Representative End User Clients
Representative Automation Clients
Representative Software Clients
Cognite, a Gold Sponsor of ARC Advisory Group’s 24th Annual ARC Industry Forum in Bengaluru, was represented by Ravi N. Mishra, Sr. Director – Sales (India), and Pawan Misra, Principal Designer for Data Integrations, Cognite India. In their presentation, Taming the Wild State of Industrial AI Starts with Great Data, they explored how contextualized industrial data can provide the foundation needed to move from basic AI assistants toward agentic workflows that support real operational tasks.
Industrial companies have no shortage of AI models, copilots, and agents to experiment with. The harder problem is giving those systems the right industrial context. Operational information is typically distributed across historians, SCADA and control systems, maintenance applications, engineering documents, enterprise platforms, 3D models, and other repositories.
Cognite’s central argument is that AI model capability alone is not enough. Industrial AI depends on bringing these different forms of information together in a way that allows AI systems to understand the relationships between assets, processes, events, documents, and operational history.
The presentation can be viewed on YouTube or here:
Context Is the Bottleneck for Industrial AI
Cognite describes its own AI journey as progressing from general-purpose chat agents toward increasingly specialized and integrated workflows.
An early phase used a single general-purpose chat agent, but the approach lacked the domain context and guardrails required for industrial applications. This progressed to task-specific agents designed for defined use cases, followed by deep research and ambient agents capable of handling more complex analytical tasks.
The next step is agentic workflows that combine agents with generated applications and retain humans in the loop. Examples include root cause analysis, work package creation, operations summarization, quality investigation, and batch analytics.
However, building the agent itself is only part of the challenge.
Cognite identifies context as the bottleneck. Industrial data remains multimodal and siloed, while document-based retrieval-augmented generation and vector search may struggle to find the specific information required for complex industrial questions. The challenge is not simply locating data, but understanding how different pieces of information relate to one another.
Cognite identifies grounding and industrial context, rather than AI model capability alone, as a key bottleneck when operational data remains distributed across multiple systems and formats
Building an Industrial Knowledge Graph
Industrial environments generate many different types of information.
An individual asset may be represented in an equipment hierarchy, produce real-time time-series measurements, appear in engineering drawings and maintenance records, have associated work orders and inspection histories, and be represented in 3D models or images.
These sources frequently use different naming conventions and schemas.
Cognite Data Fusion is designed to connect these sources and contextualize them within an Industrial Knowledge Graph. The presentation showed information spanning assets, time series, events, engineering files, maintenance work orders, 2D and 3D information, and visual data being linked into a common industrial context.
Cognite also emphasizes governance around that data foundation. Governed access controls determine who can access information, while an open, stable API allows organizations to use their own applications, agents, and AI models against the contextualized data.
According to Cognite, its industrial data foundation can connect more than 100 IT, OT, and engineering technology sources, with the knowledge graph providing the layer that ingests, contextualizes, and governs the information.
Cognite’s Industrial Knowledge Graph connects operational, engineering, maintenance, and visual information to provide applications and AI agents with a common industrial context
From AI Assistants to Agentic Workflows
Once that contextual foundation exists, AI can begin supporting more specific industrial workflows.
Cognite identifies root cause analysis, work package generation, and shift handover as three practical starting points. The company says its root cause analysis approach can reduce time-to-cause from hours to minutes, while its shift-handover application can reduce meeting duration by 25 percent or more.
The broader goal is not simply to make existing workflows faster.
Cognite argues that agentic systems can begin changing how work is organized. Rather than an operator repeatedly searching multiple systems for information, an agent can monitor assets, retrieve relevant context, assemble information, and surface the cases that require human attention.
This can also extend to application creation. Cognite Flows is designed to allow domain experts to create workflows, visualizations, and applications using AI, reducing their dependence on separate development teams for every new operational requirement.
The human remains part of the process, but the distribution of work can change as agents take on more of the information gathering and orchestration.
Root Cause Analysis in Practice
Pawan Mishra demonstrated this approach through an AI-assisted root cause analysis workflow.
The investigation began with a specific equipment problem and used multiple agents to retrieve and analyze relevant information.
The workflow brought together equipment information, failure modes, time-series data, engineering drawings, maintenance history, work orders, notifications, shift reports, and other contextual data from the industrial data foundation. The agent could then systematically evaluate potential causes, document the supporting evidence, and generate an RCA report.
Importantly, the workflow was designed to expose the evidence used in reaching a conclusion. This allows engineers to review the information and determine whether the agent’s analysis is appropriate rather than treating the AI output as an unexplained answer.
The workflow can also be adapted to an organization’s existing RCA methodology, including established investigation and reporting processes.
According to Cognite, the RCA demonstrated during the presentation could run in approximately 10 minutes. The company said the customer-specific workflow itself was built in approximately two and a half days once the underlying knowledge graph was already available.
That qualification is important: the speed of building the AI workflow depends heavily on having the contextualized industrial data foundation already in place.
Building the Foundation for Industrial AI at Scale
Cognite sees the same data foundation supporting increasingly sophisticated industrial AI.
The company is currently collaborating on the training and testing of a time-series foundation model using a large multimodal and contextualized industrial dataset. The presentation cited 93 trillion time-series data points and 4.4 billion data-modeling instances as part of that foundation.
The longer-term objective is to enable AI models that can work across different industrial assets while remaining grounded in operational context.
For industrial organizations, however, the immediate lesson is more fundamental. The effectiveness of an AI agent depends on the quality of the information available to it and the relationships established between different sources of industrial data.
Building additional copilots or choosing more capable models will not by itself solve fragmented asset hierarchies, inconsistent naming, disconnected engineering information, or inaccessible maintenance history.
Industrial AI becomes more useful when those sources are connected, contextualized, governed, and made accessible to the systems that need them.
As manufacturers and process industries move from isolated AI use cases toward agentic workflows, the competitive advantage may therefore come less from the model itself and more from the industrial knowledge foundation underneath it.