Skip to main content

Interview prep

Top 25 AI Automation Specialist interview questions

20 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. 4ConceptBeginner 1

    Why do language models hallucinate, and how do you reduce it in an app?

    Short answer

    A model generates likely text, not checked facts, so when it lacks the right information it can produce a confident, plausible answer that is false. In an app you reduce this by giving it the right source text (RAG), asking it to cite and to say when it does not know, and checking outputs.

    The explanation

    Hallucination is a side effect of how models work: they are trained to continue text fluently, and nothing in plain generation checks claims against reality. Gaps in training data, questions about recent events, and very specific details such as citations, numbers and names are where it shows most.

    Practical defences: ground answers in retrieved documents and instruct the model to answer only from them; allow and reward "I don't know"; ask for citations and verify that cited text exists; use structured output so code can validate fields; and measure hallucination rate on a test set so you know whether a change helped.

    An example you can use in the interview

    Ask a model 'What is the return policy of Sharma Electronics?' and, without access to that shop's policy, it may invent a believable '30-day return policy' because that pattern is common. In the shop's own assistant, we retrieve the actual policy and pass it in, tell the model to answer only from it and to say 'I don't know' otherwise, and check that answers quote the policy.

    They may follow up with

    • How would you detect a made-up citation automatically?
    • Can RAG remove hallucination completely?

    Avoid

    • Claiming a better prompt alone can eliminate hallucination
    • Not giving the model a way to say it does not know

    Practice this oneLesson: Use a chat assistant for a real task and check the answer

  5. 5ConceptBeginner 3

    What is a webhook, and how is it different from polling?

    Short answer

    Polling means your workflow asks a service every few minutes whether anything changed. A webhook is the reverse: the service sends an HTTP request to your URL the moment an event happens, such as a payment succeeding. Webhooks are faster and use fewer requests, but you must secure the endpoint and handle repeats.

    The explanation

    With a webhook you get a URL from your automation tool and register it in the other service. Security matters: verify the signature the sender includes, use HTTPS, and never trust the payload blindly.

    Senders retry when your endpoint fails, so the same event can arrive twice; make the workflow idempotent by checking an event ID before acting. Polling is still fine when a service offers no webhooks or near-real-time is not needed.

    An example you can use in the interview

    Polling is like phoning the courier every ten minutes to ask if your parcel has arrived. A webhook is the courier ringing your doorbell when it arrives. With Razorpay, for example, a webhook tells your system the moment a payment succeeds, so you do not have to keep asking. You should verify the webhook's signature and ignore repeats of the same event.

    They may follow up with

    • How would you stop a duplicate payment webhook from creating two orders?
    • How do you verify a webhook really came from the sender?

    Avoid

    • Not verifying signatures
    • Assuming each event arrives exactly once

    Practice this oneLesson: Webhooks

  6. 6Use caseBeginner 2

    How do you decide which process to automate first?

    Short answer

    Map the candidate processes and score them: how often they run, how much time each run takes, how many errors they cause, how rule-based and stable the steps are, and how risky a mistake would be. Start with a frequent, repetitive, low-risk process with clear inputs and outputs, so you can show a quick, measurable win.

    The explanation

    A simple scoring table works: volume times minutes per run gives the time at stake; add error cost; subtract for exceptions, unstable inputs or approvals that need judgement. Talk to the people who do the work, because the written process and the real one often differ.

    A first win builds trust. Invoice data entry or lead routing beats a messy, exception-heavy process even if the messy one looks bigger on paper.

    An example you can use in the interview

    At a distributor, we scored five processes. Order entry from emailed purchase orders ran 80 times a day at about 6 minutes each, was rule-based, and a mistake was caught at dispatch. Credit approvals ran less often and were risky. We started with order entry, showed time saved within a month, and used that win to get support for the next process.

    They may follow up with

    • How would you measure the time a process takes today?
    • What makes a process a bad candidate?

    Avoid

    • Automating a broken process without fixing it first
    • Choosing the biggest process instead of the clearest one

    Practice this oneLesson: Mapping a process

  7. 7ScenarioIntermediate 1

    A client says 'automate our sales process'. What do you do first?

    Short answer

    Turn the vague goal into a clear scope. Ask what outcome they want and how they measure it, map the current process step by step with the people who do it, find the slowest or most error-prone steps, and agree one workflow to start with, written acceptance criteria, what is out of scope and who owns it after handover.

    The explanation

    Good questions: Which systems hold the data? What volume? What happens with exceptions? Who approves changes? What does success look like in numbers?

    Write it down in a one-page scope: the trigger, steps, outputs, error handling, success measure and timeline. Saying no to the rest for now protects both you and the client.

    An example you can use in the interview

    A furniture retailer said 'automate our sales process'. I asked what they wanted to change: they said leads were going cold. Mapping the process with two salespeople showed website and Instagram leads were copied by hand once a day. We scoped one workflow: capture leads from both sources into the CRM within five minutes and send a WhatsApp reply, with written acceptance criteria, quotations and follow-ups out of scope, and the sales manager as owner.

    They may follow up with

    • What goes into acceptance criteria?
    • How do you handle a request added halfway through?

    Avoid

    • Starting to build before agreeing scope
    • Automating every step at once

    Practice this oneLesson: Scoping an automation project

  8. 8Use caseBeginner 2

    Walk me through building a simple automation in n8n or Zapier.

    Short answer

    Every workflow starts with a trigger, such as a new form submission or a new row in a sheet. Then actions run in order: transform the data, call a service, send a message. For example: when a Google Form is submitted, add the lead to a CRM, send a WhatsApp or email confirmation, and post in the sales channel.

    The explanation

    Good answers show the habits that make automations reliable: pin or save sample data while building, name each node clearly, map fields deliberately rather than by guesswork, and add an error path that notifies someone.

    Mention where it runs: Zapier and n8n Cloud are hosted; n8n can also be self-hosted, which matters for cost and data residency.

    An example you can use in the interview

    For a yoga studio I built this in n8n: the trigger is a new Google Form booking; a step formats the name and phone; then it adds the person to their Google Sheet of members, sends a confirmation email with the class time, and posts a message in the instructors' Telegram group. If the email fails, the error workflow messages the owner.

    They may follow up with

    • How would you stop duplicate leads?
    • What happens if the CRM is down when the form is submitted?

    Avoid

    • Skipping testing with real-looking data
    • No error handling

    Practice this oneLesson: Your first n8n workflow

  9. 9Use caseBeginner 2

    How would you choose between n8n, Zapier, Make and Power Automate?

    Short answer

    Match the tool to the client. Zapier is the easiest with the most app connectors; Make handles complex visual flows at lower cost; n8n is open source and can be self-hosted, which suits data control and heavy use; Power Automate fits companies already on Microsoft 365. Compare connectors needed, pricing at expected volume, data residency, and who will maintain it.

    The explanation

    Pricing models differ: some charge per task or operation, so a workflow with many steps running thousands of times a day can get expensive; self-hosted n8n trades licence cost for hosting and maintenance effort. Check prices on the vendors' pages at the time, because they change.

    The maintainer matters as much as the features: a tool the client's team can understand and fix beats a cleverer one only you can run.

    An example you can use in the interview

    A five-person marketing agency with simple flows and no developer got Zapier, because they could maintain it themselves. A healthcare startup handling patient data chose self-hosted n8n so data stayed in their own cloud and heavy volumes did not raise per-task costs. A manufacturing client already deep in Microsoft 365 and Teams used Power Automate, since their IT team already managed it.

    They may follow up with

    • How would volume change your choice?
    • When would you recommend self-hosting?

    Avoid

    • Picking a tool before knowing the apps and volume
    • Quoting prices from memory

    Practice this oneLesson: Make, Zapier and Power Automate

  10. 10ScenarioBeginner 3

    An API returns JSON in a different shape from what the next app needs. How do you handle it?

    Short answer

    Look at a real sample of both shapes first. Then map each target field from the source: rename fields, pull values out of nested objects, split or join text, convert dates and numbers, and loop over arrays. Test with edge cases such as missing fields.

    The explanation

    Typical traps: dates in different formats or time zones (store in ISO format, and use Asia/Kolkata for display in India), numbers sent as strings, optional fields that are sometimes absent, and arrays where you expected one object.

    In n8n, use the Edit Fields node, or a small Code node for anything complex. Keep transformations in one clearly named step so the next person can see what changes, and add a check that stops the workflow with a clear error if a required field is missing.

    An example you can use in the interview

    A payment gateway's webhook sent the amount in paise inside payload.payment.entity, while the accounting tool wanted rupees as a top-level 'total' and the customer name split into first and last. In n8n I mapped each field in an Edit Fields (Set) node, divided by 100, split the name on the first space, converted the Unix timestamp to a date, and tested with a payment that had no customer name. That case used 'Walk-in customer' instead of failing.

    They may follow up with

    • How do you handle an array of line items?
    • What would you do if a required field is sometimes missing?

    Avoid

    • Mapping from memory instead of a real sample
    • Ignoring time zones in dates

    Practice this oneLesson: Reading and shaping JSON

  11. 11ScenarioBeginner 3

    An API your workflow calls starts returning 429 errors. What do you do?

    Short answer

    429 means too many requests. I would read the API's rate-limit documentation and the Retry-After header, slow down with batching or a delay between calls, and retry with exponential backoff instead of hammering it. If volume truly needs more, I would ask for a higher limit or use a bulk endpoint, and make sure failed items are retried, not lost.

    The explanation

    Rate limits protect the provider's service. A workflow that loops over hundreds of records as fast as it can will hit them. The right response respects the limit rather than fighting it.

    Practical steps: honour Retry-After if present; add backoff with jitter; process records in batches with pauses; cache results you fetch repeatedly; and queue failed items for a later retry so nothing silently disappears. Log how often you hit the limit, because a steady increase usually means the business has grown and the design needs to change.

    An example you can use in the interview

    A workflow that synced contacts from a form tool to a CRM ran fine until a webinar brought in two thousand sign-ups in an hour. The CRM started returning 429s and some contacts never arrived. I changed it to batches of fifty with a pause between them, retries with backoff that read the Retry-After header, and a 'failed' sheet that a second workflow retried every hour. Everything synced that evening, and I set an alert if the failed sheet ever grows.

    They may follow up with

    • What is exponential backoff?
    • How would you make sure no records were lost?

    Avoid

    • Retrying immediately in a tight loop
    • Letting failed records drop silently

    Practice this oneLesson: Call a web API and read JSON

  12. 12ScenarioBeginner 3

    When would you use RPA instead of an API integration?

    Short answer

    Use an API whenever one exists: it is faster, more reliable and easier to maintain. Use RPA, software robots that click through screens like a person, when a system has no API, such as an old desktop app or a government portal, and the process is stable. RPA breaks when screens change, so monitor it closely.

    The explanation

    RPA tools such as UiPath or Power Automate Desktop record or script interactions with the user interface. They are a bridge, not a foundation: every layout change can break a bot.

    Good practice: confirm there really is no API or export, keep bots to stable screens, handle pop-ups and timeouts, and log screenshots on failure.

    An example you can use in the interview

    A client wanted invoices entered into both their cloud accounting tool and an old desktop inventory system. The accounting tool had an API, so we used it, which was fast and reliable. The inventory system had no API and no import, and its screens had not changed in years, so we used RPA for that part, with screenshots and an alert if a screen looked different.

    They may follow up with

    • What breaks RPA bots most often?
    • How would you monitor a bot that runs overnight?

    Avoid

    • Choosing RPA when an API exists
    • No plan for screen changes

    Practice this oneLesson: RPA basics

  13. 13Use caseBeginner 3

    A client's old billing software has no API. How would you automate data entry into it?

    Short answer

    First look for another way in: an import feature, a supported database or file import, or a newer version with an API. If none exists, use RPA to drive the screen like a person would, with a dedicated bot account, checks after each entry, screenshots and logs, and an alert when the screen changes.

    The explanation

    RPA works through the user interface, so it is slower and breaks when screens change. It is the right choice when there is truly no better option and the volume justifies it, but it needs care.

    Good RPA design uses stable selectors, waits for screens to load, verifies each saved record, runs under its own account with only the permissions it needs, and keeps a queue so failed items can be retried. Agree with the client who maintains it, and document it, because these bots need regular attention.

    An example you can use in the interview

    A clinic chain used a 15-year-old desktop billing system, and staff re-typed lab orders from a web form into it. The vendor confirmed there was no API but there was a CSV import for patients, not orders. So we used the import for new patients and Power Automate Desktop for orders, under a separate bot login. The bot checked each saved order number, took a screenshot, and moved failures to a list for the front desk. It needed fixing twice in the first year when the vendor updated a screen, which we had warned them about.

    They may follow up with

    • What makes RPA bots break?
    • How would you test it before going live?

    Avoid

    • Jumping to RPA without checking for import features
    • Running the bot under a person's login

    Practice this oneLesson: RPA basics

  14. 14Use caseIntermediate 1

    Design an automation that reads supplier invoices from email and enters them into accounting software.

    Short answer

    Trigger on new emails with attachments, extract text from the PDF or image, then ask a language model to return fixed JSON fields (supplier, GSTIN, invoice number, date, line items, tax, total) using structured output. Validate in code (totals add up, valid GSTIN, no duplicates), send failures to a person, then create the bill.

    The explanation

    The model does the reading; code does the checking. Keep a human approval step at first and measure accuracy on a sample of real invoices before letting clean ones go straight through.

    Also think about data: invoices contain business and sometimes personal data, so use an approved provider, keep files in the company's storage, and log every action for audit.

    An example you can use in the interview

    For a trading company I built this in n8n: an email trigger for the accounts inbox, a PDF-to-text step, and a model call returning JSON with supplier, GSTIN, invoice number, date, line items, tax and total. A code step checked that totals added up, the GSTIN was 15 characters in the right pattern, and the invoice number was not already in Tally. Anything failing went to the accountant's queue; the rest created a draft bill.

    They may follow up with

    • What checks would you run on the extracted totals?
    • When would you remove the human approval step?

    Avoid

    • Trusting model output without validation
    • No duplicate check

    Practice this oneLesson: Extracting data from documents

  15. 15ScenarioIntermediate 1

    The AI in your invoice automation sometimes reads the total wrong. How do you stop wrong amounts reaching the accounts?

    Short answer

    Never trust the extracted total alone. Check it: line items should add up to the subtotal, tax should match the rate, and the total should equal subtotal plus tax. Compare with the purchase order where one exists. Anything that fails a check or is above a set amount goes to a person before posting. Track the error rate.

    The explanation

    AI extraction is good but not perfect, especially with scans, handwriting or unusual layouts. Because invoices contain numbers that relate to each other, you can catch most mistakes with arithmetic rules instead of trusting the model.

    Common checks: line items times quantities add to the subtotal; GST amounts match the rates and the GSTIN format is valid; the total matches; the vendor exists in your master data; the invoice number is not a duplicate. Route failures to a review queue showing the image beside the extracted fields, and use corrected examples to improve prompts.

    An example you can use in the interview

    For a distributor processing about two hundred supplier invoices a week, the AI occasionally read ₹18,450 as ₹1,845 on faint scans. We added checks: line totals must add up to the subtotal, GST must match the stated rates, and the total must equal the subtotal plus GST. Any mismatch, or any invoice over ₹50,000, went to the accountant with the scan beside the fields. Wrong amounts stopped reaching the ledger, and only a small share of invoices needed a human look.

    They may follow up with

    • What threshold would you use for human review?
    • How would you measure accuracy over time?

    Avoid

    • Posting whatever the model extracts
    • Reviewing everything by hand, which removes the benefit

    Practice this oneLesson: Extracting data from documents

  16. 16Use caseIntermediate 1

    How would you build AI triage for a support inbox?

    Short answer

    Classify each incoming email with a model into a fixed set of categories and urgency, extract key fields such as order number, and route it to the right queue or person. Use clear category definitions with examples, ask for structured output, and send uncertain cases to a human. Measure accuracy on a labelled sample and review misroutes weekly.

    The explanation

    Start with the categories the support team already uses, not new ones. Add a confidence or 'other' category so the model is not forced to guess. Draft replies can come later, with a person approving them before sending.

    Track the business result: time to first response, misroute rate and how often agents change the category.

    An example you can use in the interview

    For an online shop's support inbox, I defined seven categories with two example emails each, such as 'refund request' and 'delivery delay', plus urgency. The model returned JSON with category, urgency, order number and a confidence flag. Low-confidence emails went to a general queue. On 300 labelled emails it routed most correctly, and every Monday we reviewed misroutes and updated the definitions.

    They may follow up with

    • How would you measure whether the triage is good enough?
    • What would you do with emails in Hindi or Tamil?

    Avoid

    • Letting the model invent categories
    • No fallback for uncertain cases

    Practice this oneLesson: Triage for email and tickets

  17. 17Use caseIntermediate 2

    How would you design a WhatsApp assistant for a small business, and when should it hand over to a person?

    Short answer

    Use the WhatsApp Business Platform through an approved provider, answer common questions from the business's own FAQ and catalogue with retrieval, and use templates for messages the business starts, as Meta requires. Hand over to a person when the customer asks, when the bot is unsure, for complaints, payments or anything sensitive, and tell the customer it is happening.

    The explanation

    Keep the scope narrow at first: opening hours, prices, order status and booking. Store conversation history securely and follow the DPDP Act for personal data. Meta's rules require opt-in before sending business-initiated messages.

    Measure resolution rate, handover rate and customer satisfaction, and read a sample of conversations every week to find gaps.

    An example you can use in the interview

    For a bakery taking cake orders on WhatsApp, the assistant answers prices, flavours and delivery areas from their menu sheet, and uses an approved template for order-ready messages. It hands over to the owner when a customer asks for a person, wants a custom design, complains, or when it is unsure, and it says 'I'm passing you to Anita, who will reply shortly.'

    They may follow up with

    • How would the bot know it is unsure?
    • What should never be automated in this bot?

    Avoid

    • No way to reach a human
    • Sending promotional messages without opt-in

    Practice this oneLesson: Chat assistants for websites and WhatsApp

  18. 18ScenarioIntermediate 1

    Your automation fails halfway through. How do you design for that?

    Short answer

    Assume steps will fail. Add retries with backoff for temporary errors, an error workflow that alerts a person with the details, and make steps idempotent so re-running does not create duplicates. Log each run with its input, and keep a way to replay failed items once the cause is fixed.

    The explanation

    Distinguish temporary failures, such as a timeout or rate limit, from permanent ones, such as invalid data; retry the first and route the second to a person. In n8n, an error workflow can catch failures from any workflow and post to a team channel.

    For multi-step actions that must all happen or none, record progress so you can resume, or add a compensating step, such as cancelling an order created before the payment step failed.

    An example you can use in the interview

    In a workflow that created a customer in the CRM, then an invoice, then sent an email, the invoice API sometimes timed out after the customer was created. Re-running created duplicate customers. I made each step check first, looking up the customer by email before creating, added retries with backoff on the invoice step, and an error workflow that posted the failed record to a Slack channel with a 'replay' link once fixed.

    They may follow up with

    • How do you avoid creating a customer twice when a step is retried?
    • What should the alert say?

    Avoid

    • Silent failures nobody notices
    • Retrying permanent errors forever

    Practice this oneLesson: Error handling and retries

  19. 19ScenarioBeginner 3

    Your automation created every order twice last night. How do you find and fix it?

    Short answer

    Stop the workflow, list the duplicates and clean them up with the business. Then check the run history: often the trigger fired twice, a webhook was retried after a timeout, or a step was rerun after a failure. Fix it with idempotency: a unique order ID checked before creating, and safe retries. Add an alert for unusual volumes.

    The explanation

    Duplicates are one of the most common automation bugs. Webhook senders often retry when they do not get a quick success response, and many platforms rerun failed executions. If your 'create' step does not check whether the record already exists, every retry creates another copy.

    The standard fix is an idempotency key, usually the source system's order or event ID, stored and checked before creating anything. Respond to webhooks quickly and do slow work after. Then add monitoring: a count of records created per hour with an alert when it jumps.

    An example you can use in the interview

    In an n8n workflow I built for a small online shop, Shopify webhooks created invoices in their accounting tool. One night the accounting API was slow, our workflow responded late, Shopify retried, and about sixty orders were invoiced twice. I paused it, exported the duplicates for the accountant to void, and changed the workflow to respond to the webhook immediately, then check for an existing invoice with the Shopify order ID before creating one. I added a daily count check that messages me if invoices outnumber orders.

    They may follow up with

    • How would you clean up the duplicates safely?
    • What is an idempotency key?

    Avoid

    • Deleting duplicates by hand and not fixing the cause
    • Turning off retries completely

    Practice this oneLesson: Error handling and retries

  20. 20ScenarioIntermediate 1

    Your automation handles customer personal data. What precautions do you take?

    Short answer

    Pass only the fields each step needs, use approved tools and AI providers, keep credentials in the tool's secure store, restrict who can view executions, mask personal data in logs, and delete old run data. Under India's DPDP Act, data should be used only for a lawful purpose, usually the one customers consented to.

    The explanation

    Automation tools keep execution history, which can quietly become a copy of customer data; set them to prune old executions. Check where a hosted tool stores data, and whether sending data to an AI model is allowed by the company's policy.

    Document the data flow: which systems, which fields, who can access them. That document is what a security review or a customer will ask for.

    An example you can use in the interview

    In a workflow that moved patient appointment details from a form to a clinic's calendar, I passed only name, phone and time, not the reason for the visit. Credentials sat in n8n's credential store, only two staff could view executions, and execution data was deleted after 14 days. The AI step that wrote reminder messages used the clinic's approved provider and got only the first name and time.

    They may follow up with

    • Where can personal data leak in an n8n setup?
    • What would you put in a data-flow document?

    Avoid

    • Logging full payloads forever
    • Sending personal data to unapproved AI tools

    Practice this oneLesson: Data quality and personal data

  21. 21ScenarioBeginner 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

  22. 22Use caseIntermediate 1

    How do you measure the return on an automation project?

    Short answer

    Measure the baseline before you build: runs per month, minutes per run, error rate and cost of errors. After launch, measure the same things plus the cost to build and run the automation (tools, model calls, maintenance). Hours saved times loaded hourly cost, plus errors avoided, minus running cost, gives the monthly return; payback is build cost divided by that.

    The explanation

    Example: 1,200 invoices a month at 6 minutes each is 120 hours. If automation handles 80% and a person spends one minute checking each automated invoice, the work drops to 40 hours (24 hours for the remaining 240 invoices plus 16 hours of checks), so you save 80 hours a month. At ₹500 an hour that is ₹40,000 a month before tool and maintenance costs.

    Also report benefits that are real but harder to price, such as faster response times, and be honest about time saved that is not redeployed to useful work.

    An example you can use in the interview

    For a lead-entry automation, the baseline was 600 leads a month at 4 minutes each, about 40 hours, plus a few lost leads from typos. At a loaded cost of ₹400 an hour that is ₹16,000 a month. The automation cost ₹2,500 a month in tools and model calls and ₹60,000 to build. So it saves about ₹13,500 a month and pays back in under five months, before counting the leads no longer lost.

    They may follow up with

    • How would you measure minutes per run before automating?
    • What costs do people forget?

    Avoid

    • No baseline
    • Ignoring maintenance and model costs

    Practice this oneLesson: Measuring return on investment

  23. 23Real lifePro 1

    How do you hand over an AI system to the client's team?

    Short answer

    Plan it from the start. Provide documentation for different readers: how to use it, how to run and monitor it, how it is built, and why key decisions were made. Hand over the eval set and how to run it, runbooks for common failures, and access and secrets transfer. Pair with their engineers before leaving and agree a support window.

    The explanation

    The Diátaxis approach separates tutorials, how-to guides, reference and explanation, which helps each reader find what they need.

    The real test of a handover is whether the client's team can change a prompt, re-run evals and fix a common incident without you. Run that test before the engagement ends.

    An example you can use in the interview

    At the end of a support-assistant project, I left a user guide for agents, an operations runbook covering the five most common failures, an architecture document with why we chose each component, and the eval set with a one-command script to run it. For two weeks their engineers ran deployments while I watched, and we agreed a month of email support. They ran their first model update on their own.

    They may follow up with

    • What goes into a runbook?
    • How would you check the client can run it without you?

    Avoid

    • A single giant document nobody reads
    • Leaving without handing over evals

    Practice this oneLesson: Documentation and handover

  24. 24Real lifeBeginner 3

    The client's staff are worried your automation will replace them. How do you handle it?

    Short answer

    Take it seriously and involve them early. I meet the people who do the work, map the process with them, and ask which parts they find most tedious. I am honest about what will change, show how the automation removes repetitive steps rather than judgement, give them control over exceptions, and make sure management communicates its plans for their roles.

    The explanation

    Fear of job loss is reasonable, and ignoring it leads to quiet resistance, workarounds and failed projects. The people doing the work also know the exceptions that will break your automation, so their help makes it better.

    Be careful not to make promises that are management's to make. Your part is to involve staff in design, train them, make them owners of the parts that need judgement, and share early wins that make their day easier. Encourage the client's leaders to say clearly how time saved will be used.

    An example you can use in the interview

    At a logistics client, the billing team of four were cold in our first meeting. I asked to sit with them for a morning and list every step of raising an invoice. They pointed out three exception types I had not heard about. We automated the copying between systems and left exceptions and customer calls with them. I asked their manager to tell the team directly that the time saved would go to clearing a backlog of overdue payments. By launch, the team lead was the one demoing it to other branches.

    They may follow up with

    • What if management does plan to reduce staff?
    • How do you get their help mapping the process?

    Avoid

    • Promising nobody will lose their job when you cannot know
    • Designing the automation without the people who do the work

    Practice this oneLesson: Scoping an automation project

  25. 25Real lifeBeginner 3

    Tell me about an automation you built and how much time it saved.

    Short answer

    Pick one real automation. Describe the manual process before, with a measured number such as minutes per item and items per week; what you built and with which tool; how you handled errors; and the result after a few weeks, measured the same way. Mention who uses it now and one thing you would improve.

    The explanation

    Interviewers want evidence that you understand the business, not only the tool. Start with the problem as the user felt it, and give a baseline you actually measured or carefully estimated. Explain the design briefly, including what happens when something fails, because that shows real-world experience.

    The result should be honest and specific: time saved, errors reduced, faster response. If it is a personal or volunteer project, that is fine. Being clear about your assumptions, such as how you estimated the time before, makes your number believable.

    An example you can use in the interview

    A coaching centre's office copied every new enquiry from their website form, WhatsApp and phone notes into a spreadsheet, then sent a brochure by hand. I timed it with the staff: about six minutes per enquiry, around 300 enquiries a month, so roughly 30 hours a month. I built an n8n workflow that collects form and WhatsApp Business enquiries into the sheet and sends the brochure automatically, with failures posted to a staff group. After a month it was about eight hours of checking and follow-up calls, and replies went out within minutes instead of the next day.

    They may follow up with

    • How did you measure the time before?
    • What happened when it failed?

    Avoid

    • Big savings figures with no baseline
    • Only describing the tool's features

    Practice this oneLesson: Measuring return on investment

Today's AI turns are used

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