Open AI chair explains why every business will need an AI agent

Generative AI, AI Agents, and the Future of Work: A Discussion with Brett

Key Concepts:

• Generative AI: Artificial intelligence capable of creating new content (text, images, code, etc.) rather than simply executing pre-defined rules.
• AI Agents: AI systems designed to perform specific tasks autonomously, often interacting with users in a conversational manner.
• Foundation Models: Large AI models (like GPT-4, Claude) trained on massive datasets, serving as the base for various applications.
• Hallucinations: Instances where AI models generate incorrect or nonsensical information.
• Non-Deterministic Models: AI models that produce different outputs even with the same input, posing challenges for testing and reliability.
• Defense in Depth: A security strategy involving multiple layers of protection to mitigate risks.
• Outcomes-Based Pricing: A pricing model where payment is tied to the successful delivery of results, rather than simply usage of a product.

I. The “Eureka” Moment & The Potential of Generative AI

Brett recounts a pivotal moment in 2022 when OpenAI’s demonstration of “Dolly,” a model capable of generating images from prompts (specifically, an “avocado chair”), fundamentally shifted his perspective on computing. He highlights the shift from computers as “executors of rules” to creative entities capable of generating novel ideas. This realization sparked his excitement about the potential for generative AI to unlock previously impossible inventions by making complex tasks easy. He frames this as taking things that were impossible and making them easy, creating ingredients for an “infinite number of recipes.”

II. Sierra: Building Customer-Facing AI Agents

Sierra focuses on developing AI agents for customer interactions. Brett draws a parallel to the early days of the internet, arguing that, just as every company needed a website in 1995, every company will need an AI agent by 2025. These agents will handle a vast majority of digital customer interactions, surpassing the capabilities of traditional websites. The core value proposition is improved customer experience, efficiency, and cost reduction.

III. Navigating the Hype, Hope, and Disappointment in AI Adoption

Brett acknowledges the current mixed sentiment surrounding AI – a blend of excitement and frustration. He attributes this to the nascent stage of the technology, comparing it to the early days of mainframes, PCs, and the internet. Initially, companies had to build everything from scratch. He cites a 1997-98 Wired article detailing banks spending $20-$50 million simply to add login functionality to their websites, a task now achievable in hours with modern tools like Shopify and Codeex. This illustrates the pattern of high initial costs and dissatisfaction followed by the emergence of off-the-shelf solutions. He predicts a similar trajectory for AI agents, with specialized vendors offering solutions for specific use cases within the next 4-5 years (e.g., Harvey for legal, Sierra for customer experience). Currently, many companies are struggling to integrate raw AI components, but this will evolve.

IV. Addressing Technical Challenges: Hallucinations, Security, and Robustness

Brett addresses concerns about AI limitations, specifically “hallucinations” (incorrect outputs) and the non-deterministic nature of models (producing different results from the same prompt). He emphasizes that human imperfection exists as well, drawing a parallel to regulations in financial services designed to prevent advisors from making false promises. He advocates for a proactive approach to AI control, mirroring the “defense in depth” security strategy: preventing errors, detecting them when they occur, and limiting their impact. He stresses the importance of “guard rails” and narrow use cases to improve reliability, comparing the challenge to moving from a “science problem” (AGI) to an “engineering problem” (building agents for specific processes).

V. The Frontier of AI Agent Capabilities & Industry Adoption

The current frontier lies in expanding the scope of AI agent applications. While simple tasks like return processing are relatively low-risk, more complex areas like medical triage or financial advice require caution. Brett notes surprisingly high adoption rates in regulated industries, with companies starting with lower-risk use cases to build experience. He believes AI can improve consistency and control compared to human interactions, even with inherent imperfections. He highlights the potential for AI to augment human roles, such as assisting bankers with data collection during mortgage applications.

VI. Evaluation, Pricing, and the Shift in Software Value

Brett emphasizes the importance of defining clear business outcomes when evaluating AI solutions, rather than focusing solely on technical metrics. He uses self-service resolution rate as an example, cautioning against metrics that are easily “gameable.” He advocates for coupling business metrics with customer satisfaction scores. Sierra employs “outcomes-based pricing,” charging clients only when the AI agent successfully resolves a customer issue. This reflects a shift in value from software as a productivity tool to software as a solution that delivers tangible results. He argues that companies generally prefer buying solutions to problems rather than building and maintaining software themselves.

VII. The Future Landscape: Foundation Models vs. Specialized Agents

Brett addresses the question of whether powerful foundation models will eventually render specialized AI agent vendors obsolete. He acknowledges the potential for AI-generated code to lower software development costs but believes companies will continue to prefer buying solutions over building and maintaining software internally. He predicts a thriving ecosystem of vendors offering specialized AI agents for specific tasks and departments (e.g., financial auditing, customer service, lead generation). He frames this as a shift from providing “wood” (models) to delivering “houses” (complete solutions).

VIII. AI’s Impact on Jobs and the Need for Reskilling

Brett anticipates that AI will reshape the job market, with some roles becoming automated. However, he rejects the notion of widespread job losses, arguing that new roles will emerge. He emphasizes the importance of continuous learning and adaptation, particularly for white-collar workers. He notes that the rapid pace of change in AI requires individuals to proactively acquire new skills. He highlights the opportunity for existing employees, like call center managers, to transition into roles focused on designing and managing AI agents. He believes that the current disruption is unique in that everyone, including those developing the technology, is experiencing the same uncertainties.

IX. How AI is Improving Brett’s Management Style

Brett utilizes AI as a “creative foil” for his own writing, using it to critique his ideas and identify flaws. He also leverages AI-powered summarization tools to manage information overload. He acknowledges that while AI can enhance productivity, he still values the deliberate thinking process involved in writing and prefers to generate content himself.

Notable Quote:

“We’ve taken things that were impossible, we’ve made them easy, and now those are ingredients to almost an infinite number of recipes that we can now make.” – Brett, on the transformative potential of generative AI.

Conclusion:

Brett’s perspective paints a picture of a rapidly evolving AI landscape. While acknowledging the current challenges and uncertainties, he remains optimistic about the long-term potential of AI agents to transform businesses and improve customer experiences. He emphasizes the importance of focusing on business outcomes, building robust solutions, and embracing continuous learning to navigate this disruptive era. The future, according to Brett, is one of “agents everywhere,” offering specialized solutions to a wide range of problems.