Industrial & physical AI · 2 of 3

Manufacturing AI needs a verification layer between insight and control

August 13, 2026 · by Kate Sai Kishore · 10 min read

Factory and IIoT flow diagram where sensor evidence passes through verification before a bounded action

Industrial systems are very good at executing valid commands. They are not designed to decide whether an AI agent had a good reason to issue one.

A PLC can enforce interlocks, ranges, and sequences. A historian can record values. An MES can track production. The missing layer appears when a probabilistic system interprets that context and proposes a change: which evidence supports the proposal, and does this identity have authority to make it?

Opportunity, not deployment claim. The patterns below are proposed use cases. qbrin's public industrial posts describe controlled experiments and simulations; they do not establish production deployment in a customer's plant.

Put the boundary above deterministic control

Fast protection loops, safety instrumented systems, and certified controller logic should remain deterministic and local. A trust layer is better suited to supervisory decisions around them: work-order creation, maintenance recommendations, quality holds, recipe selection, energy scheduling, alarm triage, and bounded setpoint proposals.

The AI can reason over context. The plant's existing controls remain authoritative. The verification layer decides whether the contextual premise is supported and whether the requested action is inside policy.

A reference flow for an IIoT decision

Bind the request to the asset, line, batch, recipe, operating mode, and time window.
Retrieve evidence from approved sources: historian, CMMS, MES, quality records, manuals, and change logs.
Verify the diagnosis and the conditions required by the procedure.
Classify the proposed action by risk and reversibility.
Apply identity, site, shift, and approval policy.
Return ADVISE, HOLD, DENY, or a narrowly parameterized action request.

High-value starting points

Use case	Why it fits	First safe mode
Maintenance triage	Evidence spans alarms, manuals, history, and parts	Rank causes with citations; human creates work order
Quality investigation	Claims depend on batch, recipe, inspection, and change context	Propose a hold with the supporting record
Energy optimization	Recommendations need production and equipment constraints	Simulate schedule changes; no direct control
Operator assistance	Procedures are useful only when version and mode match	Answer with exact step and source
Supervisory setpoints	Potential value is high; consequence is higher	Shadow recommendations and require approval

Why existing interlocks are necessary but insufficient

An interlock can stop pressure from exceeding a limit. It may still accept a command that is syntactically valid, within range, and wrong for the operating objective. qbrin's published water-treatment experiment and power-grid experiment are designed to illustrate this gap: the underlying control systems can carry out commands that fit their mechanics without judging the agent's evidence or intent.

Those are controlled experiments, not field-deployment claims. Their value is conceptual: process protection and epistemic authorization solve different problems.

Industrial evidence has shape

Plant claims are not supported by text alone. A useful evidence bundle may include a time series, event sequence, equipment hierarchy, calibration state, unit of measure, maintenance revision, operator note, and production context. The trust layer has to preserve those relationships rather than flattening every signal into prose.

Freshness should be typed. A manual may remain authoritative for months; a vibration reading may expire in seconds. A batch disposition may require two signatures. A verifier that ignores these semantics can produce a beautifully cited wrong recommendation.

Start in shadow mode

Choose one line and one decision class. Let the system observe approved data and produce a proposed action plus evidence, but do not connect it to control. Compare its proposals with operator decisions, especially during changeovers and abnormal situations. Count unsupported claims, missed conditions, false holds, and time saved.

Then add a human approval workflow with a narrow action envelope. Preserve a hard separation between the AI's reasoning, the trust decision, and the controller's safety logic. Expansion should follow measured performance and site governance—not pressure to announce autonomy.

The industrial opportunity is not an AI that can touch every control. It is an AI whose reasons, permissions, and limits are clear before it touches any.

Evaluate qbrin on your workflow →