The Three Families of AI Design Questions

"Design a recommendation system." "Design an AI assistant for customer support." "Design a coding agent." These sound alike, but they ask for different kinds of thinking. Recognizing the family in the first minute lets you reach for the right template and spend your time where it counts.

A bakery makes the same loaf at scale, and the craft is in the recipe, the oven schedule, and quality checks. A restaurant responds to a different order every time, from a pantry of ingredients. A catering crew goes out to a venue with a van of equipment and has to make decisions on site, with no way to call back to the kitchen. Classic ML is the bakery, GenAI is the restaurant, and agentic systems are the catering crew.

• -What it is: a trained model predicts or ranks something, for each user or item, at scale.
• -Examples: recommendations, feed ranking, search ranking, ad click prediction, fraud detection, harmful content detection.
• -Signature parts: labels and features, a retrieval and ranking funnel, a feature store, offline metrics and online experiments, drift monitoring.
• -Signature risks: leakage and skew, biased feedback, calibration, label delay, feedback loops.
• -Covered in: the ML systems modules and the classic design questions.

• -What it is: an LLM or another generative model produces text, images, or code in response to a request, often grounded in your data.
• -Examples: a chat assistant, an answer engine over documents, text-to-image, summarization, a drafting tool.
• -Signature parts: prompt and context assembly, retrieval, model choice and routing, streaming, guardrails, caching, evals.
• -Signature risks: hallucination, prompt injection, cost per request, latency, data leakage, unsafe output.
• -Covered in: the retrieval, inference, and evals modules, and the product designs.

• -What it is: a model takes multiple steps, calling tools and acting in the world, with some autonomy.
• -Examples: a support agent that can issue refunds, a coding agent, a research agent, a workflow automation.
• -Signature parts: tools and their permissions, a control loop with limits, memory and context management, human approval, trajectory evals.
• -Signature risks: compounding errors, excessive agency, prompt injection through tool output, runaway cost and loops.
• -Covered in: the agent design modules and the case studies.

A support product may use a classic classifier to route tickets, a GenAI answer engine over the help center, and an agent that can look up orders and issue refunds. Name the pieces and treat each with the right template. When an interviewer gives a blended question, say so out loud: "This has a retrieval part, a ranking part, and an action part. I will take them in that order."

A 45-minute answer has to fit a lot. The code prints a suggested schedule for each family. The numbers are a starting point, not a rule. Press Run and compare where the largest blocks of time fall in each plan.

In the classic ML plan the largest block, 10 minutes, goes to the model and funnel, with 8 each for the data and for evaluation. In the GenAI plan, architecture gets 10 and retrieval, prompts, and models get 9. In the agentic plan, tools and the control loop get 10, with 8 for safety and permissions and 9 for evals and reliability, since agents fail in ways that need those guards. All three keep a few minutes at the start to clarify and a few at the end for failure modes, which is where strong answers separate from average ones.