AI in IBM Maximo: Watsonx, Predictive Maintenance, and the Rise of Work Order Intelligence

For most of its history, enterprise asset management software was a system of record. It tracked what broke, when it was fixed, who did the work, and what parts were used. The intelligence lived in the maintenance planner's head, in spreadsheets, or in reliability engineering reports that arrived weeks after the failure. Artificial intelligence is changing that. In IBM Maximo Application Suite, AI is no longer a separate dashboard or an external data science project. It is becoming part of the workflow.

IBM has tied Maximo closely to watsonx, its enterprise AI platform. The integration shows up in several forms: predictive models that estimate remaining useful life, generative AI that recommends failure codes and accelerates work order approval, agentic AI that explains asset condition, and visual inspection models that detect defects from images and video. For maintenance organizations, the question is no longer whether AI is relevant. It is where to start and how to avoid pilots that never reach production.

This article covers the practical AI capabilities available in Maximo today. We will look at predictive maintenance through Maximo Predict, generative AI through Maximo Work Order Intelligence, agentic AI through Maximo Condition Insight, and the emerging role of visual inspection. The focus is on what these tools do, what data they need, how they fit together, and what it takes to make them useful in a real maintenance environment. The goal is not to hype the technology. It is to show where it fits into the daily work of keeping assets running.

Predictive Maintenance with Maximo Predict

Maximo Predict is the application in MAS dedicated to forecasting asset failures, degradation, and downtime. It builds on Maximo Health, which means it inherits health scores, condition data, and asset hierarchies. The core idea is to use historical maintenance records, sensor data, inspection reports, and environmental data to train models that estimate when an asset is likely to fail or degrade.

The workflow is collaborative. A reliability engineer defines asset groups based on common failure modes or operating contexts. A data scientist works with notebooks, either the defaults provided by IBM or custom extensions, to build and train models. The trained models are deployed to Watson Machine Learning, and the resulting predictions flow back into Maximo asset records. Each asset can then display the current probability of specific failure modes, estimated days until failure, and whether detected anomalies point toward a probable failure.

A typical integration pattern uses meter readings and time-series data from Maximo Monitor. A Python service periodically queries the Maximo API for the latest meter readings, runs inference against a trained model, and writes the predicted remaining useful life back to the asset record. When the remaining useful life drops below a threshold, an automation script in Maximo creates a preventive or predictive work order.

# Simplified inference loop against a trained scikit-learn model
import pickle
import requests
from joblib import load

MODEL_PATH = "model.joblib"
MAXIMO_URL = "https://maximo.example.com/maximo"
TOKEN = "base64…word"
ASSET_ID = "PUMP-001"

# Load the trained model
classifier = load(MODEL_PATH)

# Pull latest meter readings from Maximo
response = requests.get(
    f"{MAXIMO_URL}/oslc/os/mxasset",
    headers={"maxauth": TOKEN, "Content-Type": "application/json"},
    params={
        "oslc.select": "assetmeter,expectedlife",
        "oslc.where": f"assetnum={ASSET_ID}",
        "lean": 1
    }
)
meter_data = extract_meters(response.json())

# Predict remaining useful life
predictions = classifier.predict([meter_data])

# Update the asset record
requests.post(
    f"{MAXIMO_URL}/oslc/os/mxasset/{rest_id}",
    headers={"maxauth": TOKEN},
    data={"expectedlife": int(predictions[0])}
)

This example is intentionally simple. In production, you would add model versioning, drift detection, logging, and error handling. The important point is that Maximo Predict does not replace the data scientist. It provides the scaffolding that connects models to maintenance workflows. The model can be built in Watson Studio, deployed in Watson Machine Learning, and consumed inside Maximo without custom integration plumbing.

The business value shows up in work queue prioritization. Maximo Predict includes work queues that track assets with a high probability of failure or assets that will fail before the next scheduled preventive maintenance work order is generated. Planners can focus attention on the highest-risk items instead of treating every asset equally. That shift from calendar-based to risk-based scheduling is where predictive maintenance usually pays for itself.

Model quality depends on the data that feeds it. Assets need consistent classifications, failure records need accurate timestamps, and sensor data needs to be aligned with operating context. A model trained on dirty data will produce predictions that look reasonable but fail in production. The data science effort is important, but the data engineering effort is usually the bottleneck.

Generative AI and Work Order Intelligence

Released with MAS 9.0, Maximo Work Order Intelligence uses IBM watsonx generative AI to address one of the most common data quality problems in maintenance: poorly described work orders. A technician writes "machine broke" or "pump noisy," and the maintenance manager has to figure out what actually happened before the work order can be approved, coded, and analyzed.

Work Order Intelligence reads the description, uses a generative model trained on work order data to recommend the most likely problem code, and provides a reliability score for the recommendation. As more details are added to the work order, the recommendation is regenerated. The maintenance manager can then approve the work order faster, assign the right technician, and ensure the failure data will be useful later for reliability analysis.

The benefit is not just speed. It is data consistency. Failure codes are the raw material of reliability engineering. If they are missing, wrong, or inconsistently applied, failure mode analysis becomes unreliable. Work Order Intelligence makes it easier to capture good data at the point of creation, which improves every downstream report, dashboard, and predictive model. Over time, the organization builds a cleaner, more usable history of what actually fails and why.

IBM has also applied generative AI to failure mode and effects analysis (FMEA). In MAS 9.0, Maximo Reliability Strategies introduced the ability to access, create, import, and modify FMEAs. IBM Research collaborated with the Maximo team to use watsonx generative AI to expand the library of asset-specific failure details and mitigation activities. This capability was in technical preview at the 9.0 release and is intended to help organizations build FMEAs faster for assets that are specific to their industry.

The value of AI-assisted FMEA is speed without loss of rigor. Traditional FMEA workshops are valuable but slow. They require the right experts in a room for hours and can become inconsistent across sites. An AI-generated draft gives the team a starting point based on asset type, operating context, and historical failure data. Human experts still review and validate the draft, but the starting point is better than a blank page. The result is a more complete reliability strategy library in less time.

Generative AI in Maximo is most effective when it is embedded in a workflow rather than used as a chatbot. Work Order Intelligence appears where the manager is already approving work orders. AI-assisted FMEA appears where the reliability engineer is already building strategies. That context-aware placement is what turns a novelty into a habit.

Agentic AI and Maximo Condition Insight

In late 2025, IBM introduced Maximo Condition Insight, an agentic AI capability within Maximo Asset Performance Management. The idea is to move beyond dashboards and raw scores into plain-language explanations of asset condition. Condition Insight evaluates work orders, metrics, time-series data, meter readings, FMEA data, and alerts to summarize asset condition, highlight trends, and recommend corrective actions.

The significance of agentic AI here is not just automation. It is interpretation. A maintenance technician may see a health score of 72 and not know whether that is urgent. Condition Insight can explain what is driving the score, what trends are visible, and what actions are recommended. Because it is built on watsonx, the outputs are designed to be explainable and enterprise-grade, which matters in regulated industries where black-box recommendations are not acceptable.

This capability fits naturally into a condition-based maintenance strategy. Instead of relying solely on fixed schedules or raw sensor thresholds, teams can use AI to interpret complex, multi-source data and decide when to act. The recommendations can trigger work orders, schedule inspections, or escalate to reliability engineers for deeper review. The agentic layer does not replace the engineer. It reduces the time spent gathering and interpreting data so the engineer can spend more time deciding.

A useful way to think about Condition Insight is as a translator between raw data and maintenance action. Sensor thresholds, health scores, and work order history are valuable, but they require interpretation. Condition Insight turns that interpretation into a narrative that a planner, supervisor, or executive can understand. That narrative becomes the basis for faster, more confident decisions.

The agentic approach also reduces the cognitive load on operators. In a large plant, a single shift supervisor may be responsible for hundreds of assets. A ranked list of condition summaries, each with a recommended next step, is more actionable than a wall of charts. The AI does the reading; the human does the deciding. The combination of scaled data processing and human judgment is the practical promise of agentic maintenance AI.

Visual Inspection and AI-Augmented Field Work

Maximo Visual Inspection adds computer vision to the AI portfolio. Technicians, drones, or fixed cameras capture images and video of assets. AI models analyze the media for defects, anomalies, and critical conditions. The results can create inspection records, generate work orders, or feed health scores without a human manually reviewing every image.

The field use cases are expanding rapidly. A drone inspects a flare stack or transmission tower and flags corrosion. A mobile technician photographs a leaking seal and the model classifies severity. A production line camera spots a defect on a finished part and routes the finding to quality and maintenance simultaneously. The model training and inference happen through the Visual Inspection application, and the outcomes flow into the broader Maximo workflow.

Visual inspection is particularly useful for assets that are hard to reach, dangerous to inspect frequently, or numerous enough that manual review is impractical. It does not eliminate the need for expert judgment. It prioritizes what the expert should look at first. A model might flag a hundred images and rank them by confidence, allowing a human inspector to focus on the dozen most likely defects.

The practical deployment pattern is usually iterative. Start with a single asset class and a well-defined defect. Collect and label a few hundred images. Train a model, validate it against a held-out test set, and run it in shadow mode where the model predicts but humans still make the final call. Once accuracy is high enough, allow the model to create inspection records automatically for the highest-confidence cases while routing borderline cases to human review. This staged approach builds trust and catches edge cases before they reach production.

Visual inspection also creates a feedback loop that improves the asset record. Every image, classification, and human correction becomes training data for the next model version. Over time, the model becomes more accurate and the inspection process becomes faster. The key is to capture the feedback consistently, not just the initial prediction.

Connecting the AI Capabilities: From Data to Decision

The real power of AI in Maximo emerges when the capabilities connect. Maximo Monitor supplies sensor data. Maximo Health aggregates condition data and assigns health scores. Maximo Predict forecasts failures. Condition Insight explains what the data means. Work Order Intelligence ensures work orders are coded correctly. Visual Inspection catches defects the other data sources miss. Together, they form a closed loop from data collection to maintenance action.

The connection is not automatic. It requires deliberate integration. Health scores need to be configured to consume Monitor metrics. Predict models need to reference Health scores and asset attributes. Condition Insight needs access to work orders, FMEAs, and time-series data. Work Order Intelligence needs historical work orders with reliable failure codes. Visual Inspection needs model outputs that feed into the asset record or inspection object. Each link is a configuration and data quality project.

Organizations that approach these capabilities as a single AI strategy rather than separate pilots get more value. A predictive model is more useful when its predictions feed into Condition Insight, which then generates a prioritized work order, which Work Order Intelligence codes correctly. A visual inspection finding is more useful when it updates the asset's health score and triggers the right reliability review. The platform is designed for this kind of flow, but it takes planning to make it real.

The maturity curve is also worth understanding. Most organizations start with condition monitoring and basic dashboards. Then they add health scoring. Predictive models come next, once failure data is clean enough to train on. Generative AI for work orders and FMEAs can run in parallel because they depend on historical work order quality rather than sensor data. Visual inspection is usually a separate track driven by field inspection needs. Trying to do everything at once spreads resources too thin and undermines trust.

Data and Organizational Readiness

AI in Maximo is not a magic button. The models need clean, relevant, and accessible data. Predictive maintenance requires historical failure records, sensor data, and asset context. Work Order Intelligence requires a corpus of well-coded historical work orders. Condition Insight requires integrated data from multiple Maximo applications. Visual Inspection requires labeled images and a feedback loop to improve model accuracy over time.

The organizational readiness is just as important as the data readiness. Maintenance teams need to trust the recommendations. Planners need to understand when to override the model. Data scientists need access to Maximo data in a format they can use. Reliability engineers need to validate that predictions actually lead to better outcomes than existing schedules. Without trust, the most accurate model will sit unused.

A practical starting point is to identify one asset class with a known pain point and good data. Build a predictive or generative AI workflow for that class, measure the results, and iterate. Trying to deploy AI across every asset at once usually produces shallow, untrusted outputs that never reach production. A focused pilot with a clear metric, such as reduction in unplanned downtime or improvement in failure code accuracy, is far more likely to succeed and expand.

Another readiness factor is integration across MAS applications. Predict uses Health. Condition Insight uses APM. Work Order Intelligence is part of Manage. Visual Inspection is its own application. Licensing, data entitlements, and user access need to be coordinated across the suite. Before launching an AI initiative, confirm that the necessary applications are licensed, the data pipelines are configured, and the administrators understand how the outputs flow from one application to another.

Practical Implications

For maintenance leaders, the implication is that AI should be treated as a maintenance capability, not an IT experiment. The goal is to change decisions: which work orders to approve first, which assets to inspect, which failures to predict. If an AI project does not change a decision, it is not delivering value. Define the decision first, then choose the AI capability that supports it.

For data scientists, the implication is that Maximo is becoming a better platform for operationalizing models. The path from Watson Studio notebook to deployed model to Maximo work order is clearer than it was in earlier generations. The challenge shifts from integration to model quality and business validation. A model that is accurate on a test set but ignored by planners has not succeeded.

For Maximo administrators, the implication is that AI features require configuration across multiple applications. Predict uses Health. Condition Insight uses APM. Work Order Intelligence is part of Manage. Visual Inspection is its own application. Licensing, data entitlements, and user access need to be coordinated across the suite. Administrators need to think in terms of workflows that cross application boundaries rather than isolated application configurations.

For reliability engineers, the implication is that AI augments rather than replaces expertise. FMEA drafts, condition summaries, and failure predictions are inputs to engineering judgment. The best implementations combine AI-generated suggestions with human validation, so the organization gets speed without sacrificing accountability.

Bottom Line

AI in Maximo has moved from promise to product. Watsonx powers predictive maintenance, generative work order coding, agentic condition explanations, and visual defect detection. Each capability addresses a real maintenance problem: unexpected failures, poor work order data, complex condition interpretation, and slow manual inspections.

The organizations that will benefit most are the ones that treat AI as part of a broader reliability strategy. They clean their data, define clear use cases, validate outcomes, and integrate recommendations into daily workflows. The technology is ready. The differentiator will be how well maintenance organizations adopt it.