ML-specific project management methodologies
Instruction and application
CompleteIt is not only the model that needs tuning - your project management approach should be just as adaptive. ML involves uncertainty, iteration and changing data. This section compares practical frameworks - from CRISP-DM to Agile, Waterfall and Kanban - and shows how MLOps and iteration planning keep delivery on track.
Choosing the right project approach
Not every methodology fits ML out of the box. You need approaches that tolerate evolving data, shifting metrics and experimental phases.
Methodology snapshots
CRISP-DM (cross-industry standard process for data mining)
-
Use for: Framing the problem and structuring early ML work.
-
Best fit: Discovery, business alignment, modelling and evaluation cycles.
-
How to use: Move through business understanding, data understanding, preparation, modelling, evaluation and deployment; keep business value central.
-
Example: A telecom firm uses CRISP-DM to understand churn drivers before building models.Agile / Scrum
-
Use for: Iterative model development with frequent feedback.
-
Best fit: Goals that evolve; cross-functional collaboration.
-
How to use: Time-boxed sprints (e.g. baseline model, feature experiments); include product, data and business in planning and retrospectives.
-
Example: A healthcare team runs two-week sprints to refine a risk model based on evaluation results.Waterfall
-
Use for: Highly regulated or fixed-scope deliverables.
-
Best fit: Compliance-heavy systems where change cost is high.
-
How to use: Linear plan, build, test, deploy with signed-off requirements up front.
-
Example: A government agency delivers an audited credit-scoring workflow with strict gates.Kanban
-
Use for: MLOps, monitoring-driven work and continuous improvement.
-
Best fit: Retraining, A/B tests, production support.
-
How to use: Visual board (backlog, in progress, review, deployed); limit WIP; prioritise operational work.
-
Example: A streaming service updates recommendations weekly from behaviour shifts.
| Methodology | Best for | Strengths | Limitations |
|---|---|---|---|
| CRISP-DM | Data-heavy projects with clear business goals | Structured; keeps business needs central | Less explicit on roles and calendar planning |
| Agile/Scrum | Collaborative build-and-test cycles | Flexible; frequent stakeholder feedback | Sprints may clash with open-ended experiments |
| Waterfall | Fixed-scope or compliance-led ML | Simple milestones; predictable phases | Inflexible when data or performance shifts |
| Kanban | Ongoing MLOps and incremental updates | Visual flow; supports continuous delivery | Needs explicit prioritisation on top of the board |
Tip
In practice, combine methods: e.g. CRISP-DM for framing, then Agile for iterative build, then Kanban for production iteration. Adapt the process to ML's mix of experimentation and operations.
Go deeper on CRISP-DM
For the classic process description, see the Cross-industry standard process for data mining overview and your organisation's preferred reference (many teams still use the original CRISP-DM phase names as a shared vocabulary).
MLOps in project planning
MLOps applies DevOps ideas to ML so models stay reliable, fair and effective after launch. Planning for MLOps early reduces handoff friction between data science and engineering and avoids silent degradation in production.
MLOps helps you manage:
- Version control for data, code and models.
- Automated testing and retraining pipelines.
- Deployment paths from development to production.
- Monitoring and alerts for drift, latency and quality.
- Fairness and explainability expectations for governance.
Example
A retailer launches product recommendations. Seasonal shifts reduce relevance. The team adds data validation, a two-week retraining cadence and a dashboard that alerts when click-through drops - avoiding a full rebuild.
Key point
Treat MLOps as operational glue after launch: shared language between teams, less rework, fewer "it worked in the notebook" surprises.
Managing iteration in ML projects
ML progress is rarely linear. Small data or scope changes can move metrics a lot. Iteration is part of success, not a sign of failure.Why iteration matters
- Performance can jump or fall from modest pipeline changes.
- Experiments need time for tuning, features and assumptions.
- Failing fast on hypotheses is how you learn what works.
Practices that help
- Flexible milestones: e.g. "baseline achieved", "three modelling experiments complete", "feature set v1 for first use case".
- ML-aware frameworks: CRISP-ML(Q) extends CRISP-DM with explicit quality assurance and lifecycle checkpoints (see also the CRISP-ML(Q) paper PDF).
- Visible experiment tracking: MLflow, Weights & Biases or structured spreadsheets for what changed, what ran and what worked.
- Redefine "done": e.g. "meets accuracy and bias bars", "fallback and monitoring in place", "stakeholder sign-off for limited pilot" - not infinite polish.
Action item: Choose the right project management approach
Select the best answers below.
- A. Pure Waterfall with no iteration after requirements sign-off
- B. CRISP-DM (possibly combined with Agile for delivery cadence)
- C. Kanban only, with no explicit discovery phases
Feedback: CRISP-DM centres business and data understanding; you can layer Agile for sprint-based execution.
- A. One-shot Waterfall release only
- B. A single six-month requirements phase
- C. Kanban (or similar flow) with MLOps practices
Feedback: Operational ML benefits from continuous flow, visibility and explicit maintenance work.
- A. Ensure reliable delivery, monitoring and iteration after models ship
- B. Replace the need for any project charter or milestones
- C. Guarantee model accuracy never changes in production
Feedback: MLOps focuses on sustainable operations: pipelines, monitoring, retraining and governance - not a one-time build.