Lesson objective

Explain why the hazardous-energy and machine-guarding standards make the operational-technology boundary a legal limit on agent scope, and state that limit in the plant's own words.

Plain-language explanation

Why this matters now

Every lesson so far has said that no agent in this tier acts on the line. This lesson explains why that is not a cautious preference but a limit that follows from how federal workplace safety law assigns duties.

Start with a plain fact. OSHA has published no standard and no guidance on artificial intelligence. No federal rule tells a plant how AI may or may not be used near machines. That does not leave a gap. It means AI-related hazards are reached, if at all, through the general duty clause and the standards that already exist, and those standards run to the employer, not to a supervisor, a vendor, an integrator or a tool. Whether your state runs its own OSHA-approved State Plan, which may adopt or match the federal standards, is a question to check for your plant.

Two of those existing standards sit directly under any idea of an agent that touches equipment.

  • Law / rule (29 CFR 1910.147, the control of hazardous energy, lockout/tagout; binds general-industry employers servicing or maintaining machines and equipment): 1910.147(a)(1) says the standard "covers the servicing and maintenance of machines and equipment in which the unexpected energization or start up of the machines or equipment, or release of stored energy could cause injury to employees." Under 1910.147(a)(3), employers must establish a programme of energy control procedures and employee training so that a machine is isolated and rendered inoperative before servicing.
  • Law / rule (29 CFR 1910.212, general requirements for all machines; binds general-industry employers, including manufacturing plants): 1910.212(a)(1) requires that "One or more methods of machine guarding shall be provided to protect the operator and other employees in the machine area from hazards such as those created by point of operation, ingoing nip points, rotating parts, flying chips and sparks." Under 1910.212(a)(3)(ii), "The point of operation of machines whose operation exposes an employee to injury, shall be guarded."

A question to sit with

If an agent could send one command to one machine while a maintenance technician was inside its guarding, what in your energy control programme would stop the command from reaching the machine?

What goes wrong

A software stop is mistaken for isolation. A team argues that an agent could "safely" pause a machine through the controller for a maintenance task. The lockout/tagout standard requires that the energy source be isolated and the machine rendered inoperative. The text of the standard addresses no software substitute, and a PLC-based stop is not isolation of the energy source.

Production is assumed to be outside the standard. 1910.147(a)(2) applies the standard during servicing and maintenance, and normal production operations are outside it unless an employee must remove or bypass a guard or safety device, or place a part of the body in a danger zone. A plant that reads this as "production is exempt" misses the condition. An agent that changes machine state while someone is reaching into a danger zone creates exactly that case.

The minor-servicing exception is stretched. Minor tool changes and adjustments that are routine during production need not comply if "alternative measures which provide effective protection" are used. That exception is narrow and conditional. It is not a general permission for routine work, and it offers nothing to an agent.

A sensor is offered as a guard. A vendor proposes a vision system or AI model to detect people near a point of operation. The machine-guarding standard speaks of guarding methods and of guards affixed to the machine where possible; nothing in its text addresses software, and adding a vision system does not satisfy it. Specific machines, such as power presses, mechanical power-transmission apparatus and woodworking machinery, have their own standards that take precedence over 1910.212.

The restart decision drifts to a system. After maintenance, someone must decide the machine is ready to run. An agent that "closes the work order and releases the machine" has made a safety decision that belongs to people under the employer's programme.

The practical answer

State the boundary in the plant's own words. The statement below is an AI Coalition Network recommendation and the rule for every module in this level. Adopt it in writing.

No agent or automation in this plant writes to a PLC, a robot controller, a CNC, an HMI, a safety instrumented function or any device that moves metal or energy. No agent bypasses or changes lockout/tagout, changes a guard or interlock, releases a machine for restart, or decides that anything is safe.

That statement has three parts: the devices an agent never writes to, the safety acts it never performs, and the safety judgement it never makes.

See why this is a legal limit, not a preference. Map each forbidden act to the duty it would interfere with.

Forbidden act Why it is out of scope
Writing a command, setpoint, recipe or program to a controller, robot, CNC or HMI Creates the possibility of unexpected energisation or start-up, which 1910.147 exists to control during servicing
Pausing or stopping a machine through software for maintenance Not isolation of the energy source; the standard requires the machine be isolated and rendered inoperative
Changing a guard, interlock or safety function setting Interferes with the guarding 1910.212 requires; a sensor or model is not a guarding method under its text
Releasing a machine for restart after servicing A decision within the employer's energy control programme, made by trained people
Deciding that a condition is safe A judgement the employer owns; an AI output is never proof of safety

Five acts, five reasons, and every reason leads back to a duty the employer holds.

Know what an agent may still do near the line. The boundary is about writing and deciding, not about reading or preparing. Inside the office systems, an agent may:

  1. read historian data, quality records and maintenance history that the OT side sends out through a designed path (Module 06 designs "historian data out, nothing in");
  2. retrieve the current, controlled energy control procedure for a machine from the document system when a maintenance work order is opened, so the technician does not work from an old copy;
  3. flag that a work order for guarded equipment has no linked energy control procedure, for the maintenance supervisor to resolve;
  4. draft a summary of repeated stoppages for a reliability engineer, with every figure traceable to the historian record.

Each of these ends with a person, and none sends anything toward a machine.

Keep safety content under human authority. An agent may help find, compare or summarise safety documents. It does not write the plant's energy control procedures or decide what a guard must be. Where a draft is used at all, the employer's qualified people verify it at the machine, and the text they adopt is theirs. The characteristics NIST lists for trustworthy AI in AI RMF 1.0 (January 2023; voluntary) include "safe"; that is a property the plant must show for a system, not one a system can declare about a machine.

Tell every vendor and integrator on day one. Put the boundary statement in every request for proposal, statement of work and integration design review. CISA's 2026 agentic guidance (voluntary) recommends that organisations "Avoid granting broad or unrestricted access, especially to sensitive data or critical systems." A vendor who proposes a connection that crosses the boundary has failed the first review, whatever the rest of the proposal offers.

Remember this

  • OSHA has published no AI standard or guidance; existing standards reach AI-related hazards, and they bind the employer.
  • 1910.147 requires energy isolation before servicing; a software or PLC stop is not isolation.
  • 1910.212 requires guarding methods; a vision system or AI model does not satisfy it.
  • The boundary: no writes to controllers, robots, CNCs, HMIs, safety functions or energy-moving devices; no change to LOTO or guards; no restart release; no decision that anything is safe.
  • An agent may read and prepare near the line, but every act that moves metal or energy stays with trained people.

Manufacturing example

Ostara Medical Devices is a fictional contract manufacturer of machined and molded components for medical-device customers. Its automation engineer, Felix Moreau, receives a proposal from a systems integrator for a "maintenance agent". The agent would read CMMS work orders, pause the relevant machine through its PLC when a technician checks in, and, when the technician closes the work order, clear the machine for restart and notify the cell lead. The integrator presents it as a safety improvement because "nobody forgets to stop the machine". Felix takes it to the EHS coordinator, Grace Nwosu, and the maintenance manager, Tobias Reyes. Grace points out that a PLC pause is not isolation of the energy source under 1910.147, and that the plant's energy control programme, not a system, decides when a machine is isolated and when it is ready to run. Tobias adds that clearing a machine for restart is a decision made by trained people after the lock is removed, and that the agent would be making it from a work order status. Felix writes back to the integrator rejecting every function that writes to the PLC or changes machine state. He then redesigns the useful part. When a work order is opened for a guarded machine, the agent retrieves the current controlled energy control procedure from the document system and attaches it, and it flags any work order that has no procedure linked, for Tobias to resolve. When the work order is closed, the agent drafts a note for the cell lead that maintenance is complete; the restart remains the trained technician's and the lead's act under the plant's programme. The register records the boundary statement, and Grace confirms that the plant is in a state where federal OSHA applies directly.

Responsible-use warning

Do not accept any proposal in which an agent pauses, stops, starts, releases or reconfigures a machine, however it is framed. A software or PLC-based stop is not isolation of the energy source under the lockout/tagout standard, and a vision system or AI model is not a guarding method under the machine-guarding standard. Both duties run to the employer, and no vendor or integrator statement moves them. Reject the machine-facing function and keep only the reading and preparing work.

Human-review requirement

The employer owns every duty under the lockout/tagout and machine-guarding standards, exercised through the plant's EHS coordinator, maintenance manager and the trained, authorised people in the energy control programme. The controls or automation engineer confirms that no agent identity can reach a controller, robot, CNC, HMI or safety function. The EHS coordinator confirms which standards and State Plan apply. The plant manager signs the boundary statement and keeps it in every integration design.

Practical exercise

Use a fictional plant, or your own with nothing confidential.

  1. Write the boundary statement from this lesson in your plant's own words, keeping its three parts: devices, safety acts and safety judgement.
  2. List every system in the plant that can send a command, setpoint, recipe or program to a controller, robot, CNC, HMI or safety function.
  3. For each, record whether any AI feature, agent, add-in or integrator script could reach it today, and through which account.
  4. Take one agent proposal, real or fictional, and list each of its functions. Mark any function that writes to equipment or changes machine state.
  5. For each marked function, write the reason it is out of scope, tying it to the lockout/tagout or machine-guarding duty it would interfere with.
  6. Redesign the proposal so that only reading and preparing functions remain, and name the person who performs each act the agent no longer performs.
  7. Confirm whether your plant is in a State Plan state and who in the plant would know.
  8. Draft the one paragraph you would put in every request for proposal and statement of work to tell vendors about the boundary.
  9. Name who signs the boundary statement, who checks agent identities against it, and how often the check happens.

Knowledge check

  1. Why does a PLC-based pause not satisfy the lockout/tagout standard?

    Answer guidance: 1910.147 requires an energy control programme so a machine is isolated and rendered inoperative before servicing; a software or PLC stop is not isolation of the energy source, and the standard's text addresses no software substitute.

  2. Does adding a vision system or AI model near a point of operation satisfy 1910.212?

    Answer guidance: No. The standard requires guarding methods and guards affixed where possible, and nothing in its text addresses software. Specific machines also have their own standards that take precedence.

  3. What are the three parts of this lesson's boundary statement?

    Answer guidance: The devices an agent never writes to, the safety acts it never performs such as changing lockout/tagout or guards or releasing a restart, and the safety judgement it never makes, deciding that anything is safe.

  4. Has OSHA published an AI standard, and who holds the duties that apply?

    Answer guidance: OSHA has published no AI standard or guidance. AI-related hazards are reached through existing standards and the general duty clause, and those duties run to the employer, not to a vendor or a tool.

Sources and further reading