What Is the AWS Certified Generative AI Developer Credential?

AIP-C01 certifies that you can design, build, secure and operate generative AI applications on AWS. It spans five domains covering foundation model integration and data management, implementation and integration, AI safety and governance, operational efficiency, and testing and troubleshooting. The exam runs 75 questions in 180 minutes and costs $300.

The professional tier matters. AWS runs an AI Practitioner exam for people who need vocabulary and a Machine Learning Engineer associate exam for people who ship models. This one sits above both and assumes you have already put a generative AI application in front of real users, with all the awkwardness that implies: cost that surprised you, a model that hallucinated in production, a retrieval step that returned the wrong document.

The blueprint reflects that. Objectives are written as design decisions with constraints attached rather than as capabilities. You are asked to create resilient architectures for continuous operation during service disruptions, not to describe what a foundation model is. The full objective list is on the AIP-C01 exam page if you want to gauge the depth before committing.

How Are the 75 Questions Split Across the Five Domains?

AIP-C01 publishes weightings, and they are heavily front-loaded. Foundation model integration, data management and compliance carries 31 percent on its own. Implementation and integration adds 26 percent. Together those two account for 57 percent of the paper, roughly 43 questions out of 75.

The shape gives you a clear planning instruction. The bottom two domains together are 23 percent, about 17 questions, and they are the ones candidates most enjoy revising because they are concrete. The top domain alone is worth more than both, and it is the one that requires architectural judgement rather than recall.

A scaled score of 750 out of 1000 is not 75 percent of questions correct. Scaling adjusts for form difficulty, so read it as needing solid performance in the two heavy domains and no collapse in the others.

What Does the 31 Percent Foundation Model Domain Actually Ask?

The largest domain covers six objective groups: analysing requirements and designing solutions, selecting and configuring foundation models, building data validation and processing pipelines, designing vector store solutions, designing retrieval mechanisms, and implementing prompt engineering strategies with governance. Nearly a quarter of the exam sits in retrieval and vector work alone.

Model selection is a design decision, not a preference

You are asked to assess and choose foundation models using performance benchmarks, capability analysis and limitation evaluation, and then to build architecture that allows the model to be swapped without code changes. The named pattern uses Lambda, API Gateway and AppConfig to keep model choice configurable. Resilience is examined too, including Step Functions circuit breaker patterns and Bedrock Cross-Region Inference for models with limited regional availability.

Retrieval is the heart of it

Two full objective groups cover vector stores and retrieval. Document segmentation, embedding model selection by dimensionality and domain fit, vector search through OpenSearch or Aurora with pgvector or Bedrock Knowledge Bases, hybrid search combining keywords and vectors, rerankers, query expansion and query decomposition all appear by name. So does the Model Context Protocol as a way of exposing vector queries to a model.

The practical test is whether you know why each knob exists. Chunk size affects retrieval precision. Embedding dimensionality affects both accuracy and storage cost. Reranking fixes a specific failure where the right document is retrieved but ranked fourth. Candidates who have built a retrieval system have opinions about all three; candidates who have read about one do not.

Prompt governance, not prompt tricks

The prompt engineering objective is written as governance. Bedrock Prompt Management for parameterised templates and approval workflows, Bedrock Guardrails for responsible AI enforcement, S3 for template repositories, CloudTrail for usage tracking and CloudWatch Logs for access logging. The exam cares about who changed the system prompt and when, far more than about clever phrasing.

Which AWS Services Do You Need Hands-On With?

The AIP-C01 blueprint names roughly thirty services. Not all carry equal weight, but a core set appears repeatedly across domains and effectively defines the hands-on requirement. Reading the documentation is not enough for these, because the questions are about behaviour under constraint rather than about capability.

The services that appear in more than one domain

• Amazon Bedrock, including Guardrails, Knowledge Bases, Prompt Management, Cross-Region Inference and reranker models. The Bedrock user guide is the single most useful reading on the list.
• Amazon SageMaker AI, specifically the Model Registry for versioning, Processing for data preparation, and endpoint deployment for fine-tuned models.
• Amazon OpenSearch Service with vector search and the Neural plugin, plus sharding and multi-index strategies for scale.
• AWS Lambda and Step Functions, which carry most of the orchestration and resilience objectives.
• Amazon CloudWatch and AWS CloudTrail, which carry observability and audit across three domains.

The supporting cast

Aurora with pgvector, RDS, DynamoDB and S3 all appear as vector or metadata stores. Glue Data Quality and SageMaker Data Wrangler carry data validation. Comprehend handles entity extraction and intent recognition, Transcribe handles audio. API Gateway and AppConfig carry the swap-the-model pattern. The AWS Well-Architected Framework and its Generative AI Lens are named as the standard for reviewing a design.

Fine-tuning gets its own objective, including parameter-efficient techniques such as low-rank adaptation, plus lifecycle management: automated deployment pipelines, rollback strategies for failed deployments, and retiring models that have been superseded.

What Does AI Safety, Security and Governance Cover?

The third domain carries 20 percent, roughly 15 questions, and it is the one that most clearly separates a professional exam from an associate one. It covers guardrails and content filtering, data protection across the generative AI pipeline, access control, auditability, and the compliance obligations that attach to a system making decisions about people.

Where safety differs from ordinary cloud security

Standard cloud security questions are about who can reach a resource. Generative AI security questions add two problems on top: the model can be persuaded to do something it should not, and the data it retrieves may not belong to the person asking. Guardrails address the first. Metadata filtering and access-aware retrieval address the second, and they are architecture decisions taken at the vector store rather than at the model.

Compliance appears inside the largest domain rather than this one, which is a deliberate signal: AWS treats compliance as a data-management concern, decided when you choose where documents live and how they are tagged, not as a policy layer bolted on afterwards.

Cost and observability

Operational efficiency carries 12 percent and testing and troubleshooting 11 percent. Together they cover token cost management, latency, caching, evaluation metrics and the awkward business of debugging a system whose output is non-deterministic. Traditional testing assumes a fixed expected result. Generative AI testing does not, which is why evaluation frameworks rather than assertions dominate this material.

What Does AIP-C01 Cost and What Experience Does AWS Expect?

AIP-C01 costs $300 USD and runs 75 questions in 180 minutes, either at a Pearson VUE test centre or as an online proctored session. The pass mark is 750 on a 100 to 1000 scale, and the exam is offered in English, Japanese, Korean and Simplified Chinese.

The experience bar is the real gate

AWS recommends two or more years building production-grade applications on AWS or with open-source technologies, general AI or machine learning or data engineering experience, and one year of hands-on generative AI implementation. That third clause is the one to take literally. Everything in the two heaviest domains assumes you have made retrieval and model-selection decisions and lived with the consequences.

There is no mandatory prerequisite certification, so nothing stops you booking without the AI Practitioner or Machine Learning Engineer credentials. The recommendation is about capability rather than paperwork, and it is published openly on the AWS certification page.

Practitioner, Machine Learning Engineer or Generative AI Developer?

AWS now runs three AI-facing credentials and they describe three different jobs. AI Practitioner is a foundational credential for people who need to speak the language. Machine Learning Engineer Associate is for people who build and operate models. AIP-C01 is for people who build applications on top of foundation models they did not train.

The distinction that matters

The Machine Learning Engineer route assumes you own the model: training data, features, evaluation, deployment. The Generative AI Developer route assumes the model is somebody else’s and your job is everything around it, which is retrieval, prompting, guardrails, cost, latency and failure handling. Those are genuinely different skill sets, and holding one does not shorten the other by much.

Independent data supports taking the generative route seriously. The Stack Overflow developer survey records how quickly AI tooling has become routine in professional development work rather than experimental, which is the shift that created this credential in the first place.

If you are mapping out the whole route rather than a single exam, the AWS certification path lays out how the associate, professional and specialty tiers connect.

How Should You Build the Hands-On Year AWS Expects?

If you do not yet have the year of generative AI implementation AWS recommends, the efficient way to build it is to ship one complete retrieval application and then break it deliberately. That single project touches four of the five domains, and the debugging is where the exam-relevant judgement actually forms.

1. Stand up a Bedrock application that answers questions over a document set you own, using a Knowledge Base rather than hand-rolled retrieval.
2. Replace the managed retrieval with your own pipeline through OpenSearch or Aurora with pgvector, so you have made the chunking and embedding decisions yourself.
3. Change the chunk size twice and measure what happens to answer quality, because that trade-off is examined directly.
4. Add a reranker and confirm you can explain which failure it fixed, rather than adding it because the blueprint mentions it.
5. Put Guardrails in front of the application and try to get past them, since the safety domain is written from the attacker’s side as well as the builder’s.
6. Instrument token cost and latency through CloudWatch until you can predict the monthly bill, which is what the operational efficiency domain is really testing.
7. Swap the underlying model without changing application code, using the configuration pattern the blueprint names, and confirm nothing else breaks.

AWS publishes an official workshop repository that covers much of this ground, and the Bedrock workshop is a faster starting point than assembling the pieces yourself. Candidates who already work in this space commonly report six to ten weeks of part-time study on top of existing experience. Those coming from a general AWS background without generative AI work should expect considerably longer, and the honest read on which AWS certificates hire is worth weighing before spending that time.

Frequently Asked Questions

How many questions are on the AIP-C01 exam?

The AWS Certified Generative AI Developer – Professional exam has 75 questions, either multiple choice or multiple response, with a 180-minute limit. That is roughly two minutes and twenty seconds per question, which is generous until you meet the longer architecture scenarios in the two heaviest domains.

What is the passing score for the AWS generative AI certification?

You need 750 on a scale that runs from 100 to 1000. Because the score is scaled rather than a raw percentage, it adjusts for the difficulty of the exam form you sit. Treat it as needing strong performance in the two front-loaded domains and no collapse anywhere else.

How much does AIP-C01 cost?

The exam costs $300 USD, which is standard for the AWS professional tier. It can be taken at a Pearson VUE test centre or as an online proctored session, and it is offered in English, Japanese, Korean and Simplified Chinese.

Do you need the AWS AI Practitioner certification first?

No. AWS sets no mandatory prerequisite certification for AIP-C01. It recommends two or more years of production AWS or open-source development, general AI or data engineering experience, and one year of hands-on generative AI implementation. That experience matters far more than any prior badge.

Which domain carries the most marks in AIP-C01?

Foundation model integration, data management and compliance, at 31 percent. That is roughly 23 of the 75 questions, and it covers model selection, data pipelines, vector stores, retrieval design and prompt governance. Implementation and integration follows at 26 percent.

Does the exam test Amazon Bedrock specifically?

Heavily. Bedrock appears in every domain, including Guardrails, Knowledge Bases, Prompt Management, Cross-Region Inference and reranker models. Candidates who have only used foundation models through other interfaces will find the Bedrock-specific behaviour questions difficult to answer from documentation alone.

Is AIP-C01 a machine learning exam?

No. It assumes the model is somebody else’s and tests everything around it: retrieval, prompting, guardrails, cost, latency and failure handling. If you own training data, features and model evaluation, the Machine Learning Engineer associate exam describes your job more accurately.

What vector databases does the AWS generative AI certification cover?

Amazon OpenSearch Service with vector search and the Neural plugin, Amazon Aurora with the pgvector extension, and Bedrock Knowledge Bases as a managed option. DynamoDB and RDS appear for metadata and document repositories rather than as primary vector stores.

How long does it take to prepare for AIP-C01?

Candidates already building generative AI applications on AWS commonly report six to ten weeks of part-time study. Those coming from a general AWS background without hands-on generative AI work need considerably longer, because the two heaviest domains assume decisions you can only have made in practice.

Is the AWS generative AI certification worth it?

It is worth it if you are already shipping generative AI applications and want a professional-tier credential that says so. It is not a route into the field, because the blueprint assumes a year of hands-on implementation that no amount of reading substitutes for.

Conclusion

AIP-C01 is an architecture exam wearing an AI label. Fifty-seven percent of the marks sit in foundation model integration and implementation, which means retrieval design, model selection and orchestration decide the outcome far more than safety or testing do. The named services are not decoration either: Bedrock, OpenSearch, SageMaker AI, Lambda and Step Functions appear across domains because the exam expects you to have chosen between them under real constraints.

Treat the recommended year of hands-on generative AI work as the real prerequisite. Ship one retrieval application, break it on purpose, and instrument what it costs. Once those decisions feel routine, working practice items against the five domains is the fastest way to find what is still missing.