Problem Definition — Does AI Fit?
Asha's first pitch for Pantry Pal was "an AI cooking app". That names a technology, not a problem. Before anything gets built she needs two answers: whose problem is this, and is AI the right tool for each job inside it, or would plain code do better?
The rule
AI fits when many answers are acceptable and a wrong one is cheap to catch.
Start with a problem statement that contains no technology at all: who has the problem, what goes wrong, how often, and what they do about it today.
People who cook for themselves after work throw groceries away most weeks, because at 8 p.m. they cannot think of a dinner that uses what they already have. Today they order food or cook the same two meals.
Then test each job inside the problem. Language models are strong at fuzzy, open-ended work — suggesting, summarising, rewording, sorting messy text — where many different answers are fine. They are weak where there is one exact answer, where a rule must never bend, or where a mistake is costly and hard to notice.
Three questions:
- Is more than one answer acceptable? Dinner ideas from eggs, rice and spinach: many. The total calories in a recipe: one.
- Does a good-enough answer still help? A decent suggestion beats staring into the fridge.
- Can a person catch a wrong answer cheaply? An odd recipe is obvious and harmless to skip. A wrong medicine dose is neither.
Three yeses: AI fits. Any no: use plain code, a trusted data source or a person — or keep the AI but put a check around it.
The shortcut
Sort every job in the product into one of three bins: AI (fuzzy, many answers, cheap mistakes), plain code (exact rules, one right answer), or AI plus a guard (AI does the fuzzy part, code enforces the hard rule).
Suggest three dinners from the item list
Many good answers, and a weak one is easy to skip. AI.
The trap is starting from "where can I add AI?" Start from the problem, then give AI only the jobs it is actually good at.
Drills
Answer fast — trust the rule.
Which of these is a problem statement rather than a solution?
Which job fits AI best?
Pantry Pal must never show a peanut dish to a user with a peanut allergy. What is the soundest design?
Why is "exact calories for diabetic users" a poor first job for AI?
Remember
- Write the problem with no technology in it: who, what goes wrong, how often, today's workaround.
- AI fits when many answers are acceptable and a wrong one is cheap to catch; otherwise use plain code, or AI plus a guard.
- Trap: starting from "where can I add AI?" instead of from the problem.