Intel Foundry

Two years ago, the bottleneck in industrial AI was the model. Today, it’s what the model can see.

That’s the central finding of a new IT@Intel case study, Proving the Value of Context in AI-Driven Industrial Operations, co-authored by Intel Foundry Facilities engineers and Formant. Together, we tested how operational context affects agentic AI alarm triage for critical facility waste systems.

We kept the model and questions the same. Then we added context in four cumulative layers. The improvement came largely from making the right knowledge available—not changing the model.

Read the full case study on Intel.com →

The problem: alarms move to the ROC, but the knowledge doesn’t

Intel Foundry Facilities is moving toward centralized Remote Operations Centers (ROCs), where engineers monitor multiple facilities from one place. Bringing alarms into a shared dashboard is one part of that transition. Knowing what each alarm means—and what to do next—is another.

A single industrial system can generate several hundred alarms a year. The knowledge needed to triage them may be spread across experienced operators, globally distributed Troubleshooting Steps (TSS) documents that do not always match local equipment, and incident notes stored in a separate application.

The study asked: Given a specific alarm, can an AI agent detect and classify it, retrieve the right context, and recommend a resolution with an auditable rationale?

The test bed was the Slurry Copper Waste (SCW) ion-exchange treatment system, one of 320 systems at the host facility. The team used 540 days of historical data covering 47 alarms and 380 TSS procedures. The system had read-only access and no connection to live controls.

Same model, same questions, four layers of context

The work proceeded in four cumulative phases. Each phase used the same model and asked the same questions about the same alarms. Only the available context changed.

1. Documentation. TSS procedures, SOPs, P&IDs and equipment manuals were converted into structured artifacts. The model could answer what a document said, but it sometimes selected the wrong procedure with confidence.

2. Relational structure. Adding asset hierarchy, upstream and downstream dependencies, alarm-to-procedure mappings, setpoints and alarm frequency helped the model choose the right procedure and build an operational briefing. It also became better at flagging genuine ambiguity instead of inventing connections.

3. Operational data. Time-series trends, redundant sensor comparisons, alarm history, comments and work orders helped the agent recognize chain-reaction alarms, investigate root causes, and distinguish an instrument fault from a real process condition. This phase also exposed a practical gap: global documentation does not always match what is physically installed at a site.

4. Operator feedback. Subject matter experts reviewed and corrected the outputs, and those corrections went back into the knowledge store. Recommendations gained confidence scores and routing suggestions, while known gaps were less likely to recur.

The lesson was not that every troubleshooting task should be automated. Analytical steps—such as retrieving data, comparing sensors and reviewing alarm history—were far more amenable to automation than physical field steps such as fuse checks, nozzle inspections and permits.

The results

By the end of phase four, the study measured the following results across the historical test set:

• Document retrieval time: approximately 15 minutes to 2 minutes per alarm—about 7.5× faster

• Time to resolution: 60% lower on average across 47 alarms; approximately 54% when weighted by alarm volume

• Step automation rate: 66% across 47 alarms and 380 TSS procedures

• Technician time: approximately 393 hours saved per year (about 46 minutes per alarm, weighted by alarm volume)

The study calculated 581 hours saved over 540 days of historical data. The annual estimate assumes approximately 509 alarms per year. These figures describe the study’s scope and assumptions, not a live autonomous deployment.

The most valuable context wasn’t in a document

A one-hour walkthrough with the system’s plan-of-record owner produced more usable context than the documentation set.

Experienced operators carry knowledge about local equipment and past decisions that rarely makes it into formal procedures. The team built a repeatable way to capture that knowledge through normal review: when an expert corrected the agent’s output, the correction became part of the knowledge pipeline.

That work has value whether or not a facility deploys agentic AI. AI makes the cost of uncaptured operational knowledge easier to see.

The system ran in advisor mode, parallel to operators and separate from live controls. Within that boundary, it produced:

• A structured investigation report for every alarm

• Trending reports to keep engineering current

• Triage classifications—false positive, routine maintenance or true investigation—to help technicians focus

• Institutional knowledge captured through operator feedback

The path to autonomy can’t skip steps

Industrial operations require evidence and trust before increasing an AI system’s authority. The case study describes a staged path: start in advisor mode; move to co-pilot mode, where operators accept, modify or reject suggestions in their workflow; then consider bounded autonomy for a narrow class of alarms, with operators handling exceptions and audit.

Early observations from a second facility suggest the methodology can transfer to another site, although each facility’s specific context must be rebuilt. In physical operations, context lives where the work happens.

Formant builds the data, context and control layer that lets AI reason about real equipment, facilities and operations within customer security boundaries.

Thank you to the Intel Foundry Facilities and Intel Foundry Automation teams, especially co-authors Nicolas Alvarez Garcia and Derrick Thoma, for a partnership grounded in real operations.

Read the full IT@Intel case study on Intel.com → · Talk to our team about your facility →