THEORY · 60 minutes
Understand how components work together using simple, everyday analogies
When my support-triage assistant handles a message, it isn't doing one magic step — it's moving through a sequence of smaller ones. Understanding that sequence is what let me debug it when it went wrong, because I could ask "which step failed?" instead of throwing up my hands at a black box.
I find it easiest to think of an AI System's process like a skilled cook handling a dinner order: understand what was asked, decide on an approach, plan the steps, do the work, and get a little better for next time. We'll trace those same five steps — perception, reasoning, planning, execution, and learning — through one real customer message.
No math required. We'll follow the assistant's thinking with everyday analogies — a cook, a GPS — not equations. The goal is a mental model you can use to reason about where a system succeeds or fails.
The first step is perception — figuring out what the message actually means. This is like a server taking your order: they have to turn your words into something the kitchen can act on.
When a customer writes "my export keeps failing every morning," the assistant's perception step works out:
Get perception wrong and everything downstream is wrong — the system would be solving the wrong problem confidently. That's why I treat it as the most important step to test.
Once the assistant understands the message, it has to reason about the best response. This is like a GPS weighing routes — there's more than one way to get there, and it picks based on what it knows.
For the export-failure message, the reasoning step considers: Is this a known issue in the knowledge base? Is it urgent? Am I confident enough to answer, or should a human take it? The output isn't an answer yet — it's a decision about how to proceed.
AI Systems are strong at recognizing patterns, the way you spot a friend's face in a crowd without consciously analyzing it. The assistant recognizes that "export fails every morning" matches the shape of a known scheduled-job problem — even if no customer ever phrased it that exact way before.
Pattern example: if many "fails every morning" reports have historically traced back to an overnight job, the system can lean toward that explanation — while still leaving room to be wrong.
After reasoning about how to proceed, the assistant lays out a concrete plan — like a recipe with steps in the right order. For our message, the plan might be: search the knowledge base for "export failure" → if a strong match is found, draft an answer → check confidence → send or escalate.
A good plan respects order. You can't draft an answer before you've searched for one, just as you can't frost a cake before it's baked. The planning step is where the system decides what must happen before what — and that ordering is exactly what the workflow engine from Module 2 enforces.
This is where the plan becomes real work. The assistant runs the steps: it calls the knowledge base through its connector, gets results, and drafts a candidate reply. Execution is the "hands" doing what the "brain" decided.
Like a steady project manager, the system has to cope when a step doesn't go to plan. If the knowledge-base search returns nothing useful, the assistant shouldn't invent an answer — it should fall back to escalation. The way a system behaves when execution fails tells you more about its quality than how it behaves when everything works.
The last piece is learning from what happened — the way you get better at cooking by paying attention to which dishes land. For an AI System, "learning" mostly means humans feeding signals back in: which answers were right, which escalations were unnecessary, which were caught just in time.
I want to be careful here, because this is where a lot of AI writing oversells. The assistant doesn't magically improve on its own overnight. It improves because someone reviews where it was wrong and adjusts — the knowledge base, the escalation threshold, the prompts.
Where the signal comes from:
Honesty check: be skeptical of any tool that claims an AI System silently "learns" and keeps getting better with no human in the loop. Real improvement comes from real feedback and deliberate tuning.
Put the five steps together on our one message: the assistant perceives a recurring export failure, reasons that it matches a known issue and it's confident, plans a search-then-answer sequence, executes by pulling the known fix and drafting a reply, and — over time, with human feedback — the team learns to refine when it answers versus escalates.
Before the quiz, reason about a failure: if the perception step misreads "my export keeps failing" as a billing complaint, what happens to every step after it? Trace it through — this is exactly the kind of cascading error that makes perception the step I test hardest.
1. Why does Brandon call perception 'the most important step to test'?
Correct Answer: B
Perception feeds everything downstream. A misread message means the system reasons, plans, and acts on the wrong problem — a cascading error.
2. In the reasoning step for the export-failure message, what is the actual output?
Correct Answer: B
Reasoning produces a decision about approach, not the final answer. The answer comes later, during execution.
3. The planning step insists on searching the knowledge base before drafting an answer. What principle is this?
Correct Answer: B
Planning orders steps by dependency — you can't draft from search results you haven't fetched yet. The workflow engine then enforces that order.
4. During execution, the knowledge-base search returns nothing useful. What is the well-designed behavior?
Correct Answer: B
When execution can't find a reliable answer, escalation is the safe fallback. How a system handles failure reveals its real quality.
5. Brandon is skeptical of tools that claim an AI System 'learns on its own with no human in the loop.' Why?
Correct Answer: B
Genuine improvement comes from reviewing errors and adjusting the knowledge base, thresholds, and prompts — humans in the loop, not magic.
Excellent work. You can now trace an AI System's thinking step by step — and you know which step to suspect first when something goes wrong. Next, we'll look honestly at the benefits these systems bring.