AWS AI services interview questions in 2026 test whether you can pick the right managed service for a job, know which services are still open to new customers, and put them together into a secure, affordable pipeline, not whether you can recite a product list. This guide has 60 commonly asked questions with model answers. It covers Rekognition, Textract, Comprehend and Comprehend Medical, Transcribe, Polly, Translate, Lex, Personalize, HealthScribe, Amazon Nova, Bedrock Data Automation, OpenSearch and S3 Vectors, the rename and retirement changes behind Kendra, Amazon Q Business, Q Developer, Forecast, Fraud Detector and the Lookout services, and scenario rounds on document processing, contact centres, search assistants and migrations.
How to use this guide
This is the breadth guide: it covers the whole AWS AI services portfolio and how the services fit together. Three areas have their own deep guides, and this page links to them rather than repeating them: Amazon Bedrock interview questions (models, Knowledge Bases, Guardrails, evaluation), Bedrock AgentCore interview questions (agent runtime, gateway, identity, memory) and SageMaker AI interview questions (training, endpoints, MLOps). What interviewers typically probe at each level:
- Freshers and cloud engineers moving into AI: what each service does, which API fits which input, and the difference between a task-specific AI service and a foundation model.
- Mid-level engineers: async jobs, confidence thresholds, IAM service roles, encryption, quotas, event-driven pipelines and how each service is priced.
- Senior engineers and architects: managed service versus GenAI decisions, lifecycle risk (services closed to new customers), migration plans, data residency and cost per business transaction.
Service status changes often on AWS. Every status in this guide was checked against AWS documentation in October 2026; in an interview, say "as of the current documentation" and you will sound like someone who checks.
- Fundamentals and service status (Q1โQ10)
- Vision, documents, speech and language (Q11โQ20)
- Conversation, search, personalisation and forecasting (Q21โQ28)
- GenAI building blocks and retired services (Q29โQ36)
- Security and compliance (Q37โQ43)
- Cost, architecture and operations (Q44โQ50)
- Real-world scenarios (Q51โQ60)
- Key takeaways
- Interview preparation checklist
- FAQ
Fundamentals and service status
1. How would you describe the AWS AI stack, and where do the "AI services" sit in it?
Answer: Three layers. At the top are pre-trained, task-specific AI services: Rekognition (images and video), Textract (documents), Comprehend (text analytics), Transcribe (speech-to-text), Polly (text-to-speech), Translate, Lex (chatbots), Personalize (recommendations), plus domain services such as Comprehend Medical and HealthScribe. You call an API and get structured output without ML expertise. In the middle is Amazon Bedrock: foundation models including Amazon Nova, with Knowledge Bases, Guardrails, Data Automation and AgentCore. At the bottom is SageMaker AI for training and hosting your own models. Applications such as Amazon Quick and Kiro sit on top and are used rather than built.
Apps: Amazon Quick, Kiro, Connect AI agents
AI services: Rekognition Textract Comprehend Transcribe
Polly Translate Lex Personalize HealthScribe
GenAI: Bedrock (Nova + partner models), BDA,
Knowledge Bases, Guardrails, AgentCore
Build-own: SageMaker AI, EC2/EKS with GPUs
2. When do you choose a managed AI service over a GenAI model on Bedrock?
Answer: Choose the managed AI service when the task is narrow, well defined and high volume, and you need predictable output: OCR and table extraction, speech-to-text with word timestamps, PII entity detection with character offsets, face comparison, machine translation with fixed terminology. These services return typed, deterministic structures with confidence scores, are priced per page, minute or character, and are easy to test. Choose a foundation model when the task needs reasoning, flexible output or a long tail of variations: summarising a claim, answering questions over policies, classifying into categories that change monthly, or extracting fields from documents with no stable layout. Many designs use both: the AI service does perception (text, speech, entities) and the model does judgement.
| Factor | Managed AI service | Foundation model |
|---|---|---|
| Output | Fixed schema, offsets, confidence | Flexible text or JSON you define |
| Change a behaviour | Custom model, vocabulary or adapter | Change the prompt |
| Testing | Field accuracy, WER, precision/recall | Needs an evaluation set and judging |
| Cost shape | Per unit (page, minute, character) | Per token, varies with prompt size |
| Risk | Narrow errors, easy to bound | Hallucination, prompt injection |
Interview tip: Add lifecycle as a factor. In 2026 AWS has moved several narrow services into maintenance and points customers to Bedrock instead, so "will this service still accept new accounts next year?" is now part of the decision.
3. Which AWS AI services are retired or closed to new customers, and what does AWS recommend instead?
Answer: As of October 2026, AWS documentation shows the following. "Closed to new customers" means existing customers carry on; "end of support" means the service stops.
| Service or feature | Status | Recommended direction |
|---|---|---|
| Amazon Kendra | Maintenance from 30 June 2026; closed to new customers from 30 July 2026 | Amazon Bedrock Managed Knowledge Base |
| Amazon Q Business | Maintenance; closed to new customers since end of July 2026 | Amazon Quick |
| Amazon Q Developer | CLI rebranded as Kiro; IDE plugins end of support 30 April 2027 | Kiro |
| Amazon Forecast | Closed to new customers (since 29 July 2024) | SageMaker Canvas forecasting |
| Amazon Fraud Detector | Closed to new customers since 7 November 2025 | AutoGluon with SageMaker AI; AWS WAF Fraud Control for sign-up and login abuse |
| Lookout for Vision | End of support 31 October 2025 | SageMaker AI (including JumpStart models) or Bedrock |
| Lookout for Metrics | End of support 10 October 2025 | Rebuild anomaly detection with other AWS services |
| Lookout for Equipment | Closed to new customers 7 October 2025; end of support 7 October 2026 | AWS IoT SiteWise anomaly detection |
| Amazon Monitron | Closed to new customers since 31 October 2024 | AWS Partner solutions |
| Comprehend topic modeling, event detection, prompt safety classification | Not available to new customers (other Comprehend features unaffected) | Bedrock models; Bedrock Guardrails prompt-attack filter |
| Rekognition people pathing | End of support 31 October 2025 | Other Rekognition Video features or a custom model |
| Amazon Augmented AI (A2I) | Maintenance; closed to new customers from 30 July 2026 | Build human review into your own workflow |
| Bedrock Agents (now Agents Classic) | Closed to new customers from 30 July 2026 | Bedrock AgentCore |
Interview tip: Precision matters. "Kendra is shut down" is wrong; it is in maintenance with existing customers supported. "Use Kendra for a new search project" is also wrong for a new account. Saying both halves correctly is a strong signal.
4. What do "maintenance", "sunset" and "closed to new customers" mean in AWS lifecycle terms?
Answer: AWS's lifecycle page defines three stages. Maintenance: new customers cannot onboard, existing customers continue, and AWS keeps operating and supporting the service but adds no new features. Sunset: existing customers should plan to migrate, and there is a published sunset date (typically about a year out) when AWS ends operations and support. Full shutdown: the service is removed entirely. "Closed to new customers" is the practical effect of maintenance. Accounts with recent usage are usually allowlisted, so whether "you" are an existing customer is decided per account, which catches teams who create a new AWS account for a new project.
5. What are the key differences between Amazon Lex, Amazon Polly and Amazon Transcribe?
Answer: They cover different steps of an interaction. Transcribe is speech-to-text, batch or streaming, with timestamps, speaker labels and PII redaction. Polly is text-to-speech from text or SSML, with Standard, Neural, Long-form and Generative engines. Lex is the conversational layer: it recognises intent, fills slots through dialogue and calls Lambda for fulfilment, with its own speech handling for voice bots, so you do not wire Transcribe and Polly into it yourself. Lex integrates natively with Amazon Connect. Use Transcribe alone for recordings, Polly alone for announcements and accessibility, and Lex for two-way task-oriented conversations.
Interview tip: Mention that for open-ended voice conversations, a speech-to-speech model such as Amazon Nova Sonic on Bedrock is now an alternative to the classic Lex intent model. Our voice AI interview questions go deeper on latency and barge-in.
6. How does Amazon Rekognition handle face detection, comparison and search?
Answer: Rekognition separates three operations. Detection (DetectFaces) returns bounding boxes, landmarks, pose, quality and attributes such as an estimated age range. Comparison (CompareFaces) returns similarity scores for one-to-one verification, such as a selfie against an ID photo. Search works against a face collection you own: IndexFaces stores feature vectors (not images), and SearchFacesByImage or SearchFaces does one-to-many matching above a threshold you choose. Stored video uses async jobs; streaming video uses stream processors. Face Liveness checks that a real person is in front of the camera, to deter spoofing with photos, replayed video or masks.
Production consideration: Thresholds are business decisions. A high threshold for unlocking a bank account, human review below it, and never treat a match as identity proof without a second factor. Face data is biometric personal data, so consent, retention and deletion of collection entries need a documented policy.
7. Which Textract API would you use for which kind of document?
Answer: DetectDocumentText for plain OCR of printed and handwritten text. AnalyzeDocument when structure matters, with feature types FORMS (key-value pairs), TABLES, QUERIES ("What is the policy number?" answered with its location), SIGNATURES and LAYOUT. AnalyzeExpense for invoices and receipts. AnalyzeID for identity documents, documented for US government-issued IDs, so test carefully before using it on Indian documents. Analyze Lending to classify and route mortgage package pages. Custom Queries adapters tune Queries on your samples. Use synchronous calls for single pages and the async Start/Get APIs for multi-page PDFs.
8. What does Amazon Comprehend do today, and what changed in 2026?
Answer: Comprehend is managed NLP for text analytics: language detection, entities, key phrases, sentiment and targeted sentiment, syntax, PII detection and redaction, toxicity detection, and custom classification and custom entity recognition trained on your data, run synchronously, as async jobs or behind a real-time endpoint. The 2026 change: topic modeling, event detection and prompt safety classification are no longer available to new customers; accounts that used them in the last 12 months keep access, and other features are unaffected. AWS's migration guides replace topics and events with Bedrock model prompts and prompt safety with the Bedrock Guardrails prompt-attack filter.
Interview tip: The lesson interviewers want: AWS itself is moving open-ended text understanding to foundation models, while keeping narrow, offset-precise features like PII detection in Comprehend.
9. How do you get consistent, domain-correct translations from Amazon Translate?
Answer: Amazon Translate is a neural machine translation service that supports a long list of languages, including Hindi, Telugu, Tamil, Kannada, Malayalam, Marathi, Bengali, Gujarati, Punjabi and Urdu. For consistency, use custom terminology: a CSV or TMX list that forces brand names, product names and regulated terms to translate a fixed way (or not at all). For domain style, use Active Custom Translation with parallel data, which adapts batch translations to examples of your own translated content without training a model. Settings such as formality, profanity masking and brevity are supported for some language pairs. Use real-time translation for chat and short text and batch jobs for document sets in S3.
10. What are Amazon Q Business and Amazon Q Developer in 2026?
Answer: Both names have changed. Amazon Q Business is in maintenance and no longer open to new customers; existing customers get bug fixes and security updates only. AWS recommends Amazon Quick, which it calls the next evolution of Q Business, with Quick Sight, Flows, Automate, Research, Spaces and Apps, connecting to Slack, Teams, Outlook, CRMs, databases and documents. The Amazon Q Developer CLI has been rebranded as Kiro, the Kiro console is the rebranded Q Developer console, and the Q Developer IDE plugins reach end of support on 30 April 2027. Q Developer in the AWS Management Console is still documented. In Amazon Connect, what was Amazon Q in Connect is now one of the Connect AI agents.
Vision, documents, speech and language
11. Beyond faces, what does Rekognition offer, and what has been retired?
Answer: Label detection for objects, scenes and concepts; text in images; content moderation with a hierarchical unsafe-content taxonomy; celebrity recognition; PPE detection; image properties; and video segment detection (black frames, end credits, shots). Custom Labels trains your own classifier or detector for logos, products or defects, and adapters tune moderation and labels on your samples. People pathing reached end of support on 31 October 2025, so designs that depended on it need another approach.
Production consideration: For brand-new visual tasks where you need explanation or free-form description ("is this shelf display compliant with the planogram?"), a multimodal model or Bedrock Data Automation for images may beat a custom classifier. For high-volume moderation with a fixed taxonomy, Rekognition is simpler and easier to audit. See our computer vision interview questions for the model-level details.
12. How do you process large multi-page PDFs with Textract reliably?
Answer: Use the asynchronous APIs. Upload to S3, call StartDocumentAnalysis (or the text, expense or lending equivalents) with an SNS notification channel, and on completion call the matching Get operation, paging with NextToken. Track each job in a DynamoDB table (document ID, status, attempts, output location), because notifications and consumers both fail sometimes. Respect the quotas on concurrent jobs and transactions per second: queue submissions through SQS instead of firing thousands of Start calls at once, and back off exponentially on throttling.
S3 upload -> EventBridge -> SQS -> Lambda: StartJob Textract -> SNS -> SQS -> Lambda: Get results -> normalise -> validate -> S3 / DynamoDB failures -> DLQ -> human review queue
13. Textract, Bedrock Data Automation or a multimodal model directly: how do you choose for document extraction?
Answer: Textract when you need OCR geometry, tables, forms and per-field confidence for stable layouts at high volume. Bedrock Data Automation (BDA) when you want classification, extraction, normalisation and validation defined as blueprints in one managed step, with visual grounding and confidence scores, and the same interface for images, audio and video. A multimodal model on Bedrock when extraction needs reasoning or is part of a larger prompt, at the cost of your own schema validation and evaluation. Start with the most constrained tool that meets the accuracy bar and move up only for the documents that fail.
What I would check:
- How many layouts and how stable they are.
- Whether downstream systems need bounding boxes for reviewers.
- Field-level accuracy on a labelled sample of real documents, including bad scans.
- Cost per document at expected volume for each option.
Our explainer on document parsing for RAG covers the parsing side for retrieval use cases.
14. What is Amazon Bedrock Data Automation, and how is it organised?
Answer: Bedrock Data Automation uses generative AI to turn documents, images, video and audio into structured output through one API. Standard output gives per-modality results, such as a document's text, layout and summary or a video's scene summaries and moderation signals. Custom output uses blueprints: a schema of typed, described fields that BDA fills for matching documents. Blueprints and settings are grouped into a project you call at runtime. Visual grounding and confidence scores let you route weak results to people, and BDA can also be the parser for a Bedrock Knowledge Base. Still validate output against business rules before writing to a system of record.
15. How do you use Comprehend for PII detection and custom models?
Answer: DetectPiiEntities returns PII types with character offsets and confidence; ContainsPiiEntities returns only which types appear, which is enough for routing; async PII jobs write redacted copies to S3, masking or replacing by type. Custom classification trains a multi-class or multi-label document classifier from labelled examples, and custom entity recognition learns your own entity types from annotations or entity lists. Custom models run as async jobs or behind a real-time endpoint you pay for while it runs, so size endpoints carefully and delete them when idle.
16. What does Amazon Comprehend Medical add over Comprehend?
Answer: Comprehend Medical is a separate HIPAA-eligible service tuned for clinical text in US English. DetectEntities finds medical conditions, medications with dosage, strength and frequency, tests, treatments, anatomy and protected health information; DetectPHI focuses on PHI only. Its ontology linking operations map entities to standard codes: InferICD10CM for conditions, InferRxNorm for medications and InferSNOMEDCT for conditions, anatomy, tests and procedures. AWS states it does not persistently store customer content, and that output must be reviewed by trained professionals before use in patient care.
Interview tip: Point out the English-only limit. For Indian hospitals with notes in mixed English and regional languages, you need a language strategy (translation, or a model evaluated on that data) before Comprehend Medical can help.
17. Which Transcribe features matter in production?
Answer: Batch jobs for recordings and streaming for live audio. Custom vocabularies for product, drug and brand names; custom language models for heavier domain adaptation; vocabulary filters; speaker diarisation or channel identification; automatic language identification; and PII redaction of transcripts (Call Analytics batch jobs can also produce redacted audio). Transcribe Call Analytics adds sentiment, talk time, interruptions, issue detection, categories and generative call summaries. Measure word error rate on your own recordings, split by channel quality and accent.
Production consideration: Telephone audio is often 8 kHz; test with real call recordings. Region support differs between Transcribe, Transcribe Medical and Call Analytics, so check the region table before promising a Mumbai deployment.
18. Transcribe Medical or AWS HealthScribe: what is the difference?
Answer: Transcribe Medical is speech recognition tuned for medical vocabulary in US English, batch or streaming, producing a transcript. AWS HealthScribe adds generative AI: from a patient-clinician conversation it produces a transcript with speaker roles and dialogue classification, extracts medical terms, and drafts a clinical documentation file whose summaries link back to evidence in the transcript. It is HIPAA-eligible, US English, documented in US East (N. Virginia), and AWS states its output is assistive and must be reviewed by clinicians. AWS has also introduced an ambient documentation agent in Amazon Connect Health.
Interview tip: The evidence links are the feature to talk about. They let a doctor check each sentence of a note against what was actually said, which is the control that makes a generated note acceptable.
19. How do you choose and control Amazon Polly voices?
Answer: Polly has four engines: Standard (concatenative, cheapest), Neural (more natural, the usual default), Long-form (designed for long narration such as articles and training content) and Generative (most expressive and conversational). Not every voice or language is available on every engine, so pick the language first. Control pronunciation and pacing with SSML (breaks, emphasis, say-as for dates and numbers, phonemes) and with lexicons for consistent pronunciation of names and acronyms. Use SynthesizeSpeech for short real-time audio and StartSpeechSynthesisTask for long text written to S3. Speech marks give word and sentence timings for lip-sync or highlighting.
20. How would you design multilingual support for an Indian customer base using AWS AI services?
Answer: Check language support per service first. Translate covers major Indian languages; Transcribe, Polly and Lex support a narrower, different set of Indian locales; Comprehend Medical and HealthScribe are English-only. For text chat, detect language, translate with a glossary if needed, process, and translate back. For voice, use natively supported languages and fall back to a human queue or DTMF for others. Code-mixed speech such as Telugu-English is the hard case; test it explicitly, and evaluate a multilingual foundation model against the translate-then-process chain.
Conversation, search, personalisation and forecasting
21. Explain the core concepts of an Amazon Lex V2 bot.
Answer: A bot has one or more locales (language variants). Each locale has intents (what the user wants, such as CheckBalance), defined by sample utterances, and slots (the parameters the intent needs, each with a slot type such as a date or a custom list). Lex runs the dialogue to elicit and confirm slots, and calls a Lambda function as a code hook for validation during the dialogue and for fulfilment at the end. Session attributes carry context. Builds produce versions, and aliases point to versions so you can promote dev to prod and roll back. Channels include Amazon Connect for voice and web or messaging integrations.
22. What generative AI features does Lex V2 have, and how do they use Bedrock?
Answer: Lex V2's generative features run on Bedrock models you select per bot locale. The documentation lists: the descriptive bot builder (create a bot's intents and slots from a natural-language description), utterance generation, assisted slot resolution (use a model to resolve awkward slot values), assisted NLU for intent classification and slot resolution, intent disambiguation for ambiguous input, an AI-powered bot analyzer, AMAZON.QnAIntent for answering questions from a knowledge source, and a BedrockAgentIntent for calling a Bedrock agent. Features are switched on per locale and billed as Bedrock usage.
Production consideration: BedrockAgentIntent depends on Bedrock Agents, which is now Agents Classic and closed to new customers, so for a new account plan agentic steps through a Lambda code hook that calls your own agent (for example on AgentCore) instead. The Bedrock Agents to AgentCore migration guide covers that move.
23. What was Amazon Kendra used for, what is the GenAI index, and what is its status now?
Answer: Kendra is a managed intelligent search service: connectors crawl sources such as SharePoint and S3, semantic ranking returns passages for natural-language queries, and results can respect document permissions. It offers the GenAI Enterprise Edition index (the GenAI index, recommended for RAG and usable from Bedrock Knowledge Bases and Q Business), plus Basic Enterprise and Basic Developer editions. Status: maintenance mode from 30 June 2026 and closed to new customers from 30 July 2026; existing customers remain supported with bug fixes and security updates. AWS recommends Amazon Bedrock Managed Knowledge Base for migrations and new search applications.
Interview tip: Know the feature gaps AWS itself lists for the migration: query suggestions, faceted search, custom synonyms, spell checking, incremental learning from feedback and document enrichment hooks all need workarounds. That is the difference between "just switch" and a real migration plan.
24. A team built on Amazon Q Business. What is the migration path?
Answer: AWS documents a phased move to Amazon Quick. First, Bring Your Own Index (BYOI): connect the existing Q Business index to Quick as a knowledge base while Q Business keeps running. Limits: at most two Q Business indexes per Region; Q Apps, Actions and chat guardrails do not carry over; and without IAM Identity Center all Quick users can reach connected indexes, so per-user access must be rebuilt with separate knowledge bases. Then connect sources natively in Quick (MCP integrations where no connector exists), migrate identities, recreate document ACLs, rebuild Q Apps as Quick Flows, compare answer quality side by side, and cut over.
Production consideration: Quick will not ingest S3 documents that lack an ACL entry when ACLs are enabled, while Q Business allowed access to prefixes missing from the ACL file. Check this before migrating, or content will disappear without warning.
25. How does Amazon Personalize work?
Answer: Create a dataset group and import item interactions (the core signal), optional items and users metadata, and for action recommendations, actions and action interactions, in bulk from S3 and in real time through an event tracker. Then use domain recommenders for retail or media ("recommended for you", "frequently bought together") or train custom solutions with a recipe (user personalisation, similar items, personalized ranking, trending, user segmentation). Serve through a recommender or campaign, or run batch recommendations and segments. Filters apply business rules such as excluding purchased or out-of-stock items.
Interview tip: Cold start is the classic follow-up. Item metadata, popularity-based recipes and real-time events help new users and new items. Our recommendation systems interview questions cover the modelling side.
26. Would you use Personalize or build recommendations with embeddings and an LLM?
Answer: Personalize when you have steady interaction data and want a managed recommender with real-time updates, filters and batch jobs without owning the model. Embeddings plus a vector index when content similarity matters more than behaviour, as with new catalogues or long-tail items with rich descriptions. An LLM is rarely the recommender itself; it helps explain recommendations, write personalised copy, or turn a conversational request into a structured query. Mature systems combine candidates from both, apply business rules, then rank, and prove value with an A/B test against the current baseline.
27. Amazon Forecast is closed to new customers. How do you build demand forecasting on AWS now?
Answer: AWS closed Forecast to new customers in July 2024 and points to SageMaker Canvas, which builds time-series forecasting models through a no-code interface or AutoML APIs, with a model leaderboard. For more control, train forecasting models in SageMaker AI, for example with AutoGluon's time-series module, and run batch predictions on a schedule. Foundation models do not replace a numeric forecasting model, though they help explain forecasts. The fundamentals stay: clean history, related series such as price and promotions, festival calendars, backtesting and a simple baseline to beat.
See our time series forecasting interview questions for the modelling questions.
28. What replaces Amazon Fraud Detector for a new customer?
Answer: Fraud Detector stopped accepting new customers on 7 November 2025; existing customers continue. AWS recommends AutoGluon, an open-source AutoML library, for training fraud models on tabular event data, with deployment on SageMaker AI, and provides a reference architecture for real-time payment processing. For account-creation and login abuse, AWS points to AWS WAF Fraud Control, which uses managed rules (not ML models) against credential stuffing and fake sign-ups. Migrating teams should export stored event data first, then rebuild the model, rules and decision thresholds, and run both in shadow mode before switching.
Interview tip: Fraud is an imbalanced-class problem. Mention precision-recall curves, cost-weighted thresholds and a manual review band, not accuracy.
GenAI building blocks and retired services
29. What is Amazon Nova, and which models are in the family?
Answer: Amazon Nova is Amazon's own foundation model family on Bedrock. The first generation has understanding models (Nova Micro, text-only, then Nova Lite, Pro and Premier, which accept text, images, video and documents), creative models (Nova Canvas for images, Nova Reel for video) and Nova Sonic for speech-to-speech. Nova 2 adds Nova 2 Lite with extended thinking and built-in web grounding and code interpreter tools, Nova 2 Sonic with more languages, and Nova Multimodal Embeddings for cross-modal search across text, documents, images, video and audio. Nova 2 Lite supports fine-tuning, and Amazon Nova Forge is a service for building your own models on Nova.
30. How do you decide between a Nova model and a third-party model on Bedrock?
Answer: The same way you choose any model: build a small evaluation set from real tasks, then compare candidates on quality, latency, cost per successful task and operational fit. Nova's strengths are tight AWS integration, low-cost tiers for high-volume simple tasks, and fine-tuning support; third-party models on Bedrock may be stronger for particular reasoning, coding or language tasks. Region availability and cross-Region inference matter for data residency. Many systems route by task: a small model for classification and extraction, a stronger one for complex reasoning. Because Bedrock's Converse API is consistent across models, switching is mostly configuration plus re-evaluation.
Model selection, inference profiles and cost levers are covered in depth in our Bedrock interview guide.
31. Which vector store options does AWS offer, and how do you choose?
Answer: Bedrock Managed Knowledge Base: Bedrock runs the vector store and hybrid search for you; choose it when you want managed RAG and do not need direct index access. Amazon OpenSearch Service (managed clusters) and OpenSearch Serverless vector collections: choose them for high query rates, low latency, hybrid keyword-plus-vector search, aggregations, faceting and rich filtering. Amazon S3 Vectors: vector buckets and indexes inside S3 for large, cost-sensitive collections that are queried less often. Aurora PostgreSQL with pgvector: when vectors belong next to relational data and transactions. Choose on query rate, latency target, filter complexity, scale, operational effort and cost of idle capacity. Our explainer on vector databases covers the indexing basics.
| Need | Typical choice |
|---|---|
| Managed RAG, minimal ops | Bedrock Managed Knowledge Base |
| High QPS, hybrid search, facets | OpenSearch Service or Serverless |
| Very large corpus, infrequent queries | S3 Vectors |
| Vectors joined with business tables | Aurora PostgreSQL + pgvector |
32. What is Amazon S3 Vectors, and when would you not use it?
Answer: S3 Vectors adds a vector bucket type to S3, holding vector indexes of vectors with filterable metadata, queried through a dedicated API with no infrastructure to manage. AWS describes sub-second latency for infrequent queries, strongly consistent writes, a separate s3vectors IAM namespace and Block Public Access always on. It works as a Bedrock Knowledge Base vector store and integrates with OpenSearch, including exporting a snapshot to OpenSearch Serverless for higher query rates. Do not make it the only store for a high-traffic, latency-critical assistant, or where you need hybrid search, aggregations or faceting; use it for the large, cold corpus.
33. Where do AgentCore and Bedrock Agents Classic fit in an AWS AI services design?
Answer: AI services and models are capabilities; an agent is the orchestration that decides which capability to call next. On AWS, new agent work goes to Amazon Bedrock AgentCore, a set of services for running agents built with any framework: a runtime, a gateway that exposes APIs and Lambda functions as tools (including over MCP), identity, memory, built-in tools, observability and policy. Bedrock Agents, now called Agents Classic, is closed to new customers from 30 July 2026, with existing users continuing. In a breadth design, Textract, Comprehend or Transcribe usually become tools behind the AgentCore gateway, wrapped in Lambda functions with narrow IAM roles.
For the component-level questions, see our AgentCore interview questions.
34. How does SageMaker AI fit alongside the AI services, and which SageMaker features changed status?
Answer: SageMaker AI is where you go when no managed service or foundation model meets the need: a custom model trained on your data, a specific open-weight model you must host, or full control over training and inference. It is also the recommended path for retired services (Lookout for Vision, Forecast via Canvas, Fraud Detector via AutoGluon). In the June 2026 lifecycle announcement, AWS listed several SageMaker AI features as moving to maintenance, including Augmented AI, Clarify, Debugger, geospatial capabilities, Ground Truth, Model Monitor, Role Manager and Studio Lab, and Ground Truth Plus reached end of support. Core training, hosting, Pipelines and the other main features continue.
The SageMaker side is covered in depth in our SageMaker AI interview guide.
35. Amazon A2I is in maintenance. How do you design human review for AI services output now?
Answer: A2I gave managed human review loops for Textract, Rekognition and custom models. For a new account, build the loop with standard services: send low-confidence or rule-failing items to a review queue, let Step Functions wait with a task token, show reviewers the page image with highlighted fields (Textract geometry or BDA visual grounding), and resume the workflow on their decision. Record who changed what, and feed corrections into your evaluation set.
Production consideration: Set the review threshold per field, not per document. A low-confidence policy number needs review; a low-confidence "remarks" field may not.
36. Scenario: a manufacturing team still runs Lookout for Equipment and Lookout for Vision models. What do you tell them?
Answer: Lookout for Vision reached end of support on 31 October 2025, so anything still depending on it has stopped and must be rebuilt; AWS points to SageMaker AI (including JumpStart models) or Bedrock, reusing exported datasets and manifests. Lookout for Equipment reaches end of support on 7 October 2026, after which its console and resources are inaccessible. AWS recommends AWS IoT SiteWise, which has native multivariate anomaly detection supporting Lookout for Equipment-style modelling, and provides migration scripts.
What I would check:
- Which production lines and alerts still depend on each model, and who acts on the alerts.
- Whether training data, labels and manifests were exported, since resources disappear after shutdown.
- The historical alert precision, to set a bar the replacement must meet.
Production consideration: Run the replacement in shadow mode against live sensor or camera data and compare alerts before switching off the old path; an anomaly detector that floods operators with false alarms is switched off by the plant within weeks.
Security and compliance
37. How do you protect data in transit and at rest across an AI services pipeline?
Answer: In transit, AI service APIs use TLS; keep traffic private with VPC interface endpoints (AWS PrivateLink) where a service supports them, and S3 gateway endpoints. At rest, most data sits in your own S3 buckets: inputs, async job output, transcripts and redacted copies. Encrypt them with SSE-KMS customer managed keys, enforce encryption and TLS in bucket policies, pass a KMS key to services that accept one for output, limit decrypt rights in key policies, and log key use in CloudTrail. The biggest exposure is usually your own output bucket, not the AI service.
38. How do you design IAM for AWS AI services?
Answer: Two kinds of identity. Callers (a Lambda function or container) get a role with only the actions they need, such as textract:StartDocumentAnalysis and textract:GetDocumentAnalysis, not textract:*. Service roles: async services that read and write your S3 data (Comprehend jobs, Personalize imports, Textract publishing to SNS) assume a data-access role, so the caller needs iam:PassRole limited to that role. Add aws:SourceAccount conditions on trust policies against confused-deputy problems, separate roles per environment, and SCPs to block unapproved AI services or Regions.
Real-world example: Consider a GCC platform team that lets product teams use Textract and Comprehend but not Rekognition face APIs. An SCP denies rekognition:IndexFaces and rekognition:SearchFaces* everywhere except one approved account with a documented biometric policy.
39. Do AWS AI services use my content to improve their models, and how do you control that?
Answer: AWS documents that some AI services may use and store customer content for service improvement, including model training, and may store it in a different Region for that purpose. You opt out with an AI services opt-out policy in AWS Organizations, either for specific services or for all supported services, applied across accounts. Opting out deletes the historical content stored for improvement, though not content needed to provide the service. Check the current list of covered services in the Organizations documentation. Bedrock has its own data-handling terms, documented separately: it does not use prompts and completions to train models.
40. How do you handle compliance requirements such as HIPAA, the DPDP Act and data residency?
Answer: Start with a data inventory: which personal or health data enters which service, in which Region, and where outputs land. For HIPAA, use only HIPAA-eligible services under a Business Associate Addendum; Rekognition, Comprehend Medical, Transcribe Medical and HealthScribe are documented as eligible, but check the current list. For India's DPDP Act, focus on purpose limitation, consent, retention, deletion and processor obligations; see our article on the DPDP Act for AI applications. For residency, confirm each feature exists in Asia Pacific (Mumbai) or (Hyderabad), and watch Bedrock cross-Region inference, which can move data across Regions.
41. How do you handle PII across a pipeline that uses several AI services?
Answer: Redact as early as possible and keep originals in a tightly controlled store. Enable Transcribe PII redaction for audio-derived text. Run Comprehend PII detection or redaction before text reaches search indexes or models. Add Bedrock Guardrails sensitive-information filters at the model boundary as a second layer (see AI guardrails). Log metadata, not content. Detection is probabilistic, and Indian identifiers such as Aadhaar or PAN numbers may need custom regex or custom entity recognition, so test with realistic synthetic data.
42. What governance would you put around face recognition and content moderation?
Answer: For faces: a documented purpose and lawful basis, consent where required, access control on collections, automated deletion of face IDs after a retention period, per-use-case thresholds with human review, and accuracy checks across demographic groups on your own data. Never base a decision about a person on face search alone. For moderation: map Rekognition's label taxonomy to your policy, choose actions per label and confidence band (remove, blur, review), measure false positives on real uploads, and offer an appeal path. Both need an owner outside engineering.
43. How do you audit and monitor AI service usage?
Answer: CloudTrail records management and API calls for the AI services, with data events where supported, so you can answer "who called IndexFaces on which collection". CloudWatch metrics show calls, errors, throttles and latency per service; alarm on throttling and error rate. Log your own pipeline events with a correlation ID per document or call, including service, feature, confidence summary, cost estimate and outcome, without logging raw content. Use AWS Config and Security Hub for configuration drift such as unencrypted output buckets, and Cost Explorer with cost allocation tags for spend by team and use case.
Cost, architecture and operations
44. How are AWS AI services priced, and what are the common cost traps?
Answer: Mostly per unit: Textract per page and feature type, Rekognition per image or video minute, Transcribe per second of audio, Polly and Translate per character, Comprehend per unit of text. Some components bill while idle: Comprehend custom endpoints, Kendra indexes, Personalize campaigns with minimum provisioned throughput, and running Rekognition Custom Labels models. Common traps: requesting every Textract feature on every page, reprocessing documents on retries, leaving custom models running after a demo, and re-synthesising the same Polly prompt. Check each pricing page for current rates.
Interview tip: Express cost per business transaction ("cost per processed claim"), not per API call. That is how a finance team reads it. Our guide to cloud cost optimisation for AI covers the wider levers.
45. Scenario: finance asks whether replacing Textract with a multimodal LLM would cut document processing cost. How do you answer?
Answer: With a measured comparison, not an opinion. Cost per page for Textract depends on feature types; cost per page for a model depends on image tokens, prompt size and output length, plus retries for invalid JSON. Accuracy and rework cost often matter more than the API bill: a cheaper option that sends more documents to manual review costs more overall.
What I would check:
- A labelled sample of a few hundred real documents across layouts and scan quality.
- Field-level accuracy and the share routed to human review for each option.
- Full cost per document: API, retries, review time, and engineering to maintain prompts or blueprints.
- Whether a tiered design (Textract first, model only for failures) beats both.
46. How would you use Lambda, Step Functions and EventBridge in an AI services pipeline?
Answer: EventBridge or S3 notifications react to new objects. Step Functions orchestrates the flow: start an async Textract or Transcribe job, wait for the callback with a task token, run Comprehend, call a Bedrock model, validate, and branch to auto-approve or review, with retries and catches per step; its direct service integrations remove many glue Lambdas. Lambda does small stateless pieces such as normalising output and applying rules. Keep steps idempotent and pass S3 references, not payloads, in the state.
S3 -> EventBridge -> Step Functions 1 classify (BDA or rules) 2 extract (Textract async, wait token) 3 PII redact (Comprehend) 4 reason/summarise (Bedrock model) 5 validate rules -> ok? write : review
47. How do you prepare data for AI services such as Personalize, Comprehend custom models or Custom Labels?
Answer: Each service expects a specific format, so the work is mostly data engineering. Personalize interactions need user ID, item ID, timestamp and event type matching a schema, with duplicates and bot traffic removed; AWS Glue jobs or SageMaker Data Wrangler can transform raw clickstream from S3. Comprehend custom classification needs consistent labelled text; Custom Labels needs labelled images. Always keep a fixed train and test split, a data-quality report, versioned datasets in S3 and a record of which version trained which model. Our data engineering interview questions go further.
48. How would you build real-time analytics on audio or video streams with AWS AI services?
Answer: For audio, stream from Amazon Connect, a WebRTC client or a telephony bridge into Transcribe streaming or Call Analytics streaming, and publish transcript segments to Kinesis Data Streams or EventBridge for agent-assist, sentiment alerts or compliance checks. For video, Kinesis Video Streams carries the stream and a Rekognition stream processor runs face search or label detection, writing results to a data stream or SNS. Store events in S3 for Athena. Consumers must tolerate partial and out-of-order results, and streams should be stopped when idle because they bill while running.
49. How do you handle quotas, throttling and scaling with AI services?
Answer: Each service has quotas on transactions per second per API, concurrent async jobs, and payload or file sizes; some are adjustable through Service Quotas, others (such as file size and supported languages) are fixed. Design for them: buffer bursts with SQS, control concurrency in the consumer (Lambda reserved concurrency or a Step Functions Map with a concurrency limit), use exponential backoff with jitter on ThrottlingException, prefer async batch APIs for bulk work, and request increases well before a launch with a load test as evidence. Watch the CloudWatch throttle metrics and alarm on them, since throttling under load is a predictable, avoidable outage.
50. How do you evaluate AI service output before and after go-live?
Answer: Build a labelled golden set from real, anonymised inputs, including skewed scans, accents, noisy calls and mixed languages. Use task metrics: field-level precision and recall for extraction, word error rate for transcription, per-entity precision and recall for PII, intent accuracy and containment for bots, online conversion for recommendations. Set confidence thresholds from the precision-recall trade-off and the cost of each error. After go-live, review a weekly sample, watch for drift such as new layouts, and re-run the golden set after every change. For model-based steps, see our LLM evaluation interview questions.
Real-world scenarios
51. Scenario: an insurer wants to automate motor claims intake, covering claim forms, repair estimates, photos and FIR copies. Design it.
Answer: A document and image pipeline with human review where confidence or rules fail. Uploads land in an encrypted S3 bucket; Step Functions classifies each file, extracts fields from forms and estimates (Textract FORMS, TABLES and QUERIES, or a BDA blueprint per type), checks photos with Rekognition and a multimodal model for a damage description, redacts PII with Comprehend, and asks a Bedrock model to summarise the claim and flag inconsistencies such as an estimate dated before the accident. Business rules decide auto-route or review; reviewers see fields highlighted on the page image.
What I would check:
- Document types and volumes, and how many layouts each type has.
- Which fields drive decisions and the cost of getting each wrong.
- Integration with the claims system: API, queue or file drop, and its validation rules.
Production consideration: Keep the model's summary advisory, with every extracted value traceable to its source page. Our enterprise document intelligence project walks through a similar build.
52. Scenario: a bank wants to modernise its contact centre with AI. Which AWS services do you use and how?
Answer: Amazon Connect as the platform. A Lex V2 bot handles authentication and common intents such as balance, card block and statements, with Lambda fulfilment calling core banking APIs through narrow, audited interfaces. Connect AI agents (previously Amazon Q in Connect) assist human agents with answers from approved knowledge. Contact Lens provides transcripts, sentiment, categories, redaction and post-contact summaries for quality assurance. Transfers carry context so customers do not repeat themselves.
Caller -> Amazon Connect -> Lex bot (auth, intents) -> Lambda -> core APIs -> human agent + Connect AI agent assist -> Contact Lens: transcript, redaction, QA
What I would check:
- The top call reasons by volume and which are safe to automate.
- Identity verification design: OTP, device or knowledge checks handled in the flow, never passed through a model prompt.
- Containment, transfer rate and repeat-call metrics as success measures.
Production consideration: Start with one or two intents, measure, then expand. See our call centre QA AI project for the analytics side.
53. Scenario: a GCC in Bengaluru wants an internal search assistant over SharePoint, Confluence and policy PDFs, starting in a new AWS account in 2026. What do you build?
Answer: Not Kendra or Q Business: both are closed to new customers. Options: Amazon Quick if a ready-made employee assistant with connectors and access control is acceptable; Bedrock Managed Knowledge Base behind your own app if you need a custom UI or control over prompts and citations (it has SharePoint, Confluence, S3, Google Drive, OneDrive, web crawler and custom connectors, hybrid search, and Retrieve or RetrieveAndGenerate); or a customer-managed OpenSearch stack if you need faceting, synonyms or autocomplete that the managed option lacks.
What I would check:
- Permission model: users must only see documents they can open in the source system.
- Search features users depend on today: facets, filters, synonyms.
- A golden question set with expected sources from each department.
Production consideration: Retrieval quality decides user trust; our search relevance interview questions cover how to measure it.
54. Scenario: an existing Kendra customer asks you to plan the migration to Bedrock Managed Knowledge Base. Walk through it.
Answer: Kendra is in maintenance, not sunset, so there is no forced deadline, but new features will only come to the Bedrock side. Inventory indexes, connectors, metadata, ACLs and features in use. AWS's guide maps the APIs: create_knowledge_base and create_data_source replace index and data source creation, start_ingestion_job replaces sync, retrieve and retrieve_and_generate replace query and retrieve, and attribute filters translate to the new filter syntax (no substring or starts-with). Sources without native connectors are exported to S3 on a schedule with metadata sidecar files.
What I would check:
- The feature-gap list and a workaround for each one in use: synonym expansion before the query, OpenSearch for autocomplete, metadata filters for facets, a pre-ingestion pipeline for enrichment.
- Access control equivalence for every source.
- Relevance parity: run both side by side on a golden query set (NDCG or MRR, plus latency percentiles) before switching.
Production consideration: Cut over per application or department, keep Kendra as a fallback until parity is shown, then decommission indexes, since they bill while provisioned.
55. Scenario: a retailer used Comprehend topic modeling for weekly customer feedback themes and wants to set it up in a new account. What do you do?
Answer: Topic modeling is not available to new customers (accounts that used it in the past 12 months keep access). Follow AWS's migration guidance: define the topic list in a system prompt, ask a Bedrock model to return JSON with topics and a primary topic per piece of feedback, and run it as a Bedrock batch inference job over a JSONL file in S3 for the weekly load. Keep Comprehend for what still fits, such as sentiment and PII redaction, run before the model step.
What I would check:
- Whether the business wants fixed categories (classification) or discovered themes (clustering), which need different prompts or an embeddings-plus-clustering step.
- A labelled sample to measure agreement with human tagging.
- Prompt stability across runs, so week-on-week trends reflect customers, not prompt changes.
56. Scenario: a fintech built its new onboarding risk checks assuming Fraud Detector, then found it cannot sign up. How do you recover?
Answer: Split the problem. For sign-up abuse (bots, credential stuffing, fake accounts) put AWS WAF Fraud Control managed rules in front of the onboarding endpoints. For transaction or application risk scoring, train a model with AutoGluon on historical labelled events, deploy it on a SageMaker AI endpoint or serverless inference, and keep a rules layer for known patterns and regulatory checks. Add a review band between auto-approve and auto-decline, and log features and scores for audit.
What I would check:
- Label quality and delay: chargebacks and confirmed fraud arrive weeks later.
- Class imbalance and the business cost of false positives versus missed fraud.
- Explainability needs for compliance and customer complaints.
57. Scenario: a hospital group in Hyderabad wants AI-generated clinical notes from consultations. What do you advise?
Answer: HealthScribe is built for this, but it is documented as US English and in US East (N. Virginia), which raises residency questions for Indian patient data and rules it out for Telugu, Hindi or code-mixed consultations. The alternative is a custom pipeline in Indian Regions: speech recognition for supported languages, then a Bedrock model drafting notes with references to transcript turns, evaluated with clinicians. Either way, notes are drafts a doctor reviews and signs, recording consent is explicit, and audio retention is short.
What I would check:
- Languages actually spoken in consultations, measured from a sample.
- Legal view on cross-border transfer of health data under the DPDP Act and sector rules.
- Clinician review time per note, the real success metric.
Our article on AI in Indian healthcare covers the wider context.
58. Scenario: an e-commerce marketplace wants to moderate seller images, improve product search and personalise the home page. Which services do you combine?
Answer: Rekognition content moderation on every upload, with labels mapped to marketplace policy and borderline cases queued for review; Custom Labels or a multimodal model for marketplace-specific checks such as watermarks or wrong-category images. For search, OpenSearch with hybrid keyword and vector search over product text and image embeddings (Nova Multimodal Embeddings lets a photo query match product images), with S3 Vectors for the long-tail catalogue if cost demands it. For the home page, Personalize recommenders for "recommended for you" and "frequently bought together", with filters for stock and seller rules, and Personalize re-ranking of search results.
59. Scenario: a media company must check thousands of hours of video for unsafe content and make it searchable by scene. Rekognition Video or Bedrock Data Automation?
Answer: Use each where it is strongest. Rekognition Video gives timestamped moderation labels with a stable taxonomy, label and celebrity detection, on-screen text and segment detection: predictable output for compliance reports. BDA's video output adds scene summaries, on-screen text, ad and brand classification and moderation signals, which suits semantic search. Index scene summaries and Transcribe transcripts in OpenSearch as text and embeddings.
What I would check:
- The compliance standard the moderation must map to, and who signs off.
- Cost per hour of video for each option at full archive volume, and whether a sample-first approach (sampled frames, then full analysis only on flagged segments) is acceptable.
- Search quality on real editor queries.
60. Scenario: extraction accuracy on supplier invoices fell sharply after several suppliers changed their invoice templates. How do you respond?
Answer: Contain, diagnose, fix, prevent. Contain by routing affected suppliers' invoices to review. Diagnose by grouping failures by supplier and field and comparing confidence before and after; a layout change usually breaks rules that assumed positions, or Queries phrasing that no longer matches labels. Fix by updating queries or adapters, adding a BDA blueprint or model fallback for new layouts, and adding the samples to the golden set.
What I would check:
- Whether failures cluster by supplier, field or scan source.
- Whether validation rules (totals equal line sum, tax number format) caught the errors or let them through.
- How long the problem ran before anyone noticed, which tells you whether monitoring is adequate.
Production consideration: Add drift monitoring: per-supplier confidence and review-rate dashboards with alerts, so the next template change is caught on day one, not at month-end close.
If you want guided practice building pipelines like these, from Textract and Transcribe through Bedrock, security and cost control, Cloudsoft's APEX AI, ML, Cloud and Cyber Security program covers the engineering foundations behind production AI systems, in Ameerpet or live online.
Key takeaways
- AWS AI has three layers: task-specific AI services, Bedrock for foundation models, and SageMaker AI for building your own. Production systems combine them.
- Choose a managed AI service for narrow, high-volume tasks with fixed output, and a foundation model for reasoning and flexible output. Lifecycle risk is now part of the decision.
- Know the current status: Kendra, Q Business, A2I and Bedrock Agents Classic are closed to new customers; Forecast, Fraud Detector and Monitron are closed too; Lookout for Vision and Metrics have ended and Lookout for Equipment ends on 7 October 2026; Q Developer moves to Kiro.
- Bedrock Data Automation, Nova models, Bedrock Managed Knowledge Base, OpenSearch and S3 Vectors are the newer building blocks that many retired capabilities now point to.
- Security lives mostly in your own buckets and roles: KMS on outputs, narrow service roles with PassRole, private endpoints where supported, and an organisation-wide AI services opt-out policy.
- Express cost per business transaction and test options against a labelled golden set before choosing.
- For a migration, the plan is the feature-gap list, the access-control mapping and a side-by-side quality comparison, not just the API mapping.
Interview preparation checklist
- Draw the three-layer AWS AI stack from memory and place each service correctly.
- Learn the status table in Q3 and the recommended alternative for each retired or closed service.
- Build one async Textract pipeline with SNS, SQS, Step Functions and a review queue, and measure field accuracy on twenty real documents.
- Transcribe a few real phone recordings with and without a custom vocabulary and compute word error rate.
- Build a small RAG search over a document set with a Bedrock knowledge base, and compare it with an OpenSearch or S3 Vectors setup.
- Write least-privilege IAM policies for a Lambda that calls Textract and Comprehend, including the PassRole statement.
- Prepare a cost comparison method (per page or per call) with clearly labelled illustrative inputs.
- Rehearse one migration story aloud: Kendra to Bedrock Managed Knowledge Base, Q Business to Quick, or Fraud Detector to AutoGluon on SageMaker AI.
- Check the language and Region support of the services you mention for an Indian deployment.
- Read the deep guides for Bedrock, AgentCore and SageMaker AI so you can go one level deeper when asked.
FAQ
What skills are needed for AWS AI services roles?
You need Python, AWS fundamentals (IAM, S3, Lambda, Step Functions, KMS, networking), working knowledge of the main AI services and Bedrock, and evaluation skills such as building labelled test sets. Integration with business systems, cost estimation and security design matter as much as knowing individual APIs.
How should I prepare for an AWS AI interview in 2026?
Build one or two small end-to-end pipelines, for example document extraction with review and a knowledge-base search assistant, and measure their accuracy and cost. Learn which services are closed to new customers and their alternatives, and practise explaining service choices aloud.
Is Amazon Kendra still available?
Kendra is in maintenance mode. Existing customers continue to receive support, bug fixes and security updates, but it is no longer open to new customers, and AWS recommends Amazon Bedrock Managed Knowledge Base for new search and RAG applications.
What happened to Amazon Q?
Amazon Q Business is closed to new customers, and AWS recommends Amazon Quick instead. The Amazon Q Developer CLI was rebranded as Kiro, and the Q Developer IDE plugins reach end of support on 30 April 2027. The Amazon Q in Connect functionality is now part of Amazon Connect AI agents.
Do I need machine learning knowledge to use AWS AI services?
Not to call them, because the services are pre-trained APIs. You do need basic ML literacy to work well with them: confidence scores, precision and recall, thresholds, data quality and evaluation. Those topics come up in almost every interview round.
Should I learn the AI services or Amazon Bedrock first?
Learn the AWS fundamentals first, then Bedrock, because most new AI work involves foundation models. Then learn the main AI services, Textract, Transcribe, Comprehend and Rekognition, since real pipelines combine them with models for extraction, speech and PII handling.
Which AWS certification helps for AI roles?
AWS offers AI-related certifications such as AWS Certified AI Practitioner and AWS Certified Machine Learning Engineer โ Associate; check AWS's certification site for the current list. A certification helps structure your learning, but interviewers weigh a working project and clear trade-off reasoning more heavily.
Are AWS AI services a good career focus in India?
They are a useful focus for cloud and AI engineers, because many Indian services firms and GCCs run AWS workloads and need people who can combine AI services, Bedrock and core AWS securely. Pair the AWS skills with general AI engineering so they transfer to other clouds too.
AWS AI work rewards engineers who can choose the right service, prove it works with real data, and keep it secure and affordable. To build that skill set across AI, ML, cloud and security, explore Cloudsoft's APEX program. If you want hands-on depth on the generative side, the Amazon Bedrock and GenAI course covers building with foundation models on AWS. Both run in Ameerpet beside the metro or live online; call +91 96660 19191 for a free demo.
