Build an AI demo that shows when the workflow should stop

Define the task and expected evidence

Choose one bounded task, such as extracting dates from a short fictional event note. Write the expected answer yourself before running the workflow. State which fields are required and which may legitimately remain unknown. Keep the example small enough for a viewer to inspect without trusting your narration. Do not use a real customer message or private document to make the demonstration feel realistic. The test is about behaviour under a clear contract, not about exposing more data or presenting one impressive response as a complete product assessment.

Include ordinary and difficult inputs

Prepare several variations: a clear note, a note missing the date and a note containing two contradictory dates. Decide how each should be handled before observing the output. A successful result may be a request for clarification rather than a completed form. Record the exact input, prompt or configuration and output for each run so that a later improvement can be compared fairly. Avoid silently replacing a failed attempt with another run while describing the first as reliable. A single favourable example cannot establish behaviour across every user, language or document type.

Make the stop path visible to the audience

Show how the workflow labels uncertainty and keeps an incomplete result from becoming an automatic external action. For the fictional date example, that could mean leaving the event unscheduled until a person confirms the missing detail. Explain which parts you checked manually and which remain outside the test. Do not claim a safety guarantee or a universal accuracy rate from a few cases. End with the next test you would add, such as an unusual date format. A demonstration becomes more informative when viewers can see the limits and the decision boundary as clearly as the successful output.

• Write expected results before running examples.
• Use synthetic inputs with no private customer information.
• Include missing and contradictory details alongside ordinary cases.
• Show a clear human clarification or stop path before external actions.