AI Inside Maximo: From Predictive Maintenance to LLM-Powered Agents

Artificial intelligence in enterprise asset management has moved from marketing slides to production deployments. IBM Maximo Application Suite (MAS) 9 delivers AI capabilities that are not theoretical. They are running in production environments at organizations that have invested in data quality, sensor infrastructure, and model training. The AI journey in Maximo has evolved through three distinct phases: condition-based monitoring with IoT sensors, machine learning predictive models trained on historical failure data, and now generative AI powered by IBM watsonx and Granite foundation models.

This article traces that evolution, explains what each AI capability actually does, and provides practical guidance for organizations that want to move from reactive maintenance to AI-assisted asset management.

Phase One: Maximo Health and Condition-Based Monitoring

The foundation of AI in Maximo is condition data. Without real-time visibility into asset condition, predictive models have nothing to analyze. Maximo Health provides that foundation by ingesting IoT sensor data, meter readings, work history, and environmental data to provide a consolidated view of asset status.

Maximo Health enables organizations to monitor critical asset performance using data from sensors and other sources such as weather, asset records, and work history. The system calculates health scores for individual assets based on condition indicators, performance metrics, and degradation trends. A health score is not a simple threshold alarm. It is a composite metric that combines multiple data points into a single indicator that maintenance planners can use to prioritize work.

For example, a centrifugal pump might have vibration sensors, temperature sensors, and flow rate meters. Maximo Health combines these data streams with historical performance data and maintenance records to calculate a health score. A score of 85 means the pump is operating within normal parameters. A score of 62 means something is changing, even if no individual sensor has crossed an alarm threshold. That early indication is where condition-based maintenance adds value over traditional time-based or run-based preventive maintenance.

The practical implementation of Maximo Health requires several prerequisites. First, assets must have sensors installed and connected to Maximo through IoT integration. This is the most common barrier to adoption. Many organizations have assets with no sensors, or sensors that are connected to a SCADA system but not integrated with Maximo. The integration typically uses IBM Maximo Health and Predict Utilities (MHPU) or a custom integration through the MAS REST API.

Second, the asset hierarchy must be accurately modeled in Maximo. Health scores are calculated at the asset level, but they roll up to parent assets and locations. If the asset hierarchy is incorrect, health scores will not aggregate properly, and planners will not see an accurate picture of system-level condition.

Third, maintenance teams need to understand what health scores mean and how to act on them. A health score is not a work order. It is an indicator that helps planners decide whether to generate a work order. Training maintenance teams to interpret health scores and incorporate them into planning workflows is a change management exercise, not just a technical implementation.

# Example: Querying asset health scores via MAS 9 REST API
import requests

base_url = "https://mas-host.company.com"
headers = {"Authorization": "Bearer ...", "Accept": "application/json"}

# Get health scores for assets in PLANT-A
url = f"{base_url}/api/os/mxasset"
url += "?oslc.where=location=\"PLANT-A\""
url += "&oslc.select=assetnum,description,location,status,healthscore"
url += "&lean=1"

response = requests.get(url, headers=headers)
assets = response.json().get("member", [])

# Filter assets with declining health
at_risk = [a for a in assets if a.get("healthscore", 100) < 70]
print(f"Assets at risk: {len(at_risk)}")
for asset in at_risk:
    print(f"  {asset['assetnum']}: Health={asset['healthscore']} - {asset['description']}")

Phase Two: Maximo Predict and Machine Learning Models

Maximo Predict builds on the condition data from Health to forecast future asset behavior. It uses machine learning and data analytics to predict downtime, degradation, and failures. Predict works along two complementary paths that together provide a comprehensive view of asset reliability.

The first path is prediction based on historical failure data. By analyzing corrective maintenance work orders, failure codes, root causes, repair frequency, and mean time between failures (MTBF) patterns, Predict learns how assets have failed in the past. This works especially well for repetitive assets and known failure modes. It helps answer questions such as: which assets are statistically most likely to fail next, and which failure patterns tend to repeat under similar conditions. This approach is powerful, but only when failure data is consistent and meaningful. If technicians have been coding every failure as "OTHER" for ten years, the historical data path will not produce useful predictions.

The second path is prediction based on condition and sensor data. Here, Predict learns from vibration, temperature, pressure, runtime, and other condition indicators, combined with degradation trends derived from Health. Instead of asking what failed before, this approach asks what behavior usually precedes a failure. This is where early detection becomes possible, long before thresholds are crossed. A bearing might show increasing vibration at a frequency that correlates with early-stage spalling, even though the overall vibration level has not exceeded any alarm threshold. The predictive model recognizes the pattern and flags the asset for inspection.

The workflow for implementing Predict involves data scientists working alongside maintenance engineers. Assets are grouped by type, criticality, and operating context. For each group, a data scientist trains a predictive model using either the default notebooks provided by IBM or custom notebooks. The default notebooks cover common prediction types: current failure probability, predicted failure date, and remaining useful life. Custom notebooks can be extensions of the defaults or entirely custom, but all models must be deployed in Watson Machine Learning.

Once a model is trained and deployed, predictions are populated for each asset in the group. A Predictions section appears on the asset record, showing the probability of failure, the predicted failure date, and the confidence interval. Maintenance planners can use work queues to track assets that have a high probability of failure or assets that will fail before the next scheduled preventive maintenance work order.

The critical insight about Predict is this: the real value is not accuracy, it is prioritization. Predict does not tell you what will fail. It tells you where to look first. A planner sees rising failure probability, declining health, and known historical failure behavior. That combination enables confident, explainable decisions: planning maintenance instead of reacting, intervening before performance drops, and avoiding both panic and complacency. When AI is grounded in failure history and health insight, prevention becomes realistic rather than speculative.

Phase Three: Work Order Intelligence and Generative AI

MAS 9.0 introduced Maximo Work Order Intelligence, which uses IBM watsonx generative AI capabilities to enhance asset lifecycle management. This is a significant shift from the predictive analytics of Health and Predict. Work Order Intelligence uses large language models to understand and enrich work order data.

The first capability is failure code recommendation. When a technician completes a work order and enters a description of the problem, the AI model trained on work order descriptions provides recommendations for the most likely problem code. This addresses a long-standing data quality problem in Maximo deployments: technicians often select generic failure codes or leave failure coding incomplete because they are focused on completing the repair, not doing data entry. With AI-assisted failure code recommendation, the technician writes a description and the system suggests the appropriate code. This speeds up work order approval, improves data quality, and provides reliable failure data for the predictive models in Predict.

The second capability is work order data enrichment. Generative AI can enhance insufficient work order data by inferring missing fields from the description and historical patterns. If a work order description mentions "replaced V-belt on conveyor C-12" but does not specify the failure code, the AI can infer that the failure mode is likely "MECHANICAL WEAR" and the failure code is "WORN/REPLACED." This enriched data improves the quality of reports, KPIs, and predictive models that depend on accurate failure coding.

The third capability is Maximo Reliability Strategies, which provides organizations with the ability to analyze failure modes and access a comprehensive library of asset-specific failure details and mitigation activities. This solution makes it easier to create and optimize highly customized maintenance reliability strategies. For MAS 9.0, IBM collaborated with IBM Research and employed watsonx generative AI to develop more FMEAs (failure mode and effects analyses) and expand the library. This capability is available for technical preview and is being fully released in subsequent MAS updates. The AI-generated FMEAs enable organizations to build failure mode analyses faster and more effectively for any asset specific to their industry, without requiring a reliability engineer to start from scratch for every asset type.

Phase Four: Maximo Assistant and LLM-Powered Agents

The most recent AI advancement in Maximo is the introduction of LLM-powered agents that can interact with Maximo's data and functionality through natural language. IBM Research has integrated AI agents into the Maximo Application Suite to make tracking the condition of physical assets easier, more efficient, and more economical.

Maximo Assistant is a new LLM-powered interface that was recently integrated into the platform. Users can now query Maximo's databases through a chat window, and an IBM Granite-powered agent fetches the information. Users no longer have to call APIs or know how to write SQL to interact with Maximo's relational databases. A maintenance manager can type "Show me all work orders for pump P-101 in the last 90 days with failure code WORN" and the assistant returns the results in a readable format. This democratizes access to Maximo data for non-technical users, including operations managers, reliability engineers, and maintenance planners who may not be proficient in SQL or API queries.

The agent underpinning Maximo Assistant will be upgraded to a Granite 4.0 model in the near future, which will bring improved reasoning capabilities and better handling of complex multi-part questions. IBM's time-series foundation models will also be integrated into Maximo to help the Condition Insights agent pick out trends and other meaningful patterns in sensor data.

Maximo's Condition Insights agent, planned for release by end of year, will gather statistics and operational data for each asset a company tracks. The agent will estimate the asset's current condition and projected replacement date as part of a broader move from regular check-ups to more sporadic visits, only when Maximo detects signs of trouble. By shifting the focus from corrective maintenance to servicing only when needed, researchers expect that Maximo's AI capabilities can reduce unnecessary labor, help enterprises meet their sustainability goals by keeping machines at peak performance, and ultimately extend the useful life of expensive, multi-year investments.

The Condition Insights agent consolidates data silos into one view for a holistic assessment of an asset's health. Currently, asset condition data is often scattered across Maximo records, sensor databases, inspection reports, and external monitoring systems. The agent brings these together, applies analytics, and produces a single condition assessment that maintenance teams can act on.

Maximo's Asset Investment Planning agent, planned for next year, will go beyond the default optimizer to allow users to set the conditions for replacing a piece of equipment. Whether the priority is operating costs, budgeting constraints, or sustainability targets, the agent will calculate the optimal replacement timing based on the organization's specific criteria. This moves AI from descriptive and predictive analytics to prescriptive recommendations: not just telling you what might fail, but telling you what to do about it.

Here is an example of how the Maximo Assistant changes the user experience:

User: "Which assets in Plant A have a health score below 70 and no 
       open work orders?"

Maximo Assistant: "I found 3 assets in Plant A with health scores 
below 70 and no open work orders:
1. PUMP-P-101 — Health: 52 — Last PM: 45 days ago
2. MOTOR-M-204 — Health: 61 — Last PM: 120 days ago  
3. HEAT-EX-HX-08 — Health: 67 — Last PM: 90 days ago

Would you like me to create work orders for any of these assets?"

User: "Create a work order for PUMP-P-101 with priority 2"

Maximo Assistant: "Work order WO-2026-0892 created for asset PUMP-P-101 
with priority 2. The work order is in WAPPR status and assigned to 
supervisor SUP002. Would you like to assign a technician?"

Practical Implications

For organizations at the beginning of their AI journey, the most important step is data quality. Every AI capability in Maximo, from Health scores to predictive models to generative AI recommendations, depends on the quality of the underlying data. If work order descriptions are vague, failure codes are generic, and asset hierarchies are incomplete, AI will not fix those problems. It will amplify them. Invest in data cleanup before investing in AI.

For organizations that have solid data and are ready for predictive maintenance, start with Maximo Health before Maximo Predict. Health gives you the condition baseline. Predict builds on that baseline with forecasting. Trying to implement Predict without Health is like trying to forecast weather without current temperature data. Start with a pilot group of critical, well-instrumented assets, and expand from there.

For organizations already running Health and Predict, the generative AI capabilities in MAS 9.0 and beyond represent a significant productivity opportunity. Work Order Intelligence reduces the data entry burden on technicians while improving data quality. Maximo Assistant makes Maximo data accessible to non-technical users. The Condition Insights agent will consolidate monitoring data across systems. Each capability addresses a specific pain point in the maintenance workflow.

The AI capabilities in Maximo are not a replacement for maintenance expertise. They are tools that help maintenance teams make better decisions faster. A predictive model that says a pump has a 78% probability of failure in the next 30 days is not telling you to replace the pump. It is telling you to inspect the pump, review its history, and make an informed decision. The AI provides the signal. The human provides the judgment.

Bottom Line

Maximo's AI journey has moved from condition monitoring to predictive analytics to generative AI and LLM-powered agents. Each phase builds on the previous one. Health provides the condition baseline. Predict adds forecasting. Work Order Intelligence improves data quality with generative AI. Maximo Assistant democratizes data access through natural language. The Condition Insights agent will consolidate monitoring across systems. The Asset Investment Planning agent will optimize capital replacement decisions. The progression is clear: from knowing what is happening now, to predicting what will happen next, to recommending what to do about it. Organizations that invest in data quality, sensor infrastructure, and model training at each phase will see compound returns as each new AI capability builds on the foundation of the previous one.