Can an AI Agent Decide Which Wafers to Scrap?
Agentic AI is moving into fab workflows. An agent can prepare a wafer disposition recommendation, while the final call stays with MES rules and engineers. This article explains why.
The useful role for AI in wafer disposition sits before the final decision. An agent can assemble inspection evidence and recommend an action. But the decision authority remains with the fab's MES rules and engineers. Reliable classification evidence is what makes those recommendations worth considering.
Can an AI Agent Decide Which Wafers to Scrap?
An AI agent is software that plans and takes steps toward a goal, such as gathering data or calling other systems and proposing an action. In a fab, it can assemble the evidence for an action within the fab's established decision-making framework. The final release, hold, rework, or scrap decision stays with MES rules and engineers in most fabs.
Wafer scrap decisions should remain subject to the fab's established controls because of accountability. A scrap decision destroys value that cannot be recovered, and a release decision can send a defect into a customer's product. Both must be explainable afterwards, down to the evidence used and the person who owned the call. An agent that acts on its own can make that chain harder to reconstruct without appropriate controls and logging.
An agent can save engineers time in disposition. It can pull the inspection images, classification results, tool history, and the lot context into one view. It can then route the case to the right engineer.
How Do Semiconductor Fabs Keep Wafer Disposition Consistent Across Tools?
Wafer disposition is the decision on a wafer or lot after inspection or test: release, hold, rework, or scrap. Fabs keep it consistent with one defect taxonomy across tools, automated defect classification (ADC) that applies those classes the same way on every tool, MES rules that act on the classes, and a defined route for uncertain cases.
The weak point is often the classification layer. Two tools of the same type can label the same defect differently when their optics, camera settings, illumination, or recipes differ. Manual review adds variation by shift and reviewer. When the classes underneath the MES rules are inconsistent, the rules apply consistently to inconsistent inputs.
In practice, consistent classification across a tool fleet starts with validated classification models and consistent defect definitions across tools. Each classification is tied to the model version that produced it. Uncertain cases are flagged as uncertain and go to an engineer. With those in place, the MES rules on top can do their job.
What Is the Visibility Gap in Defect Disposition?
The visibility gap is the part of a device's condition that no single measurement method captures. In "Closing The Visibility Gap" (October 8, 2026), Semiconductor Engineering describes how modeling, inspection, electrical test, and field monitoring each give a partial picture, and argues for connecting their data.
For wafer disposition, a nearer gap sits inside inspection: classifications that differ between tools or carry no record of the model behind them. An agent cannot close that gap on its own, because it can only combine what the underlying systems provide. If the classification it reads is confidently wrong on one tool and uncertain on another, the agent's recommendation inherits both problems and makes them harder to spot.
Classification evidence matters more once agents enter the disposition workflow. Ranked classifications with an explicit unknown answer, supported by model version and training data records, give an agent and an engineer evidence they can check. Without that layer, agentic workflows speed up decisions that rest on weak evidence.
Who Should Do What in AI-Assisted Wafer Disposition?
AI-assisted disposition requires a clear division of responsibilities between classification, agentic workflows, and disposition authority. Keeping the roles separate keeps every decision explainable and supports accountability.
| Role | What it does | What it should not do |
|---|---|---|
| Classification (AI ADC) | Classifies defects and ranks the result, then returns an unknown result when classification criteria are not met | Decide release or scrap |
| Agent or workflow | Gathers evidence and context, proposes a disposition, routes the case | Override MES rules or act without an owner |
| MES rules and engineers | Apply approved disposition rules and resolve cases requiring engineering judgment. | Decide without the evidence trail |
In Indurion, the Robovision platform for fabs, AI ADC supplies the classification layer, and unknown-rate monitoring shows how often it answers unknown against the target set during threshold optimization. Every trained model shows the frozen, dated datasets it was trained and tested on. Robovision's Drift Detection is part of the Indurion licence. Set up per project on the Hub, it compares live inference data with selected reference data and reports a drift score against a configured threshold.
What Should a Fab Check Before AI Influences Disposition?
A fab should check five things before letting AI influence disposition. Each classification should be traceable to its model version and the datasets used to train and test that model. The classifier can answer unknown. A person starts every retraining run. Drift in the input data becomes visible. The decision trail can be reconstructed after the fact.
The five disposition checks apply whether the fab builds its own agents and models or buys them. A fab's own AI team can build the models and agents. The governance underneath them is where a dedicated platform can help: versioned datasets and controlled retraining, as well as mechanisms for monitoring uncertainty and data drift.
Teams that build their own applications take Indurion as a Platform subscription. Fabs that want Robovision's team to keep the models performing in production add AI Operations. Indurion does not release or scrap product on its own.
You take the decisions. We supply the ranked, auditable evidence behind them.
To explore the costs of manual defect review and uncertain decisions, see our defect review opportunity article. For the potential financial impact of reducing unnecessary scrap at the bond interface, see our hybrid bonding business case.