categories, while an insurance customer can request a policy change without locating the correct application screen.
• AI agents can understand language, generate responses, reason, and make decisions. These combined capabilities distinguish them from static websites, social-network interactions, and conventional mobile applications, creating a new way for businesses to communicate with customers and help them complete practical tasks.
• Customer experience improves when customers do not need detailed knowledge of a company’s catalog or internal processes. Sierra’s approach allows the agent to interpret a person’s intended outcome, connect it with relevant business information, and guide the interaction through ordinary conversation.
• Additional AI can solve some problems created by AI. Bavor says large language models are better at detecting errors in their own output than at preventing those errors during the original generation, making AI-based evaluation and checking important parts of reliable agent systems.
• Customer service is a strong early business application for generative AI because existing solutions combine high costs with low satisfaction and substantial room for return on investment. The central deployment challenge is allowing useful customer interactions while controlling the errors and hallucinations associated with language models.

Questions & Answers

Q: What is Sierra’s customer-facing AI platform?

Sierra enables companies to create branded AI agents that interact directly with customers for purposes ranging from customer service to commerce. The agents are designed to represent a company’s voice, values, policies, and internal processes. They provide a conversational alternative to navigating websites, mobile applications, product catalogs, and complicated organizational procedures.

Q: How do customer-facing AI agents improve customer service?

Customer-facing AI agents let people state their goals in ordinary language instead of learning how a company organized its website, application, catalog, or internal procedures. An agent can interpret the request, reason about relevant information, make decisions, and continue the conversation until the customer receives help or completes an intended business outcome.

Q: Why is customer service an early application for generative AI?

Customer service is an early business application for generative AI because the costs of existing solutions are high, customer satisfaction is low, and the potential return on investment is broad. Conversational agents can handle customer interactions more naturally, although businesses must address the errors and hallucinations that large language models can produce before trusting them with customers.

Q: Why can more AI help correct problems caused by AI?

More AI can help because large language models may detect errors in their own generated output more effectively than they prevent those errors during the initial response. This means an agent system can use additional model-based checking or evaluation to identify potential mistakes. Bavor describes this as a notable and somewhat unintuitive lesson from building AI systems.

Q: What can AI agents do that traditional software cannot?

AI agents can accept requests in natural language, generate appropriate language, reason about the customer’s situation, and make decisions. Traditional interfaces generally require users to follow menus or structures chosen by a designer. An agent instead allows the customer to explain a desired result conversationally and lets the software determine how to address the request.

Q: How can an AI agent simplify online shopping?

An AI agent can let a shopper describe the desired product directly, including qualities such as being a lightweight running shoe or resembling something purchased previously. The shopper does not need to navigate categories designed around the company’s product catalog. The agent can interpret the description and help identify suitable products through conversation.

Q: What convinced Clay Bavor that large language models were different?

Bavor observed early Google language models respond effectively to requests involving humor, metaphor, and analogy. One model described black holes in three words and connected the 2008 financial crisis with the nested structure in the movie Inception. These responses suggested that the models could connect concepts and reason in ways that felt fundamentally different.

Q: What business information should a branded AI agent represent?

A branded AI agent should capture the company’s voice, values, nuanced policies, and internal processes so that its customer interactions genuinely represent the business. Sierra’s approach extends beyond producing conversational text. The agent can use those business-specific instructions while helping with support, commerce, policy-related requests, and even customer-retention situations described in the episode.

Summary & Key Takeaways

• Clay Bavor co-founded Sierra with longtime friend Bret Taylor after spending 18 years at Google across Search, ads, Workspace, virtual and augmented reality, Google Lens, and Google Labs. Early language models demonstrated reasoning through metaphor and analogy, convincing Bavor that conversational AI represented a fundamentally new foundation for software and business interaction.

• Sierra enables companies to create branded, customer-facing AI agents for activities ranging from customer service to commerce. Instead of forcing customers to understand product catalogs, application menus, or corporate processes, these agents let people describe what they want conversationally, reason about the request, make decisions, and help produce the desired outcome.

• Deploying generative AI in customer interactions requires sophisticated engineering because large language models can produce errors and hallucinations. One important lesson is that additional AI can address some AI problems because models may detect errors in generated output more reliably than they avoid those errors initially. Sierra packages these capabilities through AgentOS and outcome-based pricing.