Agentic AI in Maximo: How Condition Insight and watsonx Are Changing Maintenance Decisions

# Agentic AI in Maximo: How Condition Insight and watsonx Are Changing Maintenance Decisions

Artificial intelligence has been attached to asset management for years, but most early implementations were narrow: anomaly detection on a single sensor, failure probability for one asset class, or a dashboard that highlighted outliers. The hard part was never the algorithm. It was the integration, the explanation, and the action. Data scientists built models that operations teams did not trust, and maintenance planners could not act on scores they did not understand.

That is the gap IBM is now trying to close in Maximo Application Suite. With the introduction of Maximo Condition Insight in late 2025, MAS moves beyond raw predictions toward explainable, prescriptive maintenance recommendations delivered in plain language. The capability is agentic in the sense that it interprets multiple data sources, maps condition to failure modes, and suggests corrective actions. It is built on IBM watsonx and embedded directly into Maximo workflows, so the insights appear where maintenance teams already work.

This article examines how AI is currently being delivered in MAS 9, what Maximo Condition Insight does differently from traditional predictive models, and what organizations should do before they turn these features on. The goal is to separate the genuinely useful from the merely hyped and to give technical readers a realistic map of where AI helps and where it still requires human judgment.

The Two Paths of Predictive Maintenance in Maximo

Maximo Predict, the predictive maintenance module in MAS, typically operates along two complementary paths. 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 failure patterns, the model learns how assets have failed in the past. This works best for repetitive assets with known failure modes and consistent failure coding.

The second path is prediction based on condition and sensor data. Here, the model learns from vibration, temperature, pressure, runtime, and other condition indicators combined with degradation trends from Maximo Health. Instead of asking what failed before, this approach asks what behavior usually precedes a failure. That is where early detection becomes possible, often before any threshold is crossed.

The real value is not accuracy in isolation. It is prioritization. A planner who sees rising failure probability, declining health, and a known historical failure pattern can make a confident, explainable decision: plan the maintenance, allocate parts, schedule labor, and avoid both panic and complacency. As one IBM community contributor put it, Predict does not tell you what will fail. It tells you where to look first.

A typical manufacturing example illustrates the difference. A food packaging line might experience bearing failures every few months. A calendar-based PM schedule replaces bearings on a fixed interval, regardless of condition. A Predict model that analyzes vibration trends, runtime hours, and operating temperature can flag the specific bearing that is degrading, allowing a targeted replacement during the next planned outage. The outcome is fewer unplanned stops, less collateral damage, and lower parts usage because good bearings are no longer replaced prematurely.

In utility settings, the same prioritization logic applies to transformers. A model that combines age, loading history, DGA results, and weather exposure can rank transformers by 12-month failure probability. Capital planners then use that ranking to target replacement investments, replacing the highest-risk units rather than the oldest units on a schedule. One cited utility case achieved 30 percent better targeting of replacement investments after connecting condition data to the asset management system.

What Maximo Condition Insight Adds

Announced in December 2025, IBM Maximo Condition Insight is described as an agentic AI capability within Maximo Asset Performance Management. It evaluates work orders, metrics, time-series data, meter readings, failure mode and effects analysis (FMEA), and alerts to evaluate asset condition, uncover performance patterns, and provide clear, actionable recommendations communicated in plain language.

The key difference from traditional dashboards is explanation. Instead of presenting a health score of 72 and leaving the user to interpret it, Condition Insight returns a summary of what is happening, why it matters, and what should be done next. For example, it might note that a transformer shows increasing dissolved gas trends, map those trends to a specific failure mode, and recommend scheduling an internal inspection before the next peak load period.

The architecture connects Maximo Manage to Maximo AI Service through the Maximo Manage extensible framework. Large language models from watsonx.ai handle natural language generation and interpretation, while metadata coordinates model configuration, binds models to Maximo business objects, and manages the inferencing process. Object structures, query definitions, query templates, endpoints, and invocation channels all participate in the data flow.

This design matters because the AI is not bolted on. It is wired into the same object model that already defines assets, work orders, locations, and job plans. That means recommendations can reference real work orders, real failure codes, and real asset hierarchies. The explanations are grounded in data the organization already trusts.

Key components of the integration, as described by IBM, include:

- Object structures: Define the data shape and permissions for AI features.
- Object structure query definitions: Select training and inference data sets.
- Object structure query templates: Define schemas for training, inference, and configuration.
- Endpoints: Communicate with external services such as Maximo AI Service APIs.
- Invocation channels: Map inbound and outbound data and process responses.

The agentic aspect is important. Condition Insight does not just summarize data; it proposes next steps. Those steps might include creating an inspection work order, adjusting a PM interval, or escalating an alert to a reliability engineer. The user approves or overrides the action, keeping a human in the loop while removing the friction of interpreting raw scores.

The Maximo AI Assistant and Conversational Queries

MAS 9.1 introduced a generative AI assistant built on watsonx.ai. The assistant allows maintenance managers to query asset intelligence in plain English without navigating complex screens or running manual reports. A user can ask how many open high-priority work orders exist for a particular site, what the recent failure trends are for a pump class, or which assets have overdue inspections. The assistant interprets the question, queries the relevant Maximo objects, and returns a summarized answer.

The significance is not that it replaces reports. It is that it lowers the threshold for asking questions. A floor supervisor who would never run a SQL query or build an ad hoc report can now ask a natural language question during a morning briefing. The assistant is also useful for onboarding: a new reliability engineer can explore the asset base conversationally instead of memorizing dozens of application screens.

However, the assistant is only as good as the underlying data quality. If failure codes are inconsistent, asset hierarchies are incomplete, or work order statuses are not maintained, the answers will be misleading. Organizations that get the most value from the AI assistant are usually the same ones that invested in data governance before turning on the AI features.

The assistant also points to a future where natural language becomes a legitimate interface for operational systems. That future is closer than many organizations think, but it depends on the same data foundations that make dashboards useful. There is no shortcut around clean, well-modeled asset data.

Visual Inspection: AI That Reads Images Instead of Technicians

Maximo Visual Inspection uses deep learning to analyze images and videos for classification and defect detection. It allows users to train, validate, and deploy AI models without requiring data science expertise, and it integrates into inspection workflows so that detected issues can be routed directly into work management.

The nybl deployment is a useful example. By integrating its n.vision platform with Maximo Visual Inspection and watsonx.governance, nybl performs asset-based inspections that reduce the need for manual access to hard-to-reach areas. A national grid operator in the Gulf Cooperation Council launched an initial 1,000-kilometer pilot for high-voltage and medium-voltage line inspections. Following the pilot, the operator awarded nybl a contract to scale inspections across 400,000 kilometers of power infrastructure.

Documented outcomes include a 50 percent reduction in inspection costs and safety incidents, a 30 percent decrease in inspection time, a 20 percent reduction in outage and emergency repair costs, and a 20 percent increase in grid uptime. The underlying technology is Maximo Visual Inspection, which uses deep learning to analyze images and videos for defect classification and detection, allowing users to train, validate, and deploy models without requiring data science expertise.

For maintenance organizations, visual inspection is one of the lowest-friction AI entry points because it augments an existing process. Inspectors already take photos. The AI simply classifies them faster and more consistently, then feeds the results into the same work order system that schedules repairs. The safety benefit is also significant: fewer climbs, fewer bucket-truck deployments, and fewer exposures to hazardous environments.

What Needs to Be in Place Before AI Delivers Value

AI in Maximo is not a magic switch. It requires a foundation that many organizations are still building. The prerequisites fall into four categories: data, integration, governance, and skills.

Data is the most common blocker. Historical work orders must have meaningful failure codes, root causes, and asset references. Sensor data must be aligned to the correct asset and time zone. Inspection records must be complete enough to train or validate models. If the input data is garbage, the AI output will be garbage dressed in a clean interface.

Integration is the second requirement. Condition Insight and Predict draw from multiple sources: Manage for work order and asset context, Monitor for sensor and time-series data, Health for degradation trends, and external systems for lab results or weather data. Each connection must be reliable, secure, and documented.

Governance is the third. AI recommendations that affect maintenance scheduling and safety must be traceable. Who approved the model? What data was used to train it? When was it last validated? IBM addresses this through watsonx.governance, which provides model lifecycle tracking, drift detection, and explainability features.

Skills are the fourth. Maintenance teams need enough literacy to interpret AI outputs and enough skepticism to question them. Data scientists or AI engineers may be needed for custom model development, but many Maximo AI features are designed to be operated by domain experts after initial setup.

A Realistic Implementation Sequence

Organizations that succeed with AI in Maximo usually follow a sequence rather than attempting everything at once. The first step is to cle

...