Data Availability and Quality for Feasibility

Overview
In the PMI Cognitive Project Management for AI (PMI-CPMAI) methodology, Data Availability and Quality for Feasibility is a core part of Phase I: Business Understanding (Identify Business Needs and Solutions). Before an organization commits money, people and time to an AI initiative, it must confirm that the data the solution needs actually exists, can be accessed, and is good enough to train, test and operate the model. CPMAI is a data-centric methodology. Its central idea is that AI projects are data projects first. A brilliant business case cannot succeed if the data cannot support it.

Why It Is Important
1. Most AI projects fail because of data, not algorithms. Industry research and the CPMAI material repeatedly point to poor, missing or inaccessible data as a leading cause of AI project failure. Checking data feasibility early prevents expensive failures later.
2. It drives the go/no-go decision. In CPMAI, feasibility is judged on three fronts: business feasibility, data feasibility and implementation/technology feasibility. If the data is not feasible, the project should be re-scoped, postponed or cancelled. It should not push ahead on hope.
3. It reduces risk and rework. CPMAI is iterative. Finding data gaps in Phase I is far cheaper than discovering them in Phase III (Data Preparation) or Phase IV (Model Development).
4. It protects ROI. Data acquisition, labeling and cleansing can cost more than model building. Feasibility analysis shows these costs early, so the business case reflects reality.
5. It supports trustworthy AI. Biased, unrepresentative or non-compliant data leads to unfair or illegal outcomes. Early review of data rights, privacy and representativeness supports responsible AI.

What It Is
Data availability and quality for feasibility is the structured assessment of whether the data needed for the proposed AI pattern exists and is usable. It answers questions such as:
- Does the data exist? Is it internal, external, purchasable, or does it need to be generated or collected?
- Can we access it? Consider ownership, permissions, licensing, contracts, silos, APIs, and legal or regulatory limits such as GDPR, HIPAA and consent.
- Is there enough of it? Is the volume sufficient for the chosen approach? Deep learning typically needs far more data than simple rules or classical ML.
- Is it the right data? Is it relevant to the business problem, and does it contain the signal needed to predict or classify the target?
- Is it labeled? Supervised learning needs ground-truth labels. If labels are missing, labeling cost, effort and expertise must be estimated.
- Is it of acceptable quality? Check accuracy, completeness, consistency, timeliness, validity, uniqueness and representativeness.
- Will it remain available in production? The same data must flow at inference time, at the needed frequency and latency, not just for a one-time training set.

Key Dimensions (often summarized with the V's of Big Data)
- Volume: the quantity of data.
- Variety: structured, unstructured, semi-structured, and multiple sources and formats.
- Velocity: how fast data arrives and must be processed (batch vs. streaming).
- Veracity: trustworthiness, accuracy and noise.
- Value: whether the data actually contributes to the business objective.

Data quality dimensions to remember: accuracy, completeness, consistency, timeliness/currency, validity (conforms to rules and formats), uniqueness (no duplicates), and representativeness/bias.

How It Works (Step-by-Step in CPMAI Phase I)
Step 1: Clarify the business problem and AI pattern. Identify which of the seven patterns of AI applies: recognition, conversational/human interaction, predictive analytics and decision support, patterns and anomalies, personalization/hyperpersonalization, goal-driven systems, or autonomous systems. Each pattern implies different data needs. For example, recognition needs labeled images or audio, and personalization needs user behavior histories.
Step 2: Define data requirements. List the features, target variable, granularity, history length, refresh rate and format needed.
Step 3: Inventory data sources. Catalog internal systems (ERP, CRM, logs, sensors) and external sources (third-party, open data, synthetic data). Identify the data owners and stewards.
Step 4: Assess access and rights. Confirm legal rights, consent, privacy constraints, security classification, cross-border restrictions and licensing. Involve legal, compliance and data governance teams.
Step 5: Assess quantity and labeling. Estimate whether the volume is sufficient. If labels are lacking, plan human labeling, crowdsourcing, weak supervision, transfer learning or a different approach, such as unsupervised learning.
Step 6: Assess quality. Sample and profile the data for missing values, outliers, duplicates, inconsistent formats, staleness and bias. At this stage it is a high-level check, not full data preparation.
Step 7: Assess operational availability. Confirm the data will be available at inference time in production, with pipelines that are reliable and sustainable.
Step 8: Estimate effort and cost. Include acquisition, cleansing, labeling, integration and ongoing maintenance in the business case.
Step 9: Make the feasibility decision. The options are proceed, re-scope (a smaller problem, a different AI pattern, or a narrower minimum viable product), acquire or generate data first, or stop. Document the assumptions and risks.
Step 10: Iterate. CPMAI phases loop back. Findings in Phase II (Data Understanding) may send the team back to Phase I to adjust scope.

Relationship to Other CPMAI Phases
- Phase I (Business Understanding): high-level data feasibility, covering whether the data exists and is accessible and adequate.
- Phase II (Data Understanding): deeper exploration and profiling of the identified data.
- Phase III (Data Preparation): cleaning, transforming, labeling and augmenting.
- Phases IV-VI: model development, evaluation and operationalization, all of which depend on the earlier data work.

Common Remedies When Data Is Insufficient
- Narrow the scope to a smaller, well-defined problem: think big, start small, iterate often.
- Use a simpler AI or non-AI approach that needs less data.
- Acquire third-party data or form data partnerships.
- Start data collection now and delay model building.
- Use data augmentation, synthetic data or transfer learning and pretrained models.
- Invest in labeling through internal experts, vendors or active learning.
- Improve data governance and integration to break down silos.

Red Flags in Scenarios
- The team plans to start model building without confirming data access.
- Data is owned by another department or vendor with no agreement in place.
- Historical data does not reflect current conditions (concept drift or changed processes).
- Only a tiny labeled sample exists for a complex supervised task.
- The data contains personal information with no consent or legal basis.
- The training data is unrepresentative of the population the model will serve.
- The data available at training time will not be available at prediction time (data leakage or availability mismatch).

Exam Tips: Answering Questions on Data Availability and Quality for Feasibility
1. Think data first. When a scenario asks what the project manager should do first or next before committing to an AI solution, the answer is usually to assess data availability and quality, not to select an algorithm, hire data scientists or build a prototype.
2. Know the phase. High-level data feasibility belongs in Phase I (Business Understanding). Detailed profiling belongs in Phase II (Data Understanding). Cleansing and labeling belong in Phase III (Data Preparation). Pick the answer that matches the phase described.
3. Prefer re-scoping over forcing. If data is insufficient, the best answer is usually to re-scope, iterate or acquire data. Avoid answers that proceed anyway, lower quality standards, or assume the model will compensate for bad data.
4. Remember garbage in, garbage out. Answers claiming a more advanced algorithm will fix poor data quality are almost always wrong.
5. Access is not the same as existence. Data that exists but cannot legally or practically be accessed is not available. Watch for privacy, consent, ownership and licensing clues.
6. Check labels. For supervised learning scenarios, ask whether labeled data exists. Lack of labels is a feasibility issue with cost and time impacts.
7. Consider production availability. Correct answers often consider whether data will be available continuously in operation, not just for training.
8. Look for bias and representativeness. If a scenario mentions a model serving a diverse population trained on narrow data, the issue is data quality and representativeness, which is also a trustworthy AI concern.
9. Involve the right stakeholders. Data owners, data stewards, governance, legal and subject-matter experts are typical correct stakeholders for data feasibility questions.
10. Make go/no-go explicit. CPMAI expects a documented feasibility decision. Answers that include making a go/no-go decision based on data findings are strong.
11. Match the data to the AI pattern. If a question names a pattern, such as recognition, link it to the data type needed. For example, recognition needs large volumes of labeled unstructured data.
12. Eliminate extremes. Discard answers that ignore data entirely or that propose cancelling immediately when a reasonable re-scope exists. CPMAI favors iterative, small, achievable steps.
13. Watch for keywords. Terms like sufficient, accessible, representative, labeled, timely, consent, silo and ownership signal a data feasibility question.

Quick Summary
Data Availability and Quality for Feasibility asks whether we have, can get, and can trust the data needed for this AI solution, both now and in production. In CPMAI it is a gatekeeping activity in Phase I that shapes scope, cost, risk and the go/no-go decision. On the exam, prioritize early data assessment, legal and ethical access, adequate volume and labels, and quality dimensions. Favor re-scoping and iteration over pushing forward with inadequate data.