A company says it has “AI agents.” Look closer and it may have a chat tool that drafts emails. There is nothing wrong with that. The problem is not knowing what capability we actually have.
Four levels of action
- Model: takes an input and produces an output. It is an engine; on its own it does not know your objective or have permission to act.
- Assistant: a model with instructions, context and perhaps tools. It helps when a person asks for something; that person drives the process.
- Workflow: a predefined sequence of steps. AI may be used within it, but the route has already been designed.
- Agent: receives an objective, observes the state of a task, chooses the next step within agreed limits, acts and checks the result.
Requests arrive in a service team. A model classifies a message. An assistant drafts a reply when someone asks it to. A workflow classifies, assigns and prepares a draft according to fixed rules. An agent reviews open cases, prioritizes those requiring attention, consults permitted information, proposes actions and checks what remains unresolved.
This is not a ranking of prestige. Choose the simplest design that solves the problem well and makes important decisions visible.