How I Passed AWS Certified AI Business Strategist (AB1-C01 Beta)

I passed the AWS Certified AI Business Strategist (AB1-C01) beta on 29 Sep 2026, the day it launched, with a score of 743 / 1000. It's my 15th AWS certification. This guide covers everything I used and everything I'd revise if I took it again.

AB1-C01 exam at a glance

AB1-C01 exam domains

Study resources: what I actually used

I used Stephane Maarek's course and the Skill Builder practice set, and nothing else. The Skill Builder learning plan is optional.

AB1-C01 vocabulary cheat sheet

Most of this exam is vocabulary. Know each of these terms in one line.

AB1-C01 exam traps and decision rules

The ML chain, in order

• Data → algorithm → training → model → inference → prediction. Know this order.
• The algorithm is the recipe, and the model is the result. Inference is the action, and the prediction is its output.
• There's no model without training, and no inference before a model exists.
• Training cost is paid at launch and again each time you retrain. Inference cost is paid on every use, so it grows with usage.

Drift: when the model sees new data

• A model can be right at launch and wrong six months later. Nothing crashes and no error is raised.
• Watch four kinds of drift: data, model quality, bias and feature attribution.
• Capture a baseline at training time. Compare live inputs and outputs to it on a schedule, then alert and retrain or fix the data source when a threshold is crossed.
• Some problems look like drift but aren't: a capacity limit, a bad prompt, or normal day-to-day variation.

Tokens and the context window

• Tokens decide what fits and what you pay.
• Everything counts against the context window: the question, chat history, retrieved documents, examples and the answer.
• A short question can reach the model as thousands of tokens.

Prompt, RAG or fine-tune

• Always start with the prompt, because it's the cheapest.
• If the model needs new knowledge, use RAG. Stale facts are a retrieval problem.
• If it has the right facts but the wrong tone or format, fine-tune.

Which kind of AI, if any

• An assistant gives an answer, a dashboard gives a view, a predictive model gives a score, and an agent takes an action.
• Only an agent acts on its own. Choose one only when the task truly needs autonomous action.
• Some things that sound like AI only need rules. Rules are cheaper and don't drift.

Build, buy or partner

• Build when your advantage lies in your own data or logic.
• Buy when you need speed and the capability doesn't set you apart.
• Partner when you lack expertise you can't hire in time.
• The answer depends on each use case, not on a company-wide policy.

Envision → Experiment → Launch → Scale

• Envision: exploring opportunities, with nothing in production.
• Experiment: controlled pilots, where success means learning.
• Launch: proven pilots in production, with value measured.
• Scale: AI across functions, with governance and continuous improvement.
• Some AWS material calls stage two "Align." The exam uses "Experiment."
• Readiness is set by your weakest dimension, not by an average.

Prove the value

• Record a baseline before launch. Without one, an improvement is a claim, not evidence.
• Attribution means showing the AI caused the change, not just that the number moved.
• Every initiative ends in one of three decisions: scale, pause with a fix and a date, or terminate.
• Money already spent is never a reason to continue.

Governance and risk

• Risk class = severity of a wrong outcome × likelihood. Regulated uses never sit below high.
• Reassess the risk when the use changes, for example when an internal tool starts facing customers.
• Confidence-based routing lets the system decide when it's confident and sends the rest to a person.
• ISO 27001 and SOC 2 are security certifications, not AI governance. A vendor's "Responsible AI" page is marketing, not proof.
• When you find shadow AI, give people an approved path first. Punishment doesn't fix it.

AWS CAF, ROI and cost tools

• AWS Cloud Adoption Framework (CAF): AWS guidance for assessing readiness across six perspectives: Business, People, Governance, Platform, Security and Operations. Its phases are Envision, Align, Launch and Scale.
• ROI: (value gained − total cost) ÷ total cost. Count the running costs, such as inference, monitoring and people, not just the build.
• AWS Pricing Calculator: estimates the cost of a proposed solution before you build it. Use it for the business case.
• AWS Cost Explorer: shows what you are actually spending. Tag each AI initiative so its cost can be tracked on its own.
• Savings Plans: a 1- or 3-year spend commitment in exchange for lower prices. They suit steady, predictable workloads, not early pilots.

More AB1-C01 key concepts, by domain

Domain 1: AI fundamentals

• Four data quality checks: completeness (fields present), consistency (one entity, one record), currency (the data reflects today's process), and representativeness (every group the model serves is covered).
• Bad data needs a data fix. Fix it at the source, then retrain. A smarter model, more data of the same quality, or more dashboards won't fix it.
• The rule test: if the same input always gives the same correct output, write a rule instead of using AI.
• "The answers are bad" has four causes, each with its own fix:

• "Give it more data" is never the fix by itself.
• Signs of a capacity problem: it forgets earlier turns, answers badly only on long documents, or cuts answers off mid-sentence.
• AWS service map:
  • Amazon Bedrock: you choose a foundation model, you never train one.
  • Amazon Quick: a ready-made AI business assistant on company data that respects each user's access rights.
  • Amazon SageMaker AI: for building and training your own models.
  • AWS Marketplace: third-party AI software, models, datasets and partners, billed through your AWS account. It's also a way to offer approved tools instead of shadow AI.

Domain 2: Strategy and business value

• A use case has three parts: the AI capability, the mechanism (how the work changes), and the business outcome. If one is missing, it's a wish, not a use case.
• Start from the outcome, not the technology. "Add a chatbot" is a technology looking for a problem.
• Decide in this order: outcome, then feasibility, then budget. Reversing the order is the classic mistake.
• Five things that can override build, buy or partner: budget, timeline, internal skills, strong vendor offers, and regulation.
• Missing skills mean buy or partner. "Delay until we hire" and "terminate" are both distractors.
• Rank a portfolio on four criteria: business value, feasibility, sustainability (can it be run and funded for its whole life), and strategic alignment.
• Four portfolio traps:
  • Sunk cost: money already spent is gone, so it's no reason to continue.
  • The quiet drain: an initiative nobody reviews keeps eating budget.
  • The platform switch: moving to a new platform changes the bill, not the outcome.
  • Data not ready: that's a feasibility gap, so pause and fix the data.
• KPIs are chosen before launch so the baseline can be recorded. Use tangible KPIs (convert to money) and intangible ones (satisfaction, productivity). Leaving out the intangibles understates value.
• A good KPI measures the outcome and the quality of the process. It doesn't measure inputs like headcount or budget, or infrastructure like uptime or latency.
• Lagging vs leading indicators: cost and revenue impact shows up months later. Adoption, quality trend and delivery milestones in months 1–3 predict it.
• Four kinds of AI initiative:
  • Scaling: more of the business
  • Optimisation: a better decision in an existing process
  • Process improvement: a redesigned process
  • Transformation: new customers, revenue or market position. Only transformation creates value outside the company.
• What isn't a durable advantage: being first, spending more, or building rather than buying. Your own data is what can make it durable.
• Switching platforms is an iceberg. You see the new feature and the better price. You pay for data migration, rebuilt integrations, rewritten prompts and retrained people.
• GenAI is billed per token for as long as it runs. The right spend is whatever the KPIs justify, not the lowest possible.

Domain 3: Governance and responsible AI

• The eight dimensions of responsible AI: fairness, explainability, privacy and security, safety, transparency, veracity and robustness, controllability, and governance.
• Governance starts at planning, when changes are still cheap.
• When a human must be in the loop: the decision affects rights, health, safety, money or access to a service; it's hard to reverse; or it's regulated. AI can decide alone when the stakes are low, the action is easy to undo, and the model is confident.
• Five safeguards sit between the model output and the final answer: a confidence threshold, hallucination detection, guardrails, written escalation criteria, and an audit trail.
• Risk comes from the use, not the technology. One model can carry two risk classes for two different uses.
• Obligations attach to the process. Replacing a person with a model keeps every obligation and adds some.
• Bias can enter through the framing of the problem and through the data. Retraining on skewed data reproduces the skew. The fix is ongoing fairness monitoring by group, with a threshold and an owner.
• Control harmful content at the boundary, on both inputs and outputs: filter, deny topics, mask data, refuse.
• Three families of data security: encrypt data at rest and in transit, mask or remove fields the task doesn't need, and log access. No log means no proof for the auditor.
• Least privilege: give each person and system only the access it needs.
• Shared responsibility model: AWS handles security of the cloud, and you handle security in the cloud (your data, access, configuration, and what you send to a model). A managed service reduces your work, never your responsibility.
• AWS Well-Architected lenses: the Responsible AI Lens and the Generative AI Lens.

Domain 4: Readiness and transformation

• Readiness has five dimensions: leadership, data, culture, infrastructure and governance. Your readiness is your lowest bar. Name the exact gap ("no owner for the returns data"), not a vague score.
• Four capability gaps to invest in: people, process, technology and governance.
• Prioritise by dependency (people before process), align with the strategy, and move one stage at a time.
• Four ways to build AI skills:
  • A proof of concept, when you need evidence
  • A hackathon, when you need ideas and energy
  • A training programme, when many people share a known skill gap
  • Responsible AI training, for anyone who builds, approves or relies on AI
• Starting from zero, with fear in the workforce? Run a company-wide awareness programme first.
• Roll out in waves and keep a way back. Run the manual process in parallel until each wave meets its criteria. The wrong answers are switching everyone at once, reacting to complaints, and waiting for a guarantee.
• Short-term wins build credibility for the next wave.
• AWS CAF is the answer for an organisation-wide readiness assessment.

AB1-C01 practice questions

These are my own practice questions, not real exam questions, written to test the concepts above.

1. A retailer wants a chatbot to answer from its return policy, which changes every week. What's the best approach?
A) Fine-tune the model every week B) Use RAG over the policy documents C) Train a new model D) Paste the policy into every prompt by hand

Answer

B. New or changing knowledge is a retrieval problem, so use RAG. Fine-tuning changes behaviour, not facts.

2. A loan model's overall accuracy is unchanged, but approvals for one age group have gotten worse since launch. What is this?
A) Data drift B) Bias drift C) A capacity limit D) A bad prompt

Answer

B. Accuracy holds overall but drops for one group, which is bias drift. SageMaker Clarify with Model Monitor can detect it.

3. A company needs GenAI text summarisation quickly. It doesn't set them apart from competitors, and they don't want to manage infrastructure. What fits best?
A) Build and train a custom model B) Use Amazon Bedrock C) Hire an ML research team D) Wait for the market to mature

Answer

B. It's buy, not build, because the capability isn't a differentiator. Bedrock gives managed access to foundation models.

4. A pilot claims a 20% drop in call handling time, but no one measured handling time before launch. What's the main problem?
A) The model is too small B) There is no baseline C) The pilot ran too long D) There was no fine-tuning

Answer

B. Without a baseline, the improvement is a claim, not evidence.

5. A pilot missed its targets. The sponsor wants to scale anyway because $500K has already been spent. What should happen?
A) Scale it B) Decide on evidence: pause with a fix and a date, or terminate C) Double the budget D) Extend the pilot indefinitely

Answer

B. Money already spent is never a reason to continue.

6. An internal HR summarisation tool is about to become customer-facing. What's the first thing to do?
A) Nothing, because it's already approved B) Reassess its risk classification C) Fine-tune it D) Turn off logging

Answer

B. A change in use changes the risk, so reassess the risk class.

7. Employees are pasting customer data into unapproved public AI tools. What's the best first response?
A) Discipline everyone involved B) Give them an approved, secure alternative and a clear policy C) Ignore it D) Block the internet

Answer

B. Shadow AI is fixed with an approved path. Punishment alone just pushes it underground.

8. You need to estimate the monthly cost of a proposed GenAI solution for a business case before building anything. Which tool?
A) AWS Cost Explorer B) AWS Pricing Calculator C) AWS Budgets D) SageMaker Clarify

Answer

B. The Pricing Calculator estimates cost before you build. Cost Explorer shows actual spend after.

9. An inference workload has run at steady, predictable usage for a year. How can you cut its cost?
A) Savings Plans B) Fine-tune it C) Use Cost Explorer only D) Move it to a pilot

Answer

A. Savings Plans trade a 1- or 3-year commitment for lower prices. They suit steady workloads, not early pilots.

10. A vendor offers its ISO 27001 certificate as proof of responsible AI governance. Is that enough?
A) Yes B) No. It's a security certification, so ask for AI governance evidence such as ISO/IEC 42001. C) Yes, if they also have SOC 2 D) Yes, if their website says "Responsible AI"

Answer

B. ISO 27001 and SOC 2 cover security, not AI governance.

11. In long chat sessions, the assistant starts "forgetting" what was said earlier. What's the likely cause?
A) Data drift B) The context window limit C) Bias drift D) Too much fine-tuning

Answer

B. Chat history uses up tokens in the context window. Once the window is full, earlier content drops out.

12. The model gets the facts right, but even with careful prompting its output never matches the brand's tone and format. What's next?
A) RAG B) Fine-tuning C) A bigger context window D) Rules-based logic

Answer

B. Right facts with the wrong tone or format is the case for fine-tuning.

Exam-day strategy: option elimination

1. Read the question for the business goal first: cost, speed, risk, or value.
2. Rule out the two options that clearly don't fit that goal.
3. Between the last two, pick the one that solves the business need with the least effort or risk.
4. Flag it and move on. At about 2 minutes per question, don't get stuck.
5. Save energy for the last hour. That's when every option starts to look right.

AB1-C01 FAQ

How hard is the AWS AI Business Strategist exam?
It's harder than it looks. There's no deep tech, but a lot of vocabulary, and many options that sound right. If you have an AI background, Domain 1 will be easy.

Do I need a technical background?
No. It's a business-level exam. Non-tech candidates should start with the ML chain and the vocabulary table above.

Is the Skill Builder learning plan required?
No. I passed with Stephane Maarek's course and the Skill Builder practice set only.

What's the best course for AB1-C01?
Stephane Maarek's course was the only course I used.

How long should I study?
A few focused days on the course, the slides and the practice questions. The Skill Builder learning plan is optional for anyone.

Who should take it?
Consultants, pre-sales, solution architects, product managers, and anyone who talks to business stakeholders about AI.

My AB1-C01 study plan

• [ ] Watch Stephane Maarek's course (about 3.5 hrs)
• [ ] Rewatch the review section and reread the slides
• [ ] Do Stephane Maarek's practice questions
• [ ] Take the 20-question Skill Builder practice set
• [ ] Optional: the Skill Builder learning plan (12–13 hrs)
• [ ] Eliminate options on exam day and keep a pace of about 2 minutes per question

---

Written by Vishnu Rachapudi, AWS Community Builder. More hands-on AWS guides at vishnurachapudi.com.