PMI CPMAI Practice Questions with Explanations

Correct answer: B

Model versioning is the practice of keeping every trained model as an immutable, retrievable artifact rather than overwriting it. CPMAI Phase VI treats a model as a released asset: when a new version underperforms for some consumers, the previous version must still be available so the team can roll back while they investigate. Overwriting the original destroyed the only copy of a working model, which is why the request cannot be satisfied.

Why the other options are wrong:

• A. Data governance concerns how data is classified, secured and accessed. No data was mishandled here - a model artifact was destroyed.
• C. A retraining pipeline governs when a model is refreshed. The team did retrain; the failure was not retaining the earlier result.
• D. The team did iterate - they produced a second model. Iterating is correct CPMAI practice, so this describes what they did right, not the error.

Correct answer: D

Phase III of CPMAI is Data Preparation. This is where raw data collected in Phase II is cleaned, de-duplicated, corrected, standardized and enriched, and where feature engineering and exploratory analysis take place. Enhancing and cleaning are the defining activities of this phase, and they must be complete before Phase IV Data Modeling can begin, because a model can only be as good as the data prepared for it.

Why the other options are wrong:

• A. Phase VI is Model Operationalization - deploying, monitoring and versioning a finished model.
• B. Phase I is Business Understanding - defining the problem, the ROI and the AI Go/No Go decision.
• C. Phase V is Model Evaluation - measuring whether the model meets the business goal set in Phase I.
• E. Phase II is Data Understanding - identifying what data exists, where it lives and whether it is sufficient.
• F. Phase IV is Data Modeling - selecting algorithms and training the model on data already prepared.

Correct answer: D

Operationalization has to match the environment the model actually runs in. The stated constraint is that the device is frequently offline, so any deployment that depends on a network round trip will fail exactly when the user needs it. Placing the model on the edge device itself means inference happens locally and the application keeps working regardless of connectivity, which is the standard CPMAI answer for intermittently connected edge deployments.

Why the other options are wrong:

• A. Whether the solution is generative is unrelated to the problem; the constraint is connectivity, not model type.
• B. A hybrid environment still assumes some connectivity for the cloud half, so the offline case is left unsolved.
• C. A cloud-based API is precisely the wrong choice here - it is unreachable whenever the device has no internet.

Correct answer: D

CPMAI is data-centric, but its point is that value comes from an organization's ability to manage data, not from the raw volume of it. Data only becomes useful when people with the right experience, tooling and understanding can collect, clean, govern and engineer it. Capability is the constraint that determines whether a data asset can be turned into a working model, which is why staffing for that capability is the most important consideration.

Why the other options are wrong:

• A. Collect-now-organize-later produces large, ungoverned data swamps that cost more to untangle than they return.
• B. More data is not automatically better; unlabeled, irrelevant or poor-quality data adds cost without adding signal.
• C. Algorithm selection is a Phase IV concern and is secondary to whether the data can be prepared at all.

Correct answer: D

The Predictive Analytics and Decision Support pattern covers exactly this description: learning from past behavior and existing data to forecast future outcomes or to help a human make a better decision. It is one of the seven patterns of AI, and it is distinguished by its forward-looking, human-supporting purpose rather than by the specific algorithm used.

Why the other options are wrong:

• A. Goal Driven Systems learn optimal actions by trial and reward, as in reinforcement learning and game playing, not by projecting past behavior forward.
• B. Patterns and Anomalies is a separate pattern concerned with finding groupings and outliers in existing data; the description given is prediction and decision support alone.
• C. The Recognition pattern identifies and classifies objects in images, audio or text - it labels what something is, not what will happen next.

Correct answer: A

Regulatory exposure follows the data and the users, not the office address. A system built today will still be in operation when the expansion happens, and retrofitting privacy and AI compliance into a deployed model is far more expensive than designing for it. The defensible approach is to meet the laws that apply now in the US and to design against the requirements of the markets the organization plans to enter.

Why the other options are wrong:

• B. Scoping compliance to NYC alone ignores US federal and state law that already applies and guarantees rework on expansion.
• C. US-only compliance covers today but not the planned expansion, which is the specific detail the question supplies.
• D. Data protection and AI regulation apply regardless of company size; being small is not an exemption.

Correct answer: C

Transfer learning reuses a model already trained on one task as the starting point for a related task, so the general visual features learned while classifying cars - edges, shapes, textures - carry over to classifying animals. Only the later layers need retraining on the new labeled set, which dramatically reduces the data and compute needed compared with training from scratch.

Why the other options are wrong:

• A. Reinforcement learning trains an agent through reward signals from an environment; it is not a way to reuse a trained image classifier.
• B. Generative AI creates new content rather than shortening the training of a classifier.
• D. A pre-trained model is the input to transfer learning, not the technique itself - the question asks what the technique is called.

Correct answer: A

With an inexperienced team on a first AI project, the right choice is the simplest algorithm that fits the problem. Naive Bayes is a supervised classifier, is a long-established baseline for spam detection, trains quickly, needs little tuning and produces a result the team can interpret and explain. CPMAI favours starting simple and iterating, because an understandable baseline tells you whether the problem is tractable at all.

Why the other options are wrong:

• B. A neural network adds tuning, compute and opacity that a first-time team cannot yet manage, for a problem a simple classifier already solves.
• C. An ensemble compounds the complexity of several models and makes failures harder to diagnose.
• D. Gaussian mixture models are an unsupervised clustering method and do not fit a supervised binary classification task.

Correct answer: C

CPMAI Phase I is business-first: define the problem and the return you are trying to achieve, then let that determine which of the seven patterns of AI applies. Choosing a pattern or a technology first is technology-led thinking, and it produces projects that demonstrate a capability without delivering a business outcome. The objective drives the pattern, never the reverse.

Why the other options are wrong:

• A. Selecting the pattern first inverts CPMAI and risks fitting a business case to a chosen technology.
• B. This still starts from the pattern; involving stakeholders afterwards does not correct the order.
• D. ROI is validated with stakeholders, but a number handed down without a defined problem gives the team nothing to build against.

Correct answer: C

Training, validation and test sets must each be representative of the same underlying distribution, and random subsampling is what achieves that. Any ordered or patterned split risks concentrating a class, a time period or a source in one subset, which makes evaluation results meaningless because the test set no longer resembles what the model will see in production.

Why the other options are wrong:

• A. Alphabetical order is a systematic bias - it can group related records and skew the class balance of each subset.
• B. Patterned subsampling imposes structure on the split and can align with structure already in the data, reintroducing bias.
• D. Reusing the same data for all three sets means the model is evaluated on what it memorized, so the scores are worthless.

Correct answer: B

AutoML is well suited to the mechanical parts of model building - trying candidate algorithms, tuning hyperparameters and scoring the results - because those are search problems a machine performs faster than a person. What it cannot take over is the iterative modeling and evaluation loop itself, which is where humans judge whether the model actually answers the business question defined in Phase I, whether its errors are acceptable and what to change next.

Why the other options are wrong:

• A. Automatic model assessment computes metrics, a mechanical task AutoML handles well.
• C. Hyperparameter tuning is a systematic search and is one of AutoML's strongest use cases.
• D. Automatic model selection is a standard, appropriate AutoML function.
• E. Algorithm selection is exactly the kind of comparison AutoML is built to automate.

Correct answer: A

When data problems threaten a project, CPMAI's guidance is to narrow the iteration, not widen it. Expanding scope adds new data requirements on top of requirements the team already cannot satisfy, so it multiplies the problem and pushes value further away. The workable responses are to reduce scope, use data already on hand, or acquire the missing data.

Why the other options are wrong:

• B. Checking whether existing data is sufficient is a sound way to keep the iteration moving.
• C. Adjusting the scope of the iteration is the standard CPMAI response to a data shortfall.
• D. Purchasing data is a legitimate way to close a specific gap when the budget allows.

Correct answer: B

Overfitting is when a model learns the noise and idiosyncrasies of the training data rather than the underlying pattern. It shows exactly as described: high sensitivity to small fluctuations in the data and poor generalization to examples it has not seen. The usual remedies are more or more varied training data, fewer features, regularization, or a simpler model.

Why the other options are wrong:

• A. Underfitting is the opposite - the model is too simple to capture the pattern and performs poorly on training data as well.
• C. A wrong algorithm can contribute, but the symptom described names a specific, well- defined condition.
• D. Wrong data would produce irrelevant predictions rather than oversensitivity to fluctuation.

Correct answer: B

Neural networks are the most expensive option on the list against every constraint named. They require large volumes of labeled data, long training runs on specialized hardware, and considerable expertise to design and tune, and their behaviour is harder to explain afterwards. When training time, cost and complexity are the governing factors, a neural network is the wrong choice.

Why the other options are wrong:

• A. Support vector machines train efficiently on moderate data sets and need comparatively little tuning.
• C. Naive Bayes is among the cheapest and fastest algorithms available.
• D. Gaussian mixture models are computationally modest compared with deep networks.

Correct answer: D

A model that tests well is not finished. CPMAI Phase V requires continuous evaluation, retraining and operationalization, because performance decays as real-world data drifts away from the training distribution. Treating the test result as a finish line and moving on leaves the deployed model unmonitored, with no plan for the retraining it will need.

Why the other options are wrong:

• A. The AI Go/No Go questions belong to Phase I, before the project begins; the gap here is at the evaluation stage.
• B. There is a real problem: the workflow ends evaluation instead of making it continuous.
• C. Management sign-off is an organizational policy, not the methodological failure the question describes.

Correct answer: A

Unattended bots run on their own schedule without a person present, which is what back- office work such as bulk data entry requires - the process is high volume, rules-based and does not need a human at the keyboard. Attended bots, by contrast, are triggered by and assist a user during their own work.

Why the other options are wrong:

• B. Business process outsourcing hands the work to an external provider; it is not a type of bot.
• C. An attended bot needs a human working alongside it, which defeats the purpose of automating unattended back-office processing.
• D. Rules-based, repetitive back-office data entry is the textbook case for RPA.

Correct answer: C

Extracting competitive value from big data is a leadership problem before it is a technical one. If senior leadership is data literate - understanding the volume, velocity, variety and veracity characteristics of the organization's data, how those connect to their own function, and what value can be drawn out - then investment, priorities and governance line up behind the data strategy. Without that, technical capability alone has no mandate to act on.

Why the other options are wrong:

• A. Technical depth is necessary but cannot set organizational priorities or unlock investment on its own.
• B. Understanding held only at the top does not translate into practice unless it reaches the functions that own the data.
• D. Broad team awareness is valuable, but the question asks what keeps the organization competitive, and that requires literate senior leadership.

Correct answer: A

The goldilocks approach means gathering the amount of data that is right for the specific problem - enough to represent the pattern the model must learn, without collecting volumes the team cannot clean, label, store or govern. CPMAI ties data volume to the requirements identified in Phase II rather than to an absolute rule.

Why the other options are wrong:

• B. Deliberately minimizing data risks a training set too small or too narrow to represent the problem.
• C. More is better ignores the cost of preparing and governing data, and adds noise as readily as signal.
• D. There is a correct approach - matching data volume to the identified requirements.

Correct answer: C

The AI Go/No Go decision requires a yes to every business feasibility question before the project proceeds, and a clear problem definition is the most fundamental of them. Without it there is no way to judge the data requirements, choose a pattern or measure success. The correct action is to hold the project at Phase I until the problem is defined - not to abandon it, since the definition can be produced.

Why the other options are wrong:

• A. Cancelling outright discards a project whose only defect is a missing definition that the team can still write.
• B. The Go/No Go chart is not advisory; proceeding with an unanswered feasibility question is what it exists to prevent.
• D. A problem definition that emerges later means data and modeling decisions were made with no target to aim at.

Correct answer: A

Continuous monitoring with an instant response requires real-time prediction, where each observation is scored the moment it arrives and the result is returned immediately. This is the operationalization mode built for low-latency, always-on inference such as machinery condition monitoring, where a delayed prediction has no value.

Why the other options are wrong:

• B. A web service is a delivery mechanism that can serve either real-time or batch requests; it does not by itself describe the prediction approach.
• C. Batch prediction scores accumulated data on a schedule, which cannot meet a continuous, instant requirement.
• D. Stream learning describes updating the model as data arrives, which is about training rather than the prediction approach required here.

Correct answer: C

Augmented intelligence is the use of AI to improve and accelerate human work while the human keeps the decision and the output. Taking a generative model's results and using them to do your own job faster is precisely that arrangement - the system amplifies the person rather than replacing them.

Why the other options are wrong:

• A. Hyperpersonalization tailors content or offers to an individual customer, which is a different use of generative models.
• B. An autonomous system removes the human from the loop; here the human is the one consuming and applying the output.
• D. Programmatic automation executes a defined process without judgement, whereas this describes assisting a person's own work.

Correct answer: C

Responsible AI is the dimension concerned with whether a system should be built and used at all - its purpose, its potential for misuse, and the guidelines that keep it aimed at a beneficial outcome. Setting rules for proper use of facial recognition and insisting the application serve a positive purpose is a responsibility question, not a technical disclosure one.

Why the other options are wrong:

• A. Transparent AI is about disclosing that AI is in use and how it operates, not about governing the purpose it serves.
• B. Governed AI concerns the policies, controls and oversight structures applied to a system once it exists.
• D. Explainable AI is about making individual model decisions interpretable, which is separate from whether the use itself is proper.

Correct answer: C

A spam filter answers a two-way question: is this message spam or not? That makes it a binary, or binomial, classification problem, where a supervised model is trained on labeled examples of both classes and assigns each new email to one of exactly two categories.

Why the other options are wrong:

• A. Clustering is unsupervised and groups similar messages without knowing which group is spam, so it produces no decision.
• B. Regression predicts a continuous numeric value, not a yes-or-no category.
• D. Multiclass classification is for three or more categories; spam detection has only two outcomes.

Correct answer: F

Phase V, Model Evaluation, is where the model is judged against the business objective set in Phase I, not merely against technical accuracy metrics. A model can score well and still fail to deliver the intended outcome, so this phase deliberately checks both: does it perform, and does it accomplish what the business asked for?

Why the other options are wrong:

• A. Phase IV is Data Modeling - building and training the model, before it is evaluated against the business goal.
• B. Phase I defines the business objective; it is where the target is set, not where it is measured.
• C. Phase II is Data Understanding, which precedes any model existing.
• D. Phase VI is Model Operationalization, which follows a successful evaluation.
• E. Phase III is Data Preparation, again before a model exists to evaluate.

Correct answer: B

Data labeling sits in Phase II, Data Understanding, where the team establishes what data exists, whether it is sufficient, and whether it carries the labels a supervised approach requires. Identifying that the plant images need labels, and arranging for that labeling, is part of understanding whether the data can support the project at all.

Why the other options are wrong:

• A. Phase I is Business Understanding and concerns the problem and its ROI, not the state of the data.
• C. Phase III prepares data that has already been understood and acquired - cleaning, transforming and engineering features.
• D. Phase IV trains models on data that is already labeled and prepared.
• E. Phase V evaluates a trained model.
• F. Phase VI operationalizes and monitors the model in production.

Correct answer: D

Outpainting is the generative technique that extends an image beyond its original borders, generating new content that continues the scene outward while remaining consistent with what is already there. It is the direct answer to producing content past the current image size.

Why the other options are wrong:

• A. Prompt engineering for new image generation creates a fresh image rather than extending the existing one.
• B. Inpainting fills or replaces regions inside the existing frame; it does not enlarge it.
• C. Super-resolution increases pixel density within the same boundaries, adding detail rather than area.

Correct answer: C

Transparent AI covers the visibility a person has into an AI-driven decision that affects them - being told that AI was used, what it considered, and having a route to question or contest the outcome. Giving applicants the ability to challenge a hiring decision is an obligation of the transparency layer.

Why the other options are wrong:

• A. Responsible AI concerns whether the system should be built and used for its stated purpose.
• B. Ethical AI addresses the moral principles behind the system rather than the individual's recourse against a decision.
• D. Explainable AI makes the internal reasoning of a model interpretable, which supports contesting but is not the same as providing the right to contest.
• E. Governed AI is the policy and oversight structure applied to the system as a whole.

Correct answer: B

Automation repeats defined tasks by machine and needs no learning to work; autonomous and AI systems handle situations that cannot be fully specified in rules. Applying AI methodology to a purely rules-based automation project adds cost and complexity for no gain. The conflict is avoided when the whole team can tell the two apart and scope the project accordingly.

Why the other options are wrong:

• A. Treating automation as if it were an autonomous project is what created the overkill in the first place.
• C. Escalating to senior management addresses the symptom without correcting the shared misunderstanding that caused it.
• D. A hybrid approach still applies AI practices to work that does not need them.

Correct answer: D

Synthetic training data is generated rather than collected, which makes it the right answer when the real-world data does not exist in sufficient volume. For pose estimation it is particularly suitable: poses can be generated systematically across the full range needed, they arrive already labeled, and no personal data is captured, avoiding the privacy exposure of using real people.

Why the other options are wrong:

• A. Posing employees is slow, produces a small and demographically narrow set, and raises consent issues.
• B. Data from other departments will not contain the specific pose imagery the project lacks.
• C. Surveillance footage carries serious privacy and consent problems and still would not cover the required range of poses.

Correct answer: B

Phase II starts with a full inventory of the data the organization already holds, then selects what is relevant to the defined problem. Taking inventory first means the team knows what it has before it spends on acquisition, and filtering for relevance keeps irrelevant data out of a pipeline that would otherwise have to clean and govern it.

Why the other options are wrong:

• A. Buying external data before knowing what is already available spends budget on data the company may already own.
• C. Using all existing data without filtering pulls irrelevant records into preparation and modeling.
• D. Scoping the inventory to the team's own data misses relevant data held elsewhere in the company.

Correct answer: D

AI earns its place when the problem is probabilistic - when the answer depends on patterns and likelihoods that cannot be written down as rules in advance. Deterministic problems have a known rule for every input and are better and more cheaply served by conventional code or automation. Testing whether the pattern is probabilistic is the practical rule of thumb for AI fit.

Why the other options are wrong:

• A. Statistics appear in almost every data problem, so this test does not separate AI- suitable work from ordinary analysis.
• B. This has the test backwards - deterministic problems are the ones that should not use AI.
• C. If simple rules work, the correct conclusion is to use those rules rather than to add an AI solution.

Correct answer: C

K-nearest neighbor is a supervised algorithm: it is trained on labeled examples and classifies a new observation by the labels of the points closest to it. That makes it a valid choice for a supervised classification project.

Why the other options are wrong:

• A. Gaussian mixture is an unsupervised clustering method and needs no labels.
• B. Q learning is a reinforcement learning algorithm that learns from rewards, not from labeled training data.
• D. K-means is unsupervised clustering - the similar name to K-nearest neighbor makes it a common trap, but it does not classify against labels.

Correct answer: A

Digital transformation stalls on people and process far more often than on technology. Having the right culture for AI - leadership that understands what AI can and cannot do, tolerance for iterative work, data literacy, and willingness to change how decisions are made - is what lets AI actually be adopted. Without it, the technology is bought and never used.

Why the other options are wrong:

• B. Buying technology without the culture to absorb it is exactly how the logjam formed.
• C. Pattern selection is important within a project, but it does not address the organizational blockage being described.
• D. Basic automation may help with individual tasks but leaves the cultural barrier untouched.

Correct answer: B

AI is something an organization does, not something it purchases, because value depends on the organization's own data, problems and iteration. The effective way to shift that belief in senior management is evidence rather than argument: describing a comparable project you have run, how it was approached and what result it produced, gives them a concrete alternative to buying a product.

Why the other options are wrong:

• A. Staying silent leaves a known misconception to drive the company's first AI project.
• C. Researching vendors reinforces the buy-it assumption instead of correcting it.
• D. Leading with data quality problems raises an obstacle without showing the successful approach that answers management's question.

Correct answer: A

The AI Go/No Go feasibility questions belong to Phase I, Business Understanding. This is the gate where the team confirms there is a defined problem, a business case, sufficient data prospects and organizational readiness before any data or modeling work begins.

Why the other options are wrong:

• B. Phase II is Data Understanding and follows a Go decision.
• C. Phase III is Data Preparation, well past the feasibility gate.
• D. Phase IV is Data Modeling.
• E. Phase V is Model Evaluation.
• F. Phase VI is Model Operationalization.

Correct answer: D

Identifying the pattern of AI is what tells the team which data to collect, which skills the team needs and which algorithms are appropriate. Candidate matching maps to a recognizable pattern, and settling that question is a required step before the project can be resourced properly. The project manager should determine the pattern, not skip it or infer it after the fact.

Why the other options are wrong:

• A. Proceeding without the pattern leaves data requirements and staffing undefined.
• B. Choosing an algorithm first and back-filling the pattern inverts the method and lets the tool define the problem.
• C. A straw poll substitutes opinion for analysis; the pattern follows from the problem, not from a vote.

Correct answer: B

When a model produces systematically different outcomes for one demographic while every other factor is equivalent, the cause is almost always bias in the training data, which the algorithm then reproduces and applies at scale. This is algorithmic discrimination: the model has learned a historical pattern of unequal treatment and is now enforcing it consistently.

Why the other options are wrong:

• A. Privacy and data sharing concern who may access data, not whether decisions are made unfairly.
• C. Hallucination is a failure mode of generative models producing invented content; this is a classification model producing consistent, measurable bias.
• D. A systematic pattern against one demographic is not ordinary model error, and treating it as acceptable is a legal and ethical exposure.

Correct answer: A

Descriptive analytics summarizes what has already happened - aggregating historical data, reporting prior performance metrics and surfacing patterns in the record. Summarizing and highlighting patterns in past purchasing data is precisely that, with no forecast or causal claim involved.

Why the other options are wrong:

• B. Predictive analytics forecasts future outcomes rather than describing past ones.
• C. Diagnostic analytics investigates why something happened, which goes beyond the summary requested.
• D. Projective analytics is not one of the standard analytics categories.

Correct answer: B

Retraining pipelines are the part of the data engineering stack built for model iteration. They automate the collection of fresh data, its preparation, the retraining of the model and the promotion of the new version, so that a model can be refreshed on a schedule or in response to drift rather than rebuilt by hand each time.

Why the other options are wrong:

• A. Data acquisition brings raw data in but does not by itself close the loop back to a new model version.
• C. Feature engineering derives model inputs and is one step within the pipeline, not the mechanism for iteration.
• D. An ELT pipeline moves and transforms data for storage and analysis; it is not specific to model iteration.

Correct answer: C

Getting users to label data works only where there is a large, engaged user base whose labeling is a natural by-product of using the product. A small startup has neither the volume of users nor the traffic to produce a usable labeled set that way, so it is the one option here that would not help.

Why the other options are wrong:

• A. Third-party labeling firms are a standard route and are accessible to small companies at modest volume.
• B. A contractor workforce gives direct control over labeling quality and scales to the size the startup needs.
• D. Finding data that is already labeled is the cheapest option of all when a suitable source exists.

Correct answer: A

The gap is a class the project has almost no examples of, and techniques that multiply or resample existing data cannot create examples of something that is barely represented. Buying labeled data of that vehicle type from a third party supplies genuinely new examples with the labels already applied, which is the fastest way to close a coverage gap of this kind.

Why the other options are wrong:

• B. Transformation and multiplication augment images the team already has, so a class with almost no examples stays under-represented.
• C. Sampling changes the balance of the existing set without adding any new information.
• D. Anonymization protects identity in data already held and adds no new examples.

Correct answer: D

Retraining pipelines are the part of the data engineering stack built for model iteration. They automate the collection of fresh data, its preparation, the retraining of the model and the promotion of the new version, so that a model can be refreshed on a schedule or in response to drift rather than rebuilt by hand each time.

Why the other options are wrong:

• A. Feature engineering derives model inputs and is one step within the pipeline, not the mechanism for iteration.
• B. Data acquisition brings raw data in but does not by itself close the loop back to a new model version.
• C. An ELT pipeline moves and transforms data for storage and analysis; it is not specific to model iteration.

Correct answer: B

CPMAI is iterative and does not require every data source to be secured before work begins. The team can proceed on what it has, delivering a first iteration while access to the remaining data is pursued, provided everyone accepts that the project may have to pause at a phase boundary until that access arrives. Proceeding cautiously with that expectation set is the balanced course.

Why the other options are wrong:

• A. Moving ahead and hoping treats the data gap as somebody else's problem and sets no expectation that a pause may be needed.
• C. Prioritizing the schedule over the data gap is how projects reach modeling with nothing to train on.
• D. Waiting for complete data before starting is a waterfall stance and forfeits everything the team could learn from a first iteration.

Correct answer: C

Hyperpersonalization depends on a profile that is unique to each individual and that keeps changing. To be useful the profile has to learn and adapt from ongoing behaviour, and it also has to be programmable so the organization can direct it at different purposes - marketing, service, recommendations - rather than a single fixed use. Both properties together are what makes the pattern work.

Why the other options are wrong:

• A. Learning and adapting alone leaves no way to target the profile at specific business purposes.
• B. Manual updates cannot keep pace with individual behaviour at scale, which defeats the point of the pattern.
• D. A profile per type of individual is segmentation, not personalization, and freezing it means it never reflects how the user actually changes.

Correct answer: B

The data preparation pipeline is designed in Phase II, Data Understanding, because it has to be built against the data requirements, sources and formats identified there. Phase II establishes what data exists, where it comes from and what condition it is in, and the pipeline is the mechanism defined to move it. Phase III is then where that pipeline is run and the data itself is cleaned and transformed.

Why the other options are wrong:

• A. Phase I is Business Understanding and settles the problem and ROI, before any data source is known.
• C. Phase III executes the preparation - it is where data is cleaned and enriched, not where the pipeline requirements are established.
• D. Phase IV trains models on data the pipeline has already delivered.
• E. Phase V evaluates a trained model.
• F. Phase VI operationalizes and monitors it.

Correct answer: B

A data preparation pipeline is shaped by both the pattern of AI and the specific needs of the model being built - the features required, the labels, the formats, the volumes and the transformations. Two projects rarely share all of those, so a pipeline lifted wholesale from a previous project will prepare the wrong data in the wrong shape.

Why the other options are wrong:

• A. Operationalization needs affect how a model is deployed, not how the training data is prepared.
• C. Sharing a pattern is not sufficient; two projects on the same pattern can still need entirely different features and data sources.
• D. Reuse without adaptation is precisely the problem - components may be reusable, but the pipeline as a whole is not.

Correct answer: C

The decision is not ideological. The lead should assess the data first and determine whether a pre-trained model genuinely fits the problem, and if it does, have the team research the market to find the model best suited to the project. That keeps the choice evidence-based rather than settling an argument by preference.

Why the other options are wrong:

• A. Ruling out pre-trained models outright discards a legitimate option that often saves substantial data and training effort.
• B. Restricting the search to models the company already owns arbitrarily narrows the field and may force a poor fit.
• D. Building the same thing twice in parallel doubles the cost and delays the iteration to settle a question that assessment can answer.

Correct answer: A

The diagnosis is that the project is being run as a software project when it is a data project, so the correction is to apply data-centric methodology practices - iterative phases driven by data understanding, preparation and evaluation rather than by feature delivery. That addresses the actual cause and can be done with the team already in place.

Why the other options are wrong:

• B. Agile practices are useful but were designed around software delivery; layering them on top does not supply the data-centric structure the project is missing.
• C. Replacing the team discards two months of domain knowledge to solve a methodology problem that retraining fixes.
• D. Outsourcing the technical work removes the team from the very practices they need to learn, and leaves the project dependent on a third party.

Correct answer: B

Data requirements are determined by the problem and the data itself - how much data is needed, how good it has to be, and where it will come from. The composition of the data team affects how the requirements are met and how quickly, but it does not change what the data requirements actually are.

Why the other options are wrong:

• A. Quality directly sets the requirement - poor source data means more must be collected or cleaned to reach a usable standard.
• C. Volume is a data requirement in its own right and drives storage, labeling and processing needs.
• D. Source and location determine format, availability and which privacy regimes apply, all of which shape the requirements.

Correct answer: A

Recommending products based on what similar customers bought is a nearest-neighbor problem: find the customers closest to this one in behaviour, then surface what they purchased. K-nearest neighbor does exactly that, using labeled purchase history to score similarity and produce a recommendation.

Why the other options are wrong:

• B. K-means groups customers into clusters without producing a per-customer recommendation, and the similar name makes it a common trap.
• C. A neural network could be trained for this but is heavier than the problem requires and is not the best choice given a direct similarity method.
• D. Hyperpersonalization is a pattern of AI, not an algorithm.