Five lessons, about 55 minutes. AI will draft your SQL, your formulas, your Python — the skill this module teaches is catching what it gets confidently wrong.
An AI tool is excellent at pattern-completing syntax — drafting a plausible SQL query, a plausible Excel formula, a plausible pandas snippet, instantly, from a rough description. It is not reliably good at knowing whether that syntax is correct against your specific schema, your specific data's quirks, or the business rule Chidi actually meant.
| Genuinely strong at | Drafting syntax fast, explaining unfamiliar code, suggesting approaches you hadn't considered, catching typos in code you show it. |
| Unreliable at | Knowing your actual schema without being shown it, knowing which of two plausible interpretations of a vague request is right, catching its own logical errors, admitting uncertainty unprompted. |
The tool is fast and confident regardless of whether it's right. Confidence was never the signal to trust — checking always was, and that hasn't changed.
An AI-drafted query for "total revenue by branch this month" might genuinely be perfect — or it might quietly omit the VOID filter from lesson 3.2, or join to promotions the naive way from lesson 3.3. It reads exactly the same either way: clean, confident SQL.
AI drafts: SELECT branch_id, SUM(revenue) FROM transactions GROUP BY branch_id;
This is syntactically correct and will run without error. It is also missing the transaction_type filter entirely — a mistake this whole course has spent modules 3 and 4 training you to catch. The AI didn't know Sahara Mart has VOID rows unless told; it drafted the most generic reasonable-looking version of the question.
The discipline: treat every AI-drafted query or formula exactly as you'd treat one copy-pasted from a stranger online — plausible until independently checked against a small, known-correct case you can verify by hand.
Before trusting any AI-assisted number that's going to Chidi, check it three ways: does it match a rough mental estimate, does it match a simpler alternative calculation, and does it match what a small manual sample of the raw data shows?
Three independent checks that agree is genuinely strong evidence. One AI-generated number that merely looks plausible is not evidence at all — it's a guess dressed in confident formatting.
This is the use case AI tools are genuinely most reliable at: given a query, formula, or script you didn't write, ask it to explain what the code actually does, step by step. You're not trusting it to be correct about the business — you're using it to shortcut reading unfamiliar syntax, which is a much lower-stakes, more verifiable task.
Even here, check the explanation against a small piece of actual output — if it claims a query filters out VOID transactions, confirm that by looking at the actual WHERE clause yourself, rather than trusting the paraphrase blind. It's usually right about mechanical explanation; it's worth the ten seconds to confirm anyway.
Sahara Mart's raw transaction data likely contains nothing individually sensitive — but a customer loyalty export with names, phone numbers and purchase history is a different matter entirely, and pasting a sample of it into a public AI chat tool to "just check the formula" is a real data exposure, not a hypothetical one.
Instead of pasting real customer rows to debug a VLOOKUP, build three or four fake rows with the same column structure and plausible-but-invented values, and debug against those. The formula logic is identical; the actual data never left your machine.
"It's just for a quick formula check" is exactly how sensitive data ends up somewhere it was never supposed to be. The habit of building a small fake sample takes thirty extra seconds and removes the risk entirely.
The worksheet pack has the AI-drafted queries you'll be reviewing — open it before you start.
The worksheet pack contains ten AI-drafted SQL queries and Excel formulas against Sahara Mart's data. Seven are correct. Three contain real errors of exactly the kind this module named — a missing filter, a fan-out join, a wrong aggregation choice, a sign error, or similar.
Whether you can catch an error in code you didn't write, under the exact condition that makes it hardest — when the code looks completely reasonable and runs without any error message at all. This is the daily reality of working alongside AI-assisted drafting, and it's precisely the skill the rest of this course has been quietly building toward.
Module 9 turns to communicating findings — including how to present a number you've verified this carefully without burying Chidi in the verification process itself.