MLA-C01 to MLA-C02: What Machine Learning Candidates Need Now

The original framing of this page is now outdated. MLA-C01 was the first AWS Certified Machine Learning Engineer – Associate exam, but English delivery ended on September 28, 2026. AWS opened registration for the MLA-C02 beta on September 1 and began English beta delivery on September 29. MLA-C01 remains available in Japanese, Korean, and Simplified Chinese until the updated exam reaches general availability.

The important point is that the certification itself is not disappearing. It is being updated to reflect a machine-learning engineering role that now includes generative AI, foundation models, large language models, agentic workflows, Amazon Bedrock, and responsible AI in addition to traditional ML engineering. Candidates preparing in English should therefore study MLA-C02, not continue following an MLA-C01-only plan.

A useful supporting overview is the AWS Machine Learning Engineer Associate transition to MLA-C02. This article goes deeper into what the beta changes and what preparation should look like now.

MLA-C02 keeps the four-domain structure but broadens what “ML engineer” means

AWS did not replace the entire exam structure. MLA-C02 still has four domains: Data Preparation for ML and AI; ML Model and Foundation Model Development; Deployment and Orchestration of ML and AI Workflows; and Operating, Monitoring, and Securing ML and AI Solutions. The published weights are 28%, 24%, 24%, and 24% respectively.

That continuity matters because traditional ML engineering remains central. You still need to reason about data quality, feature preparation, model development, deployment, monitoring, security, and operational workflows. The update expands those responsibilities into generative and agentic AI rather than abandoning them.

If your preparation treats MLA-C02 as “an LLM exam,” it will be unbalanced. If it ignores foundation models and AI workflows because you studied MLA-C01 material, it will also be incomplete.

The beta has different logistics from a standard AWS exam

AWS lists the MLA-C02 beta at 170 minutes with 85 questions and a beta price of $75 USD. The beta is English only. AWS notes that beta exams contain additional items used for statistical evaluation and that these additional items do not affect the candidate’s score.

Do not use the longer duration as a reason to slow down excessively. More items mean more opportunities to spend too long on ambiguous questions. A practical strategy is to maintain a steady first pass, mark a small number of genuinely uncertain items, and preserve time for review rather than repeatedly reconsidering questions already answered with good evidence.

Also remember that beta content is validating an updated role. Questions are likely to reward architectural understanding and operational judgment more than memorization of every parameter in every AI service.

Data preparation is still the largest domain

At 28% of scored content, data preparation remains the largest domain. That is a useful corrective to AI hype. Production ML and AI systems still depend on data quality, appropriate transformation, governance, access, storage, and reproducible pipelines.

Practice identifying missing values, leakage, biased sampling, inconsistent schemas, class imbalance, and inappropriate transformations. Understand how batch and streaming data differ operationally. Be able to reason about data catalogs, lineage, permissions, and the separation of raw, curated, and model-ready datasets.

The broader machine learning engineering skills is helpful here because it frames data work as part of an end-to-end production system rather than preprocessing that happens once before training.

Model development now includes foundation models and generative AI

MLA-C02 explicitly expands the model-development domain to include foundation models and generative AI. AWS highlights model selection, customization, operationalization, and Amazon Bedrock as important additions.

Prepare to reason about whether a problem needs a traditional supervised model, a foundation-model workflow, retrieval augmentation, fine-tuning or customization, or a simpler rules-based approach. The exam can test whether the chosen technique fits the data, latency, cost, governance, and quality requirements.

Do not reduce this to memorizing model names. A good answer should explain the engineering trade-off. Retrieval can improve grounding without retraining a foundation model. Fine-tuning changes model behavior but introduces new data, evaluation, and lifecycle responsibilities. Agentic workflows add orchestration and tool-use concerns that must be observed and secured.

Responsible AI belongs inside engineering decisions

AWS’s update gives responsible AI greater visibility. That does not mean memorizing a list of principles and then forgetting them during architecture questions. Fairness, privacy, transparency, safety, and governance influence dataset selection, evaluation, monitoring, access controls, and the decision to automate an action.

The responsible AI and governance discussion is useful supporting material even though that article targets a different AWS certification. The concepts matter because MLA-C02 engineers are expected to operationalize AI, not merely produce a promising model in a notebook.

Practice scenarios where model quality is high but the deployment is still unacceptable because sensitive data leaks, monitoring cannot detect drift, outputs are unsafe, or the process cannot be audited.

Deployment and orchestration are where models become production systems

A model that works offline is only the beginning. MLA-C02 expects you to think about endpoints, batch inference, pipelines, orchestration, versioning, scaling, rollout, rollback, and integration with applications or data workflows.

Ask how the inference pattern changes the architecture. A real-time endpoint has different latency and scaling needs from asynchronous or batch inference. A foundation-model application may depend on retrieval, guardrails, tools, and external data sources. A multi-step agent workflow introduces state and failure points that a single model call does not.

Use the same operational habits that help in observability: define what success looks like before deployment, monitor the signals that reveal user impact, and preserve enough context to diagnose failures.

Monitoring must cover the model and the system around it

ML systems fail in more ways than traditional applications. Infrastructure can fail, but so can data quality, feature distributions, model performance, retrieval relevance, output quality, and cost assumptions. A healthy endpoint does not prove that predictions remain useful.

Build a monitoring model that separates platform health from model behavior. Latency, error rate, throughput, and resource utilization describe the serving system. Prediction quality, drift indicators, bias metrics, retrieval quality, and safety signals describe the AI behavior. Business metrics may reveal whether either layer is actually delivering value.

MLA-C02 readiness improves when you can decide which signal belongs at which layer and what action should follow a threshold breach.

Security expands when models can access tools and sensitive context

Traditional least privilege still applies. Workloads should use scoped roles and temporary credentials, data should be encrypted appropriately, secrets should be protected, and logs should not expose sensitive values. Generative and agentic systems add new boundaries because models may receive private context, invoke tools, or act on behalf of users.

That makes permission design and data boundaries central. An agent that can call an internal API should not automatically inherit unlimited privileges. Retrieval sources should enforce authorization rather than treating every indexed document as universally accessible. Logging should capture useful evidence without reproducing sensitive prompts or records unnecessarily.

A review of secrets management reinforces the basic control pattern: credentials and sensitive configuration need lifecycle management, not just storage.

Build a preparation plan around production questions, not feature lists

For each domain, create scenarios that force a decision. How would you prepare a noisy dataset? Which model approach fits the requirement? How will the workflow be deployed? Which metrics reveal drift? How is access controlled? What changes when generative AI is added to an existing ML pipeline?

Use diagrams and small labs to connect services, but keep the question centered on behavior. If you can explain the data flow, model choice, deployment path, monitoring plan, and security boundary without relying on a memorized console sequence, your knowledge will transfer better to unfamiliar items.

The AWS AI certifications can also help distinguish MLA-C02 from foundational AI awareness and the more developer-focused professional AI credential.

Evaluation deserves more attention in the updated exam because an AI system can fail even when the infrastructure is healthy. For traditional ML, choose metrics that fit the prediction problem and class distribution. For generative AI, evaluation may include factual grounding, relevance, safety, task completion, latency, and cost. No single metric proves quality across every use case.

Practice distinguishing offline evaluation from production monitoring. Offline tests help compare candidate models or prompts before release. Production signals reveal changing data, changing user behavior, retrieval failures, new safety issues, or cost growth after deployment. A mature workflow connects the two: define acceptance criteria before launch, then monitor whether real traffic continues to satisfy them.

Data lineage and reproducibility also matter. If a model behaves differently, an engineer should be able to identify which dataset, feature transformation, model version, configuration, or retrieval source changed. That is why versioning and metadata are operational controls, not documentation chores.

For beta preparation, avoid chasing every new AI announcement. Freeze your study against the MLA-C02 exam guide and use AWS’s in-scope service list as a boundary. The fastest way to dilute readiness is to study interesting generative-AI topics that the blueprint does not require while leaving core data preparation, deployment, monitoring, and security weak.

A final note for candidates who already earned MLA-C01: AWS says the credential remains active for its normal validity period. The exam-code transition does not invalidate an existing certification. The practical question is therefore about what to study next, not whether the credential suddenly disappears when the beta begins.