The CPMAI Approach: An Overview of Six Phases Explained

The CPMAI Approach: An Overview of Six Phases

The CRISP-DM framework provides six AI and machine learning project management phases that can be adapted to project development with CPMAI. The phases are Business Understanding, Data Understanding, Data Preparation, Model Development, Model Evaluation, and Model Operationalization. Data science teams engage in an iterative process throughout the six phases. This means that teams can go back to the earlier phases if the original assumptions of the business change as a result of new data or new model outcomes.

What Is Underlying CPMAI?

CPMAI builds on the CRISP-DM framework and was designed as a supplement to CRISP-DM to address gaps in enterprise-level AI and machine learning project management. Most data science frameworks are designed to generate a working model from unstructured data. CPMAI addresses project management, business alignment, and AI ethics in addition to the modeling.

The six phases outlined in CPMAI can be used to structure an approach to project development, although this does not necessarily mean a Waterfall approach to project management. Teams may find it necessary to go back from Phase 5 (Model Evaluation) to Phase 3 (Data Preparation) and even Phase 1 (Business Understanding) to continue with the project.

How Does Each of the Six CPMAI Phases Connect to the Rest?

Each phase provides a unique output that helps progress to other phases. Unlike the CPMAI-specific phases, the PMI-CPMAI examination grouping domains include only five key areas as opposed to treating each phase in equal measure. Within the most weighted domain of the examination, Phases 2 and 3 illustrate differing levels of maturity of data and AI model development, respectively.

.	Phase	Name	Core Deliverable	Related PMI-CPMAI Exam Domain
.	1	Business Understanding	Business case and project scope statement	II. Identify Business Needs and Solutions (26%)
.	2	Data Understanding	Data source map and availability assessment	III. Identify Data Needs (26%)
.	3	Data Preparation	Clean, model-ready dataset	III. Identify Data Needs (26%)
.	4	Model Development	Trained model prototype	IV. Manage AI Model Development and Evaluation (16%)
.	5	Model Evaluation	Model evaluation and bias audit report	IV. Manage AI Model Development and Evaluation (16%)
.	6	Model Operationalization	Production deployment and monitoring plan	V. Operationalize AI Solution (17%)

For the CPMAI examination, governance and responsible AI practices (Domain I, 15% of the exam) span all six phases, as opposed to being in just one. In all six phases, throughout the entire lifecycle, bias checks, privacy protections, and actions to aid accountability are included, from Business Understanding to the Operationalization of Models.

What Is Each Phase of the AI Project Management Process?

Phase 1: How Do You Manage the Phase of an AI Project Called Business Understanding?

During this phase, business teams document a business issue that an AI solution may address — covered in full depth in Phase 1: Business Understanding in CPMAI. Teams perform an assessment to justify the use case for employing AI.

Document the business issue and the stakeholders.

Assess the case and consider if AI or ML is utilized to address the issue, as opposed to using a simple set of business rules.

Consider the issue within the boundaries of business, finance, and operations.

Prepare the project scope statement with success criteria and KPIs, and project scope statement with success criteria and KPIs.

Calculate the projected return on investment (ROI) and present the business case to the management team.

A common problem during this phase occurs when teams execute Data Understanding in prior phases. Teams often learn mid-project the AI issue is in fact an issue of poor data quality, or a due process issue.

Phase 2: How Do You Manage the Phase of an AI Project Called Data Understanding?

During this phase, business teams outline the location and mapping of data that an AI solution needs. Teams also determine if the project may continue, if the project needs to be rescope due to the data available, or if the project should be put on hold.

Describe the types of data, formats, and volumes needed to solve the business problem.

Describe internal data, third-party data, and public data that can be used to solve the problem.

Describe the availability of subject matter experts that can explain the data.

Describe the AI workspace, storage, and infrastructure for the data pipeline build.

Describe the data gaps and assess the data for quality and availability.

Typical mistake: Teams assume data “is somewhere in the organization” and skip steps to confirm data is being stored, if it is stored in the correct format, and/or if historical data meet the minimum volume to train a model.

.Phase 3: What Activities are included in the Preparation of Data?

Data Preparation is the most time-consuming activity during the CPMAI process, accounting for approximately 80% of the total time expected and includes activities focused on removing data outliers, readiness data model transformations, and data labeling. The readiness of data is the largest determinant of the quality of models compared to the selection of algorithms in Phase 4. Consequently, an appropriate selection of algorithm trained on poor data would also lead to a poor model.

Clean data of duplicate, incorrect, or missing data.

Transform data to a format and structure that the model will accept.

Validate the accuracy and consistency of each dataset label.

Assess the data quality against the standards set during Data Understanding.

Record the data lineage and data protection and privacy compliance prior to model construction.

Teams generally think this step takes far less effort than it actually does, especially during the planning stage. The closer they get to the deadline, the more this step gets compressed which results in a hit to the model’s overall performance in Phase 4 and 5..

Phase 4: Explaining the Development of Models

We can break the development of models into choosing the correct algorithms and training the models. Hyperparameter tuning falls into this category as well. While you are developing your models, your project manager has to ensure the project is aligned with the company’s interests. Most people initially contacting us regarding an AI project think this is the primary focus of the project, but at CPMAI, we define this as one of the many phases of an AI project.

Select the business problem related algorithms.

Get your models trained after defining a set of tracking experiments for different configurations.

Refine the hyperparameters to a satisfactory level for the success metrics.

Track your model’s development and training with an emphasis on the cost of getting this computing resource.

Keeping focused on the original business problem will help maintain the model’s scope.

Teams lose the business goal the model was designed for by focusing too hard on improving the accuracy of the models.

Phase 5. What does model evaluation involve?

This phase assesses if the model constructed in Phase 1 satisfies the Performance Criteria. Education will take place if there are issues with poor performance of the model on unseen data or if there is bias, or there are no explanations for the model prior to deployment. For CPMAI, this phase acts as a gate of quality. If the model is not evaluated, this phase leads to the preparation of either the data or model.

Evaluate the model’s accuracy on unseen data.

Evaluate if there is bias in the model considering demographic or business segmentation.

Assess explanations for high-risk or high-stakes model outputs.

Evaluate the model’s performance to the criteria in the Business Understanding phase.

Evaluate if the model is prepared for deployment.

For most teams, the focus is on the model’s accuracy, and they evaluate for bias or explainability in isolation, which can result in breached compliance or a major trust deficit.

Phase 6. What does model operationalisation involve?

A validated model is deployed to production, and solutions for monitoring model drift, automating model retraining, and integrating model services with the existing components of the business are put in place. Unlike traditional software projects that end upon deployment, CPMAI views operationalization as a subsequent phase; a deployed model requires active oversight.

Deploy the model, and design the model’s integration with the existing business processes.

Develop a model governance policy and a framework for managing model versions.

Create model drifting and deterioration monitoring.

Automate or schedule retraining depending on changes made to the data.

Analyze performance metrics to determine the outcome.

Organizations tend to stop monitoring once they ‘deploy’ their model. As a result, many models gradually and irreversibly start to lose their accuracy.

What is the Reason Teams seem to Return to early CPMAI Phases?

The requirements for each subsequent phase in a CPMAI Methodology project are seldom fully known. For instance, a second phase assessment of data can reveal that the objectives of the first phase were set too optimistically, or a fifth phase assessment can show that data of the third phase was considerably missing certain transformations. The iterative nature of the CPMAI Methodology is perhaps the largest deviation from traditional predictive project methods, which, after the scoping of a project, assume that the requirements then remain constant.

Where Can I Obtain Six CPMAI Phase Trainings?

PMTI offers instructor-led live training for the CPMAI Certification Training, which includes working through all six phases of the CPMAI Methodology, along with structured exercises for each phase. .

CPMAI Methodology Frequently Asked Questions

Are the six CPMAI Phases sequential?

There is no order. Usually, if teams go back to any of the six phases, this is done once a gap is identified in one of the later phases and the team returns to the Data Preparation phase.

Which of the six phases is the most time consuming?

Data preparation (Phase 3) usually takes up over 80% of the entire project, placing this phase as the longest phase of any AI project.

How do the six phases of the PMI-CPMAI certification intersect?

No. The six phases are separated into five weighted domains by PMI with Identify Business Needs and Solutions and Identify Data Needs accounting for 52% of the exam questions.

Which phase is most often omitted or skimmed?

Teams tend to under-invest in this stage which results in technically correct models that resolve the wrong business requirement.

.

.

.