AI Inside Maximo 9.2: Condition Insight, Agentic Workflows, and the End of Gut-Feel Maintenance

For years, AI in enterprise asset management was a promise wrapped in a demo. Vendors showed dashboards with predictive models that looked impressive in conference rooms but fell apart when confronted with the messy, incomplete, often contradictory data that real maintenance organizations actually produce. IBM Maximo Application Suite 9.2, released in June 2026, represents a genuine shift away from that pattern. AI is no longer a separate layer bolted on top of Maximo. It is embedded directly into the workflows that reliability, maintenance, field service, safety, and operations teams use every day.

This is not a cosmetic distinction. When AI sits outside the workflow, it requires someone to check the AI output, translate it into action, and update the system of record. That translation step is where most AI initiatives stall. The AI model predicts a failure, the prediction sits in a dashboard, nobody creates the work order, and the asset fails anyway. MAS 9.2 closes that loop by putting AI where the work happens, so that insights translate into actions within the same system that manages the work.

This article examines the specific AI capabilities in MAS 9.2, how they work, what data they require, and where the limitations are. The goal is to give practitioners a clear-eyed assessment of what AI in Maximo can and cannot do today, based on IBM's documentation, community experience, and the architectural realities of enterprise asset data.

Maximo Condition Insight: Pattern Recognition for Asset Health

Maximo Condition Insight is the flagship AI capability in MAS 9.2. It brings together work orders, inspections, meter readings, and reliability strategies to identify patterns in asset behavior and recommend what to do next. The intent is to help teams diagnose issues earlier, make more consistent decisions, and reduce dependence on fragmented systems or individual expertise alone.

Condition-based maintenance has traditionally required significant data analysis by experienced specialists. A reliability engineer would pull work order history, cross-reference it with meter readings and inspection results, look for patterns in failure data, and make a judgment call about what to do. This process is slow, inconsistent across engineers, and does not scale to organizations with thousands of assets. Condition Insight aims to simplify and scale that analysis.

The way Condition Insight works architecturally is important to understand. It does not replace the existing Health and Predict applications. It sits on top of them, adding a layer of pattern recognition and recommendation generation that the individual applications cannot produce on their own. The data flow looks like this:

• Maximo Health calculates health scores for assets based on meter readings, work order history, and asset attributes. Health scores are numeric values (0-100) that represent the current condition of an asset relative to its expected baseline.
• Maximo Predict applies machine learning models to historical failure data to generate failure probability scores and estimated failure dates. These models are trained using Jupyter notebooks provided with the platform, and data scientists can customize the models for specific asset types or failure modes.
• Maximo Monitor collects IoT sensor data from connected assets, providing real-time condition data that feeds into health calculations and anomaly detection.
• Condition Insight synthesizes all of the above, plus inspection results and reliability strategy data, to identify patterns that no single data source would reveal on its own. It then generates recommendations for action.

The recommendation engine is where the AI adds the most value. Rather than simply flagging assets as "unhealthy" or "likely to fail," Condition Insight generates specific, actionable recommendations. For example, instead of saying "PUMP-101 has a 72% probability of failure in the next 30 days," it might recommend: "Inspect PUMP-101 seal assembly. Similar vibration patterns in PUMP-103 and PUMP-107 led to seal failure within 15 days. Create work order with priority 2 and assign to seal repair crew."

The quality of these recommendations depends heavily on the quality of the underlying data. This is the fundamental limitation of AI in asset management, and it is worth spending time on. Maximo's AI capabilities rely on three categories of data:

Asset hierarchy data: The asset record, its parent-child relationships, its location in the site and organizational hierarchy. This is usually well-maintained in organizations that have been on Maximo for years, but newly implemented systems often have incomplete or inconsistent hierarchies.

Historical maintenance data: Work order history, including failure codes, remedies, labor transactions, and material issues. This is where most organizations have problems. Work orders are often closed with generic failure codes ("MECHANICAL FAILURE"), no remedy codes, or missing labor transactions. AI models trained on this data will produce unreliable predictions because the patterns they identify are based on incomplete information.

Condition data: Meter readings, IoT sensor data, and inspection results. This is the newest data category for most organizations, and the quality varies widely. IoT sensor data is usually high quality (sensors produce consistent readings), but manual meter readings and inspection results often suffer from data entry errors and inconsistencies.

Organizations considering Condition Insight should conduct a data quality assessment before deployment. Check the last 12 months of work order data for failure code completeness, remedy code usage, and labor transaction accuracy. If more than 30% of work orders have missing or generic failure codes, address that issue first. No AI model can compensate for training data that does not accurately capture what actually happened.

Agentic Workflows and the MCP Server

MAS 9.2 introduces agentic workflows, which represent a step beyond the recommendation model described above. Instead of generating recommendations that a human must act on, agentic workflows allow AI agents to participate in operational processes directly, taking actions within Maximo through the MCP Server.

The distinction between recommendations and agentic actions is significant. A recommendation says "you should create a work order." An agentic action creates the work order, routes it through the appropriate approval workflow, and notifies the assigned technician. The human's role shifts from being the executor of AI recommendations to being the approver of AI actions.

This shift has implications for organizational design and trust. Organizations that have spent years building trust in human-driven maintenance processes are now being asked to trust AI-driven actions. The transition needs to be gradual, with clear guardrails and audit trails. MAS 9.2 provides the technical infrastructure for this through the MCP Server's governance capabilities, but organizational change management is the harder problem.

A practical example of an agentic workflow in action:

• A reliability agent monitors asset health scores every hour via the MCP Server.
• When PUMP-101's health score drops below 60, the agent queries the last 90 days of work order history for that asset.
• The agent analyzes the pattern (vibration increasing, temperature trending up, last inspection 45 days ago) and determines that a seal inspection is needed.
• The agent creates a work order through the Maximo REST API with the appropriate asset, job plan, priority, and safety plan.
• The work order enters the normal approval workflow. A human supervisor reviews and approves it.
• The agent monitors the work order status and, when it is completed, updates its own model with the outcome data.

This is not science fiction. Every step of this workflow is technically possible with MAS 9.2. The question is whether your organization is ready for it, and whether your data quality supports reliable AI decision-making.

Maximo Assistant on Mobile: Natural Language in the Field

Field service technicians are the unsung heroes of asset management. They are also the people most likely to be frustrated by enterprise software that requires navigating through five screens to find the information they need. Maximo Assistant on Mobile, introduced in MAS 9.2, addresses this by enabling technicians to use natural language to find asset information, review history, and complete work efficiently in the field.

The assistant is not a chatbot. It is a natural language interface to Maximo's data and workflows. A technician can ask "What work orders are assigned to me today?" and get a structured response with the relevant work orders, their priorities, locations, and asset details. They can ask "Show me the last five work orders on PUMP-101" and get the history without navigating through the asset application. They can complete work orders by saying "Close work order 1234, set status to COMP, enter 3.5 hours labor."

The technology behind this combines natural language processing (to understand the technician's intent), Maximo's REST API (to retrieve and update data), and generative AI (to format responses in a way that is easy to read on a mobile device). The assistant runs on the mobile device, which means it works in disconnected environments. The AI model is served from the MAS cluster, but the natural language understanding and response generation happen on-device when connectivity is available, with fallback to simpler text-based interactions when offline.

The practical value of the mobile assistant is in reducing the time technicians spend on administrative tasks. In field service operations, technicians typically spend 30-40% of their day on administrative work: finding asset information, reviewing work order history, entering labor hours, updating work order status. The assistant compresses these tasks into natural language interactions that take seconds rather than minutes. Over a day, this adds up to 1-2 hours of additional productive time per technician.

Maximo Visual Inspection: AI-Driven Quality Checks

Maximo Visual Inspection enables AI-based visual inspection with local inference directly on the device. This capability has been available in earlier MAS versions but receives significant enhancements in 9.2, including improved model accuracy, support for more device types, and better integration with Maximo Manage workflows.

The use cases for visual inspection in asset management are broader than they might appear at first:

• Corrosion detection: A technician photographs a pipe section, and the visual inspection model identifies corrosion, estimates its severity, and flags the asset for follow-up inspection or repair.
• Equipment condition assessment: A technician photographs a motor or pump, and the model identifies visual indicators of wear (oil leaks, mounting bolt condition, housing damage) that might not be captured by sensor data.
• Safety compliance verification: A technician photographs a locked-out/tagged-out energy source, and the model verifies that the lockout device is properly installed and the tag is present and legible.
• Inventory verification: A technician photographs a storeroom shelf, and the model counts the items visible and compares against the expected inventory in Maximo.

The local inference capability is important for field operations. Running the visual inspection model on the device means the technician gets results immediately, without uploading the image to a server and waiting for a response. This is critical in environments with poor connectivity (underground facilities, remote field sites, inside large metal structures where cellular service is unreliable). It also addresses data privacy concerns, since images of sensitive equipment do not need to leave the device.

The models are trained using IBM Maximo Visual Inspection's training pipeline, which allows organizations to upload labeled images and train custom models for their specific equipment. Pre-trained models are available for common use cases (general corrosion, surface defects), but most organizations will need to train custom models on their own equipment images to achieve acceptable accuracy. This requires a library of labeled images, typically 200-500 images per defect type, which can be a significant upfront investment for organizations starting from scratch.

What AI in Maximo Cannot Do (Yet)

Honesty about limitations is essential for setting realistic expectations. Here are the things that AI in Maximo 9.2 cannot do, despite what marketing materials might imply:

It cannot replace experienced reliability engineers. AI can identify patterns and generate recommendations faster than a human can, but it cannot replicate the contextual knowledge that an experienced engineer brings. The engineer knows that PUMP-101 has had a recurring seal issue because of a design flaw in the piping layout, something that is not captured in any data field. AI recommendations should be treated as input to the engineer's decision-making process, not as a replacement for it.

It cannot compensate for poor data quality. If your work order history is full of generic failure codes, missing labor transactions, and incomplete asset hierarchies, AI models will produce unreliable predictions. This is not a limitation that future AI advances will overcome. The models can only work with the data they are given, and garbage in still produces garbage out.

It cannot predict novel failure modes. Machine learning models are trained on historical data. If an asset fails in a way that has never been seen before, the model will not predict it. This is particularly relevant for new equipment that has no failure history, or for assets operating under new conditions (different load profiles, new environmental conditions, modified operating procedures).

It cannot manage organizational change. Implementing AI-driven maintenance requires changes to how work is prioritized, how decisions are made, and how teams collaborate. These are organizational challenges, not technical ones, and AI cannot solve them. Organizations that deploy AI capabilities without addressing the organizational change dimensions will see poor adoption and limited value.

Practical Implications

For organizations on MAS 8.x or earlier, Condition Insight and the agentic workflow capabilities in 9.2 are strong reasons to upgrade. But before upgrading, conduct a data quality assessment. The single highest-ROI activity for most organizations is not deploying new AI features. It is cleaning up work order data, improving failure code compliance, and ensuring that asset hierarchies are accurate and complete. This data quality work is the foundation that AI capabilities build on.

For organizations on Maximo 7.6.x, the path to AI requires migrating to MAS first. The AI capabilities are not available in 7.6.x, and there is no backport path. Use the migration as an opportunity to establish good data practices: enforce failure code usage, require labor transactions on all work orders, and implement mobile solutions that capture data at the point of work rather than relying on after-the-fact data entry.

For teams building custom AI models using Maximo Predict's notebook environment, invest in understanding the data pipeline. The default notebooks provided by IBM are starting points, not production-ready models. Real-world deployment requires custom feature engineering, model validation against held-out data, and ongoing monitoring of model performance. A model that was accurate when deployed may degrade over time as asset conditions, operating patterns, or maintenance practices change.

Start with Condition Insight and the mobile assistant. These capabilities deliver immediate value with relatively low implementation effort and do not require custom model development. Agentic workflows and custom AI agents are more powerful but require significant investment in data infrastructure, model development, and organizational change management. Walk before you run.

Bottom Line

MAS 9.2 represents the first time that AI in Maximo is genuinely embedded in operational workflows rather than being a separate analytical layer. Condition Insight, the mobile assistant, and visual inspection deliver tangible value today. Agentic workflows and the MCP Server open the door to more advanced AI-driven automation, but realizing that potential requires investment in data quality, model development, and organizational change. The organizations that will benefit most are not the ones with the most advanced AI models. They are the ones with the cleanest data, the most disciplined maintenance processes, and the strongest commitment to using AI as a tool that supports human decision-making rather than replacing it.