The AIF-C01 blueprint is split across 5 domains. Weights are the official share of the exam; see the official exam guide for the authoritative breakdown.
What you must be able to do. Place AI, ML, and deep learning in relation to one another, match a business problem to the right capability, and walk the ML lifecycle as a loop that includes monitoring and retraining.
In one sentenceThe shared vocabulary the rest of the exam assumes: what AI, ML, and deep learning are, the three learning styles, and the lifecycle from data to deployment to monitoring.
Recall check: answer these from memory first
- Place AI, machine learning, and deep learning in relation to one another in one sentence.
- Give a one-line distinction between supervised, unsupervised, and reinforcement learning, with an example use case for each.
- Name the stages of the ML lifecycle in order, and say why monitoring and retraining belong in it rather than after it.
What it tests. The vocabulary and mental model: what AI, machine learning, and deep learning are and how they relate, the difference between supervised, unsupervised, and reinforcement learning, and the stages of the ML lifecycle from data preparation through training, evaluation, deployment, and monitoring. It also checks that you can match a business problem to an appropriate AI capability, including recognising when simple rules or analytics beat machine learning.
How to study it. Get the terms exact first, because later domains build on them. Be able to say in one sentence each what training, inference, a model, and a dataset are. Then learn the lifecycle as a loop, not a line: models degrade as data shifts, so monitoring and retraining are part of the job. Drill use-case questions until you can spot the distractor that proposes machine learning for a problem a lookup table would solve.
Easy to confuse
- AI versus machine learning versus deep learning. They nest: AI is the broad field, machine learning is the subset that learns patterns from data, and deep learning is the subset of machine learning that uses many-layered neural networks. The exam tests whether you put the broader term where a narrower one belongs.
- Supervised versus unsupervised learning. Supervised learning trains on labelled examples to predict a known target; unsupervised learning finds structure in unlabelled data, such as clusters. If the scenario already has labelled outcomes it is supervised; if it is grouping or finding patterns with no labels it is unsupervised.
- A problem for machine learning versus a problem for plain rules. Machine learning earns its keep when the pattern is complex and learned from data; a fixed lookup or a simple rule wins when the logic is known and deterministic. The exam plants a machine-learning distractor on a problem a lookup table would solve.
Worked example from the AIF-C01 bank
lock_openFree sampleFundamentals of AI and MLeasy
A retail company wants a system that improves its product recommendation accuracy over time by learning from customer purchase history, without being explicitly programmed with new rules each season. Which AI discipline best describes this approach?
- ARule-based expert systems, which encode domain knowledge as hand-crafted if-then rules.
- BMachine learning, which trains models on historical data so that predictive behaviour improves with experience.check_circle Correct
- CBusiness intelligence reporting, which aggregates and visualises historical sales data for human analysts.
- DRobotic process automation, which executes predefined workflow steps to mimic human repetitive tasks.
Understand that machine learning is the AI discipline in which models improve from data without explicit rule updates. Machine learning algorithms discover patterns in training data and adjust internal parameters so that predictions become more accurate as more data is seen. This contrasts with rule-based systems, which require a human to codify every new condition. The self-improving property from historical purchase data is the hallmark of ML.
Why A is wrong: Tempting because rule-based systems do drive recommendations, but they require explicit programming of every new rule. They do not learn autonomously from data, which is the defining requirement in the stem.
Why B is correct: Machine learning is the discipline in which algorithms learn patterns from data and improve performance on a task without being explicitly re-programmed, matching the scenario precisely.
Why C is wrong: Tempting because it uses historical purchase data, but BI reporting surfaces information for human decision-making rather than training a model that autonomously improves its own predictions.
Why D is wrong: Tempting because RPA is often grouped with AI tooling, but it follows a fixed script and does not learn from data. It cannot improve recommendation accuracy through experience.
What you must be able to do. Explain generative AI at a working level, predict how temperature and top-p change output, and judge when a probabilistic model is the wrong tool for a task that needs a guaranteed answer.
In one sentenceHow generative AI actually works: tokens, embeddings, prompts, and foundation models, the sampling knobs that trade determinism for variety, and the limits that decide when not to use it.
Recall check: answer these from memory first
- Explain in one line each what a token, an embedding, and a foundation model are.
- Say what temperature 0 does to the output and why an audit summary would want it.
- Name three limitations of generative AI that would make it the wrong tool for a task needing a guaranteed-correct answer.
What it tests. How generative AI works at a useful level: tokens, embeddings, prompts, and foundation models, and how a large language model predicts the next token. It tests the practical effect of sampling parameters such as temperature and top-p, and a clear-eyed view of what generative AI is good at and where it fails, including hallucination, cost, latency, and inconsistency.
How to study it. Anchor the sampling concepts to outcomes you can reason about: low temperature gives reproducible, on-message output, which is what an audit summary needs; higher temperature widens variety. Learn the limitations as decision criteria, so that when a scenario needs a guaranteed-correct factual answer you can explain why a raw model is the wrong choice. This domain carries real weight, so do not skim it.
Easy to confuse
- Temperature versus top-p. Both control randomness, but temperature scales how flat the probability distribution is, while top-p limits sampling to the smallest set of tokens whose probabilities sum to p. Temperature 0 forces greedy, reproducible output; the exam pairs the requirement (reproducible versus varied) with the setting.
- Generative versus discriminative models. A generative model produces new content such as text or images; a discriminative model classifies or predicts a label for existing input. If the task is creating or drafting it is generative; if it is sorting or scoring it is discriminative.
- Hallucination versus a wrong retrieval. A hallucination is the model inventing plausible content with no grounding; a retrieval error is a grounded system fetching the wrong document. The fix differs: grounding and evaluation reduce hallucination, better retrieval fixes the second, so the exam wants the cause named correctly.
Worked example from the AIF-C01 bank
lock_openFree sampleGenerative AI fundamentalsmedium
A support team sends the same prompt to a hosted LLM many times a day and gets noticeably different wording each time. They want answers to stay closely on-message and vary far less, without demanding byte-for-byte identical output. Which change best achieves this?
- ALower the temperature toward 0.2 so the model concentrates on the highest-probability tokens.check_circle Correct
- BRaise the temperature toward 1.0 so the model considers a wider spread of next tokens.
- CIncrease the maximum output-token limit so each answer has room to finish completely.
- DRaise top-p toward 1.0 so the model samples from the full probability distribution.
Lowering temperature concentrates output on the most-probable tokens, reducing run-to-run variation. Temperature scales how sharply the model favours its most-probable tokens. Lowering it toward 0 concentrates probability mass on the top tokens, so the same prompt produces closely similar, on-message wording each time without forcing fully identical output.
Why A is correct: Correct. A low temperature sharpens the distribution toward the likeliest tokens, so wording stays consistent run to run while still allowing slight variation.
Why B is wrong: Higher temperature flattens the token distribution and increases variation, which moves away from the on-message consistency the team asked for.
Why C is wrong: The token limit governs how long an answer can be, not how much its wording varies, so consistency is unchanged.
Why D is wrong: A top-p of 1.0 keeps the entire distribution in play, which preserves or increases variation rather than reducing it.
What you must be able to do. Design an application on a foundation model under real constraints, and make the retrieval-versus-fine-tuning decision automatically from the signals in the scenario.
In one sentenceThe largest domain: building on foundation models, the retrieval-versus-fine-tuning choice, the training and tuning process, and how to evaluate a model honestly.
Recall check: answer these from memory first
- Given a knowledge source that changes weekly, say whether to use retrieval or fine-tuning and why in one line.
- Name four design constraints you weigh when building on a foundation model.
- Distinguish pre-training, fine-tuning, and instruction tuning in one line each.
What it tests. Designing applications on top of foundation models: model selection, context windows, cost and latency, and guardrails, plus the big architectural choice of retrieval-augmented generation versus fine-tuning. It also covers the training and fine-tuning process at a conceptual level and the methods used to evaluate model performance.
How to study it. This is the largest domain, so spend the most time here. Make the retrieval-versus-fine-tuning decision automatic: a source that changes often wants retrieval and a re-index, not a fresh fine-tuning run; fine-tuning bakes knowledge and behaviour into the weights. Learn evaluation as a theme that recurs across the exam: a single metric rarely tells the truth, and the evaluation has to reflect the real task.
Easy to confuse
- Retrieval-augmented generation versus fine-tuning. Retrieval grounds the model in current documents at query time without changing the weights; fine-tuning bakes knowledge and behaviour into the weights through further training. A frequently changing source points to retrieval and a re-index, which is the exam's most common tell on this domain.
- Context window versus fine-tuning. The context window is how much text the model can read in a single request; fine-tuning changes the model itself. Putting more reference text in the prompt uses the context window and needs no training, so the exam tests whether you reach for retraining when supplying context would do.
- Pre-training versus fine-tuning. Pre-training builds a general model from a large unlabelled corpus at great cost; fine-tuning adapts that model to a narrower task with a much smaller dataset. The exam contrasts the scale and cost of the two and asks which step a given scenario actually needs.
Worked example from the AIF-C01 bank
lock_openFree sampleApplications of foundation modelsmedium
A developer is building a customer-support chatbot that must answer questions about a company's product catalogue, which is updated weekly. The catalogue contains roughly 200,000 tokens of text. Which design consideration most directly determines whether a single foundation model call can process the entire catalogue without retrieval augmentation?
- AThe model's training data cutoff date, because any catalogue content newer than the cutoff will be ignored by the model during inference.
- BThe model's output token limit, because generating long catalogue summaries requires a high output ceiling to avoid truncation.
- CThe model's context window size, because the full catalogue must fit within the input limit for a single inference call.check_circle Correct
- DThe model's temperature setting, because a lower temperature ensures the model attends to all catalogue entries rather than sampling selectively.
Understand that a model's context window size is the primary architectural constraint when deciding whether retrieval augmentation is needed. Foundation models have a fixed context window - the maximum number of tokens they can process in a single inference call, covering both input and output. When source material exceeds this window, the application must use retrieval-augmented generation (RAG) or chunking strategies to surface relevant content before the model call. Training cutoff, temperature, and output limits are real model properties but do not govern input capacity.
Why A is wrong: Training cutoff affects factual knowledge baked into weights, not the ability to process input tokens supplied at inference time. Providing the catalogue as context bypasses the cutoff problem entirely, so this is not the limiting factor here.
Why B is wrong: Output token limits affect response length, not input capacity. The question asks about processing the catalogue as input, so output limits are irrelevant to this constraint.
Why C is correct: Context window defines the maximum tokens a model can accept per call. If the catalogue exceeds that limit, the model cannot process it in one shot, making retrieval augmentation necessary.
Why D is wrong: Temperature controls output randomness and has no bearing on how many input tokens the model can accept. It is a tempting distractor because 'attending to all entries' sounds plausible, but attention and context length are separate concepts.
What you must be able to do. Match a prompting technique to the task it helps, and evaluate prompt quality against a rubric rather than trusting one happy example.
In one sentenceChoosing and improving prompts: zero-shot, few-shot, and chain-of-thought, and treating prompt quality as something you measure against a rubric, not eyeball once.
Recall check: answer these from memory first
- Say when you would reach for few-shot prompting and when for chain-of-thought.
- Explain why a prompt that works on one example still needs an evaluation set before you trust it.
- Name the three levers in a prompt, beyond the question itself, that change the result.
What it tests. Choosing and improving prompts: zero-shot, few-shot, and chain-of-thought prompting, how instructions, examples, and output formats change results, and how to evaluate prompt quality against a rubric and iterate. The theme is that a prompt which works on one example can fail on the next, so evaluation sets matter.
How to study it. Match each technique to the kind of task it helps: few-shot for format and tone, chain-of-thought for multi-step reasoning. The smaller weight makes this a good domain to secure with focused practice rather than long study. Treat prompt evaluation as the same discipline as model evaluation: measure against a rubric, compare variants, do not trust a single happy example.
Easy to confuse
- Zero-shot versus few-shot prompting. Zero-shot gives the instruction with no worked examples; few-shot includes a handful of input-output examples to fix the format and tone. If the scenario needs a specific output shape or style, the examples in few-shot are the lever the exam is pointing at.
- Few-shot versus chain-of-thought prompting. Few-shot shows examples of the answer; chain-of-thought asks the model to reason through intermediate steps before answering. Use few-shot for format and tone, chain-of-thought for multi-step reasoning, which is the distinction the exam tests.
Worked example from the AIF-C01 bank
lock_openFree samplePrompt engineering and evaluationmedium
A team is building an evaluation set to judge whether changes to a customer-support summarisation prompt genuinely improve quality rather than just helping on a few familiar cases. Which two characteristics make the evaluation set a trustworthy basis for deciding that a prompt change will generalise to unseen production inputs? Select TWO.
- AIt contains a varied spread of input types and difficult edge cases that mirror the distribution of real production traffic.check_circle Correct
- BIt is held fixed across every prompt variant being compared so a change in score reflects the prompt and not the data.check_circle Correct
- CIt is small enough for a reviewer to re-run by hand within a minute after each wording tweak to the prompt.
- DIt is limited to the inputs that the current production prompt already summarises correctly and cleanly today.
- EIt is regenerated from the newest variant's strongest sample outputs immediately before each comparison run.
A trustworthy prompt evaluation set must be representative of real input variety and held fixed across variants so score differences reflect the prompt alone. A prompt that works on one example can fail on the next because the example understated the variety of real inputs. An evaluation set predicts generalisation only when it samples that variety, including hard edge cases, and stays constant across variants so any score movement is attributable to the prompt change rather than to a shifting or self-selected dataset.
Why A is correct: Representative, varied coverage is what lets a score on the set predict behaviour on unseen inputs rather than on a narrow slice.
Why B is correct: Holding the set constant isolates the prompt as the only variable, which is required for a fair like-for-like comparison.
Why C is wrong: Convenience and speed are practical but do not make the set representative, so a fast tiny set can still give a misleading signal.
Why D is wrong: Including only inputs the prompt already passes inflates the score and hides the failures iteration is meant to surface.
Why E is wrong: Rebuilding the set from a variant's own best outputs biases the test toward that variant and breaks comparability.
What you must be able to do. Recognise what documents and discloses a model responsibly, and apply the AI-specific and general security controls a governed AI system needs.
In one sentenceBuilding AI responsibly and securely: bias and fairness, model cards, explainability, and the security controls including prompt-injection defence.
Recall check: answer these from memory first
- List the four things a model card discloses about a model.
- Name the AI-specific security threat the exam expects you to defend against, and how.
- Explain why explainability matters for trust and compliance, not just for debugging.
What it tests. Building AI responsibly and securely: bias, fairness, and safety, the artefacts that document a model such as model cards, why transparency and explainability matter for trust and compliance, and the security and governance obligations on AI systems, including least-privilege access, data protection, and prompt-injection defence.
How to study it. Learn what a model card discloses and why, because it ties together intended use, limitations, training data, and evaluation. Keep the security items concrete: prompt injection is the AI-specific threat to be ready for, and least privilege and data protection are the general controls that still apply. This is the smallest domain but the questions are usually straightforward marks once the concepts are clear.
Easy to confuse
- Bias versus variance. In responsible-AI terms the exam means bias as unfair or skewed outcomes against a group, not the bias-variance trade-off of model error. Read the context: a fairness scenario is about discrimination and representativeness, not about underfitting.
- Interpretability versus post-hoc explainability. An interpretable model is transparent by design so you can read its logic directly; post-hoc explainability applies a separate method to explain an opaque model after the fact. The exam asks which approach a given trust or compliance need calls for.
- Prompt injection versus a general security control. Prompt injection is the AI-specific attack where crafted input hijacks the model's instructions; least privilege and data protection are the general controls that apply to any system. The exam wants the AI-specific threat named as such, not folded into ordinary access control.
Worked example from the AIF-C01 bank
lock_openFree sampleResponsible and secure AIeasy
A data science team notices that a credit-scoring model approves loans at a significantly lower rate for one demographic group compared to all others, even though that group has similar repayment histories. Which practice most directly addresses this disparity during model development?
- AEvaluating disaggregated fairness metrics across demographic subgroups and retraining with corrected data or constraintscheck_circle Correct
- BIncreasing the model's overall accuracy on the full training dataset
- CDeploying the model behind an API so that end users cannot inspect its decision logic
- DRemoving all demographic attributes from the training features before the initial model fit
Understand how disaggregated fairness evaluation and targeted mitigation correct demographic disparity in AI model outputs. Responsible AI development requires measuring model behaviour separately for each subgroup rather than relying on aggregate metrics. Disaggregated fairness metrics (such as equalised odds or demographic parity difference) reveal where a model fails specific groups, enabling corrective actions such as data augmentation, re-weighting, or in-training fairness constraints that reduce the disparity directly.
Why A is correct: Disaggregated evaluation exposes differential performance per subgroup, allowing targeted mitigation such as re-sampling, re-weighting, or fairness constraints during retraining.
Why B is wrong: Tempting because higher accuracy sounds like a better model, but aggregate accuracy can improve while bias against a subgroup worsens if that group is a small fraction of the dataset.
Why C is wrong: Hiding decision logic via an API does nothing to correct the underlying bias; it only reduces transparency, which compounds responsible AI concerns.
Why D is wrong: Removing protected attributes is a common first step but does not eliminate bias because correlated proxy features (postcode, occupation) can still encode demographic information.