Skip to main content

Interview prep

Top 25 AI Product Manager interview questions

17 of these 25 are scenario, use-case or real-life questions, the kind that decide an interview. Each has a short answer to say out loud, the explanation behind it, and an example you can tell.
Practice out loud
  1. 1ConceptBeginner 1

    What is the difference between AI, machine learning and generative AI?

    Short answer

    AI is the broad goal of computers doing tasks that seem to need intelligence. Machine learning is the main way we build AI today: systems learn patterns from data instead of following hand-written rules. Generative AI is a kind of machine learning that creates new content, such as text, images or code, for example ChatGPT, Gemini or Claude.

    The explanation

    A simple picture is three circles, one inside the other. A spam filter that learned from labelled emails is machine learning. A chess program that follows rules written by people is AI but not machine learning. A chatbot that writes an email for you is generative AI.

    In interviews for non-technical roles, the point is to show you can explain this clearly to a colleague, with an example from daily life or from the company's own work.

    An example you can use in the interview

    I use a cricket example. AI is the big goal: a computer that can, say, act as a match commentator. Machine learning is how we usually get there: show a model thousands of past matches so it learns to predict a team's score. Generative AI is the part that creates new content, such as writing a match report or a commentary line from the scorecard.

    They may follow up with

    • Is a calculator AI? Why or why not?
    • Give an example of machine learning you used today.

    Avoid

    • Treating the three terms as the same thing
    • Giving definitions with no example

    Practice this oneLesson: What AI, machine learning and a language model are

  2. 2ConceptBeginner 1

    What is a large language model, and how does it produce an answer?

    Short answer

    A large language model is a neural network trained on a huge amount of text to predict the next token. It writes an answer one token at a time, each time choosing a likely next token given everything before it. It has no lookup table of facts; it has patterns learned during training.

    The explanation

    Training happens in stages. Pre-training teaches the model to predict the next token on a very large text corpus, which is where its knowledge of language and the world comes from. Later stages, such as instruction tuning and learning from human feedback, teach it to follow instructions and answer helpfully.

    At inference time the model takes your prompt as tokens, computes a probability for every possible next token, picks one (how it picks depends on settings such as temperature), appends it, and repeats until it stops. That is why the same prompt can give different answers, and why a fluent answer can still be wrong: the model optimises for likely text, not for checked truth.

    An example you can use in the interview

    It is like a very well-read autocomplete. Type 'The capital of Karnataka is' and it predicts 'Bengaluru' because that pattern appeared often in its training text, then predicts the next token after that, and so on. It does not look anything up, which is why it can also confidently complete a sentence with something wrong if the pattern seems likely.

    They may follow up with

    • What does pre-training give the model that instruction tuning does not?
    • Why can the same prompt give different answers?
    • Where does a model's knowledge cut-off come from?

    Avoid

    • Saying the model searches a database of facts when it answers
    • Treating fluent output as evidence that the answer is correct

    Practice this oneLesson: What AI, machine learning and a language model are

  3. 3ConceptBeginner 1

    What is the difference between training and inference?

    Short answer

    Training is when a model learns: its internal weights are adjusted using data, which is slow and expensive. Inference is when a trained model is used to make predictions or generate text, with the weights fixed. Most GenAI engineers work at inference time, calling a trained model through an API.

    The explanation

    Training a large language model needs huge datasets, many accelerators and weeks of time, so few companies do it from scratch. Fine-tuning is a smaller round of training on top of an existing model. Inference is each request a user makes; its cost is per call and depends on input and output size.

    This split explains common interview points: the model does not learn from your chats during inference unless the provider later trains on data, its knowledge stops at the training cut-off, and the cheapest way to change behaviour is usually prompting or retrieval at inference time rather than training.

    An example you can use in the interview

    Training a large model is like the years a doctor spends in medical college: slow, expensive and done once. Inference is the doctor seeing a patient: using what was learned, quickly, many times a day. When I call a model API in my app, I am only doing inference; the provider did the training.

    They may follow up with

    • Does a model learn from the prompts you send it?
    • Where does fine-tuning sit between the two?

    Avoid

    • Saying the model updates itself after every conversation
    • Confusing fine-tuning with prompting

    Practice this oneLesson: What AI, machine learning and a language model are

  4. 4ConceptIntermediate 1

    What is a context window, and what happens when a conversation gets too long?

    Short answer

    The context window is the maximum number of tokens the model can consider in one request: system prompt, conversation history, documents and its own reply together. When you go past it, the request fails or something must be cut, so apps trim, summarise or retrieve only the relevant parts.

    The explanation

    Everything the model sees in one call shares the window. In a long chat, each turn resends the earlier turns, so the input grows. Common strategies are to drop the oldest turns, summarise earlier parts, or store information outside the prompt and retrieve only what the current question needs.

    A bigger window is not free: more input tokens cost more and take longer, and models can pay less attention to material buried in the middle of a very long input. So even with a large window, sending only what is relevant usually gives better and cheaper answers.

    An example you can use in the interview

    In a customer chat app, a long conversation plus a 30-page policy document can exceed the model's limit. In one app I worked on, we kept the last ten messages, replaced older ones with a running summary, and sent only the three most relevant policy sections retrieved for the current question. The chat could go on for an hour without failing or slowing down.

    They may follow up with

    • If the window is huge, why use RAG at all?
    • How would you handle a chat that runs for hundreds of turns?

    Avoid

    • Thinking only the latest message counts towards the limit
    • Assuming a larger window always gives better answers

    Practice this oneLesson: Tokens and the context window

  5. 5ConceptPro 3

    When would you use prompting, RAG or fine-tuning?

    Short answer

    Start with prompting: it is fastest and cheapest to change. Add RAG when the model needs knowledge it lacks, especially private or changing information. Consider fine-tuning when you need a consistent behaviour, style or format that prompting cannot reach, have many good examples, and can evaluate the result.

    The explanation

    The approaches solve different problems. RAG adds knowledge at question time and keeps it current. Fine-tuning changes behaviour: tone, output format, a narrow skill, or getting a smaller, cheaper model to match a bigger one on one task. It is a poor way to add facts that change, because you must retrain to update them.

    They also combine: a fine-tuned model can still use RAG. A strong answer mentions evaluation first (you cannot choose without a test set), the cost of maintaining training data, and the order: prompt, then retrieval, then fine-tuning, moving on only when measurements show the simpler step is not enough.

    An example you can use in the interview

    For a bank's FAQ assistant: prompting got the tone and format right in a day. The model did not know the bank's current interest rates, so we added RAG over the rate pages, which update often. Much later, to classify complaints into 40 categories the regulator required, we fine-tuned a smaller model on thousands of labelled complaints, because prompting alone could not hit the accuracy we needed.

    They may follow up with

    • Can you combine RAG and fine-tuning? Give an example.
    • What do you need before fine-tuning is worth trying?

    Avoid

    • Proposing fine-tuning to teach a model facts that change weekly
    • Choosing an approach without mentioning how you would measure it

    Practice this oneLesson: Prompting, RAG or fine-tuning

  6. 6ScenarioPro 3

    When is fine-tuning worth the effort?

    Short answer

    When a narrow, repeated task needs behaviour that prompting cannot reliably produce, you have hundreds to thousands of high-quality examples, and an eval set to prove the gain. Common wins: a strict output format, a house style, a specialised classification, or making a smaller model match a larger one to cut cost and latency.

    The explanation

    Fine-tuning has real costs: collecting and cleaning data, training runs, evaluating, hosting or paying for the tuned model, and redoing it when the base model improves. A newer base model may make your fine-tune unnecessary.

    Before fine-tuning, try better prompts with examples, structured output, and retrieval. If you do fine-tune, keep a held-out test set, compare against the best prompted baseline, and watch for losing general abilities the task still needs.

    An example you can use in the interview

    A company classifying support tickets into 60 categories found that even a large model with detailed instructions was inconsistent on similar categories, and each call was expensive. They had 20,000 tickets labelled by agents over two years. Fine-tuning a small model on them gave higher accuracy on a held-out set and a fraction of the cost and latency, which made it worth the effort.

    They may follow up with

    • What would your baseline be before fine-tuning?
    • What happens to your fine-tune when the provider releases a better base model?

    Avoid

    • Fine-tuning before trying a well-prompted baseline
    • Ignoring the ongoing cost of maintaining training data

    Practice this oneLesson: Prompting, RAG or fine-tuning

  7. 7ConceptPro 1

    How would you evaluate an LLM application?

    Short answer

    Build a test set of realistic inputs with what a good answer needs. Check what you can with code (format, required facts, forbidden content), use human or model review for subjective quality, and track scores for every change. In production, add user feedback and monitoring, and turn failures into new test cases.

    The explanation

    A layered approach, as Hamel Husain describes: level 1 is fast unit-test style assertions run on every change; level 2 is human and model-based review of samples, with the judge checked against human labels; level 3 is A/B testing with real users for product-level impact.

    Start by looking at real outputs and writing down how they fail, then design checks for those failures. Keep the eval set versioned, add cases from production bugs, and report results by category, so a gain in one area does not hide a loss in another.

    An example you can use in the interview

    For an app that writes product descriptions for an online store, I would build 100 real products with what a good description must include: the material, size and price, and no claims we cannot back. Code checks length, format and banned words. A judge model, checked against human ratings, scores readability. Every prompt change runs this set, and descriptions the store team rejects become new test cases.

    They may follow up with

    • Where do your first test cases come from?
    • How do you evaluate subjective quality such as tone?

    Avoid

    • Relying on vibe checks of a few examples
    • Using only generic benchmarks instead of task-specific tests

    Practice this oneLesson: Test sets and evals

  8. 8ConceptBeginner 1

    What is bias in AI? Give an example and how to reduce it.

    Short answer

    AI bias is when a system treats some groups unfairly, usually because it learned from data that reflects past unfairness or leaves groups out. For example, a hiring model trained on past hires may rank women lower if most past hires were men. Reduce it by checking results per group, fixing the data and keeping humans accountable.

    The explanation

    Bias can come from the data, from how the problem is framed, from what is measured, and from who tested the system. In India, speech and language tools that work well in English and poorly in other languages or accents are a common example.

    Fairness is not one number: you decide which kind of fairness matters for the decision, measure it for each group, and keep checking after launch.

    An example you can use in the interview

    A clear Indian example is a loan model trained on past approvals from city branches. It may score rural applicants lower, not because they repay worse, but because few were approved before, so the data has little about them. Checking approval rates and repayment by region would show it, and the fix could be better data, removing proxies like PIN code, and human review of borderline cases.

    They may follow up with

    • How would you test a loan approval model for bias?
    • Can removing the gender column remove gender bias?

    Avoid

    • Saying AI is neutral because it is maths
    • Thinking removing a sensitive column fixes bias

    Practice this oneLesson: Bias and fairness

  9. 9ConceptIntermediate 1

    How is an AI product manager's job different from a regular PM's?

    Short answer

    The core is the same: find valuable problems and ship solutions. The differences: output is probabilistic, so you define quality with evals rather than fixed specs; you design for errors and trust; costs grow with usage, so unit economics matter; data and feedback loops are part of the product; and responsible AI and regulation need explicit attention.

    The explanation

    Show you have done it, or practiced it: describe an eval you wrote, a trade-off between model quality and cost, or a design choice you made because the model is sometimes wrong.

    Interviewers also look for enough technical depth to talk with engineers about prompting, retrieval and fine-tuning trade-offs, without pretending to be an engineer.

    An example you can use in the interview

    As a regular PM for a payments screen, my spec could say 'the button shows the amount'. As an AI PM for a 'categorise my spending' feature, I cannot spec every output, so I define quality as 'at least 90% of transactions correctly categorised on our 500-transaction test set', design an easy way to fix a wrong category, watch cost per user, and plan how corrections improve the model.

    They may follow up with

    • Give an example of an AI-specific trade-off you made.
    • How technical does an AI PM need to be?

    Avoid

    • Saying it is the same job
    • Claiming deep ML expertise you cannot back up

    Practice this oneLesson: AI PM interviews

  10. 10Use caseIntermediate 1

    How do you decide whether a problem is a good fit for AI?

    Short answer

    Start from the user problem, not the technology. A good fit is a task that is frequent and valuable, tolerates some errors or lets a person check the output, has data or examples to learn from, and is hard to solve with simple rules. If a rule, a better form or a search box solves it, do that instead.

    The explanation

    Ask: what does a wrong answer cost, and who catches it? Drafting a reply that an agent edits tolerates errors; approving a loan does not. Check feasibility with a quick prototype on real examples before committing a roadmap.

    Also weigh the four classic product risks: value, usability, feasibility and viability. AI adds a fifth in practice: trust, meaning whether users will rely on output that is sometimes wrong.

    An example you can use in the interview

    A hospital asked whether AI could 'improve patient experience'. Of their ideas, appointment reminders were better solved with a simple SMS rule, and diagnosis suggestions were too risky to start with. Discharge summaries were a good fit: written many times a day, time-consuming for doctors, checkable by the doctor before signing, and with thousands of past examples. We started there.

    They may follow up with

    • What would make you say no to an AI feature request?
    • How would you test feasibility in a week?

    Avoid

    • Starting from 'we need AI'
    • Ignoring the cost of wrong answers

    Practice this oneLesson: Choosing a problem for AI

  11. 11ScenarioIntermediate 2

    Should we build our own AI feature or buy a vendor product?

    Short answer

    Buy when the need is common and not what makes you different, such as meeting notes or a generic support bot, and a vendor meets your security and data rules. Build when the feature is core to your product, needs your own data, or vendors cannot meet your quality, cost or compliance needs.

    The explanation

    Compare total cost over two or three years: licences or API spend, engineering time, maintenance, evals and support. Check lock-in: can you export data and switch? Run both through the same eval on your real cases before deciding.

    Most teams today 'buy' the model through an API and 'build' the product around it: prompts, retrieval over their data, workflows and UX.

    An example you can use in the interview

    At a healthtech company we needed meeting notes for internal calls and a symptom-triage feature for patients. For notes we bought an enterprise tool that met our data rules, in a week. For triage we built, because it used our own clinical protocols and needed strict testing and control. That split saved months and kept our effort on what made us different.

    They may follow up with

    • How would you compare a vendor's quality claims with your needs?
    • What lock-in risks would you check?

    Avoid

    • Deciding on gut feel
    • Ignoring maintenance cost of building

    Practice this oneLesson: Feasibility and build versus buy

  12. 12ScenarioPro 3

    Models now have very large context windows. Why not put all the documents in the prompt instead of using RAG?

    Short answer

    For a small, stable document set, putting it all in the prompt can be simpler and work well, especially with prompt caching. For large or changing collections, RAG is cheaper, faster and scales: you send only relevant text, can enforce access control per user, and can cite sources. Models can also miss details buried in very long inputs.

    The explanation

    Cost and latency grow with input tokens on every request, so a full-corpus prompt is expensive at volume, though caching a fixed prefix helps. Very long inputs can reduce attention to material in the middle. And a corpus of millions of documents will not fit anyway.

    A practical answer is "it depends, and here is how I would decide": if the knowledge base fits comfortably and rarely changes, try long context with caching and measure; otherwise use retrieval, perhaps retrieving larger sections to give the model more context.

    An example you can use in the interview

    A client with a 40-page product manual asked whether they needed RAG. For that, I put the whole manual in the prompt with caching, and it worked well and was simple. Another client had 30,000 documents with different access rights per department; there, RAG was the only practical option, because we could filter by permission, keep cost per question low and cite the exact page.

    They may follow up with

    • How does prompt caching change this decision?
    • How would you handle per-user document permissions?

    Avoid

    • Saying RAG is obsolete because of long context
    • Ignoring cost per request at scale

    Practice this oneLesson: Prompting, RAG or fine-tuning

  13. 13Use caseIntermediate 3

    What does a good data strategy for an AI product look like?

    Short answer

    Know what data you need for evals, retrieval and possibly fine-tuning; how you collect it with consent; how you label and keep quality high; who can access it; and how product usage creates new data, such as corrections and feedback, that improves the product over time. Treat a well-maintained eval set as one of your most valuable assets.

    The explanation

    Many teams think 'data strategy' means training a model, but for most AI products the data that matters is: documents for retrieval, examples for evals, and logs of real usage. Design feedback capture into the product from day one: thumbs, edits and escalations are labels.

    Respect limits: consent and purpose under the DPDP Act, customer contracts, and retention rules. A data flywheel only works if users trust you with their data.

    An example you can use in the interview

    For a resume-feedback product I would start by building an eval set of 300 resumes with expert feedback, collected with the users' consent. Every time a user edits the AI's suggestion, we log the before and after as a signal of what it got wrong. Access to resumes is restricted and deleted after the retention period. Over six months that loop gives us better evals and, if it ever makes sense, data to fine-tune on.

    They may follow up with

    • What product signals can act as labels?
    • How would you keep the eval set from going stale?

    Avoid

    • Equating data strategy with training a model
    • Collecting data without consent or purpose

    Practice this oneLesson: Data strategy for AI products

  14. 14Use caseIntermediate 2

    How do you design a product experience when the AI is sometimes wrong?

    Short answer

    Set expectations honestly, make output easy to check and edit, show sources or reasoning where possible, and keep the user in control of anything consequential. Offer clear ways to correct, undo or give feedback, and design graceful fallbacks when the model is unsure, such as asking a clarifying question or saying it does not know.

    The explanation

    Patterns from Google's People and AI Guidebook and Microsoft's guidelines for human-AI interaction: explain what the system can do, show confidence where it is meaningful, make it easy to dismiss suggestions, and learn from corrections.

    Match autonomy to the cost of errors: suggest rather than act for high-stakes tasks. Then measure: how often users edit, accept or reject outputs tells you where quality falls short.

    An example you can use in the interview

    In an AI feature that drafts insurance claim summaries, we labelled the draft clearly, showed the document line behind each fact so the officer could click and check it, made every field editable, and kept 'Submit' as a human action. When the model was unsure about a date, it left the field blank and said why, instead of guessing.

    They may follow up with

    • When would you show a confidence score, and when not?
    • How would you use user corrections?

    Avoid

    • Hiding that output is AI-generated
    • Designing only the happy path

    Practice this oneLesson: UX for uncertain output

  15. 15Use caseIntermediate 2

    When should a product keep a human in the loop?

    Short answer

    When mistakes are costly, hard to undo, or affect people's rights, money, health or safety; when the model's accuracy is not yet proven; and when regulation or accountability requires a named person to decide. Design the review so it is real, not rubber-stamping: show the evidence, keep the queue manageable and track how often reviewers change outputs.

    The explanation

    Human review costs time, so use it where it buys the most: high-risk items, low-confidence outputs and a random sample of the rest for quality monitoring. As measured accuracy improves, you can move from approving everything to reviewing exceptions.

    Watch for automation bias, where reviewers approve whatever the AI says. Rotating samples of known-wrong outputs into the queue is one way to check reviewers are paying attention.

    An example you can use in the interview

    In an AI tool suggesting insurance claim approvals, small, clear claims under ₹10,000 that matched policy rules were approved automatically. Larger or unusual claims went to an officer, who saw the documents and the AI's reasoning side by side. We tracked how often officers changed the AI's recommendation; when it stayed low for months, we raised the automatic limit slowly.

    They may follow up with

    • How would you know reviewers are actually reviewing?
    • When would you reduce human review?

    Avoid

    • Review that is a rubber stamp
    • Reviewing everything forever

    Practice this oneLesson: Designing human in the loop

  16. 16Use caseIntermediate 2

    As a PM, how do you work with engineers on evals?

    Short answer

    Own the definition of good: write down what a good answer must do and avoid for the main user tasks, help collect realistic examples with domain experts, and agree on the pass bar for launch. Engineers automate the checks; you review failures with them regularly, decide trade-offs and make sure new production failures become test cases.

    The explanation

    Evals are the AI version of acceptance criteria. A PM who can read twenty outputs and say why each is good or bad is far more useful than one who asks only for an accuracy number.

    Keep the eval set tied to product priorities: weight the cases users care about most, add segments such as language or customer type, and track scores on a simple dashboard reviewed in every planning meeting.

    An example you can use in the interview

    On an AI feature that answers sellers' questions about marketplace fees, I wrote what a good answer must do, such as quote the exact fee for that category, link the policy and never guess, and gathered 120 real seller questions with the support team. Engineers turned these into automated checks. Every Thursday we reviewed the failures together, and I decided that reaching 95% on fee questions was the bar for launch.

    They may follow up with

    • What is the PM's role versus the engineer's in evals?
    • How do you set a launch bar?

    Avoid

    • Leaving quality definitions to engineers alone
    • Asking for one accuracy number

    Practice this oneLesson: Test sets and evals

  17. 17Use casePro 1

    How would you build a test set for an LLM feature when you have no labelled data?

    Short answer

    Start from real or realistic inputs: logs, support tickets, or examples written with domain experts. Cover common cases, edge cases and known failure types. For each, write what a good answer must include or avoid. Begin small, around 30 to 50 cases, review them with people who know the domain, and grow the set from production failures.

    The explanation

    Synthetic examples generated by a model can fill gaps, but review them, because they tend to be easier and more uniform than real inputs. Balance the set across categories that matter to the business, and include adversarial inputs such as prompt injection attempts and off-topic questions.

    Write criteria, not just one reference answer, since many good answers are possible. Version the set, and keep a held-out part you do not tune against, so you can confirm improvements are real.

    An example you can use in the interview

    When we started an AI assistant for a telecom's support team with no labelled data, I took 200 recent tickets, worked with two senior agents to pick 50 that covered common and tricky cases, and for each wrote what a correct reply must say and must not say. Every week we added the cases where the assistant went wrong in production. Within two months the set had about 150 cases.

    They may follow up with

    • How do you stop your test set from becoming too easy?
    • How large should it be at first?

    Avoid

    • Using only synthetic, model-written examples
    • Writing one exact reference answer and checking for an exact match

    Practice this oneLesson: Test sets and evals

  18. 18Use caseIntermediate 2

    How would you measure the success of an AI feature?

    Short answer

    Use three layers. A business outcome, such as tickets resolved without escalation or time saved per user. Product usage, such as adoption, acceptance or edit rate of suggestions, and retention. And quality and safety guardrails from evals and monitoring, such as accuracy on a test set, harmful output rate, latency and cost per task. Agree targets before launch.

    The explanation

    Model metrics alone mislead: a more accurate model that nobody trusts does not help. Usage alone misleads too: people may use a feature that quietly gives wrong answers.

    Pick one primary metric tied to the user's goal, a few guardrails that must not get worse, and run an experiment or a staged rollout to attribute the change to the feature.

    An example you can use in the interview

    For an AI reply-suggestion feature in a support tool, I would agree three layers before launch. Business: average handling time down by 15% without a drop in satisfaction. Usage: share of tickets where agents accept or lightly edit the suggestion. Guardrails: accuracy on a 200-ticket test set above an agreed bar, no suggestions containing another customer's data, and cost per ticket under ₹2.

    They may follow up with

    • What is a guardrail metric?
    • How would you know users trust the feature?

    Avoid

    • Only model accuracy
    • No targets agreed before launch

    Practice this oneLesson: The KPI that matters

  19. 19ScenarioPro 3

    Your AI feature launched, usage is high, but customer satisfaction dropped. What do you do?

    Short answer

    High usage with falling satisfaction suggests people are forced through it or it is failing them. I segment the data to see who is unhappy and on which tasks, read real conversations, and check quality metrics and escalations. Then I fix the biggest failure, make it easy to reach a person, and measure outcomes, not just usage.

    The explanation

    Usage alone is a weak metric for AI features, especially if the feature sits in front of what users actually wanted, such as a chatbot before human support. Satisfaction, task completion, repeat contacts and escalations tell you whether it helps.

    Start with evidence: sample conversations from unhappy users, tag the failure types, and quantify them. Common causes are wrong answers, no way to reach a human, slow responses or a tone that feels dismissive. Fix the top cause, run a controlled test, and report the trade-off honestly to leadership.

    An example you can use in the interview

    We launched an AI assistant in a bank app's help section, and within a month most help sessions went through it, but app-store ratings mentioning support fell. Reading 200 conversations, we found a third were about card blocks and disputes, which the assistant could only explain, not do, and there was no button to reach an agent. We added a 'talk to a person' option for those topics and measured resolution rate instead of sessions. Ratings mentioning support improved over the next two months.

    They may follow up with

    • Which metrics would you have tracked from launch?
    • Would you roll the feature back?

    Avoid

    • Celebrating usage numbers
    • Changing the model before understanding the failures

    Practice this oneLesson: The KPI that matters

  20. 20Use casePro 1

    How would you price an AI feature whose cost grows with usage?

    Short answer

    Start from unit economics: estimate the model and infrastructure cost per task and per active user at realistic usage, including heavy users. Then choose a model: include it in existing plans with fair-use limits, charge per seat with usage caps, sell credits, or price on outcomes. Protect margin with caps and cheaper models for simple steps.

    The explanation

    The risk with flat pricing is a small group of heavy users costing more than they pay. Look at the usage distribution, not the average.

    Value matters as much as cost: if the feature saves an hour a day, price against that value. For Indian customers, price sensitivity is high, so a generous free tier with clear limits and a cheap upgrade often works better than metered billing that feels unpredictable.

    An example you can use in the interview

    For an AI resume-writer in a jobs app, the cost was about ₹3 per resume, but some users generated 50 versions. We included five generations a month in the free plan, unlimited with a fair-use cap of 100 in the paid plan, and used a smaller model for simple formatting steps. Heavy users still had a good margin, and the free limit was enough for most people to try it.

    They may follow up with

    • How would you find your heaviest users' cost?
    • When does usage-based pricing make sense?

    Avoid

    • Pricing on average usage only
    • Ignoring falling model prices

    Practice this oneLesson: Pricing AI features

  21. 21Use casePro 2

    How would you keep the cost of an LLM feature under control?

    Short answer

    Measure cost per request and per user, then cut it: use the cheapest model that passes your evals, shorten prompts and outputs, cache repeated prefixes and answers, batch work that can wait, and stop runaway usage with per-user limits and a provider spending cap. Recheck when traffic or prices change.

    The explanation

    Cost is tokens times price times volume, so attack each factor. Routing simple requests to a small model often gives the biggest saving. Trimming retrieved context and conversation history cuts input tokens. Batch APIs are cheaper for offline jobs.

    Guardrails stop surprises: rate limits per user, maximum steps for agents, a hard monthly cap, and alerts when daily spend jumps. For a product in India, cost per active user has to fit low price points, so model the unit economics before launch.

    An example you can use in the interview

    For a document Q&A feature costing about ₹4 per question, I logged cost by step and found the long system prompt and ten retrieved chunks were most of it. Caching the fixed prompt prefix, cutting to five reranked chunks, and moving simple greeting messages to a smaller model brought it to about ₹1.20 with no drop on our eval set. We also added a per-user daily cap.

    They may follow up with

    • How would you calculate cost per active user?
    • What would you do if one user generated a large share of the cost?

    Avoid

    • Only switching to a cheaper model without checking quality
    • Having no hard spending limit

    Practice this oneLesson: Cost control

  22. 22Use caseIntermediate 3

    What responsible AI checks would you run before launching an AI feature?

    Short answer

    Identify who could be harmed and how; test quality and fairness across user groups and languages; red-team for misuse, prompt injection and harmful output; check privacy and data handling against the DPDP Act and your policies; make sure users know they are dealing with AI; set up monitoring, an incident process and an owner; and document decisions.

    The explanation

    Frameworks help structure this: the NIST AI Risk Management Framework's govern, map, measure and manage functions, and the EU AI Act's risk tiers if you serve EU users. Scale effort to risk: a writing helper needs less than a credit decision.

    A launch checklist with named sign-offs turns principles into practice, and a post-launch review after a few weeks catches what testing missed.

    An example you can use in the interview

    Before launching an AI loan-eligibility explainer, we listed who could be harmed: applicants misled about eligibility. We tested answers in English, Hindi and Tamil, checked results across income groups, red-teamed for prompts asking it to promise approval, confirmed only needed data was used under our privacy policy, labelled it as AI, named a product owner, and set up a kill switch and an incident channel.

    They may follow up with

    • How would you scale these checks for a low-risk feature?
    • Who should sign off?

    Avoid

    • Treating responsible AI as a one-time checkbox
    • No owner after launch

    Practice this oneLesson: AI risk frameworks

  23. 23ScenarioBeginner 1

    What information should you never paste into a public AI chatbot at work?

    Short answer

    Anything confidential or personal unless your company has approved the tool for it: customer details such as names, phone numbers, Aadhaar or account numbers, passwords and keys, unreleased financials, contracts and source code. Check your company's AI policy and use approved tools, and remove or mask personal details first.

    The explanation

    Free consumer tools may keep conversations and, depending on settings, use them to improve models. Business plans usually offer stronger terms, which is why companies approve specific tools.

    India's Digital Personal Data Protection Act makes organisations responsible for how they handle personal data, so a careless paste can become a compliance problem as well as a security one. A safe habit is to replace real names and numbers with placeholders before asking for help.

    An example you can use in the interview

    A colleague once pasted a customer complaint into a free chatbot, including the customer's full name, phone number and account number, to get help drafting a reply. I suggested redoing it with the details replaced by placeholders such as [Customer] and [Account], which gave just as good a draft, and pointed them to our approved company assistant for anything containing customer data.

    They may follow up with

    • How would you get help drafting a reply to a customer complaint without sharing their data?
    • What should a company AI policy say about this?

    Avoid

    • Assuming a chat is private because only you can see it
    • Forgetting that code and contracts are confidential too

    Practice this oneLesson: Use AI safely at work

  24. 24ScenarioIntermediate 3

    You launched an AI tool internally and nobody uses it. What do you do?

    Short answer

    Find out why before pushing harder: talk to intended users, watch them work and check usage data. Common causes are a weak fit with their real workflow, lack of trust in output quality, no time to learn, or fear about jobs. Fix the biggest cause, find champions, train on real tasks and have leaders use it visibly.

    The explanation

    Change management models such as ADKAR (awareness, desire, knowledge, ability, reinforcement) are a useful checklist: people need to know why, want to, know how, be able to and have it reinforced.

    Measure adoption by team and task, not just log-ins, and close the loop: when users report a problem, fix it quickly and tell them, which builds trust faster than any announcement.

    An example you can use in the interview

    We launched an AI writing assistant for a 50-person sales team and usage stayed near zero. I sat with five reps for a day and saw why: it lived in a separate website, while they wrote everything in the CRM. We added it as a button inside the CRM, built three templates for their real emails, and asked the top seller to share how she used it. Weekly users rose to most of the team within a month.

    They may follow up with

    • How would you find out why people are not using it?
    • What role do managers play?

    Avoid

    • Sending more announcement emails
    • Blaming users instead of the fit

    Practice this oneLesson: Change management and training

  25. 25Real lifeIntermediate 1

    Tell me about an AI product decision you made and how it turned out.

    Short answer

    Pick a real decision with a trade-off, such as build versus buy, whether to launch with known errors, or where to put a human in the loop. Explain the options, the evidence you gathered, why you chose, how you measured the result, and what you learned, including anything you would change.

    The explanation

    Good AI PM answers show judgement under uncertainty. The decision should involve something AI-specific: error rates, evaluation, cost per use, data, risk or user trust.

    Show your process: you defined what good looks like, worked with engineers on an eval, talked to users, estimated cost, and decided with clear criteria. Then share the outcome honestly, including if it partly failed. If you have not worked as a PM yet, a decision from a college project, a startup or your current role, told with the same structure, works.

    An example you can use in the interview

    At an ed-tech startup I had to decide whether to launch AI-graded essay feedback for Class 10 students. On our test set of 300 teacher-graded essays, the AI's scores matched teachers closely for most essays but were badly off for some answers in Hinglish. Instead of a full launch, we shipped feedback comments without scores and kept teachers grading. Students used the feedback a lot, teachers said grading went faster, and we added Hinglish essays to the eval set before considering scores again.

    They may follow up with

    • What would you have done if the eval scores were lower?
    • Who disagreed, and how did you handle it?

    Avoid

    • A decision with no trade-off
    • Taking credit for the whole team's work

    Practice this oneLesson: AI PM interviews

Today's AI turns are used

Lessons, roadmaps, videos, practice cards and the code playground stay free with your account, with no limit.