What are the different MLOps maturity levels (Level 0, Level 1, and Level 2)?

MLOps maturity levels describe how much of a team's machine learning workflow is automated—from manually training and releasing models to automatically testing and delivering entire ML pipelines.

A commonly used framework from Google Cloud describes three levels: Level 0, Level 1, and Level 2. These levels are a useful guide, not a universal certification or a requirement that every company must reach Level 2.

A practical example

Imagine an online store that recommends products. Customer preferences change with seasons, new products, and special offers. The store wants its recommendation model to stay useful as those changes happen.

Level 0 — Manual ML process

A data scientist downloads customer data, prepares it in a notebook, trains a model, and sends the model file to an engineer for deployment.

When recommendations become outdated, someone must repeat most of those steps manually.

• What is manual? Data preparation, training, evaluation, and handoffs between people.
• Main challenge: Updates can be slow, difficult to reproduce, and dependent on individual team members.
• When it may be enough: A small proof of concept or a model that rarely changes.

Level 1 — Automated ML pipeline

Instead of running every step separately, the store builds a training pipeline that connects data validation, preparation, training, and model evaluation.

For example, a scheduled run can use the latest approved customer data, train a new recommendation model, and check whether it meets the release criteria. With suitable validation and deployment controls, approved versions can be delivered to the application.

• What improves? Repeatable training, fewer manual handoffs, and more frequent model updates.
• What still needs work? Changing the *pipeline code itself* may still require manual testing and deployment.
• Key idea: Automate the flow that creates new models. This is often called Continuous Training (CT).

Level 2 — Automated ML pipeline plus CI/CD

Now the store also automates how changes to the ML pipeline are tested and released.

Suppose an engineer improves the code that prepares customer data. When that code is committed:

1. Continuous Integration (CI) checks the updated code and pipeline components.
2. Continuous Delivery (CD) packages and delivers tested pipeline changes through the release process.
3. The updated training pipeline can run when triggered and produce a candidate model.
4. Model validation and release controls determine whether the candidate should serve predictions.

• What improves? Both the *model-building process* and the *delivery of changes to that process* become more reliable.
• Key idea: Automate training pipelines and their software release process.

Remember the difference

• Level 0: People manually run the ML steps.
• Level 1: A pipeline automatically runs the ML steps.
• Level 2: CI/CD also tests and delivers changes to the pipeline.

Important: Automation does not eliminate review, monitoring, or safeguards. A newly trained model should not replace a production model simply because its training finished.

In short: MLOps maturity is about making the journey from data to a reliable model increasingly repeatable, testable, and safe—not about buying more tools.