Learning Path
AI Agents Explained Path
The term AI agent is used for everything from a simple FAQ bot to a complex autonomous system — and that dilution is the problem. This path explains it plainly: what an agent is, how it differs from a chatbot, a workflow and robotic process automation, what agents can actually do, when they fit, how to measure their return honestly, and how to govern them so they do not act without you.
How this path works. Each lesson below is a short summary of the topic it links to, with an exercise that proves you can do the work. Work through them in order, do the exercise, and mark each lesson complete when you have read the guide and done the work. The two-question check under each lesson is for yourself — it is not scored and it is not a certificate. Your progress is stored only in your own browser, in local storage — no account, nothing sent to us. Completing this path gives you no certificate and no qualification, and it promises no income: the proof is the finished work from the capstone lesson, not a certificate.
What you will learn
- What an AI agent is, in plain English.
- Agent vs chatbot vs RPA vs workflow: the comparison table.
- What agents can actually do — perceive, reason, act, verify, learn.
- Use cases: good fit versus bad fit.
- Honest ROI and governance: permissions, escalation, logging, cost caps.
Your progress
0 of 6 lessons complete.
The lessons
Read each lesson summary, then open the full guide it links to. The summaries are written for this path — they do not repeat the guides, they prepare you for them.
Lesson 1: What an AI agent is — the plain-English definition
An AI agent is software that works toward a goal with tools: it perceives a situation, reasons about what to do, acts — by looking something up, running a process, or making a change — verifies the result, and learns from it. The difference between an agent and a language model answering questions is the doing: a model answers, an agent acts, and reports back. And the first warning is that the term is diluted by overuse — applied to everything from a simple FAQ bot to a complex autonomous system — so the plain definition is the one that matters.
This lesson fixes the definition with examples: an agent that triages support tickets, an agent that processes invoices, and a system that only looks like an agent because someone labelled it one. You will learn the test: does it perceive, reason, act and verify? If it only answers, it is a chatbot, however the marketing describes it.
Exercise: Write the plain-English definition of an AI agent, with one example that fits it and one that does not.
Check yourself — two questions on this lesson
1. What does an AI agent do?
Perceive, reason, act, verify and learn — with tools and a goal.
2. What is the test?
Does it act and verify? If it only answers, it is a chatbot.
Lesson 2: Agent vs chatbot vs RPA vs workflow — the comparison table
Four tools are routinely confused, and the comparison table sorts them: a chatbot answers questions in conversation; robotic process automation repeats fixed steps — it does not decide; a workflow is a defined sequence of steps that runs when triggered; an agent decides what to do next, uses tools, and acts toward a goal. Same word, different job — and buying the wrong one is how a small business pays agent prices for a chatbot's work.
This lesson builds the table on one example — a support inbox — and shows what each tool would do with it, so the differences are visible instead of theoretical. You will learn the buying rule: name the job first, then match the tool, and never buy an agent for a job a workflow does better and cheaper.
Exercise: Fill in the comparison table for one task: what a chatbot, RPA, workflow and agent would each do with it.
Check yourself — two questions on this lesson
1. What does RPA do?
It repeats fixed steps — it does not decide.
2. What is the buying rule?
Name the job first, then match the tool.
Lesson 3: What agents can actually do — perceive, reason, act, verify, learn
An agent's loop is five steps: perceive — read the situation, the ticket, the invoice; reason — decide what to do; act — use a tool to do it; verify — check the result; learn — remember whether it worked. The loop is what separates the agent from the chatbot: the chatbot's loop ends at the answer, and the agent's loop ends at the outcome. But the loop is also the risk: an agent that acts without verification is a machine that makes mistakes at machine speed.
This lesson walks the loop on a real task — triaging a support inbox — and shows what each step looks like in practice, including the step that matters most for a small business: verification, and the human who checks it. You will learn why the learning step is the slowest to trust, and why an agent that has not yet been verified is an agent that should not act alone.
Exercise: Walk the five-step loop for one task in your business, naming the verification step.
Check yourself — two questions on this lesson
1. What is the agent's loop?
Perceive, reason, act, verify, learn.
2. Which step matters most for a small business?
Verification — and the human who checks it.
Lesson 4: Use cases — good fit vs bad fit
Agents fit narrow, repeatable work with a clear result: support triage, invoice processing, lead qualification, document extraction. They do not fit open-ended judgement, decisions with real consequences, or anything where an error is expensive and hard to undo. The fit question is not whether the agent is clever — it is whether the task is narrow, repeatable, and cheap to check, so that a mistake costs little and is caught fast.
This lesson sorts use cases into fits and misfits on the test above, using real small-business tasks. You will learn the good-fit pattern — a task a junior staff member could do with a checklist — and the bad-fit pattern — a task that needs judgement, trust, or a decision you would not delegate to a stranger. If a task fits the junior-with-a-checklist test, it fits an agent with a human checking.
Exercise: Sort five tasks in your business into good fit and bad fit for an agent, with one line each.
Check yourself — two questions on this lesson
1. What makes a task a good fit?
Narrow, repeatable, and cheap to check — a mistake costs little.
2. What is a bad fit?
Judgement, trust, or consequences you would not delegate to a stranger.
Lesson 5: Honest ROI — baseline, pilot, measured KPI (no vendor percentages)
Vendor ROI figures for agents are never sourced — one vendor claims invoice processing eliminates 80% to 90% of manual data entry, and this path does not republish that as a fact. The honest method is simpler and slower: measure a baseline before, run a small pilot, measure the same number after, and compare. The KPI is counted, not claimed, and the comparison is between your own two numbers, not between your number and a vendor's brochure.
This lesson builds the honest measurement: what to count before the pilot, how long to run it, and how to read the result — including the result that says the agent did not help, which is a success, because it stopped a bad purchase. You will learn why a pilot that proves nothing is worth more than a brochure that proves everything.
Exercise: Define your pilot: the baseline, the KPI, the length, and the decision the result will drive.
Check yourself — two questions on this lesson
1. Why not use vendor ROI figures?
They are never sourced — they are brochures, not evidence.
2. What is the honest method?
Baseline, pilot, measured KPI — your own two numbers compared.
Lesson 6: Governance — permissions, escalation, logging, cost caps
An agent that can act needs rules about what it may touch. Governance is the design of those rules: permissions that give the agent the least it needs — read, not write, where read is enough — escalation that puts a human in the loop before any consequential action, logging that records everything the agent did, and a cost cap that stops the bill from running away. Never give an agent write access to live business systems without human approval — that sentence is the whole of the rule in one line.
This lesson builds the governance checklist for one real small business: what the agent may read, what it must ask about, what it may never touch, how its actions are logged and reviewed, and where the cost cap sits. It ends with the capstone: a use-case picker result for one business, and the governance checklist for it.
Exercise: Capstone: complete a use-case picker result for one real small business, with the governance checklist for it.
Check yourself — two questions on this lesson
1. What is the first governance rule?
Least privilege — give the agent the least access it needs.
2. What must never be given without human approval?
Write access to live business systems.
Try the tools
These free tools let you practise what the lessons describe — everything runs in your browser, and nothing you type is sent anywhere.
- Startup Cost Calculator — total the cost of piloting an agent
- ROI Calculator — judge a pilot by its measured return
- Business Model Comparison — compare agent-led models with others
Check what you learned
6 questions, all based on the lessons above. Answer them all, then check your score — each answer explains itself, so a wrong answer teaches as much as a right one. The quiz is practice, not a test: there is no pass mark and no certificate.
Answer all 6 questions, then check your score.
Download your notes
Download a plain-text summary of this path: the lesson list, which lessons you marked complete, and your quiz score. The file is generated on your device — nothing is uploaded.
Key terms in this path
The terms this path uses are defined in the Glossary, each with an example. If a lesson uses a word you do not know, look it up there first.
Frequently asked questions
What is an AI agent?
Software that perceives a situation, reasons about it, acts, verifies the result, and learns — with tools and permissions, not just words. A chatbot answers; an agent does, and reports back.
How is an AI agent different from a chatbot?
A chatbot answers questions. An agent takes actions — it can look up data, run a process, and complete a task. The comparison lesson in this path lays the full table out.
Are AI agents the same as robots?
No. In software, an agent is code with tools and goals, not a physical machine. Robotic process automation is a different tool again — it repeats fixed steps, it does not decide.
Can a small business actually use AI agents?
Yes, for narrow tasks — support triage, invoice processing, lead qualification, document extraction — with a human in the loop and permissions that limit what the agent can touch.
How do you measure whether an agent is worth it?
By a pilot: a baseline before, a measured result after, and a KPI that is counted rather than claimed. Vendor percentages are not evidence — the ROI lesson shows the honest method.