Root cause analysis
Root cause analysis, done by the agent in your data
A real root cause analysis in a factory is not a clever model. It is the patience to walk through alarms, telemetry, work orders, and shift notes at the same time, find the thread that connects them, and explain it in a way a person can act on. Digel does that walk for you, grounded in your plant's actual data, with the chart attached and the evidence cited.
Why root cause analysis is hard in manufacturing
The cause of a plant incident is almost never in one system. It is a vibration trend in the historian, a skipped preventive maintenance job in the CMMS, a feedstock change in the ERP, and a sticky note on the HMI that said "the left pump runs hot, use the right one". A person who is good at RCA spends most of their time assembling that picture, not analyzing it.
Every tool in your plant makes that assembly harder. The CMMS is a silo. The historian is a silo. The MES is a silo. The documentation is on SharePoint. The knowledge is in the head of the senior operator who already left for the weekend. Running a real RCA means a senior person opening four tabs and paging the on-call engineer. That does not scale, and it is why most incidents get closed with a guess and a shrug.
How the agent does root cause analysis
Everything in Digel lives in the same industrial context graph: sensor history, equipment relationships, work orders, procedures, shift reports, and operator notes. When the agent investigates, it walks the graph outward from the symptom. It pulls live telemetry, fetches the charts that matter, checks neighboring equipment, reads the related work orders, and cites the documents it used. For deeper cases it delegates to a research subagent that assembles a multi-source picture before it answers.
You get back a plain-language explanation with the data attached. Not a score. Not a probability. A paragraph a maintenance lead can read and act on, or argue with.
A triage item, verbatim
Pump 4 vibration is drifting. RMS has been climbing for 36 hours and is now 18% above its normal band. The same pattern preceded the bearing failure on Pump 2 last October. There is an open work order on Pump 4 from three weeks ago that was never closed out. Want to open a maintenance issue?
What AI troubleshooting looks like in production
Six concrete moves the agent makes during an investigation, in use today at our pilot customers:
Anomaly detection on live signals
The agent watches every connected tag continuously and flags drift that looks like the pattern preceding a past failure on similar equipment.
Walk from symptom to cause
Traverse the context graph from the affected asset outward: neighbors, upstream and downstream equipment, material flow path, related work orders.
Compare against history
Find the time this asset (or a similar one) behaved the same way before. Pull the work orders from that incident. Show what was done.
Cross-domain reasoning
Pull the relevant feedstock change from ERP, the skipped PM from the CMMS, the shift report note, and the historian trace into one narrative.
Arguing back in chat
Disagree with a finding and the agent re-investigates against live data. One customer caught a bird-counter bug that way.
Conversational dashboards
Ask for a chart comparing two tags over the same window and get it inline, as part of the investigation, not as a separate dashboard task.
The agent proposes. Your team decides.
A lot of AI-in-the-factory pitches end with a big red automate button and a promise. That is not how plants work. Operators have hands on the equipment and carry the consequences of decisions. They should make them.
Every investigation comes with a finished write-up and a one-click next step: open a maintenance issue, reassign it, dismiss it, or argue with it in chat. Dismissed findings teach the agent what your team cares about. Arguments teach it what it missed. Closed items become history the next investigation can draw on.
For the broader maintenance surface, see AI maintenance management. For the overarching picture, see AI for manufacturing.