Beyond Condition Insight: The 2026 Maximo AI Stack — watsonx Orchestrate Agent Control Plane, AI Service Templates, and the Practitioner Build Path

For most of 2024 and 2025, the "AI in Maximo" conversation was about the AI broker, the named templates, the Maximo Assistant, and the early agentic experiments. The May 5, 2026 Think conference and the May 27, 2026 Maximo AI Service 9.1.15 release changed the conversation. The pieces are now coherent. The build path is now clear. This article walks through the 2026 Maximo AI stack as it actually exists today, and the practitioner's build path for production agentic AI on Maximo.

The four layers of the 2026 Maximo AI stack

The 2026 Maximo AI architecture is best understood as four layers, each with a distinct purpose and a distinct deployment model. Building anything serious requires understanding all four.

Layer 1: The Maximo AI Service (the inference layer)

The Maximo AI Service is the inference layer in the Maximo namespace. As of August 1, 2025, it replaced the original AI broker. The current line is AI Service 9.1.x, with 9.1.15 shipped on May 27, 2026. The 9.1.15 release included a security fix and a migration of the RSL (Required Service License) secret creation from Ansible automation to the operator — a quiet but important reliability improvement.

The AI Service exposes a small set of named templates:

• Recommended Problem Code — given a work order description and the asset, return the most likely problem code from the failure hierarchy.
• Recommended Field Values — given a record creation context, return the most likely values for the unspecified fields.
• Similar Record Lookup — given a work order, find the most similar historical work orders and surface them for the planner.
• AI-Assisted FMEA Generation — given an asset classification, return a draft FMEA with failure modes, effects, and recommended tasks.
• Maximo Assistant Configuration — the template that backs the conversational Maximo Assistant interface.

Each template is a pre-configured inference pipeline: input contract, prompt, model selection, output contract. You enable the template, point it at the data it needs, and the AI Service does the inference. You do not have to write a prompt or select a model.

The AI Service 9.1.x includes a limited-use license to watsonx.ai and incurs an additional AppPoint cost. This is the most common point of confusion in licensing conversations. AI Service is not "included with Manage" in MAS 9.1. It is an add-on. The licensing model is per-tenant for the inference, with consumption limits on the underlying watsonx.ai calls.

Layer 2: The Maximo AI Assistant (the conversational surface)

The Maximo AI Assistant is the conversational surface in the Manage UI. It is a chat-style interface backed by NL2OSLC, the natural-language-to-OSLC translation layer. The user types "show me all open work orders on the north feeder," and the Assistant translates the request into an OSLC query, runs the query against Manage, and returns the results.

The 2026 release added tighter integration with the AI Service templates. The Assistant can now invoke the Recommended Problem Code, Recommended Field Values, and Similar Record Lookup templates directly in the conversation. The user can ask "what was the last problem on Pump 101?" and the Assistant returns the historical record, the related problem codes, and a recommended problem code for the current WO.

The Assistant also supports a skill model. A skill is a configured capability that the Assistant can invoke — a query, a template, an automation script, a custom action. Skills are configured in the AI Service and exposed to the Assistant. The 2026 release added a skill library that ships with the AI Service, including common Maximo skills ("find open work orders," "show asset hierarchy," "create a service request," "log a failure").

The Assistant is the right surface for the user who is doing a structured task and wants a faster way to get to the data. The right surface for the user who is doing an unstructured task and wants the AI to take action is layer 3.

Layer 3: Agentic workflows (the orchestration layer)

The agentic workflow layer is what was "experimental" in 2024, "early adopter" in 2025, and "production-ready" in 2026. The 2026 agentic story has three pillars:

Pillar 1: Maximo Condition Insight. Condition Insight is the agentic capability that ships inside the Maximo namespace. It reads the asset's data (sensors, meters, work history, failure history, inspection records), reasons about the asset's condition, and produces a structured output (a recommended work order, a recommended PM, a recommended escalation). Condition Insight is the production-ready pattern for the "AI decides what to do next" use case.

Pillar 2: Maximo IT + watsonx Orchestrate. Maximo IT is the IT asset management application in the Maximo namespace, and the 2026 release added an integration with watsonx Orchestrate that lets Orchestrate agents invoke Maximo IT as a tool. The use cases include IT service management (a chatbot that creates an IT ticket), IT asset lifecycle (an agent that tracks a laptop's lifecycle from procurement to retirement), and IT inventory (an agent that answers "how many Dell laptops do we have under warranty").

Pillar 3: External agent frameworks. The 2026 Orchestrate release added third-party agent registration — agents built on LangChain, AutoGen, CrewAI, or any other framework can be registered in Orchestrate and managed alongside the IBM-native agents. The Orchestrate control plane applies the same governance, audit, and observability to all of them, regardless of origin.

Layer 4: The watsonx Orchestrate Agent Control Plane (the governance layer)

The headline announcement from Think 2026 (May 5, 2026) was the next-generation watsonx Orchestrate as an Agent Control Plane — now in private preview, with general availability expected later in 2026. The framing is "the control plane for hundreds of AI agents built by different teams, on different frameworks, with consistent policy enforcement and auditability."

The four capabilities that matter most for a Maximo practitioner:

1. Cross-platform agent deployment. Agents built on any framework (LangChain, AutoGen, CrewAI, IBM ADK, or custom) can be registered in Orchestrate and managed as a single population.
2. Consistent policy enforcement. Policies (data access rules, model selection rules, escalation rules, approval rules) are defined once in Orchestrate and applied uniformly across all registered agents, regardless of which framework built them.
3. Full traceability and audit logging. Every agent decision, every tool invocation, every data access is logged in Orchestrate's audit trail. The audit trail is the source of truth for "what did the AI do, and why."
4. Integration with watsonx.data for governed data access. The agent's data access is mediated through watsonx.data, which enforces the data governance policies (row-level security, column-level masking, PII handling) at the data layer, not at the agent layer.

The Agent Control Plane is the layer that makes agentic AI enterprise-safe. Without it, you have a collection of agents built by different teams, each with their own governance (or no governance), each with their own audit (or no audit), and each with their own data access (or no data access). With it, you have a managed agent population that the security, compliance, and risk teams can sign off on.

The 2026 watsonx Orchestrate feature set

The June 2026 Orchestrate release added capabilities that the Maximo practitioner should know about:

• Python toolkits in workflows. Python logic can now be bundled into a reusable toolkit and invoked from an agentic workflow, alongside the MCP toolkits that were already supported. The use case for Maximo: a toolkit that wraps the Maximo REST API, with pre-built functions for common operations (create WO, find assets, log a failure), and the agent invokes the toolkit functions as needed.
• Custom embedded chat fonts. A small but practical improvement for the customer that is embedding the Maximo Assistant or an Orchestrate chatbot in their own application — the chat UI can be styled to match the brand.
• New workflow controls. Loops (iterate over data, poll for status), retries with backoff, optional inputs, and asynchronous execution. The use case for Maximo: a workflow that polls a work order's status every 5 minutes for up to an hour, with retries on transient failures.
• Prebuilt PDF agents. Extract, Insert, Generate, Update — the four operations on PDF documents, out of the box. The use case for Maximo: a workflow that extracts data from a P&ID PDF, generates a draft asset record, and inserts it into Maximo.
• Thumbs-up/thumbs-down feedback in embedded chat. Stored on Orchestrate servers, accessible through the Messages API. The use case for Maximo: a feedback signal that the team can use to tune the Maximo Assistant's responses.
• Dynamic knowledge base schemas (ADK 2.10.0). Knowledge bases can have dynamic input and output schemas, configurable for richer external knowledge integrations. The use case for Maximo: a knowledge base that ingests Maximo's BIRT reports, asset manuals, and PM procedures, with a dynamic schema that adapts to the document type.

The 2026 IBM portfolio context

The Maximo AI stack is part of a broader IBM portfolio that Think 2026 brought into focus. The Maximo practitioner should be aware of four pieces:

1. IBM Confluent. IBM completed the Confluent acquisition, and the Confluent streaming data platform is now branded as IBM Confluent. It is integrated into the watsonx fabric as a managed Kafka and Flink service. The use case for Maximo: a real-time event stream from sensors into Maximo Monitor, with semantic context layered on by watsonx.data, flowing into agents that act on the events in real time.
2. IBM Sovereign Core (GA). Sovereign Core is now generally available. It is built on Red Hat OpenShift and Red Hat AI, and it embeds governance, compliance, and AI execution controls into the infrastructure runtime. The use case for Maximo: a regulated-industry Maximo deployment (utility, government, healthcare) where the data and the AI must stay within a specific jurisdiction and under a specific compliance regime.
3. IBM Bob (GA across Pro, Pro+, Ultra, Enterprise SaaS). Bob is the AI-powered full-SDLC tool — code generation, code review, testing, deployment. The use case for Maximo: a developer using Bob to write the automation scripts, BIRT reports, and integration components for a Maximo project. Bob is not Maximo-specific, but it is the IBM-recommended tool for the development work that surrounds a Maximo project.
4. Docling for IBM watsonx. Docling is the document intelligence platform that converts unstructured documents (PDFs, Word, images) into structured AI-ready formats (Markdown, JSON, HTML). The use case for Maximo: an asset-onboarding workflow that ingests P&IDs, nameplate photos, and equipment lists, and produces structured Maximo records.

The practitioner build path

The right way to ship production agentic AI on Maximo in 2026 is staged. The path that has worked for the early adopters is:

Stage 1: AI Service templates (weeks 1–4)

Enable the AI Service templates that match your data. The Recommended Problem Code template is the one with the highest ROI for most customers — it pays for itself in the first month through faster WO classification and better failure history data. The Similar Record Lookup template is the second-highest ROI for the planner team. Configure the templates, train them on your data, and measure the adoption.

Stage 2: Maximo Assistant with skills (weeks 5–8)

Roll out the Maximo Assistant to a pilot user group. Configure the skills (the shipped skill library is the starting point) and add custom skills for the workflows that are unique to your operation. Measure the user adoption and the time-saved-per-task metric.

Stage 3: Condition Insight for the critical assets (weeks 9–16)

Roll out Maximo Condition Insight for the critical asset population. The 10–20% of assets that drive 70–80% of the failure cost or the operational risk. Train the models on your data, validate the recommendations against the historical record, and tune. The KPI is "fraction of Condition Insight recommendations accepted by the planner team."

Stage 4: Orchestrate agent for the high-value workflow (weeks 17–24)

Identify a single high-value workflow that is currently a multi-system, multi-step manual process. Build an Orchestrate agent that automates the workflow, with Maximo as a tool, the AI Service templates as supporting inference, and the agent's decisions audited through the Orchestrate control plane. The use cases that have worked: an "asset onboarding agent" that reads P&IDs and creates Maximo records, a "PM optimization agent" that reviews PM compliance and recommends adjustments, a "service request triage agent" that classifies and routes SRs.

Stage 5: Agent Control Plane governance (weeks 25+)

When you have multiple agents in production, register them all in the Orchestrate Agent Control Plane. Define the policies, the audit requirements, and the data access rules. The control plane is what makes the agent population enterprise-safe.

The 2026 Maximo AI stack is not a single feature. It is a layered architecture with four layers, each with a distinct purpose, a distinct deployment model, and a distinct cost. The right way to use it is to understand the four layers, build on them in the right order, and use the Agent Control Plane as the governance layer when the agent population grows beyond a single use case.