Lesson · ⏱ 7 min read
🎯 Problem framing: should AI solve this, and how?
EnglishHinglish
🏠 Real-life example
A patient walks into a clinic and says 'give me antibiotics'. A good doctor first asks what the symptoms are, since when, and what has been tried. Often the real need is different from the request. When a client says 'we need an AI chatbot', a good AI professional diagnoses before prescribing.
💡 In simple words
Start from the problem, not the technology. Ask who has the problem, how often, and what it costs today. Define success as a measurable metric ('cut average ticket handling time from 9 to 5 minutes'). Check whether the data and workflow exist to support a solution. Then pick the simplest approach that works, which might be a rule, a dashboard, a classic ML model, an LLM feature or an agent. AI is a good fit when the task is frequent, pattern-based, tolerant of occasional errors and backed by data. It's a poor fit when every mistake is costly and rules already work.
🪜 Step by step
- Write the problem in business terms: "Agents spend 6 minutes per ticket finding the right policy."
- Measure today's baseline: how long, how many, how much it costs, how often it goes wrong.
- List solutions from simplest to most complex: a better process, a rule, search, classic ML, an LLM, an agent.
- Check feasibility: is the data available, can it be measured, and what happens when the AI is wrong?
- Agree one success metric and a small pilot with a deadline.
🔑 Key terms
- Success metric
- A number that shows whether the solution worked.
- Baseline
- How things perform today, without the new solution.
- Simplest viable approach
- The least complex solution that meets the goal.
👀 See it in action
Ask: "Build an AI chatbot for our support team"
Discovery: 60% of tickets are 'where is my order?'
Baseline: 9 min average handling, 30k tickets/month
Better fix: auto-reply with live order status (API) + LLM drafts for the rest
Metric: handling time 9 → 5 min, CSAT ≥ 4.3Diagnosis changed the solution, and made it cheaper and more effective.
⚠️ Common mistakes to avoid
- Starting from the technology ("let's use agents") instead of the problem.
- No baseline, so nobody can prove the AI made things better.
✅ Quick check
A client asks for 'AI to stop customers leaving'. What's the best first step?
- Build a chatbot
- Train a deep learning model right away
- Define 'leaving', measure the baseline and check what data exists
Show answer
C. Define 'leaving', measure the baseline and check what data exists: Without a clear definition, baseline and data, any model is a guess.
🙋 Questions learners ask
How do I spot a good AI use case?
Frequent, costly, mostly text or pattern based work, where a draft or suggestion helps and mistakes can be caught.
What if the data doesn't exist yet?
Start collecting it, or begin with an LLM that needs fewer examples, and log real usage to build a dataset over time.
📌 Remember
Problem → metric → data → simplest solution. Diagnose before you prescribe.
Learn this as part of a full course
Track your progress, take the quick checks and follow the lessons in the right order. This lesson is part of:
Continue in the course