AI in Maximo: From Predictive Models to the Embedded Maximo Assistant

Artificial intelligence has become one of the most discussed and most misunderstood topics in enterprise asset management. Vendors promise predictive maintenance that eliminates downtime, autonomous assistants that replace planners, and computer vision that detects every defect. The reality is more useful but less dramatic. AI in IBM Maximo Application Suite is a set of specialized capabilities that help teams make better decisions, capture cleaner data, and prioritize work more effectively. It does not replace judgment, but it can sharpen it.

The Maximo AI portfolio has two major branches. The first is predictive analytics: models that estimate failure probability, remaining useful life, and anomaly trends based on historical and sensor data. This branch is embodied in Maximo Predict and the broader Health and Predict application set. The second is assistive and generative AI: a conversational assistant embedded in MAS that can answer questions, summarize records, suggest actions, and help technicians interact with the system using natural language. IBM announced this capability as Maximo Assistant, built on watsonx.ai, and it is positioned as a productivity tool rather than a replacement for domain expertise.

Between these two branches sits visual inspection, where AI models analyze images and video to detect defects and conditions. Drones, fixed cameras, and mobile devices can feed visual data into Maximo, where models trained with visual prompting identify cracks, corrosion, leaks, and other anomalies. This is particularly valuable for assets that are hard to reach, such as flare stacks, transmission towers, and rail tracks. It does not eliminate human inspection, but it helps teams focus limited human attention on the exceptions that matter.

The key to using AI in Maximo well is to match the technique to the business problem and the data available. A generative assistant will not fix bad failure codes. A predictive model will not work without enough historical failures. A visual model needs representative images. This article walks through each capability, its data requirements, and practical adoption guidance.

Predictive Maintenance with Maximo Predict

Maximo Predict is the core predictive maintenance application in MAS. It uses machine learning to answer two primary questions for each asset: how likely is it to fail within a defined horizon, and how much useful life remains? These answers are probabilistic, not deterministic. A prediction of 78 percent failure probability in the next thirty days means the asset probably fails, but there is still a 22 percent chance it does not. Maintenance planners use these probabilities to prioritize work, order parts, and schedule interventions before a failure forces an emergency response.

The application supports several model types. Failure probability models are classification models, often based on Random Forest or Gradient Boosting, that estimate the likelihood of failure before a date threshold. Predicted failure date models use survival analysis techniques such as Cox Proportional Hazards or Weibull regression to estimate when a failure is expected. Remaining useful life models use regression to estimate how many days or cycles an asset has left. Anomaly detection models flag operating patterns that deviate from normal behavior. Each type serves a different decision pattern and requires different data.

Training these models depends heavily on data quality and volume. IBM recommends at least twenty failures per failure mode per asset class, though more is better. Two or more years of operational history, six or more months of continuous sensor data, consistent failure coding, and accurate asset attributes all improve model reliability. A common mistake is to assume that buying Maximo Predict automatically delivers accurate predictions. The tool provides the pipeline; the organization must provide the data and the domain interpretation.

Models are typically built and trained in Watson Studio using the Maximo Predict SDK, known as pmlib. Data scientists can use default notebooks for common scenarios or build custom notebooks for specialized assets. After training, models are registered and deployed through Watson Machine Learning, then enabled for continuous scoring against asset groups. Predictions flow back into Maximo Health dashboards and asset records, where they can trigger work queues and work orders in Maximo Manage.

A concrete example helps illustrate the workflow. A chemical plant with two hundred critical pumps defines a pump asset group in Maximo Predict. The group includes attributes such as manufacturer, model, install date, operating hours, and process fluid. Work order history supplies failure codes for seal leaks, bearing failures, and cavitation damage. Vibration and temperature sensors feed Maximo Monitor with hourly readings. A data scientist uses the default failure probability notebook in Watson Studio, tweaks features based on the plant engineer's input, and trains a Gradient Boosting model. After validation, the model is deployed to score pumps daily. Pumps with a probability above 70 percent are added to a work queue for inspection. Over time, the plant tracks whether the model's high-probability alerts actually precede failures and uses that feedback to retrain the model quarterly.

The following table summarizes the model types and their typical use cases:

The Maximo Assistant and Generative AI

Maximo Assistant is IBM's generative AI interface embedded within MAS. It is built on watsonx.ai and is designed to answer natural language questions about asset data. A maintenance supervisor can ask which work orders are missing job plans, find assets at a specific site, identify open work related to a leak, or reassign a work order to another technician. The assistant interprets the intent, queries the Maximo data model, and returns a concise result within the same user interface.

The immediate value is speed. A supervisor who would otherwise navigate several screens or run a structured report can get an answer in seconds. The assistant also reduces the cognitive load on users who do not know the exact field names, query syntax, or navigation paths in Maximo. This is particularly useful for new users, occasional users, and managers who need quick situational awareness without becoming Maximo power users.

Generative AI in Maximo is also being applied to work order intelligence. The system can suggest problem codes based on work order descriptions, which improves data consistency and makes failure analysis easier. Over time, IBM has indicated that this capability will expand to surface patterns across asset types, maintenance history, and other dimensions. The goal is not to remove the human from the decision but to reduce the manual classification work that often leads to incomplete or inconsistent records.

Failure mode analysis is another promising use case. Traditionally, reliability engineers spend weeks compiling asset history, maintenance records, and operational data to understand why a class of assets fails. An AI assistant that can access the repository of historical and operational data can accelerate this by summarizing patterns and suggesting root causes. The engineer still validates and contextualizes the findings, but the time spent gathering and organizing information drops significantly.

Consider a typical supervisor scenario. A high-priority alarm arrives for a compressor at a remote site. The supervisor opens Maximo Assistant and asks for the asset's recent work history and related failure patterns. The assistant returns a summary of the last six work orders, the most common problem code, and a list of currently open work. The supervisor asks the assistant to reassign the newest work order to a technician with the right certification and then prompts it to identify all open work orders at that site missing a job plan. Tasks that would have required several queries, phone calls, and manual filtering now happen in one conversational session. The supervisor remains in control of every decision; the assistant simply removes friction.

Visual Inspection and Visual Prompting

Visual inspection in Maximo uses AI models to analyze images and video captured by drones, cameras, or mobile devices. The models detect defects, anomalies, and conditions that might indicate a problem. IBM's Visual Prompting capability allows a user to highlight a specific component in an image, such as a weld, bolt, or insulator, and teach the model to isolate and analyze that component. This reduces the need for massive labeled datasets and makes it easier to adapt models to new asset types.

The workflow typically starts with image capture. A drone inspects a transmission line, a camera monitors a production line, or a technician photographs a damaged pump. The images are fed into a model that scores each one for the presence of defined conditions. Exceptions are routed to human reviewers, who confirm the finding and create a work order if needed. This pattern removes technicians from hazardous environments for routine inspections and allows them to focus on repairs rather than rote observation.

In rail and infrastructure, visual inspection is used to detect cracked sleepers, missing fasteners, corrosion, and vegetation encroachment. In manufacturing, it can spot product defects or equipment wear. In oil and gas, it can identify leaks, flare stack conditions, and structural issues. The common thread is that the AI handles the first pass, and humans handle the judgment and action.

Adopting visual inspection requires representative images. A model trained only on daytime summer images may fail in winter lighting or nighttime infrared captures. The asset class, camera angle, lighting, and resolution all matter. Teams should start with a narrow defect class, build a diverse image library, and measure model precision and recall before scaling. Visual prompting lowers the training burden, but it does not remove the need for quality data.

A real production pattern is the drone inspection of transmission lines. A utility defines the defect classes of interest: broken insulators, corroved hardware, vegetation proximity, and damaged conductors. Field teams fly a set of test segments and capture thousands of images under varying weather, time of day, and camera angles. A domain expert uses visual prompting to mark examples of each defect. The model trains on these prompts and is then deployed to score new flights. Images flagged by the model are reviewed by engineers, and confirmed defects create inspection work orders in Maximo Manage with geolocation attached. The utility tracks precision, recall, and false positive rate by comparing model findings to manual inspections over the first six months.

Connecting Predictions to Work Management

AI produces insights, but maintenance still happens through work orders, schedules, and crews. The bridge between prediction and action is where many AI projects succeed or fail. A high failure probability alert that sits in a dashboard is less valuable than the same alert converted into a prioritized work queue with a job plan and a parts list. Maximo Predict and Maximo Manage are designed to work together so that predictions can flow directly into the maintenance workflow.

The simplest pattern is threshold-based work queue generation. A model scores assets daily. Assets whose failure probability exceeds a defined threshold are added to a queue for planner review. The planner confirms the finding, selects a job plan, checks parts availability, and schedules the work. This keeps humans in control while removing the manual task of scanning hundreds of asset health scores. Thresholds should be set by asset criticality and failure mode, not globally. A noncritical spare pump might have a higher threshold than a reactor cooling pump.

More advanced patterns combine predictions with resource optimization. A refinery might run a model that predicts heat exchanger fouling and then use Maximo Scheduler to bundle related work into the next planned turnaround. A fleet operator might combine rolling stock health scores with depot capacity and parts inventory to optimize maintenance windows. These integrations require clean data, stable interfaces, and close collaboration between data scientists, planners, and operations.

The key design principle is that the output of an AI model should be treated as a recommendation, not a command. Maintenance planners know things the model does not: upcoming shutdowns, parts shortages, contractor availability, and competing priorities. The system should make the recommendation easy to understand and act on, then capture the planner's decision for future training data. This feedback loop improves both trust and model accuracy over time.

Data Readiness: The Foundation of Every AI Project

No AI capability in Maximo works well without clean, relevant data. Predictive models need accurate failure history and sensor data. Generative assistants need well-structured records with consistent fields. Visual models need representative images with clear labels. Before launching any AI project, teams should audit the data that will feed it.

For predictive maintenance, the most common data gaps are inconsistent failure codes, missing install dates, incomplete meter readings, and sensor data that is not tied to the right asset record. A model trained on noisy failure labels will produce unreliable predictions. An asset without an install date cannot be used in age-based survival models. Sensor streams that are not mapped to Maximo asset identifiers cannot be correlated with work history. Fixing these issues is unglamorous work, but it determines whether the project succeeds.

For generative AI, the challenge is data accessibility and permissions. The assistant must be able to query the same records the user is authorized to see. If asset hierarchies are messy or field names are inconsistent, the assistant may return incomplete answers. Data governance, access controls, and metadata quality all matter. Organizations should not assume that a generative assistant will understand a poorly organized data model.

For visual inspection, the primary data concern is image diversity and label quality. A model that works in one plant may not work in another if the equipment looks different, the camera angle changes, or the lighting varies. Labels must be precise enough to train the model but consistent enough to avoid conflicting examples. A small, well-curated dataset usually beats a large, noisy one.

The following checklist can be used before starting an AI project in Maximo:

Model Governance and Lifecycle Management

AI models are production assets and should be managed like production assets. That means version control, change management, monitoring, and retirement. A model that predicts pump failures should have an owner, a documented training dataset, a validation report, and a retraining schedule. When operating conditions change, such as a new pump vendor or a different process fluid, the model should be reassessed and possibly retrained. Treating models as static artifacts is a common reason predictive maintenance initiatives lose accuracy over time.

Model drift is a real concern. Predictive models can degrade when the asset population, operating environment, or maintenance strategy changes. A model trained during a period of high utilization may underperform when utilization drops. A model trained on one vendor's pumps may not generalize to another vendor's design. Monitoring key metrics such as precision, recall, calibration, and alert rate helps teams detect drift before it becomes a business problem. Scheduled retraining, triggered retraining, and champion-challenger comparisons are all valid governance patterns.

Access control and explainability also matter. Maintenance planners need to know why an asset was flagged, not just that it was flagged. Maximo Predict surfaces prediction scores and contributing features, which helps users understand the basis for a recommendation. Data scientists should document assumptions, feature lists, and known limitations. This documentation is essential when models affect safety-critical decisions or when regulators ask for evidence that decisions were made responsibly.

The final governance consideration is retirement. A model that no longer adds value should be decommissioned rather than left running silently. This requires a defined process for identifying underperforming models, archiving their history, and notifying users. Model retirement is as important as model deployment. An outdated model that still generates alerts consumes planner attention and erodes trust in the entire AI program.

Practical Implications

AI adoption in Maximo should be staged. Start with one well-defined problem, such as predicting failures for a single critical pump class or using the assistant to identify work orders missing job plans. Measure the outcome before expanding. This approach builds organizational trust and reveals data gaps early.

Teams should also set realistic expectations with leadership. AI provides probabilities, not certainties. A 78 percent failure probability does not mean the asset will definitely fail. Visual inspection may flag false positives that require human review. The assistant may misunderstand an ambiguous question. These limitations are manageable, but they must be communicated. Overpromising leads to disillusionment and abandoned projects.

Governance is essential. Predictive models should be retrained when asset conditions, operating contexts, or failure patterns change. Assistants should have clear boundaries for what they can and cannot do. Visual models should be reviewed for bias and drift. Assign ownership for each model or assistant, document assumptions, and schedule periodic reviews. AI in Maximo is not a one-time installation; it is an ongoing capability that requires care.

Organizations should also create a feedback loop between the field and the data science team. Maintenance technicians and reliability engineers know when a prediction seems wrong or when a visual flag is a false positive. That feedback should be captured and used to improve training data, feature engineering, and model thresholds. Without this loop, models drift away from operational reality.

Bottom Line

AI in Maximo spans predictive maintenance, generative assistance, visual inspection, and the integration of predictions into work management. Each capability is powerful when matched to the right problem and supported by clean data. Maximo Predict turns work history and sensor data into failure probability and remaining life estimates. Maximo Assistant uses natural language to accelerate queries and administrative tasks. Visual prompting makes image-based defect detection more accessible. The connection to work management turns predictions into scheduled, resourced actions. None of these tools replace human expertise, but they can extend it. The organizations that benefit most are those that invest in data readiness, set realistic expectations, govern their models as production assets, and close the feedback loop with the people who maintain the assets every day.