Good agent use cases
AI prototypes are plenty, impact is rare. It’s hard to know what’s worth building — what's good for users and for the business. That's also true for AI agents: everyone is building them, but few can show real impact. So when should you (not) use AI agents? A few useful filters:
Open-ended input. If the inputs are narrow and predictable, a simple logic/heuristic may be enough. Agents shine when inputs are diverse and messy.
Complexity. A single decision is not an agent — it’s just classification. “When was Obama born?” is a lookup, not a workflow. Also if the process is deterministic, agents add no value.
On the other extreme, if the control flow space explodes (five steps with ten tools each already gives a million paths), it's unmanageable. Agents live in the middle ground: structured enough to be tractable, flexible enough to adapt.
Solution space. When there’s only one correct answer, it’s hard to find it. When there are many paths to get to the answer, agents have a higher chance. Think creative copy, sales pitches, or brainstorming ideas — where humans can weed out bad answers, but not rank many good ones.
Low cost of mistakes. The best use cases are where errors are cheap: read-only (like research) or reversible actions like versioned code. These make testing and adoption far easier.
§
Finding strong use cases that fit most of these criteria while delivering value is tricky. It requires a deep understanding of existing products and workflows.