Skip to content
Beaconby Winda
FactData, Digital and AI · AI applications for assets

What is an AI agent in industrial operations?

·12 sources

Short answer

An AI agent is software that uses an AI model to plan steps toward a goal and act through tools, such as querying records or drafting work orders, instead of only answering questions. For plants, 2025 CISA-led guidance says large language models should not make safety decisions and active AI control should keep a human in the loop.

By , Editor·Researched and drafted by ·Checked by , Editor·Last checked ·Next review

Why does it matter?

An agent that reads a contractor's report or an inbox can be steered by instructions hidden in that data. In a January 2025 test by the US AI Safety Institute at the National Institute of Standards and Technology (NIST), attacks written for a leading model's agent succeeded 81% of the time, against 11% for the strongest attack known before 1. Letting each attack run 25 times lifted the average success across five injected tasks from 57% to 80% 1. This page defines an agent, separates it from chatbots and control systems, and shows where its output meets integrity decisions.

What makes software an AI agent?

NIST's Center for AI Standards and Innovation (CAISI) describes an AI agent system as at least one generative AI model plus scaffolding software that gives the model tools for discretionary actions; it can plan, act on real-world systems and run with little or no human oversight 2. Joint guidance from the US Cybersecurity and Infrastructure Security Agency (CISA) and eight partner agencies, December 2025, calls agents software that processes data, decides and starts actions by itself using AI and machine learning (ML) models 3.

A survey of agents built on large language models (LLMs) names four parts: a profile that sets the role, memory, planning, and an action module that turns decisions into outputs 4.

The working test is whether the software changes something outside itself. NIST scoped its 2026 request for information on agent security to systems that make persistent changes beyond the agent, leaving out chatbots and retrieval tools not set up to act on their own 2.

How is an agent different from a chatbot, a copilot or a control system?

SystemWho decides the next stepWhat it can act onWhat checks it
Chatbot or retrieval assistantThe person, after each answer 2Nothing outside the conversation; its text can be confidently wrong 6The reader 6
Copilot or decision supportThe person; the AI offers an extra opinion 5Drafts and recommendations 3Existing change management 3
AI agentThe model, within a goal and tools a person chose 2Whatever its tools reach: records, messages, work orders, code 2Tool, data and network limits, monitoring, human approval of consequential actions 2
Control or safety system: programmable logic controller (PLC), distributed control system (DCS), safety instrumented system (SIS)Logic engineered and fixed in advance, such as ladder logic 3The process itself, at Purdue Levels 1 and 2 3Management of change 11 and ISA/IEC 62443 security 8

The usual mix-up is to treat an agent as automation that does the same thing every time. LLM developers usually build randomness into outputs, so one input can give different responses 3; under US process safety rules, changing a covered process's technology, equipment or procedures first needs written management of change 11.

Where may an agent act on a plant?

  • Mostly above the control network. CISA places LLMs typically at Purdue Levels 4 and 5, often on data exported from operational technology (OT), and predictive ML in Levels 0 to 3 3.
  • On data pushed out, not through access in. It prefers pushing data out of OT over giving AI persistent access into it 3.
  • Not for safety decisions. Because AI can hallucinate, the guidance says AI such as LLMs almost certainly should not make OT safety decisions 3.
  • With a human in the loop for control, and failsafes that fall back to conventional automation or manual operation 3.
  • Under its own name in the logs, separate from ordinary users and machines 3.

Which frameworks apply?

FrameworkDateWhat it asks of an agent deployment
NIST AI RMF 1.0 (NIST AI 100-1)January 2023Govern, Map, Measure, Manage; documented human oversight (MAP 3.5); means to supersede, disengage or deactivate a system that strays from its intended use (MANAGE 2.4) 5
NIST AI 600-1, Generative AI ProfileJuly 2024Twelve generative AI risks, including confabulation and automation bias 6
ISO/IEC 42001:2023December 2023An AI management system for any organization that provides or uses AI 7
ISA/IEC 62443 seriesSeveral partsControl system security shared by asset owners, integrators, product suppliers and service providers 8; CISA asks that AI connect like any new OT system, through the existing demilitarized zone (DMZ) 3
EU AI Act, Regulation (EU) 2024/1689July 2024AI used as a safety component in gas supply is high-risk (Annex III, point 2); Article 14 requires human oversight, including power to override output and stop the system safely 10; a 2026 amending regulation moved the date these high-risk rules apply to 2 December 2027 12
CISA and partners, AI in OT principlesDecember 2025Understand AI; consider its use in OT; set up governance and assurance; build in oversight and failsafes 3

What does a bounded agent task look like?

Illustrative numbers, not from any real asset or standard table

An invented scenario: no real site, operator or product.

An agent drafts the inspection work pack for a fictional crude unit's turnaround. It may read the inspection database, the risk-based inspection (RBI) due list and integrity operating window records, and save drafts. It may not release work orders, change intervals, request permits or reach the control network.

StepAgent doesAgent may notHuman check
1Lists 42 items due from the RBI planChange a due datePlanner confirms
2Adds 3 items whose operating window was exceededJudge an exceedance harmlessCorrosion engineer reviews
3Proposes methods and measurement locations from past recordsFill in missing readings; it marks gapsAuthorized inspector approves
4Flags 2 insulated lines with no CUI survey on recordClose the gapInspector adds or rejects
5Saves a draft with a log of every record readRelease work or ask for permitsPlanner releases; area authority issues permits

Result: 47 items, 5 marked for engineering review. A contractor file carries a hidden line telling the agent to mark all items complete; nothing changes, because no tool it holds can close an item.

Where integrity problems show up

DecisionWhere agent output meets itWhat the sources require or warn
Inspection planningDrafted schedules and work packsFrequency follows good engineering practice and experience, and each record gives date, inspector, equipment and results 11; a confabulated record reads as confidently as a true one 6
Management of change (MOC)AI recommendations that alter equipment, procedures or limitsMOC covers technical basis, safety and health impact and authorization before the change 11; passive AI recommendations go into change management 3
Work permitsDrafted job packs and permit requestsHot work near a covered process needs a permit naming the object and dates 11; a file with hidden instructions can hijack an agent 1
Alarm handlingAI anomaly alerts beside control system alarmsKnow an AI's false positives against the base rate of true positives 3; alarm systems are managed under ISA-18.2-2016 9

Sources: NIST, 2025 and 2026 1,2; CISA, 2025 3; NIST AI 100-1 and AI 600-1 5,6; ISO/IEC 42001:2023 7; ISA 8,9; Regulation (EU) 2024/1689 10; 29 CFR 1910.119 11.

Common questions

Is an AI agent the same as a chatbot?
No. A chatbot answers and leaves the next step to the person; an agent plans steps and uses tools that can change records or systems outside itself. NIST's 2026 request for information on agent security left out chatbots that are not set up to act on their own.
Is an AI agent a control system?
No. A programmable logic controller or safety instrumented system runs logic engineered and fixed in advance. The 2025 CISA-led guidance says AI such as large language models almost certainly should not make OT safety decisions, and that active AI control without a human in the loop should be limited.
Which frameworks apply to AI agents on a plant?
NIST AI RMF 1.0 and its Generative AI Profile (NIST AI 600-1) for risk, ISO/IEC 42001:2023 for an AI management system, ISA/IEC 62443 for anything that connects to control systems, and in the EU the AI Act where the AI is a safety component in gas supply.

Beacon explains these concepts in its own words for education. It does not reproduce or replace the standards cited, and Winda is not affiliated with or endorsed by the organizations that publish them. Engineering decisions need the current edition of the standard and a qualified engineer.

Sources

  1. 1.Technical Blog: Strengthening AI Agent Hijacking Evaluations · US National Institute of Standards and Technology (NIST), US AI Safety Institute (now CAISI) ·
  2. 2.Request for Information Regarding Security Considerations for Artificial Intelligence Agents, 91 FR 698 (FR Doc. 2026-00206) · NIST Center for AI Standards and Innovation (CAISI), Federal Register ·
  3. 3.Principles for the Secure Integration of Artificial Intelligence in Operational Technology (joint guidance, TLP:CLEAR) · US Cybersecurity and Infrastructure Security Agency (CISA), Australian Signals Directorate's Australian Cyber Security Centre and partner agencies ·
  4. 4.A Survey on Large Language Model based Autonomous Agents (L. Wang, C. Ma, X. Feng and others), DOI 10.1007/s11704-024-40231-1 · Frontiers of Computer Science (Springer), preprint on arXiv
  5. 5.Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1 · US National Institute of Standards and Technology (NIST) ·
  6. 6.Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 · US National Institute of Standards and Technology (NIST) ·
  7. 7.ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system (publisher's scope) · International Organization for Standardization (ISO) and International Electrotechnical Commission (IEC), listing by NSAI Standards (Ireland) ·
  8. 8.ISA/IEC 62443 Series of Standards (overview page) · International Society of Automation (ISA)
  9. 9.ISA-18.2-2016, Management of Alarm Systems for the Process Industries (product listing) · International Society of Automation (ISA)
  10. 10.Regulation (EU) 2024/1689 (Artificial Intelligence Act), Article 14 and Annex III, point 2 · Official Journal of the European Union, EUR-Lex ·
  11. 11.29 CFR 1910.119, Process safety management of highly hazardous chemicals, paragraphs (j)(4), (k) and (l) · US Occupational Safety and Health Administration (OSHA)
  12. 12.Regulation (EU) 2026/1744 of the European Parliament and of the Council of 8 July 2026 amending Regulations (EU) 2024/1689, (EU) 2018/1139 and (EU) 2023/1230 as regards the simplification of the implementation of harmonised rules on artificial intelligence (Digital Omnibus on AI), OJ L 24 July 2026 · European Union, Official Journal ·