AI in Maximo: From Health Scores to Agentic Workflows - A Practitioner's Guide to Maximo Predict, Health, and Visual Inspection

AI in Maximo Is Not a Single Feature

When people say "AI in Maximo," they usually mean one of three things: a health score on an asset record, a prediction that an asset is about to fail, or a defect detected in a photograph. The reality is that Maximo Application Suite (MAS) has multiple AI-capable applications, each with its own model library, its own data pipeline, and its own deployment story.

This article walks through the AI landscape inside MAS as of mid-2026. We cover the core applications, the SDK that powers them, the default model templates, and the patterns that work in production.

The AI Application Stack

The Maximo Application Suite contains four applications that have meaningful AI capabilities. Each one is a separate deployable component, but they share a common data lake, a common asset model, and a common user identity model.

Maximo Health

Maximo Health is the simplest of the four. It takes asset data (failure history, condition readings, PM compliance, age, criticality) and produces a single health score between 0 and 100 for each asset. The score is calculated by a rule-based engine that considers asset type, environment, and operating context, and it can be tuned per industry or per asset class.

Health is intentionally not predictive. It tells you the _current_ state of the asset based on what you already know. The output is a number, a color (green, yellow, red), and a list of contributing factors. The contributing factors are the most useful piece: they tell you _why_ the asset's health is what it is, so the planner can make an informed decision.

code block

Maximo Predict

Maximo Predict is where the actual machine learning happens. It is a separate deployable that runs in Watson Machine Learning, with a Python SDK called `pmlib` that handles data ingestion, model training, model registration, and scoring.

Predict includes five prebuilt model templates that cover the most common predictive maintenance use cases:

| Model Type | What It Predicts | Typical Use Case |
|---|---|---|
| Failure Probability | Probability of failure in next 30/60/90 days | Risk-based maintenance scheduling |
| Time to Failure (RUL) | Estimated remaining useful life in days | End-of-life planning, capital budgeting |
| Anomaly Detection | Operating context-specific anomalies | Early warning of unknown failure modes |
| End of Life Curve | Asset degradation curve over time | Long-term replacement forecasting |
| Failure Mode Specific | Probability of specific failure mode (e.g., bearing failure) | Targeted maintenance strategies |

Each template is a Jupyter notebook that can be customized. The data scientist extends the notebook with domain-specific features, retrains the model on the customer's data, and registers the trained model back into Maximo Predict. Once registered, the model runs on a schedule (e.g., nightly) and produces predictions for every asset in the configured asset group.

Maximo Visual Inspection

Maximo Visual Inspection is the computer vision application. It takes images (or video frames) from cameras, drones, or handheld devices and detects defects using deep learning models. The output is a list of detected defects with bounding boxes, confidence scores, and severity classifications.

The typical use cases are visual inspection tasks that are dangerous, repetitive, or require a trained eye:

- Drone-based inspection of transmission line insulators
- Camera-based inspection of conveyor belts for tears or misalignment
- Handheld inspection of pressure vessel welds for cracks
- Robotic inspection of substation equipment for oil leaks or damage

Visual Inspection is built on top of Maximo Computer Vision, which is a separately licensed IBM product. The integration with Maximo Manage is through the standard Object Structure and Integration Framework, so a detected defect can automatically create a work order, attach the annotated image, and route it to the right craft.

Maximo Assist (Agentic Workflows)

The newest addition to the AI application stack is Maximo Assist, which brings agentic AI capabilities into the platform. Maximo Assist is a chat-based interface that allows technicians, planners, and supervisors to interact with Maximo using natural language.

The agents in Maximo Assist are built on watsonx.ai and can be configured to perform specific tasks:

- Work Order Agent: Creates, updates, and queries work orders based on natural language requests. "Create a priority 2 work order for TX-2014-087 to investigate the rising DGA hydrogen."
- Asset History Agent: Answers questions about an asset's history. "When was the last time we replaced the oil in TX-2014-087?"
- Predictive Insights Agent: Surfaces and explains predictions from Maximo Predict. "What assets in the North region are most likely to fail in the next 30 days?"
- Procedure Agent: Walks a technician through a complex procedure step by step, answering questions and capturing data along the way.

The agentic workflows are still maturing, but they represent the direction IBM is heading. The vision is that Maximo becomes less of a system you _use_ and more of a system you _collaborate with_.

The pmlib SDK: How Predict Actually Works

For practitioners, the most important piece of the AI stack is the `pmlib` Python SDK. It is the API through which data scientists build, train, register, and deploy predictive models into Maximo Predict.

Installation and Setup

`pmlib` is installed in a Watson Studio notebook environment. The latest MAS releases ship `pmlib` as a zip file that can be uploaded to Watson Studio and installed with `pip install`.

code block

Creating an Asset Group

Before you can train a model, you need an asset group. The asset group defines the population of assets that the model will score. It is a logical grouping based on classification, location, or any other Maximo attribute.

code block

Loading Training Data

`pmlib` provides helper functions to pull training data from Maximo Manage and the Maximo Data Lake. The data scientist does not need to know the underlying MBO structure; the SDK handles the MBO-to-pandas translation.

code block

Training a Model

The five prebuilt model templates are exposed as `ModelPipeline` classes. You instantiate the template, fit it on your data, and evaluate the results.

code block

Registering and Deploying the Model

Once the model meets your accuracy thresholds, you register it back into Maximo Predict and enable it for continuous scoring.

code block

From this point on, the model runs on the configured schedule. For every asset in `TX_138KV_NORTH`, Maximo Predict produces a failure probability score, a time-to-failure estimate, and any detected anomalies. These predictions are visible in the Maximo Manage asset record, in the Predict UI, and in the Maximo Monitor dashboards.

The Data You Need

The single biggest predictor of model quality is data quality. Before you invest in custom models, make sure you have the data to support them.

Asset Master Data

Every asset in the model needs accurate master data: install date, manufacturer, model, rated capacity, classification, location, and criticality. Garbage in, garbage out is the rule.

Failure History

Failure history is the labeled training data. Maximo Predict learns from past failures, so the more accurate and complete your failure records, the better the model. This means capturing failure dates, failure modes, root causes, and resolution actions in Maximo Work Orders with discipline.

If your failure history is incomplete (because your team is still using spreadsheets or paper), the models will be weak. Fix the data first, then build the models.

Sensor Data

For anomaly detection and RUL models, you need time-series sensor data. This comes from your IoT platform (typically Watson IoT Platform or another MQTT broker) and lands in the Maximo Data Lake.

The minimum useful sensor coverage depends on the asset class, but the rule of thumb is:

- 12-24 months of historical sensor data for initial training
- Continuous streaming for ongoing scoring
- Sensor data at a granularity that captures the relevant failure modes (e.g., vibration data at 1 kHz for bearing analysis)

Maintenance History

The model uses PM compliance, PM frequency, and corrective work order history as features. If your PM records are accurate, the model can learn that assets with poor PM compliance are more likely to fail. If your PM records are noisy, the model will be confused.

Model Selection: Which Template to Use

Choosing the right model template is the most important decision after data quality. Here is a decision tree that works for most situations.

| Goal | Recommended Template | Why |
|---|---|---|
| Identify assets that need attention this month | Failure Probability (30 day) | Actionable, time-bounded, easy to integrate into work planning |
| Plan capital replacement over 5 years | End of Life Curve | Long horizon, useful for budget planning |
| Detect unknown issues | Anomaly Detection | Does not require labeled failure data, catches novel failure modes |
| Diagnose specific failure modes | Failure Mode Specific | Requires labeled failure mode data, very high value when available |
| Schedule PM based on condition | Time to Failure (RUL) | Directly drives PM scheduling decisions |

The most common starting point is the 30-day Failure Probability model. It is the easiest to interpret, the easiest to validate against actual failures, and the easiest to translate into work orders. Once you have it working, you can layer in the more sophisticated models.

Closing the Loop with Automation

A prediction that does not lead to action is a wasted prediction. The whole point of running a model is to influence maintenance decisions. In MAS, this is done through automation scripts that watch the prediction outputs and take action when thresholds are crossed.

code block

This pattern (model produces prediction, automation script watches prediction, action is taken) is the closed-loop workflow that makes predictive maintenance real. Without it, the predictions are just numbers on a dashboard.

Common Pitfalls

After several years of Maximo Predict deployments in production, a few common mistakes repeat.

Pitfall 1: Too few assets in the group. The default model templates need at least 50-100 assets in the group to produce meaningful predictions. If you have only 10 critical pumps, the model will overfit and produce noise. In that case, consider combining similar asset classes or using a simpler rules-based approach until you have more data. Pitfall 2: Insufficient failure history. The models need labeled failure data to learn from. If you have only two failures in the last five years, the model cannot distinguish between assets that are reliable and assets that are about to fail. Fix the failure recording process first. Pitfall 3: Ignoring data drift. Asset populations, operating conditions, and failure modes change over time. A model trained on 2022 data may be wrong by 2026. Schedule regular retraining (at least annually) and monitor prediction accuracy over time. Pitfall 4: Predictions without action. If the predictions do not flow into work orders, the i

...