AI Essentials · practical exercise
Your first AI reporting workflow: calculate, draft, check.
Practise turning two simple figures into a summary you can trust. Then change the evidence and decide when the answer should stop.
Use invented information in your approved AI tool. This lesson does not upload data, call an AI model or award a certification.
Try the evidence exercise- Read the source
- Check the calculation
- Review the draft
Start with a question you can answer.
A good first workflow is small enough to check yourself. Suppose a team recorded 24 enquiries in April and 30 in May, using the same definition across two complete months. Your task is a two-sentence summary of the change.
Calculate the result before asking an AI tool to explain it. The difference is six enquiries. Six divided by 24 is 0.25, so the increase is 25%. Keeping your own calculation gives you something to compare with the draft.
If the earlier count is zero, percentage change is undefined. Show the counts and absolute difference instead. A polished percentage is not useful when its denominator does not support it.
Your turn · all figures are synthetic
What should happen next?
Source card
April: 24 enquiries. May: 30 enquiries. Both periods are complete and use the same definition.
Give the model evidence and a boundary.
Use this practice prompt with the complete, comparable figures. Keep the result as a draft, even if it sounds confident.
Write a two-sentence draft reporting summary from these synthetic figures: April, 24 enquiries; May, 30 enquiries. The definition and reporting periods are comparable. State the absolute and percentage change. Use only these facts. Do not claim a cause. Label the result DRAFT and list what a reviewer should verify.
Does the response contain the right counts? Is the calculation correct? Has it invented a campaign, a trend or a reason for the increase? Another person should be able to trace each claim back to the source.
Test the failure, not just the happy path.
Remove the May count from the prompt and explain that its export is missing. A useful answer asks for the source or says it cannot calculate the change. It should not turn missing information into zero.
Next, restore the counts but say that May includes a category April excludes. The arithmetic is possible, but the comparison is misleading. Ask the model to identify the break in the series rather than write a confident performance claim.
These tests teach a repeatable habit: inspect the evidence before accepting the explanation. Save one deliberately failed example alongside the prompt so you can test future changes to the workflow.
Keep a small review card.
- Are the sources present and suitable?
- Do periods and definitions match?
- Can I reproduce the calculation?
- Does every factual claim have evidence?
- Has the draft mistaken a change for a cause?
- Who approves the result?
Repeat the method on a fictional meeting follow-up or short document comparison. Keep the source small enough to inspect yourself.