Comparing Available Data Against Requirements

Overview
In the PMI Cognitive Project Management for AI (CPMAI) methodology, the second phase, Data Understanding, builds directly on Business Understanding. Once the team has defined the AI pattern, the business problem and the success criteria, it must answer a critical question: Do we actually have the data needed to build this AI solution?

Comparing available data against requirements is the structured gap analysis that answers this question. It sets what the model needs (data requirements) side by side with what the organization has or can realistically get (available data). The result tells the team whether to proceed, adjust scope, acquire or generate data, or stop the project.

Why It Is Important
1. AI projects fail mostly because of data. CPMAI stresses that most AI project failures come from data problems, not algorithm problems. Insufficient, poor-quality, inaccessible or irrelevant data is the leading cause of failed or abandoned AI initiatives. Finding gaps early prevents costly late-stage failure.
2. It supports the Go/No-Go decision. CPMAI is iterative and includes explicit decision points. A data gap that cannot be closed may mean the project should not proceed, or should be re-scoped, before money is spent on modeling.
3. It keeps feasibility realistic. A great business case means nothing without data. The comparison grounds the AI ambition in reality.
4. It drives the work plan. Gaps show what data acquisition, labeling, cleansing, integration, privacy review or synthetic data work is needed. Each of these affects schedule, budget and resources.
5. It manages risk and compliance. Some data exists but cannot be used for legal, ethical, privacy, consent or contractual reasons. Comparing availability includes checking permissible use, not just existence.
6. It supports trustworthy AI. Checking representativeness and bias at this stage helps avoid models that underperform for some populations or create ethical harm.

What It Is
It is a systematic assessment that maps each data requirement to the current data landscape and labels each one as met, partially met or not met (gap). Data requirements usually cover these dimensions:
- Quantity / Volume: Is there enough data, including enough examples of each class? Rare events such as fraud often lack enough positive examples.
- Quality: Accuracy, completeness, consistency, timeliness, validity and uniqueness.
- Relevance: Does the data actually contain signals related to the target outcome?
- Labels / Ground truth: For supervised learning, are reliable labels available? Who will label, and at what cost?
- Representativeness and bias: Does the data reflect the real-world population and the conditions the model will face in production?
- Granularity and format: Is the level of detail right (transaction vs. daily summary)? Is the data structured, unstructured, images, text or time series?
- Timeliness and frequency: Is there enough history, and will fresh data arrive at the frequency needed for inference?
- Accessibility: Can the team physically and technically reach the data (systems, APIs, silos, data owners)?
- Legal, privacy, security and ethical permission: Consent, GDPR/HIPAA-type rules, licensing, data sovereignty, internal policy.
- Cost and effort: What will it cost to acquire, clean, label and maintain?
- Production availability: Will the same data be available at inference time? This prevents training on features that will not exist when the model runs.

How It Works (Step by Step)
Step 1: Derive data requirements from business understanding. Start from the AI pattern (e.g., recognition, predictive analytics, conversational, anomaly detection, hyper-personalization, goal-driven systems, autonomous systems) and the success criteria. Each pattern implies typical data needs. For example, recognition needs labeled images or audio, and conversational systems need dialogue or text corpora.
Step 2: Inventory available data. Identify internal sources (ERP, CRM, logs, data warehouses, data lakes), external sources (third-party vendors, open datasets, partners) and data that could be generated (sensors, surveys, synthetic data). Engage data owners and stewards, and review data catalogs and metadata.
Step 3: Profile and explore the data. Do initial data collection and exploratory analysis. Check volumes, distributions, missing values, class balance, data types, history and lineage. Do not assume data is fit just because it exists.
Step 4: Build a gap matrix. List each requirement against its source and status (met / partial / gap), note the evidence, and add a remediation option.
Step 5: Identify remediation options for each gap. Options include:
- Acquire external or third-party data
- Start new data collection or instrumentation
- Label data (in-house, crowdsourced, active learning)
- Augment data or generate synthetic data
- Use transfer learning or pre-trained models to reduce data needs
- Clean, integrate or transform existing data
- Obtain legal permission or consent, or anonymize data
- Re-scope the problem or choose a different AI pattern
Step 6: Assess impact on feasibility, cost, schedule and risk. Estimate the effort for each remediation and update the risk register.
Step 7: Make and document the decision. Present the findings to stakeholders and sponsors. Decide to proceed, iterate back to Business Understanding to adjust scope, or stop (No-Go). Because CPMAI is iterative, returning to an earlier phase is expected and healthy.

Practical Example
A bank wants to predict loan default.
- Requirement: five years of loan history with outcome labels.
- Available: three years, labeled.
- Status: partial.
A second requirement is income data at inference time. That data is only collected manually after approval, so it is a gap for production. A third requirement is demographic data, but using it raises fair-lending concerns, so it is restricted.
The resulting decisions: extend history from archives, redesign the features to avoid post-approval data, and conduct a fairness review. Only then proceed.

Common Pitfalls
- Assuming data exists or is usable because a system holds it.
- Ignoring label availability and labeling cost.
- Overlooking data leakage and inference-time availability.
- Skipping privacy, consent and licensing checks.
- Treating data volume as a substitute for quality and relevance.
- Not involving data owners and subject matter experts.
- Jumping to model building before the gap analysis is done.

Exam Tips: Answering Questions on Comparing Available Data Against Requirements
1. Data before models. When a scenario describes a team eager to pick an algorithm or start training, the correct answer usually involves first assessing data availability and fitness against requirements.
2. Know the phase order. The CPMAI phases are Business Understanding, Data Understanding, Data Preparation, Model Development, Model Evaluation and Model Operationalization. This comparison belongs to Data Understanding and is informed by the requirements set in Business Understanding.
3. Iteration is correct, not failure. If a significant gap appears, the best answer often is to go back to Business Understanding to re-scope, or to make a Go/No-Go decision. Pushing forward is rarely the right answer.
4. Look beyond existence. Wrong answers often stop at whether the data exists. Right answers consider quality, quantity, labels, representativeness, access, permissions and inference-time availability.
5. Privacy and ethics are data requirements. If the data exists but consent or legal clearance is missing, treat it as unavailable until the issue is resolved. Choose answers that involve legal, compliance or the data steward.
6. Choose the least risky effective remediation. For small or unlabeled datasets, look for options such as labeling strategies, transfer learning, augmentation, synthetic data or acquiring external data. Pick whichever fits the constraints in the scenario.
7. Watch for leakage traps. A feature known only after the outcome, or unavailable in production, fails the requirement even if it boosts accuracy in training.
8. Stakeholder engagement matters. Answers that involve data owners, SMEs and sponsors in confirming gaps and decisions are usually stronger than answers where the team acts alone.
9. Watch for keywords: gap analysis, data inventory, data profiling, fit for purpose, sufficient, representative, ground truth, Go/No-Go, feasibility.
10. Eliminate extremes. Be wary of absolute options such as cancel immediately or proceed and fix the data later. The balanced option is usually right: assess, quantify the gap, evaluate remediation and decide.
11. Link to business value. If closing the gap costs more than the expected value of the solution, the best answer may be to re-scope or stop. CPMAI always ties data decisions back to the business objectives.

Summary
Comparing available data against requirements is the reality check of an AI project. It turns business ambition into a clear picture of feasibility. It also surfaces gaps in quantity, quality, labels, access and permission, and it guides remediation or a Go/No-Go decision. On the exam, favor answers that are data-first, iterative, compliance-aware and involve stakeholders.