A market analysis where every number gets traced back
The situation
Students can now produce a plausible ten-page market analysis in twenty minutes. The scarce skill is no longer assembly — it is knowing which numbers are real. This workflow inverts the assignment so the research is the setup and the verification is the deliverable.
Steps
-
Assemble the picture
Open the competitor sites and any filings, then have the assistant build a comparison table across them. Fast, and deliberately unverified at this stage.
What you only learn by doing it: Have students save the raw AI output before any editing. They will need it for the comparison in step three.
-
Trace every figure to a primary source
Each number in the table gets a source, a page or section reference, and a verdict: confirmed, misattributed, or not found. Financial figures go back to the filing, not to a summary of the filing.
What you only learn by doing it: Budget more time for this than feels reasonable. It is slower than the research was, which is itself the lesson about where the value sits.
-
Run the numbers properly
Where the analysis needs actual computation, upload the data and have it run and show the Python. Students audit the generated code line by line and annotate what each step assumes — especially anything that dropped rows or chose a default test.
What you only learn by doing it: Make them find the line that handles missing data. It is almost always there, it is almost always silent, and it almost always changes the answer.
-
Make the structure visible
Turn the written analysis into a diagram, then mark every relationship the tool inferred incorrectly. It will happily render an incoherent argument as a clean chart, which is a useful demonstration that visual polish is not evidence of sound reasoning.
What you only learn by doing it: If the diagram looks fine but the argument is weak, the student has learned something more valuable than if it had failed visibly.
-
Present, with the verification appendix
The deck presents the analysis; the appendix shows every figure, its source, and what verification found. Grade the appendix.
What you only learn by doing it: Students resist this until the first time someone in the room gets caught with a fabricated figure in a live presentation. After that it sells itself.
Where this breaks down
One security point students should hear explicitly: AI browsers that act on page content are susceptible to prompt injection hidden in websites. Nobody should run an agentic browser while logged into a bank, a student information system, or anything sensitive. This is a real lesson for business students heading into operations roles, not a theoretical one.
The other trap is that AI summaries misattribute figures to the wrong source — and the wrong source is usually adjacent and authoritative-looking, which makes the error harder to spot than an invented number would be.