Key takeaways ▼
By Gleb Tsipursky
SEMICON West 2026’s smart-manufacturing agenda connected digital twins, agentic AI, physical AI, and increasingly autonomous factory operations. The final day included a session on automation and equipment for autonomous facilities. At its October CONNECT conference, PDF Solutions planned demonstrations of Exensio Aurora, an analytics architecture designed to support agentic AI across semiconductor manufacturing operations and the supply chain.
These announcements indicate the direction of investment, although product specifications and conference demonstrations do not establish performance inside a production fab. The immediate management question is practical: When may an AI-generated recommendation become an authorized action, and who is accountable for that decision?
An AI system might identify a correlation between an equipment signal and a change in yield. An engineer must still determine whether the signal reflects an instrument issue, an unusual material lot, a recent maintenance event, or a genuine process shift. In exploratory analysis, the pattern may justify further investigation. On a live production line, it cannot automatically justify a process-setting change.
Fab leaders should therefore design decision-escalation rules into workflows before granting an agent operational authority.
Start with precise permission boundaries. Separate agents allowed to retrieve records and draft explanations from those permitted to initiate holds, propose parameter changes, or alter tool instructions. An agent might recommend an action while an authorized engineer or operations leader determines whether to approve it. A general promise to keep humans in the loop offers little protection unless teams specify the decisions, reviewers, required evidence, and permissible autonomy.
Make every handoff traceable. A process engineer receiving an AI recommendation should see the source data, affected tools or lots, known uncertainties, and intended effect. Record who reviewed the evidence, which competing explanations they considered, and whether the suggestion was accepted or rejected. A fluent explanation should never replace the traceable record beneath it.
Create the exception path during testing. Include missing telemetry, contradictory process indicators, a tool returning from maintenance, an unusual material lot, and an apparent improvement followed by a downstream quality problem. Decide which exceptions suspend automation, which require a second reviewer, and who can restart a stopped workflow. People need explicit authority to challenge an agent’s recommendation without pressure to protect a rollout’s productivity targets.
Then measure decisions rather than dashboard activity. Track time to identify and investigate actionable anomalies, recommendations reversed after review, repeat exceptions, and engineering hours spent verifying suggestions. Where possible, connect those measures to controlled comparisons involving cycle time, scrap, and yield. Faster analysis can still slow the operation if it shifts hidden verification work onto engineers.
The workforce design matters too. Experienced process engineers may question an opaque recommendation because they have seen plausible correlations fail under changing production conditions. Give those engineers a formal role in constructing test cases and defining stop rules. Their objections can reveal technical failure modes that a data team overlooks.
Junior engineers face a different risk. If an agent supplies the apparent diagnosis first, they may get fewer opportunities to practice the reasoning required when the system is mistaken. Preserve structured reviews where employees inspect evidence and defend competing explanations before comparing their conclusions with the agent’s output.
The CONNECT 2026 agenda demonstrates the growing commercial focus on AI-driven manufacturing analysis. As fab operators move toward delegated decisions, leaders should be able to answer three questions about any agent-initiated action: What evidence supported it? Who could stop or override it? How will we know whether the decision improved performance?
Gleb Tsipursky, PhD, is a behavioral scientist, CEO of Disaster Avoidance Experts, and author of The Psychology of AI Adoption at Work: From Resistance to Results (Georgetown University Press, 2026).
Also Read:
The Industry Challenge: AI Is Forcing Semiconductor Innovation Beyond the Chip
TSMC’s AI Expansion Signals a Broader Shift in Computing
Terafab’s Single Roof: Why Packaging, Not Lithography, Decides High Volume
Share this post via: