Locating Domain Experts for Data Sources

Overview
In the PMI Cognitive Project Management for AI (CPMAI) methodology, Phase II: Data Understanding (closely tied to Phase I: Business Understanding) requires the team to identify what data is needed, where it lives, and whether it can be used. A key activity within Identify Data Needs is locating domain experts for data sources. These are the people who know what the data means, how it was created, how reliable it is, and how to access it.

What It Is
Locating domain experts means systematically finding and engaging the individuals who hold authoritative knowledge about each candidate data source. They typically fall into two groups:
1. Business or subject matter experts (SMEs): They understand the business process the data represents, such as claims adjusters, clinicians, underwriters, or call-center supervisors. They explain what a field actually means in practice, which edge cases exist, and how the data relates to the AI problem being solved.
2. Data or technical experts: They understand the systems that produce and store the data, such as data stewards, data owners, database administrators, data engineers, and system architects. They explain schemas, lineage, refresh frequency, known quality issues, and access mechanisms.

Other relevant people include data owners and custodians (who grant permission), legal, privacy, and compliance officers (who rule on whether data can be used), and labelers or annotators (who understand ground truth for supervised learning).

Why It Is Important
CPMAI stresses that most AI projects fail because of data problems, not algorithm problems. Domain experts reduce that risk in several ways:
- Correct interpretation: Data rarely documents itself. A column named 'status' can mean different things in different systems. Experts prevent misinterpretation that would produce a flawed model.
- Data quality insight: Experts know about gaps, biases, manual workarounds, historical system changes, and data entry habits that affect quality.
- Feasibility assessment: Early expert input shows whether enough relevant, accessible, and labeled data exists. This supports the go/no-go decision before money is spent on modeling.
- Access and permissions: Data owners must approve use. Knowing who they are early avoids schedule delays.
- Compliance and ethics: Experts flag sensitive attributes, PII, consent limits, and regulatory constraints, which supports trustworthy AI.
- Labeling and ground truth: SMEs define what a correct label is and may perform or validate labeling.
- Feature relevance: Experts suggest which variables are likely predictive and which are proxies or leakage risks.
- Stakeholder buy-in: Involving experts builds trust and adoption of the eventual AI solution.

How It Works (Typical Process)
Step 1: Start from the business problem. Using the AI pattern (for example, recognition, predictive analytics, or conversational AI) and the success criteria from Business Understanding, list the data needed to support the use case.
Step 2: Inventory candidate data sources. Identify internal systems (ERP, CRM, data warehouses, logs, documents) and external sources (third-party, public, purchased, or synthetic data).
Step 3: Map each source to its experts. For every source, identify the business SME, the technical owner or steward, and the approver. Use data catalogs, data governance offices, org charts, process owners, IT system registers, and referrals from stakeholders.
Step 4: Engage and interview. Ask structured questions: What does the data represent? How is it collected? How often is it updated? What is its volume and history? What are the known quality issues? Who can access it, and under what restrictions? Is it labeled?
Step 5: Document findings. Record them in a data source inventory or data requirements document, capturing ownership, meaning, quality, access, and constraints. Add data dictionaries and lineage notes.
Step 6: Assess sufficiency and gaps. With the experts, decide whether the data is adequate in quantity, quality, and relevance. If not, plan for acquisition, augmentation, labeling, or rescoping.
Step 7: Keep experts engaged iteratively. CPMAI is iterative and agile. Experts should be revisited during Data Preparation, Model Evaluation, and Operationalization, especially when data drift appears.

Common Challenges
- Experts are busy. Secure management sponsorship and time commitments.
- Knowledge is tribal and undocumented. Capture it in writing.
- Ownership is unclear. Escalate to data governance.
- Business and technical experts give conflicting interpretations. Reconcile them before preparing data.
- Third-party data has distant experts. Use vendor documentation and contacts.

Exam Tips: Answering Questions on Locating Domain Experts for Data Sources
1. Think data first. CPMAI holds that AI projects are data projects. When a scenario describes unclear, poor-quality, or misunderstood data, the best answer usually involves consulting the appropriate domain or data expert before modeling.
2. Choose early engagement. Prefer answers that engage experts during the Business and Data Understanding phases. Avoid answers that wait until after model building or deployment.
3. Match the expert to the question. Meaning of a business field calls for a business SME. Schema, lineage, or extraction questions call for a data engineer, DBA, or data steward. Permission to use data calls for the data owner. Privacy or regulation calls for legal or compliance. Ground truth calls for SMEs or labelers.
4. Be wary of assumptions. Wrong answers often have the data scientist guess the meaning of fields, rely only on column names, or proceed without validation. The correct answer validates with experts.
5. Watch for governance and ethics. If a scenario mentions PII, sensitive attributes, or regulated data, the best answer includes identifying owners and compliance experts and confirming permissions.
6. Look for documentation. Answers that record expert knowledge in data inventories, dictionaries, or requirements artifacts are often preferred over informal conversations.
7. Treat feasibility as a gate. If experts reveal that data is insufficient or inaccessible, CPMAI favors revisiting scope, acquiring data, or making a go/no-go decision. It does not favor forcing the model forward.
8. Remember iteration. If a question asks what to do when new data issues emerge later, returning to the experts and the Data Understanding phase is consistent with CPMAI.
9. Eliminate extremes. Discard options that skip stakeholders entirely, rely solely on automated tools to understand meaning, or immediately purchase external data without first checking internal sources and experts.
10. Look for key phrases. Correct answers often contain words like 'consult the data owner', 'engage subject matter experts', 'validate with data stewards', 'identify who owns and understands the data', and 'confirm access and permissions'.

Quick Example
Scenario: A team building a churn prediction model finds a field called 'acct_flag' with unclear values. What should the project manager do first?
Best answer: Identify and consult the business SME and data steward responsible for the source system to clarify the field's meaning and quality, then document it.
Wrong answers: Dropping the field without inquiry, letting the model decide its importance, or inferring its meaning from value distributions alone.

Summary
Locating domain experts connects raw data sources to real-world meaning, quality knowledge, access rights, and compliance constraints. In CPMAI it is a foundational step of identifying data needs. It reduces risk, informs feasibility, and enables trustworthy, successful AI. On the exam, favor early, targeted, documented, and iterative engagement with the right experts over assumptions or shortcuts.