AI Automation Project Ideas for Students
Build three testable workflows with approvals and explicit failure paths.
Part of the AI Automation track: modules, projects and first roles →
By the GullyMentor team · 5 min read · Published 2026-10-18 · Updated 2026-11-06 · Sources listed at the end. Not a promise about mentorship outcomes.
A good AI automation project is a reliable, testable process with a human approval step and a clear failure path. It is not a screenshot of connected boxes. Use fictional inputs and simulated external actions for these projects, so nothing real is sent, paid or published.
What should a portfolio project prove?
It should show that you can define a trigger, validate input, use AI only where wording varies, and stop for a person before anything consequential. It should also show what happens when something breaks. Each idea below lists the problem, data, steps and finished output.
1. Expense-description organiser
Problem: short expense descriptions arrive with unclear categories.
Data: thirty synthetic expense records with dates, amounts and short descriptions. Define an allowed category list.
Steps: 1. Use deterministic validation for the amount and date. 2. Use an AI suggestion for an unclear description, limited to the allowed list. 3. Route uncertain records to review. Do not approve or reimburse anything. 4. Test an empty description, repeated receipt ID and malformed amount. 5. Evaluate incorrect categories separately from validation failures.
Finished output: the workflow, reviewed output and an exception log.
Explain why classification is a suggestion and why financial approval remains a separate human responsibility.
2. Newsletter review queue
Problem: event submissions need a summary and a check before publication.
Data: fictional event submissions and an editorial checklist you write.
Steps: 1. Validate fields. 2. Draft a summary with an AI step. 3. Place the draft in a review queue. Keep publication simulated. 4. Use an event ID to recognise duplicates. 5. Test a cancelled event, missing venue and connection failure. 6. Check that no published state appears before a reviewer approves the final text.
Finished output: a process map, draft examples and a recovery guide.
Discuss how you would handle a correction arriving after approval without silently publishing an outdated version.
3. Knowledge-base gap report
Problem: find questions that a help document cannot answer.
Data: a small fictional help document and twenty sample questions.
Steps: 1. Have the workflow identify likely relevant sections. 2. Flag questions the document cannot answer. Do not invent answers to fill gaps. 3. Aggregate unresolved topics into a draft improvement report. 4. Test unrelated questions and contradictory passages.
Finished output: a topic table, a reviewed report and evidence linking each suggested improvement to sample questions.
Evaluate whether the report preserves uncertainty. In an interview, explain that repeated questions are evidence of a possible information gap, not proof that every suggested answer is correct.
What limits should you state?
Say plainly that the AI step can be wrong, that you reviewed its output, and that the sample is small and invented. Do not describe a workflow as "fully automated" if a person approves the result. The NIST AI Risk Management Framework is a good source for the language of risk and oversight.
What is the shared completion checklist?
Record the trigger, permissions, expected outputs and owner of every exception. Test the same event twice and interrupt one run before completion. Inspect logs for unnecessary personal data.
If you use paid connectors, estimate the number of operations and remove trial connections after the exercise. A strong demonstration includes recovery from a failure and the human decision you deliberately retained.
What mistakes are common?
- Letting the AI step send or approve things.
- No duplicate handling, so a retry repeats an action.
- Testing only the happy path.
- Logging full message text when an ID would do.
How do you check you got it?
Run the same input twice and confirm one result. Switch off the AI step and confirm items land in manual review. Ask a peer to find an input that breaks the flow.
How should you present a project?
Lead with the process map, then show one normal run and one failed run side by side. State which steps are rules, which are AI suggestions and which are human decisions. Add a short limits section. Reviewers trust a project more when it shows where the automation stops.
What to build next
Pick one project and map it first with how to map a business process before automating it. Then rehearse explaining it using AI automation interview questions. The AI Automation track shows how projects like these fit into a wider learning path.
Sources & further reading
Examples and practice recommendations are educational illustrations. Tool documentation and programme inclusions should be checked for your intended use.