CPMAI Phase II: Data Understanding

Introduction
CPMAI (Cognitive Project Management for AI) is PMI's vendor-neutral methodology for running AI, machine learning and advanced analytics projects. It is iterative and data-centric, and it builds on CRISP-DM. It has six phases:
1. Business Understanding
2. Data Understanding
3. Data Preparation
4. Model Development
5. Model Evaluation
6. Model Operationalization

Phase II, Data Understanding, sits within the domain of Identifying Data Needs. It is where the team works out whether the data needed to solve the business problem exists, can be accessed, is good enough, and is sufficient.

Why Phase II Is Important
AI projects fail far more often because of data problems than because of algorithm choice. CPMAI rests on the principle that AI projects are data projects. Phase II matters for several reasons:

• It prevents wasted investment. Finding missing, biased or inaccessible data early avoids costly rework, or abandonment, in the modeling phases.
• It validates feasibility. Phase I defines the problem and the AI pattern. Phase II confirms whether data can actually support that solution.
• It supports go/no-go decisions. CPMAI stresses asking at each phase whether the project should continue. Data Understanding is a key checkpoint.
• It reduces risk. It surfaces issues with quality, privacy, compliance, licensing, bias and ethics before they become blockers.
• It shapes later phases. What you learn here drives the scope of Data Preparation and the modeling approach.

What Phase II Is
Data Understanding is the phase in which the project team:

• Identifies the data sources needed for the AI solution.
• Collects or gains access to initial data.
• Describes, explores and assesses that data.
• Decides whether the data is adequate in quantity, quality and relevance for the chosen AI pattern and success criteria.

Its central questions:

• What data do we need?
• Where is it, and who owns it?
• Can we legally and ethically use it?
• Do we have enough of it, and is it representative?
• What is its quality?
• Is the data labeled, if supervised learning is needed?
• How will we keep getting this data in production?

How Phase II Works: Key Activities

1. Identify data requirements
Translate the business problem and AI pattern from Phase I into specific data needs. Each of the seven AI patterns has distinct data needs:

• Recognition
• Conversational and human interaction
• Predictive analytics and decision support
• Patterns and anomalies
• Hyperpersonalization
• Autonomous systems
• Goal-driven systems

For example, image recognition needs labeled images. Predictive analytics needs historical outcome data.

2. Identify and inventory data sources
Sources may be internal (databases, data warehouses, data lakes, CRM, ERP, logs), external (third-party providers, public datasets, APIs) or newly generated (sensors, surveys, synthetic data). Record who owns each source, along with location, format and access method.

3. Assess data access and governance
Determine whether the team can actually obtain the data. Consider:

• Permissions
• Data ownership
• Licensing
• Privacy regulations such as GDPR, HIPAA and CCPA
• Security constraints
• Consent
• Contractual limits

4. Evaluate data quantity
Is there enough data to train, validate and test a model? Deep learning typically needs large volumes. Check whether there are enough examples of rare classes, such as fraud cases.

5. Evaluate data quality
Check these dimensions:

• Completeness
• Accuracy
• Consistency
• Timeliness and currency
• Validity
• Uniqueness (no duplicates)

Also look for missing values, outliers, noise and errors.

6. Understand the data's characteristics (the V's of Big Data)

• Volume
• Velocity
• Variety (structured, semi-structured, unstructured)
• Veracity (trustworthiness)
• Value

7. Explore the data
Use descriptive statistics, visualization and profiling. These reveal distributions, correlations, patterns and initial insights.

8. Assess labeling needs
Supervised learning requires labeled data. Decide whether labels exist, and if not, how to create them: in-house annotators, crowdsourcing, third-party labeling services or weak supervision. Factor in labeling cost, time and quality.

9. Identify bias and representativeness
Check whether the data reflects the real-world population the model will serve. Look for sampling bias, historical bias and underrepresented groups. These feed Trustworthy and Responsible AI considerations.

10. Plan for ongoing data (production data pipeline)
Confirm that the same data will be available, in the same form, when the model is operationalized. Training data that cannot be reproduced in production is a major risk. Watch for future data drift.

11. Document findings and make a go/no-go decision
Produce data understanding artifacts:

• Data inventory
• Data quality report
• Data exploration report
• Gaps and risks

Then decide whether to proceed, loop back to Phase I to rescope the problem, or stop the project.

Key CPMAI Principles Tied to Phase II

• Iteration: CPMAI is not waterfall. Findings may send you back to Business Understanding.
• Think big, start small, iterate often: Aim for a minimum viable data set, and scope a small first iteration that can deliver value quickly.
• Data-first mindset: Data drives the solution. Do not choose the algorithm first.
• Cross-functional collaboration: Data engineers, data scientists, subject matter experts, data owners, legal/compliance and business stakeholders all take part.

Common Phase II Outputs

• Data requirements list
• Data source inventory and access plan
• Data quality assessment
• Data description and exploration report
• Labeling strategy
• Data risk register (privacy, bias, availability)
• Go/no-go recommendation

Common Pitfalls Phase II Avoids

• Assuming data exists because it 'should'.
• Ignoring data ownership and silos.
• Training on data unavailable in production.
• Underestimating labeling effort.
• Overlooking privacy and consent.
• Missing bias in historical data.
• Jumping into modeling too early.

Exam Tips: Answering Questions on CPMAI Phase II: Data Understanding

1. Know the phase order cold. Data Understanding comes after Business Understanding and before Data Preparation. If a question asks what to do before cleaning or transforming data, the answer is usually a Data Understanding activity: identifying, assessing or exploring.

2. Separate understanding from preparation. Phase II assesses and describes data. Phase III cleans, transforms, augments, labels and engineers it. Questions often test this line. Identifying missing values is Phase II. Imputing them is Phase III. Determining labeling needs is Phase II. Executing the labeling is usually Phase III.

3. Data issues trigger a go/no-go or an iteration. If the scenario says data is insufficient, inaccessible or poor quality, the best answer is usually one of these:

• Revisit Business Understanding.
• Rescope the problem.
• Find alternative sources.
• Make a go/no-go decision.

Do not push ahead with modeling.

4. Choose the 'data-first' answer. Some options jump to algorithms, tools or model selection before data is understood. Those are usually wrong.

5. Watch for production data availability. One correct answer often involves confirming that training data will also be available in the operational environment.

6. Remember governance, privacy and ethics. If personal or sensitive data is involved, the correct answer often involves legal or compliance review, consent checks or data governance. Do this before using the data.

7. Link the AI pattern to the data need. Expect questions such as 'Which data is most critical for a recognition pattern?' (labeled examples) or 'for predictive analytics?' (historical data with outcomes).

8. Quantity and quality both matter. Large volumes of bad data are not sufficient. Neither is high-quality data that is too small or unrepresentative.

9. Spot the bias trap. If a dataset underrepresents a population, the right response is to flag and address representativeness. Ignoring it is wrong, and so is relying only on model tuning later.

10. Favor collaboration answers. Engaging data owners, subject matter experts and data engineers to understand sources and meaning is typically correct.

11. Think minimum viable. CPMAI encourages starting with a small, achievable data scope. 'Collect all possible data first' is rarely the best answer.

12. Use elimination. Remove options that:

• Skip phases.
• Ignore stakeholders.
• Assume data is ready.
• Violate privacy.
• Focus on technology over business value.

Sample Question Patterns

Q: A team has defined a fraud detection use case. They discover that only 0.1% of historical transactions are labeled as fraud. What should the project manager do next?
A: Assess whether this is enough representative data. Consider sourcing more examples or plan techniques for class imbalance. If it is infeasible, revisit the scope. Do not jump straight into training.

Q: Which activity belongs to Data Understanding rather than Data Preparation?
A: Profiling data sources to assess completeness and quality. Cleaning, normalizing and feature engineering belong to Phase III.

Q: The required customer data sits in a department that has not granted access. What is the best action?
A: Engage the data owner and governance stakeholders to secure access and resolve compliance issues before proceeding.

Summary
CPMAI Phase II, Data Understanding, establishes whether the data needed for an AI solution exists and is accessible, sufficient, high quality, representative, compliant and sustainable in production. It turns business needs into data requirements, inventories and assesses sources, uncovers risks, and supports an informed go/no-go decision. For the exam, think data-first, respect the phase order, keep assessment separate from preparation, prioritize governance and ethics, and choose iteration over pushing forward when data falls short.