Work Order Intelligence and the AI Assistant: A Deep Dive into Maximo Manage 9.x

Work Order Intelligence and the AI Assistant: A Deep Dive into Maximo Manage 9.xMaximo Manage 9.0 introduced Work Order Intelligence, and 9.1 added the Maximo Assistant. Together, these are the most consequential changes to the Manage application since the move to the Maximo Application Framework. They are powered by watsonx, they run inside the Manage user interface, and they address a problem that has plagued Maximo installations for decades: the quality of failure data. This deep dive covers what each capability does, how it works under the hood, how to configure it for production, and what the 9.2 feature channel brings to the table. If you have ever looked at a Maximo report on failure trends and realized that half the work orders have no failure code or the wrong one, this is the article for you.

Work Order Intelligence: The Problem It Solves and How It Works

Failure coding is the data that powers reliability analysis. Every report on asset failure trends, every RCM study, every predictive model built on Maximo data depends on work orders having accurate failure class, problem code, cause code, and remedy code. Yet technicians have always skipped failure coding or guessed because picking the right value from a deep hierarchy is tedious, especially on a mobile device in the field. The result is a database where the failure data is inconsistent, incomplete, or misleading, and every downstream analysis built on that data is compromised.

Work Order Intelligence, introduced in Manage 9.0, attacks this problem directly. It does three concrete things. First, it speeds work order approval by surfacing what matters so planners can route work faster. Second, it improves data quality by nudging work order records toward completeness and consistency. Third, and most importantly, it recommends the problem code from the work order description. After a work order has a failure class specified, the AI reads the free-text description and suggests the most likely problem code, with a confidence score. The technician sees the top three recommendations, picks one, and the work order is updated with the chosen code.

The technical pattern is important to understand. The AI is not creating new failure codes. It is matching against the existing failure code hierarchy in your Maximo database. This means the AI is only as good as your failure code catalog. If you have a clean, well-curated failure code hierarchy, the recommendations will be useful. If your failure code catalog is a graveyard of one-off entries from 2014, the recommendations will be noisy. Before enabling Work Order Intelligence, audit your failure code hierarchy. Remove duplicates, consolidate ambiguous entries, and verify that each failure class has a reasonable set of problem codes that map to actual failure modes.

Here is what the configuration looks like in the ManageWorkspace custom resource:

apiVersion: mas.ibm.com/v1
kind: ManageWorkspace
metadata:
  name: prod-manage
spec:
  settings:
    ai:
      enabled: true
      features:
        - problemCodeRecommendations
        - duplicateDetection
        - similarWorkOrders
      model:
        inferenceModel: "watsonx-text-embeddings"
        apiKeySecret: "watsonx-api-key"

The AI broker is enabled in the ManageWorkspace custom resource, and the model is referenced by name. Inference happens at the moment the technician saves a long description, not at the moment the work order is created. The recommendations appear in a panel on the work order, and the technician can accept, reject, or override. The history of acceptance and rejection is stored on the work order and is fed back into the model, which means the recommendations improve over time as the model learns from your specific data patterns.

The second feature in the Work Order Intelligence family is duplicate detection. When a new work order is created, the system compares it against recent work orders on the same asset, the same location, or the same problem code. If a probable duplicate is found, the technician is shown the existing work order and asked to confirm. This is a meaningful reduction in duplicate work, especially in environments where field crews raise work orders independently and the visibility across crews is limited. In a utilities environment with multiple crews working in the same substation, duplicate detection can prevent two crews from independently reporting and addressing the same breaker malfunction.

A third capability is the similar work orders feature, added in 9.1. When a technician encounters a problem, they can view similar work orders for the current one. The AI finds work orders with similar descriptions, similar failure codes, or similar asset types. This reduces troubleshooting time and helps teams learn from past experiences. A technician working on a recurring pump issue can instantly see that the same pump model had the same problem three months ago at a different site, along with the resolution that was applied. This is particularly valuable for organizations with high technician turnover or distributed workforces across multiple sites.

A practical pattern for piloting Work Order Intelligence: start with a narrow scope. Pick a single crew, a single asset class, and a 60-day window. Measure two metrics: the acceptance rate (how often the technician picks one of the top three recommendations) and the override rate (how often the technician picks something else or types free text). Acceptance above 50 percent is a good signal that the feature is delivering value. Override above 30 percent is a signal that the failure code hierarchy needs work, not that the AI is broken. Document these metrics and review them with your reliability team before expanding to additional crews and asset classes.

The Maximo Assistant: Natural Language Queries Over Your Maximo Data

The Maximo Assistant was added in Manage 9.1 and expanded significantly in the 9.2 feature channel. It is a natively built generative AI chat experience delivered as part of the Maximo AI Service. Unlike Work Order Intelligence, which operates silently in the background, the Assistant is a conversational interface that users interact with directly.

In Manage 9.1, the Assistant's reach was meaningful but bounded. IBM documents default coverage for assets, work orders, and service requests. A user can type "What are my top 10 at-risk assets?" or "Show me all open work orders" and get a structured response. The Assistant understands common abbreviations like WO for work order and SR for service request. If it does not understand an abbreviation, the system administrator can complete prompt tuning to teach it.

In the 9.2 feature channel, the Assistant evolved into a context-aware, agentic teammate. It can retrieve data for any object, including custom objects, when the relevant object structures are configured for assistant access. IBM's documented defaults in that channel include work orders, assets, service requests, meters, preventive maintenance, job plans, work logs, assignments, and persons. The Assistant also gained the ability to search documentation, generate insights, and complete tasks on behalf of users through the MCP Server integration.

The configuration flow for the Assistant requires three steps. First, an administrator must set up an AI configuration through the AI Configuration application. This involves connecting to the AI Service, which requires a watsonx.ai API key. Second, the administrator defines which object structures are accessible to the Assistant. Third, the administrator can restrict what types of questions or requests the Assistant can respond to. For example, if the administrator restricted the Assistant's access to asset data, the Assistant cannot answer questions about assets.

Here is a representative configuration sequence:

1. Navigate to AI Configuration application
2. Create new AI configuration
3. Select assistant type: "Agentic Assistant" (9.2) or "Data Retrieval" (9.1)
4. Configure watsonx.ai connection:
   - API key: stored in Kubernetes secret
   - Model: gpt-oss-120b (9.2 default) or granite-3-2-8b-instruct (9.1)
5. Select enabled object structures:
   - WORKORDER, ASSET, SERVICEREQUEST, METER, PM, JOBPLAN
6. Define access restrictions (optional):
   - Restrict by site
   - Restrict by object
   - Restrict by query type
7. Save and verify connectivity
8. Test with sample prompts:
   - "Show me work orders due in the next 7 days"
   - "Which assets have the most open work orders?"

A key insight: the Maximo Assistant in 9.2 is not just a chatbot. It is an agentic capability that can orchestrate actions. It can create work orders, attach inspection results, and raise service requests through the MCP Server integration. This means the governance model for the Assistant needs to define not just what it can see, but what it can do. A practical recommendation: start with data-retrieval-only mode in 9.1, then upgrade to the agentic assistant in 9.2 only after users are comfortable with the feature and the governance model is in place.

Reliability Strategies: Where AI Meets Maintenance Engineering

Reliability Strategies was added as an included capability in MAS 8.11 and has been enhanced in 9.0 and 9.1. It is the machinery that connects how an asset fails to what you do about it, and then writes that decision into the job plans and PMs your team actually executes. In 9.1, AI was added to the workflow.

The Reliability Strategies application ships with a curated, pre-built reliability library spanning 800 asset types and 58,000 failure modes, suggesting over 5,000 preventive maintenance activities with step-by-step tasks. This library is the result of an IBM acquisition of a large body of RCM studies, and it means you can apply expert analysis rather than authoring it from scratch. For a reliability engineer starting a new FMEA on a centrifugal pump, the library provides the failure modes, effects, and recommended tasks without requiring weeks of expert workshops.

In Manage 9.0, the application provided study overviews, FMEA import and editing, bulk import of failure mode details, RPN-based action assignment, and custom strategies. The RPN (Risk Priority Number) is calculated from severity, occurrence, and detection scores, and the application lets you define RPN categories and action types through the Settings panel. Actions are then assigned based on the RPN threshold: high-RPN failure modes get immediate corrective actions, medium-RPN modes get scheduled PM tasks, and low-RPN modes are monitored.

In 9.1, AI was added to suggest boundary conditions and generate components and failure mechanisms. This is targeted, practical AI: it attacks the exact step that makes FMEAs so labor-intensive. Building out the component structure and failure mechanisms for a complex asset can take days of expert time. The AI does not replace the reliability engineer, but it accelerates the slowest part of the process from days to hours. The AI also recommends mitigation actions, closing the loop from analysis to action.

In 9.2, the Reliability Strategies gained IBM Power (ppc64le) architecture support, and the AI recommendations were enhanced to use the gpt-oss-120b model for improved accuracy. The model upgrade from granite-3-2-8b-instruct to gpt-oss-120b is a significant jump in reasoning capability, which matters for complex failure mode analysis where the AI needs to understand subtle relationships between operating conditions and failure mechanisms.

A critical point: Reliability Strategies is not FMEA with a fresh coat of paint. Its job is to make your maintenance program defensible. Every task should be traceable to a failure mode and its risk, not to habit. When an auditor asks why you perform a specific PM task at a specific interval, the answer should be "because failure mode X has an RPN of 320 and this task reduces it to 120," not "because we have always done it that way."

AI Service: The Technical Foundation

All AI features in Maximo Manage require the AI Service, which was introduced alongside Manage 9.1. The AI Service is a dedicated application within MAS that handles the connection to watsonx.ai, manages inference requests, and provides the infrastructure for AI features across Manage, Health, and other MAS applications.

Three deployment patterns are available for the AI Service. Full SaaS uses existing AppPoints with no new purchase required and no new infrastructure to manage. Hybrid AI SaaS requires SaaS AppPoints but no new infrastructure. Full Customer Managed includes a watsonx.ai license for use with AI Service only and is for organizations that need full control of their AI infrastructure.

In MAS 9.2, you can track AppPoint usage for AI Service in the MAS licensing dashboards. This is a significant addition because it gives administrators visibility into the cost of AI features and enables chargeback to departments if needed. The dashboard shows AppPoint consumption per AI feature, so you can see how much problem code recommendations cost versus the Assistant versus Reliability Strategies AI.

The AI Service supports several model templates, each mapped to specific features:

Feature Version Model Template Model

Problem code recommendations 9.0 classification watsonx text embeddings

Data-retrieval Assistant 9.1 data retrieval granite-3-2-8b-instruct

Agentic Assistant 9.2 nl2oslc, insightsgenerator gpt-oss-120b, Llama-4-Maverick

Similar work orders 9.1 similarity embedding_transformer_en_slate.125m

Reliability Strategies AI 9.1 fmea gpt-oss-120b

Condition and alert insights 9.2 insightsgenerator gpt-oss-120b

Each feature requires no training; the models are pre-trained and ready to use once the AI Service is configured. This is important because it means the time-to-value for AI features is measured in configuration hours, not in data science weeks.

A key administrative note: the AI Service requires a watsonx.ai API key. This is not a checkbox you flip during upgrade. It is an entitlement and connectivity exercise. You need the watsonx.ai service available to your environment, a valid API key configured in a Kubernetes secret, and the AI features enabled through AI Configuration. Verify connectivity before enabling features in production, and monitor the AI Service logs for inference failures, which typically indicate connectivity or entitlement issues rather than model problems.

Practical Implications

The AI features in Maximo Manage 9.x are not experimental. They are production capabilities that address real operational pain points. Work Order Intelligence fixes the oldest problem in Maximo: bad failure coding. The Assistant gives users a natural language interface that reduces training time and makes Maximo more accessible to users of all skill levels. Reliability Strategies with AI makes FMEA authoring faster and more consistent. But all of these features depend on data quality. Before enabling any AI feature, audit your failure code hierarchy, clean up your asset records, and verify that your work order data is consistent. The AI is only as good as the data it works with.

The Bottom Line

Maximo Manage 9.x has quietly become an AI-powered EAM platform. Work Order Intelligence, the Maximo Assistant, and AI-enhanced Reliability Strategies are not bolted on; they are embedded in the daily workflows that planners, technicians, and reliability engineers already use. The configuration effort is modest compared to the value, but the data quality prerequisite is real. Start with Work Order Intelligence on a single crew, expand to the Assistant for power users, and use Reliability Strategies with AI for your next FMEA study. By the time MAS 9.2 is in your production environment, you will have the data quality and user adoption patterns needed to take advantage of Condition Insight and the broader AI extensibility layer.