AI Project Boundaries and Deliverables

Overview
In the PMI Certified Professional in Managing AI (PMI-CPMAI) framework, which builds on the CPMAI (Cognitive Project Management for AI) methodology, defining AI project boundaries and deliverables is a core activity. It belongs to Phase I, Business Understanding, under the domain of identifying business needs and solutions. Here the project team works out what the AI solution will and will not do, which outputs will be produced, and how success will be measured. Only then is any data collected or any model built.

Why It Is Important
AI projects fail far more often than traditional IT projects. A leading cause is poorly defined scope. Clear boundaries and deliverables matter for several reasons:

1. They prevent scope creep and "boiling the ocean." AI invites ambitious goals such as \"let's use AI to transform customer service.\" Without boundaries, teams chase vague, unbounded problems that never deliver value. CPMAI stresses thinking big, starting small, and iterating often.
2. They manage stakeholder expectations. AI is probabilistic, not deterministic. Stakeholders must understand that a model may be 85% accurate, not 100%. Boundaries set realistic expectations about performance, autonomy and limits.
3. They enable go/no-go decisions. Clear boundaries let the team judge whether the problem truly needs AI. If a rules-based solution would do, AI may be the wrong choice.
4. They tie deliverables to business value. Every deliverable should map to a measurable business outcome, such as cost reduction, revenue increase, risk reduction or better customer experience.
5. They support governance, ethics and compliance. Defining what the system must not do is part of responsible and trustworthy AI. Examples include not making final credit decisions or not using protected attributes.
6. They control resources and timelines. Data acquisition, labeling and model training are expensive. Bounded scope keeps cost and schedule under control.

What It Is
AI project boundaries are the explicit limits of the AI initiative. They include:
- Problem scope: the specific business problem being solved. For example, \"predict which invoices are likely to be paid late\" rather than \"fix finance.\"
- AI pattern(s) in scope: CPMAI's Seven Patterns of AI help classify the solution:
  - Recognition
  - Conversational/Human Interaction
  - Predictive Analytics and Decision Support
  - Patterns and Anomalies
  - Hyperpersonalization
  - Autonomous Systems
  - Goal-Driven Systems
  Boundaries often limit a project to one or a few patterns per iteration.
- In-scope vs. out-of-scope items: processes, users, geographies, data sources, channels and decisions included or excluded.
- Level of autonomy: whether the AI only recommends (human-in-the-loop), acts with human oversight (human-on-the-loop), or acts fully autonomously.
- Data boundaries: which data sources, time periods and data types will be used, plus privacy and consent constraints.
- Performance boundaries: acceptable accuracy, precision, recall, latency and error tolerance.
- Ethical, legal and regulatory constraints: fairness, transparency, explainability, privacy (e.g., GDPR) and industry regulation.
- Technical and operational constraints: infrastructure, integration points, budget, timeline and skills.

AI project deliverables are the tangible outputs the project will produce. They may include:
- Business understanding documents: problem statement, success criteria, ROI analysis, AI go/no-go assessment
- Data deliverables: data inventory, data quality assessment, prepared and labeled datasets
- Model deliverables: trained model(s), model documentation, model cards, evaluation reports
- Operational deliverables: deployed model or API, monitoring dashboards, retraining pipelines, user training materials
- Governance deliverables: risk assessments, bias audits, explainability reports, compliance documentation
- Iteration plan: definition of the Minimum Viable Product (MVP) or pilot, plus roadmap for future iterations

How It Works
Defining boundaries and deliverables follows a structured, iterative process within CPMAI Phase I:

Step 1: Define the business problem. Start with the business need, not the technology. Ask what problem is being solved, who has it, and what it costs today.
Step 2: Determine if AI is the right solution. Use the AI go/no-go criteria, which cover three areas:
- Business feasibility: clear problem definition, willingness to change, measurable ROI
- Data feasibility: sufficient data quantity, quality and access
- Implementation feasibility: technology, skills, timeline, and ability to execute and use the model in operation
If the answer is no, or a simpler solution works, stop or redirect.
Step 3: Identify the AI pattern(s). Map the problem to one or more of the seven patterns. This clarifies the type of solution and the data needed.
Step 4: Set scope boundaries. Document what is in scope and out of scope. Narrow to the smallest valuable piece, the MVP or pilot, that can prove value quickly.
Step 5: Define success criteria and KPIs. Set both business metrics (e.g., reduce churn by 5%) and technical metrics (e.g., model F1 score above 0.80). Agree on acceptable error rates and the cost of false positives versus false negatives.
Step 6: Define the level of human involvement. Specify where humans approve, override or monitor AI outputs.
Step 7: Identify constraints and risks. Cover ethical, regulatory, data privacy, security and operational constraints.
Step 8: Specify deliverables and acceptance criteria. List each deliverable, its owner, its acceptance criteria and its timing.
Step 9: Obtain stakeholder agreement. Get sign-off from business sponsors, data owners, legal/compliance and end users.
Step 10: Revisit iteratively. CPMAI is iterative. Later phases (Data Understanding, Data Preparation, Model Development, Model Evaluation, Operationalization) may reveal that boundaries need adjusting. For example, missing data may force a scope change. Adjust through controlled change, not ad hoc expansion.

Key Concepts to Remember
- Think Big, Start Small, Iterate Often: deliver small, valuable increments rather than one large \"big bang\" AI project.
- AI projects are data projects: boundaries depend heavily on data availability and quality.
- Probabilistic outputs: deliverables must include acceptable performance thresholds, not guarantees of perfection.
- Model decay and drift: deliverables should include monitoring and retraining plans, because AI is never \"done.\"
- Avoid AI for AI's sake: boundaries must stem from business value, not technology hype.
- Responsible AI by design: ethical constraints belong in the boundaries from the start, not added later.

Common Pitfalls
- Scope defined around technology (\"build a neural network\") rather than business outcomes
- No out-of-scope list, which leads to stakeholder misunderstanding
- Combining too many AI patterns in a single iteration
- Ignoring data constraints until after scope is fixed
- Promising 100% accuracy or full automation without justification
- Omitting operational deliverables such as monitoring, retraining and user adoption
- Failing to define human oversight responsibilities

Exam Tips: Answering Questions on AI Project Boundaries and Deliverables

Tip 1: Business first, technology second. When options include \"select the algorithm\" versus \"clarify the business problem and success criteria,\" the business-focused answer is almost always correct early in the project.
Tip 2: Favor narrow, iterative scope. Prefer answers that reduce scope to a manageable MVP or pilot that delivers measurable value quickly. Be wary of answers proposing enterprise-wide, all-at-once AI rollouts.
Tip 3: Check whether AI is even needed. If a scenario describes a problem solvable with simple rules or automation, the correct answer may be that AI is not appropriate. The go/no-go assessment is a legitimate deliverable.
Tip 4: Know the go/no-go factors. Expect questions on business, data and implementation feasibility. If data is unavailable or of poor quality, the right action is usually to adjust scope, acquire data, or pause. Proceeding anyway is not the right answer.
Tip 5: Look for explicit out-of-scope statements. Answers that document both in-scope and out-of-scope items are stronger than those that only list inclusions.
Tip 6: Success criteria must be measurable. Choose answers with quantifiable business KPIs and technical metrics over vague goals like \"improve efficiency.\"
Tip 7: Address scope changes formally. When stakeholders request new features mid-project, the best answer usually involves:
- evaluating the impact
- revisiting Business Understanding
- deferring to a future iteration
It is not to absorb the request silently or reject it outright.
Tip 8: Include operational and governance deliverables. A complete deliverable set includes monitoring, retraining, documentation, and bias or ethics assessments, not just a trained model.
Tip 9: Remember human-in-the-loop. For high-risk decisions in healthcare, finance, hiring or legal, prefer answers that keep human oversight within the defined boundaries.
Tip 10: Map scenarios to the seven patterns. If a question describes the solution type, identify the pattern. For example, fraud detection maps to Patterns and Anomalies, and chatbots map to Conversational. Recommending one pattern per iteration is generally wise.
Tip 11: Watch for absolute language. Options with \"guarantee,\" \"100% accuracy,\" \"eliminate all errors\" or \"fully replace humans immediately\" are usually wrong for AI projects.
Tip 12: Recognize iteration. CPMAI phases loop back. If a later phase reveals problems, such as insufficient training data, returning to Business Understanding to re-scope is the correct move.
Tip 13: Engage stakeholders early. Answers involving business owners, data stewards, compliance and end users in defining boundaries are typically preferred over decisions made only by the data science team.

Sample Question Approach
Scenario: A retail executive asks the project manager to \"use AI to improve the entire customer journey\" within six months. What should the project manager do first?

Reasoning: The request is broad and unbounded. The best answer is to work with stakeholders to identify a specific, high-value problem within the customer journey. Examples include predicting cart abandonment or personalizing product recommendations. The PM should then define measurable success criteria, confirm data availability, and scope a pilot. Wrong answers include:
- immediately selecting a vendor platform
- hiring data scientists to build models
- promising a complete transformation in six months

Summary
AI project boundaries and deliverables turn vague AI ambitions into focused, feasible and valuable initiatives.
- Boundaries define the problem, AI patterns, autonomy, data, performance, ethics and constraints.
- Deliverables span business, data, model, operational and governance outputs.
- On the exam, choose answers that are business-driven, narrowly scoped, measurable, iterative, data-aware, ethically grounded and agreed with stakeholders.