AIB-C01 Study Guide

How to use this guide

This guide follows the four domains of the AWS Certified AI Business Strategist (AIB-C01) exam. Each domain is a section and each topic is a subsection you can open from the side menu. The exam tests business judgment through enterprise scenarios: according to AWS, its services appear as practical context, but they are not what is being tested. Knowing them at a strategic level is enough.

Technical terms are kept in English, as they appear on the exam. The markers show the origin of anything that is not official:

• (external): a real framework or law that does not appear in the official exam guide (for example EU AI Act, GDPR, NIST AI RMF, CoE models, ISO/IEC 22989). It is useful for practicing the reasoning, but do not treat it as official content.
• (heuristic): the author’s own criteria for picking answers. It is a rule of thumb, not a guarantee.
• Anything without a marker comes from the exam guide or from official AWS documentation.

The recommended order is to read a full domain, take a practice test and go back to the topics where you missed questions. The simulators on this site use 85 questions and 170 minutes, to match the beta format. Domain 2 carries the most weight (28%), so give it extra time.

Exam quick reference

Keep the format difference in mind: the current beta exam has 85 questions in 170 minutes, while the standard guide mentions 130 minutes. Prepare for the beta format if you are going to take it now.

Weight of each domain

Official practice resources

These resources require signing in and the author of this guide could not read them. Neither this guide nor the simulator questions are based on them: use them as your main reference for style and difficulty.

Domain 1: AI Fundamentals and Literacy (24%)

This domain tests whether you know which type of AI solves which problem, what it needs in terms of data and how to adapt it, without getting into technical detail.

Core concepts

AI, ML and GenAI

AI contains ML, and ML contains GenAI.

Trap: a yes-or-no prediction on tabular data is ML, not GenAI.

Data

• Quality dimensions: completeness, accuracy, consistency, timeliness (current data) and representativeness.
• Garbage in, garbage out: a bigger model does not fix bad data.
• Historical data: the model repeats the biases of past decisions. If the environment has changed, old data no longer represents the present.

Rule-based automation or AI

Best practice: hybrid. ML for the patterns and rules for the hard constraints, such as policy limits.

AI agents

• An agent plans multi-step tasks, uses tools (APIs, systems, databases) and acts with some autonomy toward a goal.
• A chatbot only answers prompts; it does not execute actions.
• Multi-agent: agent-to-agent communication with orchestration, for example a supervisor agent that distributes the work.
• Control: autonomy limits (an automatic threshold with human approval above it), action logging and oversight. Agents are not deterministic and still need supervision.

Model drift

• What it is: model performance drops after deployment because the data changes (data drift) or the relationship between inputs and outcome changes (concept drift, for example when fraudsters change tactics).
• Early signal: the distribution of the input data moves away from the training data, or KPIs fall on recent data.
• Response: continuous monitoring against KPIs and real outcomes, thresholds that trigger review or retraining, and an ongoing maintenance budget (TCO), not a one-time project budget.
• Incorrect: freezing the model, auditing only once a year or retraining “because a year has passed”.

Shadow AI

• What it is: employees using AI tools that the company has not approved or governed.
• Risks: data leakage, intellectual property exposure and regulatory non-compliance.
• Solution: a transparent tool classification (Approved, Blocked, Under evaluation), a fast track to request evaluations, approved alternatives and training.
• Banning everything pushes usage underground. The classification does not replace security review.

ISO standards

The guide cites ISO/IEC 23053 and ISO/IEC 42001 as examples (skill 1.1.6). If the question describes teams using different terms, the answer is to adopt a common vocabulary based on international standards.

Prompt engineering

• A good prompt includes: role, context, a specific instruction, audience, tone, length, output format and examples (few-shot).
• For comparable outputs, define the exact structure (for example, a table with columns) and include an example.
• Against made-up answers: “answer only with the given context and, if it is not there, say so”. It works better with RAG.
• Avoid vague prompts, unnecessary sensitive data and several tasks mixed into one sentence.

Tokens and context window

• A token is a unit of text. Many services charge for input and output tokens, so longer prompts and responses raise cost and latency.
• The context window is the maximum number of tokens per request: prompt, history and response.
• Symptoms of exceeding it: truncated responses, initial instructions ignored in long chats, parts of long documents omitted, higher cost and slowness.
• Solutions: RAG, splitting the text into fragments (chunking), summarizing the history or using a model with a larger context.

How to improve responses

• RAG reduces hallucinations but does not eliminate them, and it still requires governance and access control.
• Fine-tuning is not suitable for information that changes daily.

Domain 2: AI Strategy and Business Value Creation (28%)

This is the highest-weighted domain. It covers which use cases to choose, how to obtain them, how to prioritize them, how to prove their value and how to control their cost.

Use cases by function

To select: business problem, then measurable KPI, then value and feasibility. Never start from “the newest model” or “the most senior executive’s idea”.

Build, buy or partner

• Decision factors: budget, timeline, internal capabilities, vendor proposals and regulatory compliance (for example, data residency).
• Before signing, validate with a POC on your own data and set acceptance criteria in the contract. Never rely only on the vendor’s benchmark.
• AWS Marketplace lets you discover, try and buy third-party models and solutions with billing on your AWS account. It does not replace your own due diligence.

Prioritization

Criteria: business value, feasibility, sustainability and strategic alignment.

Sustainability means being able to operate it in the long term: ongoing cost, maintenance, data availability, talent and resource consumption.

Trap: meeting the technical goals without adoption or business impact is not success.

When AI is not the solution

• The logic is fixed and deterministic, like an overtime formula.
• There is no relevant data and no way to get it.
• Error is unacceptable and outputs cannot be reviewed.
• The root cause is process or ownership: fix the process first.
• “An AI strategy” is being sought because competitors announced one, without a defined problem.

Transitions

• From manual process to AI: assess data readiness, total cost, continuity and performance.
• Between platforms or models: run in parallel, compare, switch in phases and have a rollback plan. Re-test prompts and evaluations, because each model responds differently.
• Underestimated costs: integration, data preparation, change management and ongoing operation. The license is the visible part.

Measuring value

• A good KPI set mixes efficiency, business outcome and quality. Model accuracy alone does not prove value.
• Baseline: measure it before deploying, with the same metrics and over a representative period. If it is missing, rebuild it from historical data or compare against a control group. Never assume a baseline of zero or use the vendor’s benchmark.

ROI and payback

• ROI example: 10,000 h × $50 = $500,000 in benefit. With a cost of $200,000, the ROI is 150%.
• Payback example: with an investment of $300,000 and $25,000 of net benefit per month, the payback is 12 months.
• Include all costs: usage and inference, monitoring, integration, training and change management. Exclude sunk costs unrelated to the project.
• Time saved only counts as ROI if it is redirected to higher-value work or reduces costs.

Costs and pricing on AWS

The exam does not test AWS services in depth. It is enough to know what each pricing option is for.

AWS also lists a Reserved tier.

To optimize: use the smallest model that meets the quality bar, batch for anything not urgent, short prompts and budget alerts. Incorrect: always using the largest model or sending the full history in every prompt.

Competitive advantage

• Competitive landscape: analyze how competitors and new entrants use AI, and how customer expectations are changing. Then look for gaps. Do not copy blindly or ignore the market.
• Transforming the business model means changing how value is created and captured: from selling machines to selling guaranteed uptime, or new paid AI products. Internal efficiency does not transform the model.
• Sustainable advantage: proprietary data that improves with use (data flywheel), AI embedded in processes and organizational capabilities. Using a public model, issuing a press release or running an isolated pilot is not sustainable.

Domain 3: AI Governance and Responsible AI Leadership (24%)

The correct answer applies a control proportional to the risk, builds it in from the design stage and assigns a clear owner.

Responsible AI dimensions

These are the 8 official AWS dimensions. The guide cites fairness, explainability, privacy, safety and transparency as examples.

Trade-offs

Incorrect: deciding informally, fixing the problem once there are complaints or removing the principle from the policy.

Governance by design

• Responsible AI is built in from planning, before collecting data and choosing the model.
• Planning includes: risk classification, impact assessment, fairness metrics, human oversight requirements and acceptance criteria.
• Adding controls later forces redesign and retraining, causes outages and damages reputation.

Human oversight and guardrails

• Human-in-the-loop: a human approves before the decision takes effect. It is mandatory in health, credit, employment, insurance and rights.
• Human-on-the-loop: a human supervises and can intervene. It suits lower-risk cases.
• Criteria for escalating to a human: low confidence, a distressed customer, legal topics or amounts above a threshold.
• Amazon Bedrock Guardrails offers content filters, denied topics, word filters, sensitive information filters (PII), contextual grounding checks (against hallucinations in RAG) and automated reasoning checks. It is applied consistently across all applications.

Governance structures

• A cross-functional AI governance council (business, technology, legal, compliance, security, risk) with clear decision rights.
• A named business owner answers for the outcomes and risks of each use case. The cloud provider and “the system” are not owners.
• A RACI matrix shows who does the work, who decides, who is consulted and who is informed.
• A committee made up only of data scientists leaves legal, ethical and business blind spots.

Regulation

The guide asks you to identify and address regulatory compliance risks (skill 3.2.2), but it does not name specific laws. The EU AI Act and GDPR in this section are (external) examples for practicing the reasoning, not official exam content.

• Privacy (GDPR, external): legal basis or consent, minimization, anonymization and documentation of the processing. Responsibility remains yours, not the model provider’s.
• Data residency: choose Regions within the required jurisdiction and back it up with contractual terms. A provider having offices in the country is not enough.
• Ongoing compliance: keep an inventory of AI systems with their risk level and watch for regulatory changes in each jurisdiction.

Shared responsibility model

On Amazon Bedrock, model providers do not have access to customer prompts or responses.

• In RAG, retrieval must respect the permissions of the source documents (least privilege). Otherwise HR or finance information is exposed.
• If there is PII in prompts, combine a classification policy, guardrails that mask sensitive data, approved tools and training. A total ban with no alternative does not work.

Risk classification

• Its purpose is proportional governance: more control for high risk and less bureaucracy for low risk.
• Factors that raise risk: impact on rights or on access to services, automated decisions without human review, sensitive data and public exposure.
• Factors that lower risk: internal use, expert review before using the output and public data.
• Frameworks: ISO/IEC 42001 (in the guide) and NIST AI RMF (external), with the functions Govern, Map, Measure, Manage. Distractor: Identify, Protect, Detect, Respond belongs to the NIST Cybersecurity Framework.

Controls in production

• Continuous monitoring of performance, security and usage, with alerts.
• An incident response plan: kill switch or rollback, communication and root cause analysis.
• Prompt and response logs, protected, for traceability and audit.
• Incorrect: a single review at launch, relying on users to report problems or freezing the system.

Bias across the lifecycle

Bias drift appears when the population or the data change, and bias emerges or grows. That is why fairness is always monitored, not only at launch.

Harmful content and intellectual property

• Harmful content: guardrails and content filters, red teaming before launch and monitoring. Incorrect: relying on user reports, raising randomness or removing the system prompt.
• Intellectual property: outputs may resemble protected works. Review the provider’s terms (data use, output ownership, indemnification) and require human review before publishing.
• Pasting code or secrets into unapproved tools can expose that intellectual property or leave it retained by a third party.

Reliability

• A hallucination is a plausible but false or made-up answer.
• To mitigate them: RAG with approved sources, citations, grounding checks and human review for critical content.
• Data that must be exact, such as prices, is obtained from the system of record through a tool or API, not generated by the model.
• Degradation from input data: automated quality checks in the pipeline catch format changes immediately.

Domain 4: Business Readiness, Leadership, and AI Transformation (24%)

This domain tests whether the organization is ready and how to take AI from pilot to scale with people, data and leadership.

Readiness dimensions

Distractors that do not measure readiness: the number of vendors, the marketing budget or having “the latest model”.

Maturity

The guide does not define a maturity model with fixed stages. It only gives two examples (skill 4.1.2), and they help you recognize the starting point:

The maturity model is useful for setting realistic steps and investments from the current state. It does not guarantee success or replace strategy.

Capability gaps

• Prioritize closing the gaps that most block the strategic goals, in the order that current maturity allows. Do not skip stages: going from scattered pilots to agents across the whole company is incorrect.
• The right approach is to build the foundations (data, governance, skills) while delivering bounded use cases.
• AWS CAF organizes capabilities into six perspectives: Business, People and Governance (the business ones), and Platform, Security and Operations (the technical ones).

Data

• Silos prevent a unified customer view, which worsens personalization and predictions.
• First step: a data strategy with owners, governance and integration priorities tied to high-value use cases.
• Data owners are accountable for quality, definitions and access decisions.
• If two units do not share data, the solution is a data sharing framework with governance, access controls and benefit for both. Incorrect: copying everything without control or buying third-party data.

Infrastructure

• For real-time AI with customers: low latency, scalability at peaks and integration with channels.
• Baseline requirements: integration through APIs and security and identity.
• Without MLOps experience, managed services such as Amazon Bedrock remove the infrastructure burden.
• Incorrect: demanding a full migration, buying GPUs or replacing all legacy systems before the first pilot.

Leadership and change management

• The executive sponsor provides resources, removes barriers and signals that AI is a priority. If they leave, name a replacement quickly and realign.
• AI champions are advocates among their peers inside the units: they spread practices, drive adoption and collect feedback.
• Teams should be cross-functional (business, technical, legal, compliance, security, end users) from the start, with a RACI and a single owner for each decision.
• Communication must cover timelines, realistic expectations, impact on roles and reskilling plans, with feedback in both directions. When layoff rumors arise, respond directly and honestly with facts and the support available.

Watch out: technical barriers (quotas, context window, rate limits) are not cultural.

Training and roles

• Literacy should be tiered: basics for everyone, role-based training and Responsible AI for everyone who uses AI or makes decisions about it.
• Hackathons provide practice, uncover use cases and build commitment.
• POC programs build skills on real problems and, at the same time, demonstrate value.
• When redesigning roles, AI takes the repetitive and high-volume work. People keep judgment, critical thinking, empathy, creativity, AI oversight and exceptions.
• Delivering bad news, negotiating and making ethical decisions remain human tasks.

Scaling

1. Envision: define the business outcomes, prioritize the use cases and align the stakeholders.
2. Experiment: pilots and POCs with metrics and baseline.
3. Launch: move to production with governance, security and support.
4. Scale: extend across the whole enterprise with a platform and a CoE.

• Quick wins build credibility, learning and funding. The ideal first project has high value, low complexity, low risk and measurable, visible results.
• To track long-term value: feedback loops in the tool and periodic KPI reviews against the baseline. A one-time survey or counting deployed models is not enough.

The CoE defines standards, reusable components and best practices, and provides governance and training guidance. It does not replace business ownership. The guide only asks you to establish CoEs: the three models are common industry practice.

• Moving from pilot to production requires monitoring, incident response, security and compliance review, a support model, SLAs, documentation and scalable infrastructure. Pilot shortcuts do not scale.
• To scale to several markets, assess the regulation, language, data residency and support in each country, and deploy in waves with continuity plans. Incorrect: launching everywhere at once or waiting for perfect accuracy.

How to pick the right answer

These nine rules are (heuristic): the author’s own criteria for picking answers when you hesitate between plausible options. They help you rule out distractors, but they are not official exam content and do not replace understanding the scenario.

1. (heuristic) Start from the problem and the business outcome, not from the technology.
2. (heuristic) Prefer the iterative: pilot, metrics, phases, decision gates.
3. (heuristic) Prefer the balanced: business value together with governance. Be wary of extremes (“ban everything”, “never”, “always”, “guarantees”).
4. (heuristic) If there is high impact on people, the answer usually involves human oversight, explainability and transparency.
5. (heuristic) With a short timeline, no ML team and a non-differentiating capability, buy or find a partner before building.
6. (heuristic) Information that changes: RAG. Stable style or format: fine-tuning. Fixed logic: rules.
7. (heuristic) Without a baseline you cannot attribute the improvement or calculate ROI.
8. (heuristic) People and culture problems: transparency, training and participation.
9. (heuristic) FIRST asks for the prior step; LEAST inverts the question.

Sources

Exam and preparation

• Official certification page: beta format, price, languages, preparation resources.
• AIB-C01 exam guide: domains, weights, scoring and standard duration.
• Technologies and concepts that may appear: RAG, fine-tuning, tokens, agents, drift, shadow AI, ISO/IEC 23053 and 42001.
• AWS services in scope: consumption, instance and seat pricing structures, Pricing Calculator, Cost Explorer, Marketplace.
• AWS launch blog: what each domain tests, delivery date and the role of AWS services.
• Exam Prep Plan on AWS Skill Builder and AI Business Strategist Learning Plan: official practice resources.

Official guide by domain

• Domain 1: skills 1.1.1 to 1.3.3.
• Domain 2: skills 2.1.1 to 2.3.4.
• Domain 3: skills 3.1.1 to 3.3.4.
• Domain 4: skills 4.1.1 to 4.4.6, including the readiness dimensions, the maturity examples, the gaps (people, process, technology, governance) and the envision, experiment, launch, scale phases.

AWS documentation

• Amazon Bedrock pricing: Standard, Priority, Flex, Reserved, Batch and Provisioned Throughput tiers.
• Responsible AI on AWS: the 8 dimensions, Well-Architected Responsible AI Lens, AI Service Cards.
• Amazon Bedrock Guardrails: the six safeguards.
• Bedrock security (AWS Solutions): shared responsibility and model provider access.
• AWS CAF for AI: foundational capabilities: the six perspectives. AWS marks this whitepaper as a historical reference.
• AWS CAF: your transformation journey: Envision, Align, Launch, Scale phases.

External frameworks and practice

• EU AI Act, GDPR and NIST AI RMF (external): real frameworks not cited in the exam guide. ISO/IEC 22989 (external): the guide does not mention it.
• CoE models (centralized, hub-and-spoke, decentralized) (external): common industry practice; the guide only asks you to establish CoEs.
• ROI and payback formulas, use cases by function and leading and lagging indicators: standard business practice applied to skills 2.1.1 and 2.2.1 to 2.2.4.
• The definitions, examples and traps in Domain 1 are the author’s own explanations of skills 1.1.1 to 1.3.3.