A model takes input and generates output. An assistant combines the model with instructions, context and sometimes tools; a person directs the work. A workflow follows anticipated steps. An agent can decide what to do next within a defined goal and boundaries. These are operational distinctions for this course; market labels vary.
The important question is where the path is decided. A rule that specifies extract, validate and draft remains a predefined path even when extraction uses AI. A system that decides to consult another source, request clarification or change its plan makes dynamic choices about execution.
One problem, different solutions
Consider a weekly opportunity report. A model summarizes an supplied list. An assistant works with a manager to refine the analysis. A workflow queries records, checks fields and prepares the same report. An agent investigates exceptions and chooses steps based on findings. None of these labels grants permission to send messages or change prices.
An assistant is often enough when a person can direct the task and supply context. A workflow fits repeated processes with definable exceptions. Consider an agent when the path cannot reasonably be anticipated and tools make progress verifiable.
The cost of autonomy
Additional steps introduce coordination, delay, review and possible failures. Before expanding a design, identify the new decision the system needs to make and how its effect will be checked. Wanting an impressive label is not a functional requirement.
Map input, decision, output and owner. Use stable rules for rule-based classification. Test models with examples when interpretation is required. When an output affects business commitments, specify the approver. Architecture should make accountability visible.
Worked case
Andina Equipos holds a Monday meeting using an approved opportunity list. Its format and rules are stable. A workflow detects overdue dates and prepares drafts; the manager reviews exceptions. Calling it an agent does not add value.
If an opportunity requires investigating several documents and choosing the next source from findings, a bounded agentic capability may be useful. It needs authorized reading and an explicit stopping condition.
Practice instructions
Describe a recurring task: input, expected steps, exceptions and output.
Recommend the simplest sufficient design.
Explain which decision would justify an agent and how its result would be verified.
Do not assume write or send permissions.
Your turn
Classify and justify: A) draft a message when a salesperson asks; B) always run extract-validate-summarize; C) investigate a discrepancy and choose sources based on findings. Specify human involvement in each.
Show the commented solution
A is an assistant, with salesperson review before sending. B is a workflow, with an owner for invalid records. C is an agent if it chooses the path and has tools; a person resolves discrepancies affecting commitments. Unbounded investigation without verification is not a reliable design.