AI in Maximo: From Maximo Assistant to Visual Inspection, Predict, and the watsonx Layer

For most of its history, Maximo was a deterministic system. You configured the rules, the rules ran, and the work order got created or it did not. That world is not gone, but it is no longer the whole story. The 2026 release of Maximo Application Suite layers a substantial set of AI capabilities on top of the deterministic core, and the practical question for most Maximo teams is no longer whether to use AI in Maximo. It is which AI capabilities to use first, how to deploy them responsibly, and how to make sure the data underneath is good enough to support them.

The AI capabilities in MAS fall into four buckets. Maximo Assistant is a natural language interface that lets users query data, get explanations, and automate routine tasks using plain English. Maximo Visual Inspection is a computer vision tool that turns cameras into inspection devices, with drag-and-drop model training and edge deployment. Maximo Predict is a machine learning platform that builds failure prediction models from historical work order and asset data. Underpinning all of these is watsonx, IBM's AI and data platform, which provides the model serving, governance, and integration with external AI services. This article walks through each of these capabilities, explains what they are good for, and provides practical guidance for getting them into production.

The AI Landscape in MAS 9

The MAS 9.2 release, announced in June 2026, makes the AI positioning explicit. IBM describes 9.2 as bringing "asset-first AI directly into the workflows that reliability, maintenance, field service, safety and operations teams use every day." That phrasing matters. The AI is not a separate product that sits next to Maximo. It is embedded in the same UI, the same data model, and the same workflows that technicians and reliability engineers already use.

The product-level positioning breaks down like this:

The shift from MAS 9.1 to 9.2 is significant. In 9.1, Maximo Assistant was capable but read-only: it could answer questions, retrieve records, and compute analytical aggregates, but it could not take action. In 9.2, the Assistant evolves into a "persistent, context-aware teammate that can discuss, reason, and act across your Maximo data and documents." That means the Assistant can draft work orders, update records, schedule inspections, and trigger workflows based on natural language requests. The user reviews and approves, but the Assistant does the heavy lifting.

For mobile, the 9.2 release introduces "AI comes to the frontline, empowering technicians with AI capabilities, hands-free voice interaction and computer vision to accelerate inspections and instant data capture." This is a direct response to the field reality that most technicians work in gloves, in noise, in awkward positions, and away from keyboards. Voice-driven interaction and on-device visual inspection remove the friction that has historically limited field adoption of digital tools.

Maximo Assistant: Natural Language for Asset Data

Maximo Assistant is the most visible AI capability in MAS 9, and the easiest one to start using. It lives in the Maximo UI as a chat-style interface, and it understands natural language requests like "show me all open work orders on Pump 001", "how many inspections are overdue this week", or "create a follow-up work order for the issue we discussed yesterday."

The capabilities in 9.1 and 9.2 fall into three categories:

1. Database Retrieval: instantly display work, asset, and service data using plain language. No need to build a query in the Maximo UI or remember the field name.
2. Analytical Insights: get real-time metrics like count, sum, average, min, max, and trend analysis without writing a BIRT report or running a query.
3. Interactive Results: customize and filter data in a full-screen workspace, then drill into individual records.

In 9.2, the Assistant gains action capabilities. It can draft work orders, create service requests, schedule inspections, and update records based on natural language instructions. The user always confirms before the action is taken, but the data entry and routing logic is handled by the Assistant. For organizations that struggle with data entry quality, this is a significant productivity win.

Here is a simple example of what an Assistant interaction looks like:

User: Show me all work orders on Pump 001 from the last 90 days
       that are either closed or in progress, sorted by scheduled
       start date descending.

Assistant: I found 23 work orders matching your criteria on
           PUMP-001 from the last 90 days. Here's a summary:

           - 14 closed work orders
           - 9 in progress
           - Total labor hours: 142
           - Total cost: $18,400

           Would you like me to:
           1. Show the full list in a table
           2. Filter further (e.g., by work type)
           3. Create a follow-up work order for any of these

User: Option 3. Create a follow-up for WO-1234567, due in 30 days.

Assistant: I've drafted a follow-up work order:
           - Reference: WO-1234568 (draft)
           - Asset: PUMP-001
           - Description: "Follow-up inspection per WO-1234567"
           - Due date: 2026-08-19
           - Priority: 3 (matches parent)

           Confirm to create this work order?

The practical value here is not that the AI is clever. It is that the user does not need to navigate the Maximo UI, remember the work order tracking application, know the field name for "scheduled start date", or write a query. The friction between intent and action is dramatically reduced.

Deployment-wise, Maximo Assistant requires no special infrastructure in 9.2. It is bundled with MAS 9.2 and uses IBM's hosted models by default, with the option to deploy against an on-premises watsonx instance for organizations with strict data residency requirements. Configuration is minimal: an admin enables the Assistant for the relevant security groups, sets the data access scope, and configures any custom actions or guardrails the organization needs.

Maximo Visual Inspection: Computer Vision for Field Teams

Maximo Visual Inspection (formerly IBM Visual Insights, formerly PowerAI Vision) is a computer vision tool that puts deep learning model training and deployment into the hands of inspectors who may have no background in machine learning. The use case is straightforward: train a model to recognize defects, anomalies, or specific conditions in images and video, then deploy the model to mobile devices, edge appliances, or cloud endpoints.

The technical workflow has three phases: capture, train, and deploy. In the capture phase, the user uploads images of the object or condition of interest. For a defect detection model, the images include examples of both good and defective conditions. The labeling tool, which is web-based and drag-and-drop, lets the user draw bounding boxes around the objects of interest and assign labels. For anomaly detection, only "good" images are required. The model learns what "normal" looks like and flags anything that deviates.

In the train phase, the user clicks "Train" and the system selects an appropriate model architecture (or uses a custom model imported from an external training run), trains the model on the labeled data, and reports performance metrics. A typical model might be ready in 15 minutes to a few hours, depending on dataset size and hardware.

In the deploy phase, the trained model can be deployed as a cloud endpoint, packaged into a mobile application, or pushed to an edge appliance. Maximo Mobile includes built-in support for Visual Inspection models, so a technician in the field can capture an image with the mobile app and get an instant inference result. For example, a technician inspecting a thermal imaging camera feed can run the image through a Visual Inspection model that flags overheating components.

The Volvo case study is a classic example. Cameras mounted above a vehicle production line capture images of each car as it rolls off the line. A Visual Inspection model, trained on examples of perfect vehicles, detects imperfections that are nearly impossible for the human eye to catch. The system reports the size and location of each defect, and the car is either passed, sent for repair, or sent for repaint based on the configuration. This kind of automated visual inspection is becoming standard in automotive, electronics, and pharmaceutical manufacturing.

For infrastructure and utilities, the use cases are different but equally compelling. Drones equipped with thermal cameras can inspect transmission lines, substations, and pipelines. A Visual Inspection model trained on examples of insulator damage, oil leaks, or corrosion can flag potential issues for human review. The same approach works for PPE compliance: a model trained on hard hats, safety vests, and protective eyewear can verify that workers in a video feed are wearing the required gear.

Maximo Visual Inspection runs on OpenShift and can be deployed on a variety of hardware, including GPU-equipped nodes for training and CPU-only edge devices for inference. The edge deployment story is particularly important for organizations that need to inspect assets in remote locations without reliable connectivity. The model is trained in the cloud, packaged for the edge device, and runs inference locally. Results are cached and synced when connectivity is restored.

Maximo Predict: Machine Learning for Failure Forecasting

Maximo Predict is the AI capability that addresses the core reliability engineering question: when is this asset likely to fail? The answer drives the maintenance strategy, the spare parts inventory, and the capital replacement plan. Getting it right has measurable financial impact, and getting it wrong (either too aggressive, which wastes money on unnecessary maintenance, or too conservative, which leads to unexpected failures) is what most reliability programs are trying to fix.

Predict works by building a machine learning model from the historical asset and work order data in Maximo. The inputs include failure history, work order types, labor and material costs, downtime hours, environmental conditions, and any sensor data that has been integrated through Monitor. The output is a set of failure predictions, each with a probability and an estimated time to failure. Reliability engineers use these predictions to drive the maintenance plan.

There are three deployment patterns, depending on the asset criticality and data availability:

1. Time-based with sensor enrichment: classic preventive maintenance, but with the interval adjusted based on actual operating conditions. A pump that runs continuously gets a different maintenance schedule than one that runs intermittently.
2. Condition-based with anomaly detection: sensor data is monitored in real time, and anomalies trigger work orders. Used for rotating equipment, electrical systems, and other assets with rich sensor coverage.
3. Predictive with ML models: machine learning models trained on historical failure patterns predict failures before they occur. Used for high-criticality assets where the cost of failure justifies the investment in model development and validation.

The data quality challenge is real. Predictive models are only as good as the data they are trained on, and most Maximo customers have a mix of clean and messy data. Common issues include inconsistent failure coding, missing failure cause information, and incomplete work order closure data. Before deploying Predict, a data quality assessment is essential. The good news is that the assessment itself often surfaces improvements that pay off even without the AI.

For organizations that already have reliability programs built on failure codes, FMEA, and RCM, Predict is a natural extension. The historical analysis that reliability engineers have been doing manually for years can be automated and made more accurate. For organizations that are just starting their reliability journey, Predict can be a forcing function for the data quality work that needs to happen anyway.

watsonx: The Underlying AI Platform

Underneath Assistant, Visual Inspection, and Predict sits watsonx, IBM's AI and data platform. watsonx is not a Maximo product per se. It is the broader platform that provides model serving, data fabric, and AI governance. Maximo applications call into watsonx for inference, and watsonx provides the security, audit, and lifecycle management that enterprise AI deployments require.

The two components most relevant to Maximo deployments are watsonx.ai and watsonx.data. watsonx.ai is the model serving and development environment, with support for both IBM's foundation models and third-party models. watsonx.data is the data lakehouse that consolidates the structured and unstructured data that AI applications need.

For most Maximo customers, the practical interaction with watsonx is at the deployment level. The Maximo installer can provision a watsonx instance alongside MAS, and the Maximo applications are configured to use that instance. For organizations that already have a watsonx deployment for other AI use cases, the Maximo applications can be configured to use the existing instance. The advantage of the latter is consolidated governance, shared model catalog, and unified audit trail.

The governance features in watsonx are particularly important for regulated industries. Model cards, bias detection, drift monitoring, and explainability features help organizations comply with the EU AI Act, the US AI Bill of Rights, and various industry-specific regulations. For Maximo deployments in utilities, oil and gas, and healthcare, these governance capabilities are often a prerequisite for AI adoption, not an afterthought.

Practical Implications

For most organizations, the right starting point is Maximo Assistant. It is the lowest-risk, highest-visibility AI capability in MAS 9. It does not require new data sources, it does not require model training, and it does not require GPU infrastructure. The deployment is configuration, the user adoption is driven by the immediate productivity benefit, and the data usage is auditable. A team can have Assistant in production within a few weeks and start measuring time saved on routine queries and data entry.

Visual Inspection is the right second step for organizations with clear visual inspection use cases. Manufacturing quality control, infrastructure inspection with drones, and PPE compliance are the most common starting points. The investment is more substantial (you need images, labels, and model training), but the ROI can be large, particularly for high-volume or safety-critical inspections.

Predict is the right step for organizations that already have a reliability program and clean data. The temptation to jump straight to Predict is strong, but the failure mode is predictable: the predictions are wrong because the data is bad, the reliability team loses trust, and the AI program stalls. Investing in data quality first, then deploying Predict, is the slower path but the one that delivers durable results.

For all three, the practical advice is the same. Start with a single use case. Pick a measurable outcome. Build the data foundation before the model. Validate rigorously before scaling. And treat AI as a tool that augments human decision-making, not a replacement for it.

Bottom Line

AI in Maximo is no longer a future state. Maximo Assistant, Visual Inspection, Predict, and the underlying watsonx platform are all in production today, and the 9.2 release tightens the integration between them. The practical question is no longer whether to use AI in Maximo, but where to start and how to sequence the deployment.

The recommended sequence is Assistant first, Visual Inspection second (if the use case fits), and Predict third (after the data is ready). Each step builds organizational familiarity with AI, builds trust through measurable wins, and creates the data foundation for the next step. watsonx provides the governance and infrastructure backbone that makes enterprise AI deployment sustainable.

For organizations that are still evaluating, the time to start is now. The platform is mature, the deployment patterns are well understood, and the early adopters are already seeing measurable returns. The risk of waiting is not that the AI will not be ready. The risk is that your competitors' data will be cleaner, their predictions will be more accurate, and their maintenance programs will be more efficient than yours.