AI Inside Maximo: From Predictive Maintenance to Agentic Workflows

The AI Foundation: Predictive Maintenance with Maximo Health and Predict

AI in Maximo is not a single feature or a bolt-on module. It is a layered set of capabilities that span from basic condition monitoring to advanced predictive analytics and, as of MAS 9.2, agentic AI workflows. Understanding the progression from descriptive to predictive to prescriptive is essential for maintenance leaders who want to leverage these capabilities effectively. Each layer builds on the previous one, and attempting to skip stages typically results in unreliable predictions and frustrated users.

Maximo Health is the entry point for AI-driven asset management. It provides a consolidated, real-time view of asset condition by combining data from multiple sources: IoT sensors, meter readings, asset records, work history, inspection results, and even external data like weather conditions. The health scoring algorithm evaluates these inputs to produce a single health score for each asset, allowing maintenance teams to prioritize their attention based on actual condition rather than arbitrary schedules.

The health score is not a simple threshold. It is a composite metric that weighs multiple factors, including asset age, failure history, recent sensor readings, maintenance compliance, and criticality. An asset with a clean maintenance record but anomalous vibration readings will score differently than an asset with a history of failures but normal sensor data. The scoring model can be configured to reflect organizational priorities, such as weighting safety-critical assets more heavily than convenience equipment. This configurability is essential because every organization has different risk tolerances and operational priorities.

Maximo Predict builds on this foundation by applying machine learning algorithms to historical data. It looks for patterns in past failures, sensor readings leading up to those failures, maintenance interventions, and environmental conditions. After training on sufficient historical data, the predictive models can estimate the probability of failure within a given time window and the expected failure date for each asset. This transforms maintenance planning from a reactive activity to a predictive one, where work can be scheduled at the optimal time to prevent failure without wasting remaining useful asset life.

The architecture of Maximo Predict is worth understanding in detail. Data scientists create groups of assets with similar characteristics, then use default Jupyter notebooks provided by IBM to build and train predictive models. The notebooks are pre-configured with common machine learning pipelines, including data preparation, feature engineering, model training, and validation. The models are trained on the group's historical data, including failure events, maintenance records, and operational data. After a model is trained and deployed, predictions are populated for each asset in the group. Maintenance managers can use work queues to track assets that have a high probability of failure or assets predicted to fail before the next scheduled preventive maintenance work order.

# Simplified example: Training a predictive model for pump failures
# Maximo Predict uses Jupyter notebooks integrated with MAS

import pandas as pd
from sklearn.ensemble import RandomForestClassifier
from sklearn.model_selection import train_test_split
from sklearn.metrics import classification_report

# Load historical asset data from Maximo via REST API
asset_data = pd.read_json('https://mas-host/maximo/oslc/os/mxasset'
                          '?oslc.select=assetnum,description,status,'
                          'failurecode,installdate,location')

# Load work order history for failure events
wo_data = pd.read_json('https://mas-host/maximo/oslc/os/mxwo'
                       '?oslc.where=historyflag=true&'
                       'oslc.select=wonum,assetnum,worktype,'
                       'failurecode,actlabhrs,actmatcosts')

# Merge and prepare features
merged = asset_data.merge(wo_data, on='assetnum')
features = ['age_years', 'pm_compliance_pct', 'avg_vibration',
            'avg_temperature', 'operating_hours', 'failure_count_12mo']
X = merged[features]
y = merged['will_fail_90d']  # Binary label: failure within 90 days

X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2)

model = RandomForestClassifier(n_estimators=100, random_state=42)
model.fit(X_train, y_train)

# Validate model performance
predictions = model.predict(X_test)
print(classification_report(y_test, predictions))

# Deploy model to Maximo Predict
# Predictions will populate for each asset in the group
# Work queues will surface high-risk assets automatically

One critical caveat: predictive models are only as good as the data they are trained on. Organizations with poor historical maintenance records, inconsistent failure coding, or sparse sensor data will produce unreliable predictions. Before deploying Maximo Predict, assess your data quality by examining failure code consistency, sensor data coverage, and the completeness of work order records. Investing in data cleanup before model training will produce far better results than attempting to train models on noisy, incomplete data.

Work Order Intelligence: Generative AI for Maintenance Data Quality

MAS 9.0 introduced Maximo Work Order Intelligence, a generative AI capability powered by IBM watsonx that addresses one of the most persistent problems in maintenance management: poor data quality in work order records. Work order descriptions are often vague, failure codes are frequently missing or incorrect, and the information needed to triage and plan maintenance is buried in free-text fields that are difficult to analyze systematically.

Work Order Intelligence uses generative AI to enhance insufficient data and provide recommendations for the most likely problem code based on the work order description. When a technician enters a description like "pump making noise and vibrating," the AI model analyzes the text, considers the asset type and its failure history, and recommends the most likely failure code from the organization's configured failure code hierarchy. This helps maintenance managers quickly identify the issue, assign the right problem code, allocate technicians with the appropriate skills, and reduce troubleshooting time.

The impact on data quality compounds over time. As more work orders are enriched with accurate failure codes, the historical data becomes more reliable, which in turn improves the accuracy of predictive models that rely on failure history. This creates a positive feedback loop: better data quality leads to better predictions, which lead to more targeted maintenance interventions, which produce better data about asset behavior and failure patterns. Over months and years, this compounding effect can transform an organization's maintenance data from a liability into a strategic asset.

The generative AI model is trained on the organization's own work order history, which means it learns the specific terminology, failure patterns, and maintenance practices of that organization. A utility company's work orders will describe different issues than a food processing plant's work orders, and the model adapts accordingly. This organization-specific training is critical because generic failure code libraries cannot capture the nuances of how each organization describes and categorizes maintenance issues. The model also learns from corrections, so when a maintenance manager overrides the recommended failure code, that feedback is incorporated into future recommendations.

Maximo Assistant: Natural Language Meets Asset Management

Maximo Assistant represents a fundamental shift in how users interact with Maximo. Instead of navigating through multiple dashboards, constructing complex queries, or calling APIs, users can simply ask questions in natural language. The assistant is powered by IBM Granite, IBM's family of large language models, and it understands Maximo's query syntax, business objects, and relationships. This capability is particularly valuable for maintenance managers who need quick answers during planning meetings, reliability engineers who want to explore asset history without drilling through multiple screens, and executives who need status summaries without learning the Maximo navigation.

Users can ask questions like "What assets need my attention?" or "How is asset PUMP-4A doing?" The assistant queries Maximo's databases, analyzes the results, and returns a clear, conversational response. IBM Research has indicated that the assistant underpinning Maximo Assistant will be upgraded to a Granite 4.0 model in the near future, equipped with a native understanding of Maximo's query syntax, which will provide a smoother user experience and more sophisticated reasoning.

In MAS 9.2, Maximo Assistant evolves from a question-answering tool to a persistent, context-aware digital teammate. The assistant can now act as a single agent that analyzes requests and works across a defined set of AI tools to answer them. It can reason through multi-step questions, pull data from multiple sources, and provide comprehensive responses that combine information from work orders, asset records, inspections, and reliability analyses. As more specialized agents are added to the platform, Maximo Assistant will continue to evolve to orchestrate multiple agents and workflows, taking appropriate action rather than merely providing information.

The assistant on mobile is particularly impactful for field technicians. Technicians can use natural language to find asset information, review maintenance history, and get guidance on completing work. They can ask "What were the findings from the last inspection on this pump?" or "Show me the job plan for this work order" without removing their gloves or navigating through mobile menus. This hands-free interaction capability is especially valuable in environments where technicians wear personal protective equipment or work in confined spaces. Maximo Visual Inspection on mobile enables AI-based visual inspection with local inference directly on the device, so technicians can capture images and get immediate analysis even without cloud connectivity.

Condition Insight: Agentic AI for Condition-Based Maintenance

IBM's announcement of Maximo Condition Insight marked a significant milestone in the platform's AI journey. Condition Insight is an agentic AI capability within Maximo APM (Asset Performance Management) that works in concert with other applications and AI capabilities in the broader Maximo Application Suite. It represents a shift from periodic check-ups to condition-based maintenance, where maintenance is triggered by actual asset condition rather than fixed schedules. This shift has the potential to reduce unnecessary labor, help enterprises meet sustainability goals by keeping machines at peak performance, and extend the useful life of expensive, multi-year investments.

Condition Insight evaluates work orders, metrics, time-series data, meter readings, Failure Mode and Effects Analysis (FMEA) results, and alerts to assess asset condition. It interprets this data to explain asset condition, highlight emerging trends, and recommend corrective actions, all communicated in plain, understandable language. Unlike traditional condition monitoring systems that present raw data and leave interpretation to the reliability engineer, Condition Insight removes the barrier by analyzing asset data in seconds and returning a clear, explainable summary of condition, trends, and recommended actions.

The agentic nature of Condition Insight is what sets it apart. Rather than simply displaying a dashboard of metrics, the agent actively consolidates data silos into a holistic assessment of asset health. It evaluates the asset's current condition, estimates the projected replacement date, and recommends specific actions. The agent can detect when an asset is showing early signs of degradation and suggest that the next scheduled inspection can be deferred, or conversely, it can flag an asset that appears healthy on the surface but is showing subtle patterns that precede failure.

{
  "assetConditionInsight": {
    "asset": "PUMP-4A",
    "location": "BLDG-4",
    "overallCondition": "DEGRADED",
    "healthScore": 62,
    "trend": "DECLINING",
    "keyFindings": [
      "Vibration readings increased 15% over last 30 days",
      "Bearing temperature above baseline by 8 degrees",
      "Last PM completed 45 days ago, next due in 15 days",
      "Similar failure pattern detected on PUMP-2B in 2024"
    ],
    "recommendedActions": [
      "Schedule bearing inspection within 7 days",
      "Increase vibration monitoring frequency to daily",
      "Review lubrication records for completeness",
      "Check alignment per OEM specification document"
    ],
    "predictedReplacementDate": "2027-Q3",
    "confidenceLevel": "HIGH",
    "explanation": "The combination of rising vibration and temperature "
                   "readings, coupled with the similar failure pattern on "
                   "PUMP-2B, indicates bearing degradation. Early "
                   "intervention can extend asset life by 6-9 months."
  }
}

IBM's time-series foundation models are being integrated into Maximo to help the Condition Insights agent pick out trends and meaningful patterns in sensor data. These models are pre-trained on vast datasets and can recognize temporal patterns that would be difficult for human analysts to detect, particularly when dealing with thousands of assets generating millions of sensor readings per day. IBM combines Maximo's trusted asset data model with watsonx AI, delivering explainable, enterprise-grade AI that integrates directly into maintenance workflows, which distinguishes it from niche vendors that require custom builds.

MAS 9.2: AI Built for Asset Workflows

MAS 9.2, announced in June 2026 and now generally available, represents the most comprehensive AI integration in Maximo's history. Rather than treating AI as a separate layer, MAS 9.2 embeds AI directly into the daily work of managing assets, teams, safety, and operations. The release applies AI where work happens, helping teams identify issues earlier, make better decisions faster, and execute with greater consistency in the field.

The release expands AI across reliability insights, field execution, safety and compliance workflows, document-based information extraction, and orchestration across systems. AI-powered Field Service Management brings conversational scheduling and what-if analysis to planners and schedulers, enabling them to explore capacity changes or work prioritization using plain language. AI-enabled lease abstraction uses RAG to extract key information from lease documents, reducing manual review time. Expanded capabilities across asset-centric waste management, contractor safety oversight, and mobile-first safety workflows make it easier to manage compliance, track risk, and act in real time.

Reliability Strategies now provides expanded AI capabilities to build out failure and remediation descriptions. IBM collaborated with IBM Research to employ watsonx generative AI to develop more FMEAs and expand the library, making it easier for reliability engineers to create and optimize customized maintenance strategies. The tighter connection between Reliability Strategies, Health, Job Plans, PMs, and Meters ensures that failure analysis insights flow directly into maintenance execution.

The MCP Server: Connecting External AI Agents

The MCP (Model Context Protocol) Server in MAS 9.2 is the infrastructure that enables organizations to bring their own AI agents and integrate them directly with Maximo Manage APIs. Rather than exposing technical API endpoints directly to an AI model, the MCP Server exposes business capabilities that AI agents can understand and invoke. The AI no longer needs to understand endpoint URLs, object structure names, payload formats, or authentication mechanisms. Instead, it requests a business capability, and the MCP Server handles the translation.

This means an external AI agent, whether built on IBM watsonx, OpenAI, or any other platform, can interact with Maximo through a standardized interface. Organizations can create custom AI tools using Maximo Manage automation scripts, object structures, or workflows, and connect their own agents to the MCP server. IBM Research has also open sourced AssetOpsBench, an agentic environment and benchmark for asset management tasks, allowing Maximo partners and collaborators to develop and test their own agents in a standardized evaluation environment.

Looking ahead, IBM Research plans to introduce a third agent focused on asset investment planning. This agent will go beyond Maximo's default optimizer to allow users to set conditions for replacing equipment based on operating costs, budgeting constraints, or sustainability targets. This will give organizations a more sophisticated tool for capital allocation decisions, which have traditionally been driven by age-based replacement schedules rather than holistic cost-benefit analysis.

Practical Implications

For organizations looking to leverage AI in Maximo, a phased approach is recommended. Start with Maximo Health to establish condition monitoring and health scoring. This provides the data foundation that predictive models need. Ensure that IoT sensor data is flowing reliably and that asset records are complete and accurate. Next, evaluate Maximo Predict for assets where sufficient historical failure data exists. Not every asset class will have enough data for reliable predictions, and trying to predict failures for assets with sparse history will produce unreliable results.

Once the predictive foundation is solid, explore Work Order Intelligence to improve data quality. Better data quality improves the accuracy of predictive models, creating a virtuous cycle. Then consider Maximo Assistant for knowledge workers and mobile technicians, particularly in organizations where technicians are not comfortable navigating complex mobile applications.

For MAS 9.2 adopters, the MCP Server opens new possibilities for custom AI integrations. Start with a pilot use case, such as AI-assisted work order triage or automated condition assessment, to evaluate the MCP Server's capabilities. The agentic workflow capabilities in MAS 9.2 are designed to be extensible by IBM, customers, and the ecosystem, so early adoption positions you to benefit from future enhancements. Track AppPoint usage for AI Service on the licensing dashboards to understand the cost implications of AI features, as some capabilities consume AppPoints based on usage volume.

Bottom Line

AI in Maximo has moved from marketing slides to production capability. The progression from Maximo Health for condition monitoring, to Maximo Predict for failure prediction, to Work Order Intelligence for data quality, to Condition Insight for agentic condition assessment, to MAS 9.2's embedded AI workflows and MCP Server for external agent integration, represents a comprehensive AI strategy that touches every aspect of asset management. The organizations that will benefit most are those that invest in data quality, build their AI capabilities incrementally, and align their AI initiatives with specific operational outcomes rather than deploying AI for its own sake. With MAS 9.2 now generally available, the tools are in place. The question for maintenance leaders is not whether to adopt AI, but how quickly and how strategically they can build the data and organizational foundation to make it work.