Industrial AI at scale: Why the foundation decides the outcome

Most industrial companies have proven AI can work at a single unit or plant. Capturing value across the enterprise depends on foundational choices that most organizations make too late.

Over the past years, nearly every large manufacturer, refiner and process operator has run AI pilots. Many delivered measurable results: fewer unplanned shutdowns on a critical compressor train, tighter quality control on a single line, faster root-cause analysis for a recurring failure. Few have converted those pilots into enterprise-wide impact. The gap between proof of concept and value at scale is now the defining challenge for industrial leaders.

The cost of that gap is rising. Fluke’s 2025 survey of more than 600 manufacturing decision-makers put the average cost of unplanned downtime at $1.7 million per hour. At the same time, Deloitte and The Manufacturing Institute project that 2.4 million manufacturing roles may go unfilled by 2028. The engineers retiring now carry much of the undocumented knowledge that makes plant data usable: which tag belongs to which pump, which procedure applies, what happened the last time this alarm fired.

In our work with operators across refining, chemicals, food and beverage, and building materials, a consistent pattern separates the organizations that scale from those that stall. The ones that scale did not start with better models. They invested early in three foundations: a contextualized operational data layer, intelligence delivered inside existing workflows, and an architecture that respects data sovereignty. Organizations that defer these choices tend to rebuild each use case from scratch, site by site.

Each foundation has specific requirements, and leading operators are already putting them in place. The decisions made over the next months will determine which organizations scale and which remain in pilot mode.

1. Value at scale depends on context, not model sophistication

Industrial data is fragmented in ways that compound. The same physical pump can be P-104 in the historian, AB-0104-A in the ERP equipment master, “Charge Pump A” in the operating procedure, and an unlabeled symbol on a drawing. Historian data updates in seconds, work orders in days, and drawings whenever someone remembers. Each function owns its own data and its own access controls.

Centralizing this data in a lake solves the access problem but not the meaning problem. The pump still has four identifiers, and an engineer still reconciles them by hand for every investigation. That reconciliation is the most valuable work in the cycle, and it is typically thrown away once the question is answered.

When AI is applied before context is resolved, two things happen. Outputs are less reliable, because the model reasons over records that have not been joined. And costs scale with every new use case, because each one repeats the same integration work.

Implication for leaders: Treat contextualization as shared infrastructure, funded and governed at the enterprise level, not as a cost buried inside individual use cases.

2. Adoption depends on where intelligence is delivered

Even well-designed industrial AI often goes unused. The most common reason is not distrust of the model but friction in the workflow. When a recommendation lives in a separate dashboard, the reliability engineer or maintenance planner has to leave the tools they use all day to find it, and usage declines after the initial rollout.

The organizations that sustain adoption deliver intelligence at the point of decision. IRIS Foundry is available inside Microsoft Teams and Microsoft 365 Copilot through the Model Context Protocol, so plant teams can reach asset health, process data and agent recommendations in the tools they already use. IRIS Forge lets engineers build their own dashboards, copilots and applications on the contextualized data without a development cycle. IRIS Flows orchestrates multistep agent workflows, with human approval at decisions that carry operational or safety consequences.

Implication for leaders: Measure industrial AI by the number of operational decisions it changes, not the number of applications deployed.

3. Deployment architecture determines which options are viable

For operators in energy, refining and critical infrastructure, data residency is a regulatory and security requirement, not a preference. Approaches that depend on sending operational data to external AI services are often excluded before evaluation begins.

IRIS Foundry runs the same containerized architecture as a managed service, in the customer’s own cloud tenancy, or fully on premises, including air-gapped OT networks. The platform is model-agnostic. Where required, external model APIs are removed and replaced by a self-hosted open-weight model, so operational data and queries never leave the customer’s boundary. The platform holds no agents in the control zone and performs no control writes, which keeps it outside the safety-instrumented path regardless of how much autonomy is configured above it.

Implication for leaders: Define sovereignty, security and data-residency requirements before selecting use cases or vendors. They narrow the field and prevent costly re-architecture later.

Cortex 2.0: From a map of the plant to a model that learns and acts

Cortex is the knowledge graph at the center of IRIS Foundry. Every asset, sensor, process, event, work order, BOM line, defect, inspection, document, drawing, operator and KPI is a node. Every relationship between them is a labeled edge that can be traversed: an asset feeds an exchanger, is governed by a procedure, exhibits a failure mode, impacts a throughput KPI. Cortex 2.0 extends that foundation in five ways that matter to operations leaders.

1. Automated entity matching that improves with each pass. Matching the same physical asset across systems is the step every cross-system answer depends on. Cortex runs it as a layered pipeline: deterministic and rule-based matching clears the bulk of the estate, fuzzy and structural matching handles differences in convention and padding, semantic matching aligns tags by meaning, graph-based linking rejects candidates that contradict known structure, and anomaly detection flags bad mappings before they spread. Low-confidence matches go to human review, and every confirmation or correction trains the pipeline. The reconciliation engineers used to repeat for every investigation becomes a permanent asset.

2. Drawings and images converted into queryable relationships. AI-assisted P&ID extraction recovers assets, sensors and connections from drawings and follows lines across sheet boundaries, with each proposed connection verified before it enters the graph. Selecting a symbol on the resulting interactive diagram returns live tag values, health scores, open work orders and documents. Vision AI writes inspection, quality and safety detections into the graph against the asset, location and work order, rather than into a separate image archive, and monitors its own models for drift.

3. One graph, read from every perspective. Cortex keeps raw source data immutable for lineage and audit, and maps it onto a standards-aligned Unified Data Model based on ISA-95 and ISA-88, and OSDU in energy deployments. Because applications consume the unified model rather than source systems, a historian migration or ERP upgrade changes one set of mappings rather than every application. The same entities support an equipment view, a process-flow view and a KPI view without duplication. When a KPI degrades, impact analysis traverses its dependencies to produce a ranked set of contributing factors and the critical paths between cause and effect.

4. Relationships proposed by AI, curated by people, and grounded before generation. AI proposes typed relationships at scale as data arrives; engineers confirm, correct or re-type them, and can define custom relationship types. Every relationship is extracted from a source, inferred and then confirmed, or asserted by a user. AI is not used to invent relationships at query time. The domain-trained Industrial LLM reasons over this graph rather than its training data, so answers refer to this asset and this procedure revision, respect the user’s permissions, and link back to the tag, document and record they came from. When an engineer asks a question in plain language, an orchestrator agent and a graph agent compose an explicit execution plan, and that plan is shown next to the answer.

5. A kinetic layer that turns insight into governed action. Cortex now models what can be done to the operation as well as what it is. Creating a work order, updating a health score, recommending a setpoint, initiating an inspection or escalating to a supervisor are defined action types with parameters, validation rules and permissions. Actions write back through the systems of record, which remain authoritative. Each action is itself recorded in the graph, including what was done, by whom, on what evidence and with what outcome. The next engineer facing the same alarm can see what was done the last three times, and whether it worked. IRIS Flows applies graduated autonomy to these actions, from assistive to augmented to autonomous, as the operating record supports it.

Together, these capabilities change the economics of the program. A point integration is at its best on the day it ships and degrades as the plant changes. Cortex starts at its weakest and improves with use, because every query, confirmation and corrected match makes the next answer more reliable.

The economics: The first site pays for the model, later sites inherit it

Contextualization is an infrastructure investment that pays through the use cases built on it. Two implications follow.

First, the cost curve is front-loaded. Building the graph for an initial scope takes weeks; subsequent sites take days, because they inherit the namespace, data model and a pre-built asset library with failure-mode patterns covering more than 85 percent of common industrial assets. An organization that evaluates the program on the cost and duration of its first pilot will systematically underestimate the return, because it is measuring the least efficient instance it will ever build.

Second, use-case metrics understate the value. Customers report 30 to 50 percent reductions in unplanned downtime and three to six times ROI in the first year from reliability-led programs. Those figures capture the return on the first applications, not the lower cost of every one that follows.

Getting started: Four decisions for the next months

1. Build the contextualization layer first, on a bounded scope. Choose a unit or asset class where root-cause analysis is slow and expensive, agree baselines before work starts, and build the foundation properly rather than shortcutting to the application.
2. Sequence use cases for evidence, then scale. Root-cause analysis and predictive maintenance produce evidence quickly and expose modeling gaps early. Workforce continuity and cross-site model transfer deliver the largest structural returns but pay back over a longer horizon.
3. Embed delivery in existing tools and build internal capability. Require that new capabilities reach users in the systems they already use, and develop internal champions who can build without the vendor. Programs that skip this step tend to plateau at the use cases the vendor delivered.
4. Set governance boundaries before scaling agents. Agree on data residency, model hosting, action permissions and approval thresholds up front, and expand agent autonomy only as the operating record supports it.

The next phase of industrial AI will not be won by the organizations running the most pilots. It will be won by those that build a foundation once and keep scaling on it. The technology to do this is available today. What separates the leaders is the order in which they make these decisions.

Ready to make industrial AI an enterprise standard? Talk with the SymphonyAI Industrial team about building the foundation once and scaling it across every plant.

Further reading

Unlocking rapid refinery insights with AI-powered applications built on IRIS Foundry | Empowering plant management with digital twins, 2D/3D visualization, and AI-driven operations