PMI CPMAI cheat sheet

AI augments project management by automating scheduling, forecasting, and reporting, but accountability for scope, cost, and quality remains with the human project manager.

Predictive models estimate schedule and cost outcomes from historical project data; their accuracy depends on data volume, feature quality, and similarity between past and current projects.

Generative AI is suited to drafting status reports, meeting summaries, and risk registers; outputs are drafts requiring human verification before stakeholder distribution.

Portfolio-level AI prioritizes projects by expected value, strategic fit, and risk exposure; it does not replace governance boards that approve funding.

Rule of thumb: use AI for repeatable, data-rich, low-ambiguity tasks; reserve human judgment for novel, high-stakes, or politically sensitive decisions.

AI-assisted estimation reduces optimism bias but introduces automation bias, the tendency to accept model output without challenge; mitigation is documented challenge and sensitivity checks.

Benefit realization tracking requires baseline metrics defined before deployment; AI value claims without a pre-AI baseline cannot be validated.

Common PM use cases: resource forecasting, schedule risk simulation, sentiment analysis of stakeholder feedback, and anomaly detection in progress data.

The CRISP-DM phases are business understanding, data understanding, data preparation, modeling, evaluation, and deployment; the cycle is iterative, not strictly linear.

An ML project lifecycle adds MLOps stages beyond model build: data pipeline, training, validation, deployment, monitoring, and retraining.

Data preparation typically consumes the largest share of project effort; plans that budget only modeling time understate schedule and cost.

Train/validation/test splitting must keep test data unseen during tuning; leakage inflates reported accuracy and causes production underperformance.

Model acceptance criteria must be defined before development: metric thresholds, fairness constraints, latency, and drift tolerance.

Deployment options are batch, real-time API, and edge; choice is driven by latency requirements, data volume, and privacy constraints.

Post-deployment monitoring tracks data drift, concept drift, and performance decay; a model without a retraining trigger is an incomplete deliverable.

A minimum viable model delivered early to validate feasibility reduces sunk cost compared with a long build before first evaluation.

Data quality dimensions are accuracy, completeness, consistency, timeliness, validity, and uniqueness; each needs a defined threshold and owner.

Data lineage records origin, transformation, and movement of data; it is required to investigate model errors and satisfy audit requests.

Risk management follows identify, assess, mitigate, monitor; likelihood-times-impact scoring prioritizes the AI risk register.

Key AI risks: bias, privacy breach, hallucination, model drift, adversarial input, intellectual property exposure, and third-party dependency.

Role separation: data owner is accountable for a domain, data steward executes quality controls, and the model owner is accountable for model performance in production.

Regulatory mapping matters: the EU AI Act uses risk tiers (unacceptable, high, limited, minimal), with high-risk systems requiring documentation, human oversight, and logging.

Personal data used in training requires a lawful basis, minimization, and retention limits; anonymization must be tested, since re-identification is often feasible.

Human-in-the-loop controls are mandatory where errors cause material harm; the control must define who reviews, on what trigger, and with what authority to override.

Adoption follows stages: awareness, pilot, scaled deployment, and institutionalization; skipping pilot evidence increases failure risk at scale.

Technology acceptance is driven by perceived usefulness and perceived ease of use; both must be addressed in change plans.

Ethical AI principles commonly cited: fairness, transparency, accountability, privacy, safety, and human oversight.

Explainability requirements scale with decision impact; a credit or hiring decision needs more justification than an internal scheduling suggestion.

AI literacy programs should be role-based: executives need risk and value framing, practitioners need tooling and data handling, all staff need acceptable-use rules.

Resistance often stems from job-security concerns; transparent communication about role changes and reskilling paths reduces adoption friction.

An AI governance committee typically owns policy, use-case approval, and exception handling; it needs authority, a charter, and a defined review cadence.

Measure adoption with usage rates, time saved, error reduction, and override frequency; override frequency is an early signal of model distrust or poor fit.