The term "AI agent" is used to describe everything from a simple chatbot to a fully autonomous software engineer. That range makes it nearly useless as a description. Here is a more precise version.
The core definition
An AI agent is a program that receives a goal, observes its environment, decides what action to take next, takes that action, and then re-evaluates. The key difference from traditional automation is the "decides what to do next" part. Standard automation follows a fixed sequence of steps. An agent chooses its steps based on what it observes.
What agents can and cannot do
Agents are good at:
- Tasks that require reading and understanding unstructured text
- Multi-step workflows where the next step depends on the result of the previous one
- Processes where the inputs vary significantly between instances
- Work that requires synthesising information from multiple sources
Agents are not good at:
- Tasks requiring creative judgment (visual design, tone-sensitive writing, novel strategy)
- Workflows where legal or regulatory review is required at each step
- Anything requiring physical world interaction
- Tasks where the cost of an error is catastrophic and irreversible
A concrete example
A customer support agent receives an email. It reads the email, identifies the issue, checks the customer record in your CRM, looks up the relevant policy, drafts a response, and sends it — or escalates to a human if the situation does not fit a clear resolution path. That is an agent: it reads, reasons, decides, acts, and knows when to hand off.
The business question to ask
Before deploying an agent, ask: what does this process look like today? Who does it, how long does it take, and what does a mistake cost? If the process is high-volume, reasonably well-defined, and tolerates occasional errors, it is a candidate. If it is low-volume, highly variable, or error-intolerant, it probably is not.