3 Requirements for an AI Agent Use Case

AI agents are everywhere in today’s tech conversation. From customer service to content creation, they’re touted as the next big thing. Yet, the actual adoption of generative AI agents in commercial settings has been more measured. What’s holding back widespread deployment? And more importantly, what does it take for an AI agent to be not just a cool demo but a truly commercially viable solution?

At Decagon, we’ve had the privilege of deploying AI agents in production to many of the largest names in the tech industry (like Rippling, Eventbrite, etc) and have distilled the success of AI agents down to three essential criteria. These are the pillars on which any AI agent must stand in order to serve real customers today.

1. A Buffer for Non-Determinism

The allure of large language models (LLMs) lies in their ability to generate human-like text, but this comes with the challenge of non-determinism. Unlike traditional software, where the same input produces the same output every time, LLMs can vary in their responses. In commercial applications, this variability can be a hurdle—especially in customer-facing scenarios where consistency is key.

To make an AI agent commercially viable, you must design it to handle this uncertainty gracefully while embracing the benefits of the non-deterministic outputs. We’ve all encountered those unpredictable outputs from models like ChatGPT or Claude, which is why it’s crucial to implement robust guardrails within your AI system (a discussion for another time) and ensure a buffer to account for variations in output accuracy.

One effective approach is to incorporate a robust escalation mechanism to have humans in the loop. Take customer support, for example. If an AI agent isn’t sure how to resolve an issue, it can escalate the problem to a human agent. This safety net not only ensures customers receive accurate assistance but also builds trust in the system. In essence, a commercially viable AI agent delivers value even when it isn’t perfect, thanks to its ability to escalate when needed.

Coding co-pilots are another great example of this, which has contributed to their adoption. AI agents can write code or suggest auto-completions, but ultimately there is a human engineer reviewing it before it goes live.

In both of these cases, this is room for error because of how the use case is designed. Contrast this with an AI agent tasked with parsing cybersecurity logs to autonomously flag and review potential threats. On the surface, it seems like an ideal application for LLMs, given that logs are just blocks of text. However, the level of accuracy required—particularly in avoiding false negatives—is extraordinarily high, making it much more challenging for an AI agent to deliver meaningful value in this context.

2. Readily Available Data Sources and Tools

An AI agent’s utility is directly tied to the data it can access and the tools it can leverage. Without the right integrations, an AI agent is little more than a flashy LLM wrapper. To transition from a proof of concept to a production-ready solution, the agent must be able to interface with the systems and data sources essential to its function.

Consider a customer service AI agent: to be effective, it needs to tap into a company’s knowledge base, CRM, and other data repositories. These integrations must be straightforward to implement, and the data must be readily accessible. Moreover, your customers must be both willing and able to connect their data to your AI agent. If some of this data is already public, like a help center, that’s even better—it lowers the barrier to entry and accelerates deployment.

In short, without robust data and tool integration, an AI agent is like a car without fuel. It may look impressive, but it won’t get you very far.

3. Repeatable and Scalable

Finally, for an AI agent to be commercially viable, it must address a problem that is scalable across an organization. A solution that’s too niche, even if technically impressive, will struggle to gain traction. For a business to justify investing in an AI agent, the use case must be broad enough to impact a significant portion of its operations.

If the AI agent solves a problem that only a few employees encounter, it’s harder to justify the investment. For instance, commercial viability often hinges on the ability to scale a solution across multiple teams or departments, making it crucial that the use case is both broad and deep. If you’re talking to a 5,000-person organization and only 3 employees do the work that your agent is tackling, that’s usually not a great sign.

The View from Inside

The path to a commercially viable AI agent is not without its challenges, but the potential rewards are substantial. By embracing the uncertainty of non-deterministic systems, ensuring seamless integration with critical data and tools, and focusing on scalable use cases, you can create AI agents that do more than just impress—they deliver real, measurable value.

At Decagon, we’re fortunate to collaborate with large, sophisticated enterprises on deploying generative customer support AI agents at scale. Our insights stem from countless iterations with our customers—experimenting with ideas, some successful, others less so. This process of trial and error has shaped the lessons shared in this post.

The AI agents that succeed will not only survive the hype cycle but thrive in the marketplace, driving innovation and creating new opportunities across industries.