Inside Maximo 9.2: How AI Moved from Bolt-On to Built-In

The AI Shift in Maximo Application Suite 9.2

For years, AI in Maximo meant separate applications that you licensed, configured, and deployed alongside the core platform. Maximo Health gave you asset health scores. Maximo Predict gave you failure probability. Maximo Visual Inspection gave you computer vision. These were powerful capabilities, but they lived apart from the daily workflows of maintenance planners, reliability engineers, and field technicians. You had to leave your work order, open a different application, interpret the data, and then come back to act on it.

Maximo Application Suite 9.2, released in June 2026, represents a genuine shift away from that pattern. AI is no longer a separate layer bolted on top of Maximo. It is embedded directly into the workflows that reliability, maintenance, field service, safety, and operations teams use every day. The release expands AI across reliability insights, field execution, safety and compliance workflows, document-based information extraction, and orchestration across systems. It also introduces agentic workflows designed to help guide decisions and move work forward in operationally grounded ways.

The architectural change that enables this is the deeper integration between Maximo and the IBM watsonx platform. Watsonx provides the model serving, the data ingestion, and the agentic orchestration that the Maximo AI capabilities call into. This is not a marketing partnership. It is a technical integration where Maximo's AI capabilities use watsonx as their runtime, and the models served by watsonx are tuned for the specific tasks that asset-intensive operations require.

The five AI capabilities that most organizations will encounter first in MAS 9.2 are Maximo Condition Insight, which provides AI-driven condition-based maintenance recommendations; Maximo Predict, which builds and deploys machine-learning models that forecast days to failure, probability of failure, and remaining useful life; Maximo Monitor (formerly Maximo Health), which provides AI-enabled remote monitoring of asset condition at scale; Maximo Visual Inspection, which uses computer vision models to detect defects from images and video with local inference on mobile devices; and Maximo Assistant on Mobile, which uses natural-language AI to help technicians find asset information, review history, and complete work in the field.

Maximo Condition Insight: The Flagship

The single most important AI capability in MAS 9.2 is Maximo Condition Insight. IBM introduced it in late 2025, and it has matured into a flagship feature by mid-2026. Condition Insight is an agentic AI capability within Maximo Asset Performance Management that interprets asset data to explain asset condition, highlight emerging trends, and recommend corrective actions.

What makes Condition Insight different from the existing Health and Predict applications is that it synthesizes data from multiple sources and generates actionable recommendations. Health gives you a score. Predict gives you a failure probability. Monitor gives you sensor readings. Condition Insight takes all of those, plus inspection results and reliability strategy data, and tells you what to do next.

The data flow works in layers. Maximo Health calculates health scores for assets based on meter readings, work order history, and asset attributes. These scores are numeric values from 0 to 100 that represent the current condition of an asset relative to its expected baseline. Maximo Predict applies machine learning models to historical failure data to generate failure probability scores and estimated failure dates. These models are trained using Jupyter notebooks provided with the platform, and data scientists can customize the models for specific asset types or failure modes. Maximo Monitor collects IoT sensor data from connected assets, providing real-time condition data that feeds into health calculations and anomaly detection. Condition Insight synthesizes all of the above, plus inspection results and reliability strategy data, to identify patterns that no single data source would reveal on its own.

# Example: Condition Insight recommendation structure
# (Output from the Condition Insight engine)

recommendation = {
    "asset": "PUMP-P-101",
    "site": "BEDFORD",
    "condition_score": 42,  # 0-100, lower = worse
    "trend": "DECLINING",   # DECLINING, STABLE, or IMPROVING
    "predicted_failure_date": "2026-09-12",
    "failure_probability_30day": 0.34,
    "contributing_factors": [
        {
            "factor": "Vibration trending upward",
            "source": "Monitor (IoT sensor)",
            "detail": "RMS velocity increased 18% over 14 days, "
                      "currently 0.42 IPS (threshold: 0.45 IPS)"
        },
        {
            "factor": "Bearing temperature elevated",
            "source": "Monitor (IoT sensor)",
            "detail": "Operating temp 78C, historical avg 71C, "
                      "rate of increase accelerating"
        },
        {
            "factor": "Last PM overdue by 12 days",
            "source": "Maximo Manage (work order history)",
            "detail": "Scheduled quarterly PM was due 2026-07-20, "
                      "currently open but not started"
        },
        {
            "factor": "Similar failure pattern identified",
            "source": "Predict (ML model)",
            "detail": "3 similar assets failed within 30 days of "
                      "showing this vibration+temperature pattern"
        }
    ],
    "recommended_action": "Schedule bearing inspection within 5 days. "
                          "Consider preemptive bearing replacement "
                          "if inspection confirms wear.",
    "priority": "HIGH",
    "confidence": 0.87
}

The recommendation includes a confidence score, which reflects how much historical data the model had to work with. Assets with years of consistent meter data and failure records produce high-confidence recommendations. Assets with sparse data or recent additions to the system produce lower-confidence recommendations. The confidence score is critical for operational adoption because it gives maintenance planners a basis for deciding whether to act on a recommendation immediately or wait for more data.

Condition Insight does not replace Health and Predict. It sits on top of them, adding a layer of pattern recognition and recommendation generation that the individual applications cannot produce on their own. If you have already deployed Health and Predict, Condition Insight leverages that investment. If you have not, Condition Insight gives you a reason to, because the combined output is more valuable than the sum of the parts.

Predict and Health: The Predictive Foundation

Maximo Predict uses machine learning models to predict asset failures before they occur. The models are trained on historical data, including work orders with failure codes, meter readings, sensor data from Monitor, and asset characteristics. The output is a failure probability score for each asset, typically calculated daily, that indicates the likelihood of failure within a specified window.

The training process uses Jupyter notebooks that ship with the platform. Data scientists can customize these notebooks for specific asset types, failure modes, or operational contexts. The notebooks handle feature engineering, model training, validation, and deployment. The models support several prediction types: current failure probability (what is the likelihood this asset fails today), failure date prediction (when is this asset most likely to fail), and remaining useful life estimation (how many operational hours remain before expected failure).

# Example: Maximo Predict training notebook structure
# (Simplified - actual notebooks include data validation,
# feature engineering, hyperparameter tuning, and model validation)

import pandas as pd
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.model_selection import train_test_split

# 1. Load historical data from Maximo
work_orders = pd.read_sql("""
    SELECT wonum, assetnum, failurecode, faildate, 
           siteid, assettype, installDate
    FROM maximo.workorder 
    WHERE historyflag = 1 AND failurecode IS NOT NULL
""", maximo_connection)

meter_readings = pd.read_sql("""
    SELECT assetnum, metername, readingvalue, readingdate
    FROM maximo.assetmeter
    WHERE readingdate >= '2023-01-01'
""", maximo_connection)

# 2. Feature engineering
# Calculate rolling averages, rates of change, time-since-last-PM
features = engineer_features(work_orders, meter_readings)

# 3. Define failure labels
# Asset "failed" if it had a corrective work order within 30 days
labels = create_failure_labels(work_orders, window_days=30)

# 4. Train model
X_train, X_test, y_train, y_test = train_test_split(
    features, labels, test_size=0.2, random_state=42
)

model = GradientBoostingClassifier(
    n_estimators=200,
    max_depth=5,
    learning_rate=0.1
)
model.fit(X_train, y_train)

# 5. Evaluate
accuracy = model.score(X_test, y_test)
print(f"Model accuracy: {accuracy:.2%}")

# 6. Deploy model to Maximo Predict
# The model is registered with watsonx and deployed
# Predict scores assets daily using the deployed model

Maximo Health, now folded into what MAS 9.2 calls Maximo Monitor, provides the real-time condition monitoring layer. It collects IoT sensor data from connected assets, calculates health scores, and feeds anomaly detection. The integration between Health and Predict is bidirectional: Health provides the real-time condition data that Predict needs for feature engineering, and Predict provides the failure probability scores that Health uses to prioritize alerts.

The practical implementation sequence for organizations starting their Maximo AI journey in 2026 is to audit the data foundation first. Look at the asset hierarchy, the meter history, the work order history, and the failure records. Identify the asset classes where the data is complete and consistent enough to support AI. Start with Monitor and Predict on a single asset class, because these capabilities have the most mature data requirements and produce measurable value within the first six months. Then add Visual Inspection for high-value inspection types, pilot Condition Insight on the same asset classes where Predict and Monitor are deployed, and finally roll out Maximo Assistant on Mobile to the field technician population.

Visual Inspection and Mobile AI

Maximo Visual Inspection uses computer vision models to detect defects from images and video. The capability supports both cloud-based inference for batch processing of inspection images and local inference on mobile devices for real-time defect detection in the field. The local inference capability is particularly valuable for field operations where network connectivity is unreliable or where the inspection is time-critical.

The models are trained on labeled image datasets that organizations create by annotating inspection photos. For example, a utility might train a model to detect rust on transmission towers by labeling hundreds of inspection photos with bounding boxes around rust areas. The trained model can then analyze new inspection photos and flag areas that show similar patterns.

# Example: Training a Visual Inspection model for corrosion detection

# 1. Collect and label images
# Inspectors label photos from previous inspections:
# - Draw bounding boxes around corrosion areas
# - Classify severity: surface, moderate, severe
# - Tag location: tank floor, tank wall, pipe joint, support

# 2. Train model in Maximo Visual Inspection
# Upload labeled dataset
# Select model architecture (MobileNet, EfficientNet, or custom)
# Set training parameters:
#   - epochs: 50
#   - learning_rate: 0.001
#   - augmentation: flip, rotate, brightness adjustment

# 3. Evaluate model performance
# Test on held-out validation set
# Target metrics:
#   - Precision: >0.85 (few false positives)
#   - Recall: >0.80 (catch most real defects)
#   - mAP: >0.75 (overall detection quality)

# 4. Deploy model
# Cloud: for batch processing of uploaded inspection photos
# Mobile: export as TensorFlow Lite model for on-device inference

# 5. Field workflow
# Inspector takes photo with mobile device
# Model runs inference locally in <2 seconds
# If defect detected: auto-create inspection finding in Maximo
# If no defect: log as clean inspection

The mobile workflow is where Visual Inspection delivers the most operational value. An inspector photographs a pump bearing with a mobile device, the model runs inference locally in under two seconds, and if corrosion or wear is detected, the system automatically creates an inspection finding in Maximo with the photo, the defect location, and the severity classification. The inspector does not need to manually enter the finding, and the consistency of defect detection improves because it is not dependent on the inspector's individual experience level.

Maximo Assistant and the Agentic Shift

Maximo Assistant is the most visible AI capability in MAS. Powered by IBM watsonx, it enables users to interact with asset data using natural language. Instead of navigating through application menus and running reports, users ask questions and get answers. The assistant has evolved rapidly. As of November 2025, IBM deprecated the Granite 3.2 8B Instruct model in favor of the gpt-oss-120b model, which provides significantly better accuracy on complex queries.

# Example: Maximo Assistant interactions
# Natural language queries that the assistant handles

User: "What work orders are overdue at BEDFORD site?"
Assistant: "There are 12 overdue work orders at BEDFORD:
  - WO12347: Bearing inspection on PUMP-P-101 (14 days overdue)
  - WO12351: Transformer oil sample (10 days overdue)
  - WO12355: Monthly safety inspection (8 days overdue)
  ... [shows all 12]
Would you like me to prioritize these by criticality?"

User: "Show me assets with health scores below 50"
Assistant: "17 assets have health scores below 50:
  - PUMP-P-101: Score 42, declining trend, predicted failure
    in 38 days. Recommended action: bearing inspection.
  - MOTOR-M-204: Score 38, stable but critical. Last PM
    was 6 months overdue.
  - TRANSFORMER-T-15: Score 45, declining. Oil analysis
    shows elevated dissolved gases.
  ... [shows all 17]
The highest-risk assets are PUMP-P-101 and MOTOR-M-204."

User: "Create a work order for PUMP-P-101 bearing inspection"
Assistant: "Created work order WO12378:
  - Asset: PUMP-P-101
  - Description: Bearing inspection - Condition Insight recommendation
  - Priority: 1 (HIGH)
  - Work Type: CM
  - Scheduled Start: 2026-08-06
  - Assigned to: MECH-TEAM-1
Would you like to add any additional tasks or materials?"

The agentic shift in MAS 9.2 goes beyond the conversational assistant. Agentic workflows can autonomously guide decisions and move work forward. An agent can monitor asset conditions, generate work order recommendations, route them for approval, and track completion. The agents operate within the Maximo security framework, respecting user permissions and approval workflows. They do not bypass human oversight; they augment it.

The distinction between an AI assistant and an AI agent is important. An assistant answers questions and provides information. An agent takes action within defined boundaries. MAS 9.2 introduces agents that can operate within scoped workflows, such as monitoring a set of assets, generating recommendations, and creating draft work orders for planner review. The planner still approves the work order, but the agent has done the analysis and preparation work that previously consumed hours of manual effort.

The Watsonx Foundation

The watsonx platform provides the infrastructure that makes the Maximo AI capabilities possible. Watsonx has three components: watsonx.ai for model training and serving, watsonx.data for data ingestion and governance, and watsonx.governance for AI governance and compliance. Maximo's AI capabilities use all three.

The model serving layer in watsonx.ai hosts the large language models that power Maximo Assistant, the machine learning models that power Predict, and the computer vision models that power Visual Inspection. Models can be served from the cloud or deployed to edge devices for local inference. The flexibility is important for organizations with data sovereignty requirements or unreliable network connectivity.

Watsonx.data handles the data pipeline that feeds the AI models. It ingests data from Maximo business objects, IoT sensors, inspection results, and external systems, then makes it available for model training and inference. The data governance component ensures that the data used for AI is auditable, with lineage tracking that shows where data came from and how it was transformed.

For organizations concerned about AI governance and compliance, watsonx.governance provides the framework for monitoring model performance, detecting bias, and maintaining audit trails. This is particularly important in regulated industries where the decisions made by AI models need to be explainable and auditable.

Practical Implications

For organizations starting their Maximo AI journey, the recommended sequence is clear. First, audit the data foundation. Look at the asset hierarchy, the meter history, the work order history, and the failure records. Identify the asset classes where the data is complete and consistent enough to support AI. This audit will reveal that some asset classes have years of clean data and are ready for AI, while others have sparse data and need foundational work before AI can add value.

Second, start with Monitor and Predict on a single asset class. These capabilities have the most mature data requirements and produce measurable value within the first six months. Choose an asset class that is critical to operations, has good historical data, and has a supportive maintenance team that will act on the predictions.

Third, add Visual Inspection for high-value inspection types where the cost of missing a defect justifies the effort of training a model. Fourth, pilot Condition Insight on the same asset classes where Predict and Monitor are deployed, because Condition Insight builds on the data and models that Predict and Monitor produce.

Fifth, roll out Maximo Assistant on Mobile to the field technician population. The assistant's natural language interface is particularly valuable for technicians who need information quickly and do not want to navigate through application menus on a mobile device. Sixth, mature the agentic workflows over time, starting with scoped use cases like asset monitoring and work order recommendation generation.

Bottom Line

MAS 9.2 is the most consequential AI release in Maximo's history. Condition Insight, Predict, Visual Inspection, and Maximo Assistant are not experimental features. They are production capabilities that IBM has deployed across hundreds of organizations. The shift from bolt-on AI to built-in AI changes how maintenance, reliability, and field service teams work. Start with data quality, pick a high-value asset class, and build from there. The organizations that follow the phased adoption sequence are the ones that will see measurable results within the first year. The ones that try to deploy everything at once will struggle with data quality issues and change management fatigue that delay value realization for years.