A Framework for AI Design Answers

Use these seven steps as a starting structure for practice. The order helps you make decisions using requirements and data, rather than choosing components first.

Here is one way to divide a 45-minute interview.

These steps total 43 minutes, leaving two minutes for final questions. Follow-up questions may also arise throughout the discussion, so treat the times as a guide.

Adjust the emphasis to the question. Predictive systems may need more discussion of data and models. Generative systems may need more discussion of inference and context. Agentic systems need careful discussion of tool permissions and limits.

Cover every relevant decision, but adapt the order and depth when the interviewer asks you to focus elsewhere.

Step 1: What Are We Designing?

Before designing the system, explain your understanding of the task.

1. Name the family. Use the test from The Three Question Families. Does the system mainly produce predictions, content, or actions?

This suggests an initial architecture, such as candidate selection and ranking, a generation pipeline, or an agent with explicit limits. A mixed product may need more than one.

2. Name the known system, if you can. A question may resemble a design you already know.

For example: "This is a recommendation system where the items are podcasts." Or: "This is document question answering with stricter accuracy requirements."

Use the known design as a starting point, then check which requirements differ. Do not assume that an existing architecture fits without changes.

Next, state the scope: what you will include, what you will exclude, and why.

For example: "I'll design text-only support in English first. Voice and additional languages are outside this version."

Confirm the choice: "Does that scope match what you want to discuss?"

This reduces the risk of spending the interview designing a different system from the one intended.

Step 2: What Does Success Mean?

Turn the product goal into measurable requirements. State assumptions when the interviewer does not provide targets.

1. Quality. Choose metrics that reflect the task.

"Users find videos they enjoy" might lead to click-through rate and watch time, together with checks for satisfaction and unwanted content.

"Helpful support bot" might lead to correct resolution rate, customer satisfaction, and the rate of repeated contacts about the same issue.

2. Latency. Define how long users can wait. A feed may need to load within a few hundred milliseconds. A chatbot may have separate targets for the first visible output and the complete answer. A background batch job still needs a completion deadline and a data-freshness target.

These requirements affect what can be computed in advance and what must happen during a request.

3. Scale and cost. Estimate users, request volume, peak requests per second, and cost per request or completed task.

A free feature may need a very low cost per request. A support feature may justify a higher cost if it resolves issues reliably and reduces total support costs. Use a concrete budget rather than "near zero."

Step 3: What Data Do We Have?

Name the data sources that the design needs:

• Interaction logs, such as clicks, watches, and purchases.
• Content and catalog data.
• User profiles, where access and use are permitted.
• Documents or other information needed for generated answers.

Then ask two questions.

1. Where do labels come from? In supervised learning, a label is the target value or category for a training example. Labels can be incomplete or incorrect.

Fraud systems may use confirmed investigations and chargebacks, but chargebacks arrive late and do not all prove fraud. Feed systems may use clicks, which arrive quickly but are an imperfect signal of satisfaction.

Support systems may use reviewed conversations, verified resolutions, and user feedback. A thumbs-up or thumbs-down is useful feedback, but it is not always a reliable correctness label.

Supervised training needs suitable labels. Other approaches, such as self-supervised learning or using a pretrained model, do not require a manually labeled dataset for every new task. In all cases, explain how you will evaluate results.

2. How fresh must the data be? Some recommendations can use daily updates. Trending content or fraud detection may need much fresher signals. Confirm the requirement rather than assuming a fixed interval.

Freshness helps determine whether batch processing, streaming, or request-time lookups are needed.

Step 4: What Is the Simplest Model That Works?

Start with a baseline that can meet the basic requirements. It helps you test the pipeline and gives you a result to compare against.

Predictive family. For recommendations, popular items can be a baseline before a learned ranker. For other prediction tasks, use an appropriate rule or simple model.

Generative family. Start by evaluating an existing model with clear instructions. Add retrieval when answers require external or private information. A hosted model is one option, subject to privacy, latency, and cost requirements.

Fine-tuning means giving an existing model additional training. Consider it when evaluation shows a specific need that simpler changes do not address.

State a condition for upgrading, such as: "If the baseline misses the quality target, I will compare a tuned model with the current approach."

Agentic family. Start with a fixed workflow if the steps are known. If the model needs to choose the next step, start with one agent and a small set of narrowly permitted tools.

Multiple agents add coordination, evaluation, and operating costs. Use them only when the task benefits from that separation.

Explain the next option you would test and the evidence that would justify it.

Step 5: How Do We Serve It?

Draw the architecture and trace a request through it. This is often a large part of the discussion.

The design may use candidate selection and ranking, retrieval followed by generation, or an agent that selects tools within defined limits.

The following four serving questions apply across these designs.