Assumptions and Constraints for AI Implementation

Overview
In the PMI Certified Professional in Managing AI (PMI-CPMAI) framework, Phase I is Business Understanding, often summarized as Identify Business Needs and Solutions. One of its key outputs is a clear, documented set of assumptions and constraints for the proposed AI solution. This guide covers four things: why the topic matters, what assumptions and constraints are, how to identify and manage them, and how to answer exam questions on them.

1. Why It Is Important
AI projects fail far more often than traditional IT projects. The usual causes are:
• Unrealistic expectations.
• Poor or unavailable data.
• Unclear success criteria.
• Regulatory surprises.
• Underestimated operational effort.

Most of these failures trace back to assumptions that were never surfaced and constraints that were never acknowledged. Capturing them early gives you several benefits:
• Reduces risk: Every untested assumption is a potential risk. Making it explicit lets the team validate it before investing heavily.
• Supports the go/no-go decision: CPMAI stresses deciding early whether AI is the right solution at all. Constraints may make AI infeasible, too costly, or unnecessary.
• Sets realistic stakeholder expectations: Stakeholders often assume AI will be 100% accurate, instantly deployable, or self-maintaining. Documented assumptions correct these beliefs.
• Defines scope and feasibility: Constraints set the boundaries for data, technology, budget, timeline, ethics and regulation.
• Enables iteration: CPMAI is iterative and data-centric. Assumptions are revisited in later phases (Data Understanding, Data Preparation, Model Development, Model Evaluation, Operationalization). If an assumption proves false, the team loops back.
• Supports trustworthy AI: Ethical, privacy, transparency and explainability constraints must be identified up front, not bolted on later.

2. What They Are
Assumptions are factors believed to be true, real or certain for planning purposes, without proof or verification. In AI projects they often concern:
• Data: e.g., "We have enough labeled historical data." "The data represents current conditions." "We are legally allowed to use customer data for training."
• Patterns: e.g., "There is a learnable pattern in the data that predicts the outcome."
• Stability: e.g., "Future data will resemble past data," meaning no major data drift or concept drift.
• Stakeholders and users: e.g., "Users will trust and adopt the AI recommendations." "Subject matter experts will be available to label data."
• Technology: e.g., "Existing infrastructure can support model training and inference."
• Business: e.g., "A 10% improvement in prediction accuracy will deliver the expected ROI."
• Organization: e.g., "Leadership sponsorship will continue throughout the project."

Constraints are limiting factors that restrict the options available to the project. They are typically known facts, not beliefs. Common types:
• Data: limited volume, poor quality, missing labels, siloed sources, data residency rules, consent limitations.
• Regulation and law: GDPR, CCPA, the EU AI Act, HIPAA, industry rules, fair lending laws, right-to-explanation requirements.
• Ethics and responsible AI: fairness, bias mitigation, transparency, explainability, human oversight requirements.
• Technology: compute limits, latency requirements (e.g., real-time inference under 100 ms), edge deployment, legacy system integration, approved platforms.
• Budget: funding for data acquisition, labeling, cloud compute and talent.
• Schedule: fixed deadlines, seasonal windows, regulatory dates.
• Resources and skills: availability of data scientists, data engineers and domain experts.
• Organization and culture: change resistance, policies, governance approval processes.
• Performance: minimum acceptable accuracy, precision, recall, or tolerance for false positives and negatives.
• Operations: monitoring, retraining frequency, model governance, support model.

Key distinction
• An assumption is uncertain and must be validated.
• A constraint is a known limit that must be respected or worked within.
• An assumption that proves false often becomes either a risk or an issue.
• Constraints shape the solution design.

3. How It Works in Practice
Step 1 – Define the business problem first. CPMAI insists on starting with the business problem, not the technology. Ask what problem is being solved, why AI is needed, and what success looks like in business terms.

Step 2 – Identify the AI pattern. CPMAI uses seven patterns of AI:
• Recognition
• Conversational/human interaction
• Predictive analytics and decision support
• Patterns and anomalies
• Hyperpersonalization
• Autonomous systems
• Goal-driven systems

Each pattern carries typical assumptions and constraints. For example, autonomous systems face strict safety constraints, and recognition requires large labeled datasets.

Step 3 – Elicit assumptions and constraints from stakeholders. Use workshops, interviews, data owner consultations, legal/compliance reviews and IT architecture reviews. Include business sponsors, end users, data stewards, security, legal, ethics boards and operations.

Step 4 – Document them. Record them in an assumption log or the project charter or business case. Each entry should include:
• A description.
• The owner.
• The validation method.
• The validation date.
• The impact if false.
• The status.

Step 5 – Assess the impact on feasibility. Ask:
• Do the constraints make the AI solution infeasible?
• Would a simpler non-AI or rules-based solution work?
• Is the data available, accessible and of sufficient quality?
• Is the ROI still positive given the constraints?

Step 6 – Validate assumptions early, especially data assumptions. CPMAI's Phase II (Data Understanding) exists largely to test the data assumptions made in Phase I. Methods include:
• Data profiling.
• Proof of concept work.
• Pilot studies.
• Stakeholder confirmation.

Step 7 – Convert to risks and responses. Feed unvalidated or high-impact assumptions into the risk register with mitigation plans. Feed constraints into design decisions, for example choosing an explainable model to satisfy a regulatory explainability constraint.

Step 8 – Revisit iteratively. Review assumptions and constraints at each phase gate. If an assumption fails (e.g., the data does not contain predictive signal), return to Business Understanding and re-scope. Do not push forward.

Worked example
A bank wants an AI model to detect fraudulent credit card transactions.
• Assumptions: Historical fraud labels are accurate. Fraud patterns are reasonably stable. Fraud analysts will act on model alerts.
• Constraints: Decisions within 200 ms. Must comply with PCI-DSS and fair lending regulations. Must explain declined transactions to customers. Limited budget for cloud GPUs. Fraud cases make up less than 0.5% of data, an imbalance constraint.
• Actions:
– Validate label quality with the fraud team.
– Plan regular retraining and drift monitoring.
– Choose models that meet the latency and explainability requirements.
– Address class imbalance during data preparation.

4. Common Pitfalls
• Assuming data exists and is usable without checking.
• Treating AI as a one-time project rather than a lifecycle that needs monitoring and retraining.
• Ignoring ethical, bias and privacy constraints until late in the project.
• Assuming stakeholders understand AI limitations, such as probabilistic outputs.
• Failing to involve legal and compliance early.
• Confusing assumptions (unverified) with constraints (known limits).
• Not revisiting assumptions when new information appears.

5. How to Answer Exam Questions on This Topic
Exam questions are usually scenario-based. Typical formats:
• Classify a statement as an assumption, constraint, risk or requirement.
• Choose the best next action when an assumption proves false.
• Identify which constraint most affects solution choice.
• Decide what should happen before moving to the next phase.

A reliable approach:
• Read the scenario and find the CPMAI phase it describes. Is it still Business Understanding, or has the project moved to Data Understanding or later?
• Ask whether the statement is believed true without proof (assumption) or is a known limitation (constraint).
• Look for the answer that validates, documents, communicates or iterates. Avoid answers that ignore the problem or push forward.
• Favor answers that put the business problem and data reality ahead of technology choices.
• Remember that responsible AI, regulatory and ethical constraints are non-negotiable.

Exam Tips: Answering Questions on Assumptions and Constraints for AI Implementation
Tip 1: Words like "we believe," "it is expected," "likely," "should be available" signal an assumption. Words like "must," "cannot," "limited to," "required by law," "fixed budget" signal a constraint.

Tip 2: If an assumption proves false, the best answer is usually to reassess and iterate: return to Business Understanding, update the business case, inform stakeholders and adjust scope. Do not ignore it, and do not simply add more resources.

Tip 3: Data assumptions are the most critical in AI projects. When in doubt, choose the answer that validates data availability, quality, quantity, representativeness and legal usability early.

Tip 4: Regulatory, privacy and ethical constraints shape the solution. If a regulation requires explainability, the correct answer favors interpretable models or explainability techniques, even at some cost in accuracy.

Tip 5: Pick answers that document assumptions and constraints in the business case, charter or assumption log and communicate them to stakeholders. Undocumented assumptions are a classic wrong-answer trap.

Tip 6: If constraints make AI infeasible or too costly, the right answer may be not to use AI, or to use a simpler rules-based or analytics solution. CPMAI asks whether AI is truly needed.

Tip 7: Watch for distractors that jump straight to model building, algorithm selection or tool purchases before business needs, assumptions and constraints are clear. These are usually wrong in Phase I questions.

Tip 8: Assumptions about future data matching past data link to drift. Answers mentioning monitoring and retraining plans are strong when the question involves operational assumptions.

Tip 9: Unvalidated, high-impact assumptions should be added to the risk register with an owner and mitigation plan.

Tip 10: Include stakeholder expectations as a source of assumptions. If stakeholders expect perfect accuracy, the best answer sets realistic expectations and agrees acceptable performance thresholds tied to business value.

Tip 11: Choose answers that involve cross-functional stakeholders (legal, compliance, data owners, IT, end users) when identifying constraints. Answers relying on the data science team alone are weaker.

Tip 12: Remember the iterative nature of CPMAI. The correct answer often involves revisiting earlier decisions rather than treating the plan as fixed.

Quick Memory Aid: "DRESS-BOT"
Use this to recall constraint categories:
• Data
• Regulatory
• Ethical
• Schedule
• Skills
• Budget
• Organizational/Operational
• Technical

Summary
Assumptions are unverified beliefs that must be validated. Constraints are known limits that must be respected. In AI projects, both must be identified early in Business Understanding, documented, communicated, turned into risks or design decisions, and revisited at every phase. On the exam, favor answers that validate data early, respect regulatory and ethical limits, keep stakeholders informed, question whether AI is needed at all, and iterate when reality differs from plan.