An honest AI proposal for a business you know

The situation

A business you know — a shop, a clinic, a school, your family's work — is being told by everyone that it needs AI. The owner has asked you, because you took this course, to tell them plainly: what could actually work here, what would it need, and what would go wrong.

You will answer with evidence, not enthusiasm: tasks you have watched, a tool you have tested against facts you can verify, and a proposal that names its own limits.

What you'll be able to show

• Tell a task that suits AI from one that does not, using the three tests from the course
• Measure how often an AI tool is confidently wrong on facts you can check
• Write a one-page proposal a business owner can act on, limits included
• Say what data the business would need and whether it actually has it

What you need

• Access to any AI assistant (a free one is enough)
• One real business you can ask questions of, or observe for an hour
• A document for the proposal

Tasks

1. 1Watch the work
   Spend an hour with the business (or its owner). List every task you see or hear about — at least twelve. For each, mark whether it is repetitive, high-volume, and whether a human could check the output in seconds.
   A correct result: A table of twelve or more tasks with the three yes/no columns filled in.
2. 2Pick three and reject the rest
   Choose the three tasks that pass all three tests. For two tasks that failed, write one sentence each on why they are the wrong place to start.
   A correct result: Three candidates with reasons, two rejections with reasons.
3. 3The data test
   For each candidate, write what past examples an AI tool would need to learn from (the course's 500-examples idea) and whether the business records that data anywhere today. Be exact: a notebook, a phone, a spreadsheet, nowhere.
   A correct result: For each candidate, a named data source or the word nowhere.
4. 4Test the tool on facts you can check
   Ask an AI assistant ten factual questions about your city or the business's field that you can verify. Record each answer and whether it was right. Note whether the wrong answers sounded any less confident.
   A correct result: Ten questions, ten verdicts, a count of how many were wrong, and one sentence on tone.
5. 5Draft one workflow
   For your best candidate, draw the workflow as boxes: where the task starts, what the AI does, who checks it, what happens when it is wrong. The human check must be in the drawing.
   A correct result: A diagram with a check step and a failure path.
6. 6Write the proposal
   One page for the owner: the task, what it would need, what it would cost in time to set up, the check that keeps it safe, and three things it will NOT do. Use the numbers from tasks 3 and 4.
   A correct result: One page a non-technical owner can read in five minutes, with the error rate from task 4 quoted in it.

What to hand in

The task table, the ten-question test log, the workflow drawing (a photo of paper is fine) and the one-page proposal.

How it is marked