Phase 2: Data Understanding
Phase 2 is the point in the CPMAI methodology where the business case meets the data reality, and the project either earns the right to proceed or quietly returns to Phase 1 for rescoping.
In the last post, I covered Phase 1: Business Understanding. The job there was to confirm AI was justified, name the capability pattern, set SMART success metrics, and pass a structured Go/No-Go gate. Today, Phase 2.
A six-phase, data-centric approach for AI projects
Each phase has an exit gate. Phase 2 is the second.
Understanding
Understanding
Preparation
Development
Evaluation
Operationalisation
What Phase 2 actually does
Phase 2 answers a single, uncomfortable question:
Skip the question, or answer it with hope, and Phase 4, Model Development, becomes the moment everyone discovers the truth. By then, the budget is committed and the timeline has been communicated.
Four risks that live in the data
Every AI project carries four data risks that have no equivalent in traditional software delivery. Each of them, if surfaced late, becomes the moment the programme stalls.
Four risks that live in the data
Each one has no equivalent in traditional software delivery.
Data Dependency
The model can only learn what the data contains. Discovering this in Phase 4 instead of Phase 2 is one of the most expensive mistakes in AI delivery.
Quality Risk
AI is far less forgiving than traditional software. Inconsistencies a rule-based system can absorb become systematic failures when amplified by a learning algorithm.
Representativeness
If training data does not match the conditions under which the system will operate, performance collapses on first contact with reality. Demographic, temporal, and contextual coverage all matter.
Timeliness
Stale data produces stale models. The freshness, refresh cadence, and update reliability of the source systems must be confirmed before model work begins.
The four critical Phase 2 questions
Phase 2 must produce honest written answers to four questions. Verbal alignment is not Phase 2 closure.
Do we have data that directly captures the patterns the system needs to learn?
If not, can it be acquired, generated, or substituted?
Can we obtain the data within project timelines?
Given technical, legal, contractual, and organisational constraints.
Is the quality high enough that learning will produce reliable behaviour?
Or will preparation costs exceed project value?
Do we understand the data's known gaps, biases, and limitations?
And have we surfaced the risks they create for the system and the people it affects?
If any one of these has no clear written answer, Phase 2 is not done. The temptation to proceed on assumption is the failure mode.
The four critical Phase 2 questions
Honest written answers. Verbal alignment is not closure.
Do we have data that captures the patterns the system needs to learn?
If not, can it be acquired, generated, or substituted?
Can we obtain the data within project timelines?
Given technical, legal, contractual, and organisational constraints.
Is the quality high enough for reliable learning?
Or will preparation costs exceed project value?
Do we understand the data's gaps, biases, and limits?
And have we surfaced the risks they create for the system and the people it affects?
The AI pattern shapes the data test
The capability pattern selected in Phase 1 changes which question carries the most weight. A data assessment that does not differentiate between patterns is not an assessment. It is a checklist.
The AI pattern shapes the data test
Where each capability pattern lives or dies on its data.
The Phase 2 Go/No-Go gate
Before exiting Phase 2, run a structured readiness review against six elements. Score each Red, Amber, or Green. Same discipline as Phase 1.
- Data inventory is complete. Sources, owners, formats, volumes.
- Access is confirmed in writing, not assumed.
- Quality assessment has been performed and documented.
- Bias screening covers demographic, temporal, and contextual dimensions.
- Privacy, consent, and regulatory review is signed off.
- The data feasibility scorecard sits at Green or Amber, not Red.
Red. The data does not exist, cannot be accessed within timeline, or cannot be used legally. Stop. Return to Phase 1 and rescope, or kill the project. Pushing forward in this state is how programmes burn twelve months and a sponsor.
Amber. The data exists with known constraints. Document the constraints, assign owners and deadlines, and proceed with explicit awareness.
Green. The data is available, accessible, of acceptable quality, and reviewed. Move forward.
The Phase 2 readiness review
Six elements. Each scored Red, Amber, or Green. No negotiation.
Complete inventory of sources, owners, formats, and volumes.
If you cannot list it, you do not have it.
Access is confirmed in writing, not assumed.
A verbal "yes" from a data owner is not a Phase 2 artefact.
Quality assessment performed and documented.
Completeness, accuracy, consistency, freshness, uniqueness.
Bias screening covers demographic, temporal, and contextual dimensions.
Test for the bias you would not want a regulator to find first.
Privacy, consent, and regulatory review is signed off.
GDPR, UAE PDPL, sector regulations, contractual data rights.
Data feasibility scorecard sits at Green or Amber, not Red.
A Red scorecard returns the project to Phase 1 for rescoping.
RED
Data does not exist or cannot be used. Stop. Return to Phase 1.
AMBER
Exists with known constraints. Document, assign owners, proceed.
GREEN
Available, accessible, acceptable quality, reviewed. Move forward.
A practical default
If Phase 2 reveals that the pattern requires data the organisation does not have and cannot reasonably acquire, the responsible action is to return to Phase 1 and rescope. Not to push forward and hope Phase 3, Data Preparation, will manufacture what does not exist.
Phase 2 is iterative, too
You will return to Phase 2. When Phase 3 reveals the cleaning effort is larger than estimated. When Phase 4 shows the features cannot be learned. When Phase 6 surfaces a drift signal that means the data assumptions no longer hold. Data Understanding is a posture maintained across the life of the system, not a single deliverable.
Next: Phase 3, Data Preparation. Where the cost of the data conversation turns from theoretical to invoiced.
References
- Walch, K. & Schmelzer, R. (Cognilytica / PMI). Cognitive Project Management for AI (CPMAI) Methodology. Project Management Institute.
- Chapman, P. et al. (2000). CRISP-DM 1.0: Step-by-step data mining guide.
- Wang, R. Y. & Strong, D. M. (1996). Beyond Accuracy: What Data Quality Means to Data Consumers. Journal of Management Information Systems, 12(4), 5–33.
- Mehrabi, N. et al. (2021). A Survey on Bias and Fairness in Machine Learning. ACM Computing Surveys, 54(6), 1–35.