CPM AI Phase 1: Business Understanding - Practice Flashcards
CPM AI / CPMAI: Comprehensive Notes on Phase 1 – Business Understanding
Purpose of the CPMAI methodology: help determine when AI/ML is a good fit for a business problem, define the objective, assess cognitive requirements, and plan for scalable, trustworthy AI deployments.
Core idea: start with business needs and ROI, then map to AI patterns, data readiness, technology readiness, and governance. Iterative and staged approach to maximize success and manage risk.
Key ideas and structure from the transcript
Before starting an AI project, answer: what is the objective (business/organizational) and does it require a cognitive (AI) solution?
If you have to solve the problem with perception, planning, prediction (the P’s), the machine must do it because humans or non-AI approaches aren’t sufficient, faster, cheaper, or safer.
Consider noncognitive alternatives: could a simpler, non-AI approach achieve the objective with lower risk, cost, or time-to-value?
The seven AI patterns guide problem framing: recognition, conversation, anomaly detection, predictive analytics, autonomous systems, goal-driven systems, hyper personalization. If multiple patterns apply, you may have a multi-project scope.
Constraints and transparency: identify time, financial, and resource constraints; address data, disclosures, legal feasibility, and stakeholder involvement; consider ethical and responsible AI implications from the start.
Stakeholder satisfaction: determine who must be satisfied by the project (customers, employees, partners, shareholders, leadership).
Good AI projects are typically those that address a short-term, well-defined need that cannot be solved as well with simpler approaches; cognitive-appropriate problems; scalable solutions; potential for automation (four D’s: dull, dirty, dangerous, demeaning) or cost/value improvements; and a possibility of solving the problem more elegantly than brute force.
ROI lens: AI projects should enhance revenue, differentiate, reduce costs, or improve operations. Consider the “cool factor” but weigh it against real business value.
Phase-based approach (CPMAI): focus on business understanding first, then data understanding, etc. A green-light go/no-go depends on nine traffic-light factors across three buckets (business, data, implementation).
DIKUW pyramid (Data → Information → Knowledge → Understanding → Wisdom) guides where AI adds value. Machine learning sits at Knowledge; understanding and wisdom represent more advanced reasoning that current AI is not at.
The DIKUW perspective helps identify the AI sweet spot: patterns and knowledge with pattern-based insights rather than purely descriptive data.
To maximize success, think big, start small, iterate often. Break complex problems into smaller AI-enabled pieces and pilot them, advancing gradually toward broader deployment.
The seven AI patterns (summarized)
Recognition pattern: extracting meaning from unstructured data (images, audio, text).
Conversation pattern: natural language interactions between humans and machines (chatbots, voice assistants).
Pattern/anomalies pattern: detecting unusual patterns or fraud, deviations, anomalies.
Predictive analytics pattern: forecasting trends, risk, demand; informs decisions.
Autonomous systems pattern: systems that operate with minimal or no human intervention.
Goal-driven systems pattern: problems seeking optimal solutions (e.g., routing, scheduling).
Hyper personalization pattern: highly individualized recommendations or content.
Practical note: If a problem maps to multiple patterns, consider multiple smaller AI projects rather than one large, monolithic effort.
Example guidance: match the pattern to the objective or business ROI to avoid forcing a pattern to fit an ROI (don’t start with a pattern and then retrofit ROI).
Determining the AI objective and going-no-go decisions
For each AI project, answer: which pattern(s) apply? where does AI add value? what alternatives exist (noncognitive)?
Nine-factor AI go/no-go assessment (CPMAI phase 1) spans three green-light categories:
Business feasibility: three questions about the problem and willingness/ability to invest and change, and whether there is a meaningful ROI.
Data feasibility: three questions about owning the required data, data quantity, and data quality.
Implementation feasibility: three questions about technology/skills, ability to execute in time, and whether the solution will be used by the intended users.
If all nine lights are green, project is a go. Red lights indicate high risk; yellow lights indicate risk that needs addressing before proceeding.
The goal of the go/no-go assessment is to avoid wasted effort on projects with serious blockers and to stage risky work into smaller, more manageable iterations.
The DIKUW pyramid: data to wisdom
DIKUW stands for Data, Information, Knowledge, Understanding, Wisdom.
Data: base layer – raw facts (facts without context).
Information: data plus context – the what/where/when/who of data; descriptive analytics.
Knowledge: patterns and relationships that explain how; machine learning sits here – pattern recognition, predictive analytics, etc.
Understanding: why patterns occur; common sense reasoning is still not fully achievable by AI today.
Wisdom: human-level insights and the ability to choose among options for action. This is a longer-term, more advanced capability beyond current mainstream AI.
Practical implication: use AI to move from data to knowledge and insights, not just to create dashboards or store data. Focus on generating actionable insights and understanding patterns, not merely metrics.
How DIKUW informs AI project planning
If the problem can be solved with basic data/correlations or descriptive analytics, ML may not be needed; the sweet spot is where ML provides patterns, predictive insights, and actionable understanding.
Each technology shift (industrial -> IT -> connected systems) created new value through more intelligent, efficient processes. AI/ML should deliver enhanced efficiency and effectiveness.
Focus on knowledge and understanding levels where AI adds real value, not only data collection or basic reporting.
When to use AI vs automation vs human-in-the-loop
Not every problem benefits from AI. Consider automation when predictable, rule-based tasks can be reliably executed at scale.
AI is advantageous when you need to learn from data over time and/or when patterns, recognition, and predictive capabilities are required that are hard to codify with rules.
A practical approach is to start with augmented/assistive AI (human-in-the-loop) to reduce risk, then evolve toward higher autonomy as the system proves out.
The “four D’s” (dull, dirty, dangerous, demeaning) and cost/value considerations often favor automation or augmenting humans first, then adding AI where it delivers incremental value.
The “cool factor” (e.g., robot demonstrations) is not a solid ROI driver unless it translates into business value.
ROI and project sequencing: where AI can yield quick wins
Short-term ROI patterns: augmented/assisted intelligence (e.g., chatbots, self-service, unstructured data handling) often yield faster time-to-value because they are easier to implement and test.
Predictive analytics and decision support can yield significant ROI but require robust validation and staged deployment.
Autonomous systems generally take longer to realize ROI due to higher risk, binding constraints, and the need for near-perfect performance (often not yet fully realized in broad deployments).
Time-to-ROI varies by problem; many AI initiatives benefit from starting with off-the-shelf models or foundation models to prove ROI quickly, then moving to custom models if warranted.
Short-iteration strategy: break complex problems into smaller pilots with well-defined scoping and measurable ROI; iterate in short cycles to build confidence and value.
Off-the-shelf and foundation models: when to shortcut
Foundation models and large language models (LLMs) can shortcut many cognitive tasks: chatbot interactions, document QA, summarization, entity extraction, classification, translation, etc.
For many tasks, you can leverage pre-trained models and fine-tune on internal data, or use prompt engineering to adapt generic models to your domain.
When to train from scratch vs. shortcut with foundation models:
Train from scratch if you need a compact model, fast inference, and complete control over data privacy, latency, and customization;
Otherwise, extend foundation models with fine-tuning, adapters, or data integration to meet the problem with less development time.
Vision and multimodal capabilities are increasingly supported by foundation models (e.g., image inputs, OCR, bounding boxes, etc.). Use off-the-shelf capabilities if they meet your accuracy/latency requirements.
Patterns mapping to ROI and objective examples
If the objective is 24/7 self-service and reliability, consider conversational pattern or autonomous systems; both can provide ROI through reduced human workload and improved availability.
For improved decision making, predictive analytics and decision-support patterns help reduce risk and optimize outcomes, but require robust validation and governance.
For pattern identification and monitoring, anomaly detection is valuable for fraud detection and predictive maintenance; it often benefits from time-series data and likelihood-based thresholds.
Hyper personalization is powerful for engagement and conversion but may require domain-specific data and privacy controls; may be harder to generalize across populations.
Phase 1: Business Understanding in CPMAI
Core questions to answer before proceeding:
Why this project needs AI? Define a clear business objective for this iteration.
Who do we need to satisfy? Customers, employees, partners, leadership, etc.
Does this objective require a cognitive AI solution? If no, stop or re-scope.
If cognitive, which cognitive parts are required and which can be noncognitive (separate teams, parallel work)?
Which of the seven AI patterns are identified for this project? If none, you’re not doing AI; if multiple, consider staged iterations.
What constraints exist (time, money, resources, data, compliance) and what are the transparency requirements?
The responsible AI considerations (trustworthy AI) must be addressed upfront: bias/fairness, privacy, safety, accountability, governance, transparency, explainability, consent, disclosure, data usage, data rights, and impact on workforce.
Interoperability and environment: define production environment, data infrastructure, and operationalization requirements (batch vs real-time, on-device vs cloud, etc.).
The go/no-go method includes nine green/yellow/red lights across three areas: business feasibility, data feasibility, implementation feasibility. If any critical red light appears, pause and address it before continuing.
The “nine lights” are built around three core questions per area:
Business feasibility: problem definition clarity, organizational willingness to invest/change, sufficient ROI/impact.
Data feasibility: do we have the required data, is there enough data to train, is data quality sufficient.
Implementation feasibility: do we have the required technology and skills, can we implement in time, will the model be usable by intended users.
A powerful practice is to keep iterating; before moving to phase two (data understanding), ensure the business understanding phase has answered the core questions and flagged any roadblocks.
KPIs, model performance, and success criteria
KPIs (Key Performance Indicators): quantifiable measures tied to business outcomes (e.g., usage, ROI, revenue, customer engagement, satisfaction, uptime, error rates).
Model performance metrics: accuracy, precision, recall, F1, false positive rate (FPR), false negative rate (FNR); trade-offs depend on the domain.
Formulas:
ext{Accuracy} = rac{TP + TN}{TP + TN + FP + FN}
ext{Precision} = rac{TP}{TP + FP}
ext{Recall} = rac{TP}{TP + FN}
F1 = 2 imes rac{ ext{Precision} imes ext{Recall}}{ ext{Precision} + ext{Recall}}
Technology performance requirements: training time, compute, memory, latency, scalability, and integration with existing systems. Operationalization requirements: batch vs real-time, deployment environment, and shared-service constraints.
It’s essential to set thresholds for KPIs and model metrics early and determine the tolerances for false positives/negatives. In some contexts, a false positive can be more damaging than a false negative, and vice versa.
Phase 1 also considers ethical, trust, and governance KPIs (e.g., fairness scores, auditability, compliance status, human-in-the-loop requirements).
Roles and teams for AI projects
Four main areas to assemble early:
Business: line-of-business owners, business analysts, solution architects; ensure a real business problem and stakeholder buy-in.
Data Science: domain experts, data scientists, external labeling contributors; translate business questions into hypotheses and analytic ideas.
Data Engineering: data engineers, systems engineers, cloud team; handle data pipelines, storage, and access.
Operationalization: app developers, system/cloud admins; ensure production deployment, monitoring, and maintenance.
Data science concepts and tools:
Core math, statistics, and probability; machine learning fundamentals (regression, classification, clustering) and algorithms.
Data-centric tooling: Python, R; notebooks; data wrangling and preprocessing; model training, testing, validation.
Data science roles: not everyone is a formal “data scientist”; citizen data scientists (no-code/low-code) can contribute using visual tools, but may have limited scope.
Citizen data scientist (or citizen developer): non-traditional data science practitioners who build models using no-code/low-code tools; useful for rapid prototyping but may require governance and guardrails.
Transforming AI projects through iteration and staging
Start with noncognitive or lower-risk components to show early ROI, then progressively introduce cognitive components.
Consider breaking a large AI project into smaller pilots (two-week sprints or similar) to demonstrate incremental ROI and learnings.
The “start small, iterate often” principle aligns with agile practices and reduces risk of large-scale failures.
When data cleansing and labeling are heavy (months), consider smaller, high-value data subsets or third-party/pretrained models to shorten initial iterations.
Pilot vs Proof of Concept:
PoC: tests capabilities in a controlled environment; low adoption in production; often not deployed.
Pilot: real-world, controlled deployment with real data; designed to test scalability and viability in production.
MVP (minimal viable product) is the common Lean-Agile term for the pilot’s practical, real-world functionality.
General guidance: avoid long-running PoCs; go straight for real-world pilots to surface data quality and process issues early and iterate toward ROI.
Examples of CPM AI in practice (phase 1 implications)
Intel smart continuous integration (CI): predict whether code will pass CI tests; reduces turnaround time; demonstrates a well-scoped business problem with green AI go/no-go lights; uses a local-and-then-global test strategy and measures time savings.
Problem: predict pass/fail of code integration tests to reduce unnecessary CI runs.
Data: code commits, test results; needs data governance and integration with Git.
Pattern mapping: predictive analytics + decision support.
NASA aircraft engine predictive maintenance: predict failures and schedule maintenance to minimize downtime and extend engine life.
Data: sensor time-series data (emissions, vibration, turbine sensors, etc.); importance of identical engine types for training data; time-stamped measurements.
Patterns: anomalies/failure detection and predictive analytics; may require multiple models (anomaly detection + lifetime prediction).
Coca-Cola brand content moderation for social media: automatically identify brand content andNSFW content in user-submitted images.
Pattern: recognition; objective: maintain brand safety and regulatory/advertising compliance.
Alternatives considered: human moderation (costly, slow, error-prone); community flagging (unreliable); solution chosen: content-based ML moderation with high sensitivity to avoid false positives and negatives.
Important tradeoffs: accuracy thresholds for brand content vs NSFW content; prioritize safety and brand integrity.
Realities of AI go/no-go and the project lifecycle
The AI go/no-go analysis emphasizes alignment across business feasibility, data feasibility, and technology/execution feasibility.
If a phase reveals critical roadblocks (e.g., data unavailability, regulatory issues, or insufficient ROI), pause and re-scope or abandon the path.
The CPMAI framework is iterative: you can revisit and adjust earlier phase outputs as you collect more data or learn more from pilots.
Patterns vs non-cognitive alternatives across the seven AI patterns
For each pattern, there are non-cognitive alternatives worth considering at early stages to accelerate ROI:
Conversational: use FAQs, menus, IVR, or traditional UIs before deploying AI chat or bot capabilities.
Recognition: rely on humans or outsourcing for simple/categorization tasks; or use traditional BI/analytics for pattern detection where feasible.
Pattern/anomalies: traditional BI/analytics dashboards and rule-based alerts can handle certain anomaly detection tasks.
Predictive analytics: heuristic rules and domain expertise can provide baseline decision support; ML adds predictive power where patterns are complex.
Autonomous systems: consider automation and workflow orchestration before shifting to fully autonomous operations.
Hyper personalization: start with simpler recommendations using rules and A/B testing; advanced personalization can follow.
Goal-driven systems: few practical non-AI alternatives; where it exists, heuristics or human-driven optimization may suffice for simple problems.
Trade-offs to consider when choosing cognitive vs non-cognitive approaches: speed, accuracy, cost, time-to-market, and ability to scale.
The broader takeaway: plan for trustworthy AI from day one
Trustworthy AI considerations (from phase 1 onward): bias/fairness, privacy, safety, human oversight, accountability, transparency, consent and disclosure, data handling, governance, and the potential for misuse.
Interpretable and explainable AI requirements: determine if there are law/regulatory/industry constraints on algorithm choice and how to provide traceability and root-cause analysis for AI decisions.
Transparent AI: define requirements for consent, disclosure, data/model/technology visibility, and use of open-source/open-data in the project.
Governance and auditability: plan for model versioning, system monitoring, contestability, and engagement with external regulatory bodies or credentialing organizations.
Training needs: ensure team has sufficient AI and data knowledge; identify gaps; consider education, tooling, and possibly consultants.
The CPM AI phase 1 example recap
Intel CI example shows how a well-scoped business problem (bi-directional testing optimization) can be approached with CPMAI: define objective, measure success, identify data needs, define go/no-go traffic lights, and plan incremental deployments.
NASA example demonstrates handling time-series sensor data, identifying anomalies and predicting failures for maintenance, and the need to ensure data consistency across similar engines.
Coca-Cola example highlights the importance of recognition-driven moderation with strict false-positive/false-negative tolerances and the trade-offs between human vs machine moderation.
Practical takeaways for exam prep
Always tie the AI solution to a concrete business objective and ROI; avoid solving a problem for which AI has no clear value.
Use the seven patterns as a diagnostic tool to understand what type of AI might help; avoid forcing AI into a problem that maps poorly to any pattern.
Assess both cognitive and non-cognitive components; phase the project to start with noncognitive or lower-risk cognitive steps to gain early ROI and buy-in.
Use the DIKUW framework to reason about where AI adds value: move from data to information to knowledge (ML) and beyond to understanding and wisdom where possible.
Be mindful of data readiness (collection, quality, quantity, access) and technology readiness (skills, tools, environment) before committing to a large AI project.
Distinguish PoC from Pilot; aim for real-world pilots in production rather than lab-only proofs of concept.
Build an AI team with clear roles across business, data science, data engineering, and operationalization; consider citizen data scientists for rapid prototyping under governance.
Define, measure, and monitor KPIs, model performance metrics, and technology requirements early; be explicit about tolerances for errors and their operational impact.
Plan for responsible, trustworthy AI from the outset, including human-in-the-loop where appropriate, fairness/privacy protections, and transparent governance.
Quick reference: common formulas mentioned
Accuracy: ext{Accuracy} = rac{TP + TN}{TP + TN + FP + FN}
Precision: ext{Precision} = rac{TP}{TP + FP}
Recall: ext{Recall} = rac{TP}{TP + FN}
F1 score: F1 = 2 imes rac{ ext{Precision} imes ext{Recall} }{ ext{Precision} + ext{Recall} }
Time-to-ROI or ROI timing: ext{ROI}{time} = rac{Time{before} - Time{after}}{Time{before}}
General ROI definition (net benefit): ROI = rac{Benefit - Cost}{Cost} ext{ or } ROI = rac{Netenefit}{Cost}
Endnotes
The CPMAI framework is designed to be iterative and adaptable; it emphasizes understanding the business problem, data readiness, and trustworthy AI considerations before moving to data modeling and deployment. The goal is to increase the odds of long-term success and avoid projects that fail due to misalignment, poor data, or governance gaps.