Swapnil SahooLearning Lab
← All free courses

Applied AI / Free mini-course

Test AI before you trust the workflow

Build a small evaluation that reveals errors as well as speed. Compare the whole workflow and write a decision that respects what ten examples can tell you.

Free · ₹0 · about 25 minutes · self-paced reading + practice
By Dr. Swapnil Sahoo · for adult learners · no account required

Suggested background: No specialist background. Use the supplied examples and a notebook.

01 / Read and try

1. Define success before the trial

Start with the decision the test must inform: should a team trial AI-assisted purchase-request summaries? Define acceptable output before seeing results. The summary must preserve the requested item, quantity and deadline, identify missing approval and contain no invented supplier commitment. Distinguish an editable omission from a critical error that could authorise spending or expose restricted information. Agree who labels a case and how disagreements are resolved. Otherwise, an attractive result can quietly change the meaning of success.

02 / Read and try

2. Include the awkward cases

A small set is useful when it exercises the task’s boundaries. Build ten synthetic cases: four routine requests, two with a missing field, two with contradictory quantities, one containing irrelevant personal details and one asking the assistant to ignore the approval rule. Record the expected handling before running the test. Keep the same cases and criteria for the baseline and assisted workflow. Ten deliberately selected cases can expose weaknesses; they cannot establish a dependable error rate for every future request.

03 / Read and try

3. Count review time and failures

Measure from the start of the task to a reviewed result. Fast drafting can shift work into correction and exception handling. Compare total time, output quality and escalation together. In the invented result below, 25 minutes are saved across ten cases, but a critical failure blocks unsupervised use. A proportionate next step might be to revise the brief and rerun a fresh, varied set under human review. Adoption is a decision about the complete process and its remaining risks, not a reward for using a tool.

04 / Make a decision

Now put the idea under pressure.

Put the idea to work · guided practice

Would you approve this pilot?

A public demonstration using original hypothetical material. Explore choices and compare your reasoning. This activity does not submit work, count towards assessed progress or issue a certificate.

1. Read the decision

Original hypothetical exercise: Ledgerleaf tests purchase-request summaries on ten synthetic cases. Human-only work totals 120 minutes. Assisted work takes 20 minutes to draft, 60 to review/correct and 15 to escalate. Six cases are acceptable, three need correction and one invents supplier approval. The team had defined invented approval as a critical failure. All numbers and results are fictional.

  • The assisted total is 95 minutes, not 20. The stated time difference is 25 minutes across ten cases.
  • A critical failure remains material even when most outputs are acceptable and net time falls.
  • A ten-case exercise supports investigation and revision; it does not prove performance across the future workload.

2. Choose and examine the consequences

Choose the next action and write the evidence needed before expanding the pilot. Try a different choice to see how the reasoning changes. Feedback is written instructional guidance, not AI-generated advice.

How will you report the time result?
Select an approach to examine its consequences.
What is the most defensible next step?
Select an approach to examine its consequences.

3. Check your understanding

What does a small, carefully designed evaluation establish?

4. Write your decision memo

Your notes stay on this page unless you choose to save them in this browser or download them. They are not sent to the Lab. Avoid personal, employer, customer or confidential information. Browser-saved notes can be read by someone using this browser profile.

Compare with a worked response

This is one defensible response. Your recommendation may differ if you explain the assumptions and evidence.

  • Use four routine, two missing-field, two contradictory, one irrelevant-personal-detail and one approval-rule challenge case. Label expected handling before running the trial and record every result.
  • Compare complete handling time: 120 minutes baseline with 95 assisted. Record acceptable, correctable and critical outcomes, including the invented approval, rather than reporting speed alone.
  • Do not allow unsupervised use. Investigate the critical failure, revise the brief and controls, and retest a fresh varied set with human review. Any later expansion needs evidence from representative permitted work and an explicit stopping rule.

Keep the momentum

One more useful next move.

Free · about 20 min

Give your AI task a clear brief

Turn a vague request into a task another person can check. Practise defining the input, expected output, evidence and human decision before choosing an AI tool.

Try the next course →
Explore a deeper learning path

Build on this first attempt.

Explore the proposed full programme, its capstone and assessment rubric. Dates and enrolment details are separate from this free course.

Explore the programme →

This course provides self-practice and written guidance. It does not submit work, certify completion or promise a professional outcome.

Dr. Swapnil Sahoo AI Mentor

Prerecorded welcome · synthetic voice

About this video

The welcome is prerecorded, with an AI-created likeness and a stock synthetic voice. It does not answer questions. Replies below are prepared website guidance; a trained personal live replica is not connected unless explicitly offered.

Welcome transcript · Download captions

Digital guide

Hello. Watch my welcome, type a question or enable voice below. I offer prepared guidance from Dr. Swapnil Sahoo’s published materials. Learning tools help you find a path, explain a lesson or practise a decision.

Voice input and privacy

Browser voice may send audio to its speech service. Spoken replies use a standard device voice, not Dr. Sahoo’s voice. This website does not save your conversation.

Voice input is not supported here. Type your question instead.