Since the rise of generative AI, many manufacturers have launched a growing number of AI pilot projects. But according to Cognite, a Norwegian company born at the intersection of industry and digital technology, only a minority of these projects reach a profitable industrial-scale deployment. At the Tech For Industry Show, Emmanuelle Camus, SVP Strategic Accounts (Cognite AS) detailed the reasons behind this bottleneck and the method Cognite proposes to overcome it.

Large language models (LLMs) work by finding relationships between words, between text. But industrial data is rarely text: it takes the form of time series, 3D models, images, P&ID diagrams (piping and instrumentation diagrams) and loosely structured technical documents. A sensor reading, for example, means nothing to an AI if it isn't linked to a specific piece of equipment, process or documentation. According to Cognite, this lack of structure explains much of the hallucination risk observed when generative AI is applied directly to raw industrial data.

According to Emmanuelle Camus, most industrial AI projects remain stuck at the pilot stage for three main reasons.

Industrial data is complex, siloed across systems, and rarely linked together — a fundamental obstacle for any AI trying to derive reliable results from it.

The use cases chosen for early pilots are often selected for their demonstrative effect on senior leadership, rather than for their ability to be replicated across other sites — where IT systems often differ from one site to the next.

Getting operations, maintenance, IT and senior leadership to collaborate on the same project takes time. Two scenarios come up frequently: field-driven initiatives that later run into IT's security and infrastructure requirements when it's time to scale, or top-down initiatives that instead need to build field teams' trust in the data and the results produced.

To overcome these obstacles, Cognite recommends first building a knowledge graph — a semantic model linking the different data sources around each piece of equipment. The company says it remains agnostic about which AI model its clients use: it positions itself around structuring and exposing data, leaving clients free to choose the model (LLM or otherwise) best suited to their use case.

To illustrate contextualization, Emmanuelle Camus uses the example of booking a restaurant: the choice doesn't rest on a single piece of information, but on a set of linked data points (location, budget, reviews, transport options, travel time). Applied to industry, a pump ends up linked to engineering data, 3D models, drawings, usage data, time series, process data, repair history and work orders. Cognite says it connects to around a hundred different data sources (engineering and IT) to give this context to industrial equipment.

Cognite stresses an important distinction: a semantic model connected only to metadata isn't enough. It's also necessary to extract the information contained within the documents themselves — for example, finding that a pump referenced in a P&ID diagram is also mentioned, under a different reference, in a work order. Cognite's stated positioning is to carry out this linking work at the OT, engineering and IT levels alike — OT-IT linking being, according to Emmanuelle Camus, relatively common in the market, unlike engineering data integration, which other solutions address far less often.

Once this structured data foundation is in place, Cognite recommends focusing on high-value-added use cases. One example cited: a root cause analysis can spend 20 to 30% of its total time on data retrieval alone. With a structured, contextualized data model, this retrieval time drops sharply, freeing up capacity to develop agents capable of automating part of the analysis — these results can then feed workflows that trigger actions directly.

Cognite advises against a "big bang" approach and recommends starting with an operational pilot lasting 4 to 6 weeks, contextualizing engineering and IT data around a specific use case, deployed on a given asset. Once operational value has been demonstrated on this pilot, the deployment gradually extends to other sites, alongside the development of further use cases on a value-creation roadmap. Cognite says it works with manufacturers across various sectors: oil (its original sector), chemicals, paper, pharma and specialty machinery, with impacts generally centered on production optimization, reducing major maintenance operations, and improving productivity.

Cognite details the case of TotalEnergies, a French client that chose to entrust management of the industrial data for its upstream (exploration-production) assets to the Cognite platform. The initial problem: the Total Digital Factory had taken around 8 years to roll out part of its digital use cases across about ten sites — a pace judged too slow given the acceleration of AI.

TotalEnergies therefore chose, according to Cognite, a radical paradigm shift around data access, with a program launched around a year ago aiming to deploy digital use cases at scale across 36 assets, starting with two sites in Africa. The objective covers the entire value chain: industrial performance, production and maintenance. Unlike most Cognite clients, who typically start a pilot on two assets before considering a broader rollout, TotalEnergies moved directly into a scaled deployment approach, with an ambitious trajectory: 9 FPSOs (floating oil production units) deployed by the end of 2026, scaling up toward covering all 36 assets by 2028. This program is presented by Cognite as the most ambitious data-platform initiative TotalEnergies has ever undertaken.

Cognite structures its platform across three levels: a foundation, Cognite Data Fusion, which connects industrial data sources into a knowledge graph (making it possible to deploy a use case at a new site without having to reconnect every data source individually); an agentic layer, Atlas AI; and a workflow layer, corresponding to applications built by clients (often in low-code) to leverage the agents and improve their operations.

The central message of this session is methodological: before trying to automate industrial processes with AI, it's necessary to structure and contextualize the underlying data, linking engineering, IT and OT information. It's this foundation of trust that then makes it possible to develop high-value use cases and, as with TotalEnergies, consider an ambitious company-wide scale deployment.

Watch the full session, along with the other Tech For Industry Show replays, on our dedicated page.