Home›AI Agency Academy›Lesson 01
Module 01 · Lesson 01

What an AI agent is — and is not

Choose the smallest useful system for the work instead of calling every automation an agent.

Last updated August 5, 2026•15–25 minutes•Free AI agent course
What you will learn

Build a system-choice record

You will be able to tell when a simple rule, a human-written response, an AI assistant, or an agent is the right fit. That decision keeps the build cheaper, easier to test, and easier to explain to a client.

Why this matters

Good agent work is useful before it is impressive.

A lead-routing rule does not need a model. A high-stakes refund should not be handed to an agent. The useful middle ground is work that needs judgement from bounded context, has a repeatable outcome, and still has a named human owner.

Field note 01

Choose between a rule, assistant, agent, and human decision

Choose between a rule, assistant, agent, and human decision
Choose between a rule, assistant, agent, and human decision
Core concepts

The language that keeps the work clear.

AutomationA predictable trigger and rule: when a form says enterprise, assign the enterprise queue.
AssistantA system that drafts, summarizes, or recommends while a person makes the decision.
AgentA system that can reason through a limited job, select from allowed actions, and report what it did.
OwnerThe person responsible for the result, corrections, and exceptions after the agent is live.
The practical method

How to build a system-choice record

Name the job

Write one sentence: ‘When this happens, help this person reach this outcome.’

List the judgement

Underline the parts where rules alone break down: unclear intent, incomplete facts, or several valid next steps.

Separate actions

Mark which actions are safe to automate, which need approval, and which must stay human.

Choose the simplest design

Use a rule where a rule works. Add an assistant or agent only where its judgement has a measurable purpose.

Worked example

Worked case: what an ai agent is — and is not

A home-services business wants faster replies to web enquiries. The first version does not need an ‘AI sales employee.’

A form rule stores location and service type. An assistant turns the customer’s free-text note into a short summary. An agent may ask one approved follow-up question when the request is unclear, then routes the record to the right dispatcher.

The dispatcher still owns urgent requests and pricing. The result is a useful hybrid workflow, not a vague automation promise.

Build it in practice

Complete the working artifact

Job: [trigger] → help [user] achieve [outcome]. Judgement needed: [where rules fail]. Allowed actions: [safe actions]. Approval needed for: [actions]. Human owner: [name or role].
Architecture check

Choose the least autonomous design that can do the job.

A deterministic rule is easier to test than a model decision. A workflow can use a model inside fixed steps. An agent earns its added freedom only when the path cannot be written in advance and the value of that judgment outweighs the extra failure modes.

  • Use a rule for stable, explicit decisions.
  • Use a workflow when the steps are known but one step needs model judgment.
  • Use an agent when the system must choose its path inside a narrow job.
  • Keep decisions with material legal, financial, safety, or relationship consequences under an accountable person.

Sources used for this check

Practice

Before you move on

  • Choose one repeated workflow from your own business or a prospective client.
  • Classify every step as rule, assistant, agent, or human decision.
  • Remove any agent behaviour that has no clear reason to exist.
  • The job has one user and one intended outcome.
  • The agent is not being used where a rule is enough.
  • A human owner is named.
  • High-impact actions are not implied to be autonomous.

Lesson progress

Finished this lesson?

Save your place on this device so it is easy to pick up where you left off.

Not marked complete yet.