AI Controls In Manufacturing

Summary

This document contrasts administrative controls (human-readable policy) with engineering controls (automated enforcement), and shows a practical demo of converting a written downtime-entry policy into agentic automation that produces consistent, standards-compliant MES entries. It summarizes definitions, tools used, demo steps, best practices, business value, and speakers/sources.

Core idea

• Administrative controls are the starting point: define “what good looks like” as policies, SOPs, or data-entry standards.
• Engineering controls implement those policies in code, agents, PLC logic, or other automation so outputs are consistent and enforceable.
• Building automation before defining clear standards produces inconsistent and low-trust AI outputs. Manufacturers should first document precise, testable administrative controls, then convert them into engineering controls.

Definitions and concepts

• Administrative control: A documented policy or rule (SOP, MES data-entry standard, system prompt). Typically human-dependent and intermittently enforceable; observed human data fidelity is roughly ~60%.
• Engineering control: Technical implementations that enforce policy (machine guards, PLC logic, software/agent automation). These reduce reliance on human memory/compliance and produce consistent, standards-compliant outputs.
• System prompt: When used as rules for an agent, a system prompt functions as an administrative control.
• Agentic AI: Autonomous agents that read inputs, apply policy, and produce outputs without manual enforcement.

Practical tutorial / demo

Use case: downtime-reason entry for MES. A 7-page administrative-control document specified required fields and formats:

• Required fields: downtime category, primary cause, action taken, notified parties, contributing factor, operator ID, duration.
• Format rules and examples were included.

Platform and tools used:

• Ajanet AI (agentic AI)
• UNS Studio / UNS Browser (agent and tag browser)
• MES tags and simulators
• Tag-change triggers, get-tag tool, and script engine

Demo highlights:

1. Comment formatter (simple demo)

   • Create an agent with a system prompt:

     “Take raw comment, return only a professional formatted comment.”

   • Test: raw operator text -> agent returns a professionalized comment.

2. Downtime formatter (engineering-control demo)

   • Create an agent triggered on tag change (line status).
   • Agent reads specific MES tags (category, cause, action, operator ID, duration, etc.).
   • Agent writes a standardized, policy-compliant formatted entry into a formatted_comment tag.
   • Result: when the simulator sets line-down and populates tag values, the agent produces a complete downtime narrative (category, primary cause, action, notifications, contributing factor, status).

Implementation notes:

• Agents can both read tags and format output; alternatively, scripts can call agents (demo used the agent to read tags for clarity).
• Tag-change triggers and get-tag operations pull MES values and write payloads to the formatted comment tag.

Best practices and warnings

• Define AI strategy and administrative controls first. Precise, testable standards must exist before engineering an agent to enforce them.
• Avoid vague engineering prompts (e.g., “summarize complaints”) without acceptance criteria — this yields inconsistent, plausible-but-unreliable outputs and erodes trust.
• Use agents to augment humans (make tasks easier) or automate repetitive, low-fidelity tasks (verify vs fully replace operator input).
• Administrative controls are a prime source for selecting automation use cases; they reveal desired outcomes and the reasons behind processes.

Business value

• Converts low-fidelity, human-entered data into high-quality, consistent records.
• Improves downstream analytics (OEE, RCA, maintenance planning) by reducing “garbage in.”
• Allows operators to verify machine-crafted entries rather than manually producing them, which reduces human error and variability.

Tools and features demonstrated

• Ajanet AI / agentic AI for autonomous formatting and enforcement
• UNS Studio: agent definition, prompts, trigger types (on-demand, tag-change), and test runner
• UNS Browser / simulator: MES tag simulation and formatted-comment tag writes
• Tag-change triggers and get-tag operations for pulling MES values and writing payloads

Main speakers and sources

• Video presenter / host (author of the demo and commentary; references prior video “AI isn’t helping”)
• Ajanet AI (agentic AI product/concept used in the demo)
• UNS Studio / UNS Browser (platform used to build and run agents)
• MES system, PLCs, and shop-floor data (context and data sources used in the examples)