It is 5:40 on a Tuesday morning and the second-shift lockout has not fully cleared. A pick-and-place cell on Line 3 throws a fault, the kind that used to mean a maintenance tech walked over, racked out the disconnect, applied a personal lock, and cleared the jam by hand. This morning the cell has a new feature the integrator installed last quarter: an AI vision model that watches the infeed and, when it decides the path is clear, releases a soft restart so the line does not lose a cycle. A green-crew operator who started in March sees the model's status light turn green and reaches in. The light was wrong. The model had been trained on day-shift lighting, and at 5:40 in winter the overhead bay was darker, the part was a slightly different shade of gray after a resin change, and the model misread a partially seated nest as empty. The cell did not restart, because a hardware light curtain caught the operator's arm breaking the plane and held the machine in a safe state. Nobody got hurt. But sit with what almost happened: a software model, trained on the wrong conditions, was one design decision away from being the thing that decided whether it was safe for a human to put a hand inside a machine. That decision is the single most important boundary in this entire program, and this lesson is about where that boundary sits, why it does not move, and how Environmental Health and Safety (EHS, the plant function responsible for keeping people from getting hurt and the site within environmental and regulatory limits) has to think about AI before a vendor's demo talks anyone into crossing it.

The Line That Does Not Move

Every other lesson in this level has been about using AI to help a thinning, greener crew run a safer and higher-quality line. This one is about the exact opposite framing, and it matters that you hold both at once. AI is a knowledge multiplier for the people on the floor. It is not a substitute for the engineered safeguards that keep those people alive. The line that does not move is this: a human stays accountable for anything on the floor that can hurt someone, and a safety-critical control loop is not where AI goes, not in 2026 and not because the technology is not impressive enough.

To say that precisely you need one piece of vocabulary that controls engineers live inside and most AI vendors gloss over. A safety-critical control loop is any chain of sense, decide, and act whose failure can directly injure a person or release a hazard: a light curtain that must stop a press before a hand reaches the die, an emergency stop that must drop power across a cell, a two-hand control that forces an operator's hands away from the pinch point, a guard interlock that will not let a robot move while a door is open, an overpressure relief that must vent before a vessel ruptures, an interlock that keeps a furnace burner from firing into an unpurged chamber. These loops are designed to a discipline called functional safety, and they are governed by standards your plant already lives under: ISO 13849 and IEC 62061 for machinery, IEC 61508 as the parent standard for the electrical and electronic systems that perform safety functions, and in the United States the OSHA general duty clause plus the relevant machine-specific standards. None of those standards has a clause that says "or an AI model, if it seems confident." They require deterministic behavior, validated failure modes, and a quantified probability of dangerous failure. A modern machine-learning model offers none of those three things natively, and that is not a maturity gap that another training run closes. It is a category difference.

Here is the plain-language version. A light curtain is deterministic: the same beam break produces the same stop, every single time, and you can calculate exactly how often the circuit itself will fail dangerously, which is why a properly rated safety circuit carries a Performance Level or a Safety Integrity Level you can put a number on. An AI vision model is statistical: it produces an output that is right most of the time, with a distribution of errors that shifts when the lighting, the material, the camera angle, or the part presentation drifts, exactly the drift you have read about in the quality lessons. You can measure a vision model's false-reject and false-accept rates on a holdout set, but you cannot promise that the next frame, under conditions slightly outside what it was trained on, lands inside that measured rate. "Right ninety-nine percent of the time" is a wonderful quality inspector and a catastrophic safety device, because the one percent is not random noise you can average away; it clusters exactly around the unusual conditions, the resin change and the dark winter morning, that you most need it to handle.

A statistical system that is right most of the time is a good inspector and a fatal safety device, because the failures cluster precisely on the conditions you most need it to catch.

Advisory Versus In the Loop

The boundary becomes usable on the floor when you can sort any proposed AI use into one of two buckets and you train yourself to ask which bucket before anything else. The buckets are advisory and in the loop, and almost every safe AI deployment in this program lives in the first.

An advisory use means the AI produces information that a human reads and acts on, and a hard-engineered safeguard, not the AI, is what actually prevents the harm. The predictive-maintenance model that flags a bearing trending toward failure is advisory: it writes a work order, a planner schedules it, a tech with a lock and a tag does the repair, and if the model is wrong the worst case is a wasted inspection, not a severed finger. The vision system that grades a cosmetic defect is advisory to safety even though it touches quality, because no one is inside the machine when the camera makes its call. The chatbot that drafts a lockout-tagout procedure from the equipment manual is advisory: a qualified person verifies every step against the actual machine before anyone uses it. In every advisory case, you can ask the load-bearing question and get a safe answer: if the AI is confidently and completely wrong on its worst day, what is the most that can happen to a person? For an advisory use the honest answer is bounded, because a deterministic safeguard or a human review stands between the model's error and any human body.

An in the loop use means the AI's output is part of the chain that decides whether a hazardous motion happens or a hazardous energy is present. The vision model in the opening story was being pushed in the loop: its judgment was wired into the restart decision. A model that disables a light curtain because it "sees" the area is clear is in the loop. A model that decides a vessel is safe to pressurize is in the loop. A model that overrides an interlock to save a changeover minute is in the loop. For every one of these, run the same question and the answer turns ugly: on the model's worst day, under drifted conditions, a person can be crushed, cut, burned, or gassed because a statistical system guessed wrong about a safety state. That is the answer that ends the conversation. You do not deploy AI in the loop of a safety-critical function. The cost of being wrong is not measured in scrap dollars or downtime minutes. It is measured in an ambulance.

The economics here are not the usual yield-and-downtime math, and it helps to say the real number out loud. The Occupational Safety and Health Administration sets a statutory maximum penalty that in 2026 sits above sixteen thousand dollars for a serious violation and above one hundred sixty-five thousand dollars per willful or repeat violation, and a single willful citation is rarely the whole bill. A recordable amputation triggers a mandatory report to OSHA within twenty-four hours, an inspection, and in many states a workers' compensation experience-modifier hit that raises the plant's insurance premium for years. Industry cost models commonly put the fully loaded cost of a single serious recordable injury, counting medical, indemnity, investigation, retraining, lost productivity, and the indirect drag, in the range of forty thousand to well over one hundred thousand dollars, and a permanent disabling injury an order of magnitude beyond that. Set that against the thing the in-the-loop deployment was trying to buy in the opening story: one machine cycle, worth a few cents of throughput. The trade is not close. You would have to run the line for years of saved cycles to fund the first amputation, and no number of saved cycles brings the hand back.

Where AI Helps EHS the Most

It would be a misreading of this lesson to walk away thinking AI has no place in safety. It has a large and genuinely valuable place, and that place is everywhere except inside the control loop. Once you keep AI advisory, the same knowledge-multiplier logic that drives the rest of the program applies directly to a function that is also short-staffed and drowning in paperwork. EHS in most plants is one or two people responsible for the entire site's compliance, training, incident records, and audits, and they are exactly the kind of thin, overloaded function AI was built to help.

Drafting and maintaining procedures. A generative model can take an equipment manual, a job hazard analysis, and the plant's lockout-tagout template and produce a clean first draft of a machine-specific lockout-tagout procedure, the energy-isolation steps that must be followed before anyone services equipment. That draft is a starting point a qualified person verifies against the actual machine, the actual disconnects, and the actual stored-energy sources. Done this way it turns a two-hour writing task into a thirty-minute verify-and-correct task, and the verification, not the drafting, is the safety act. Done the wrong way, by trusting the draft, it is exactly the hallucination failure mode from earlier in this level: the model can invent an energy source that does not exist or, far worse, omit one that does, and an omitted stored-energy source in a lockout procedure is how people die. The rule is the same one that runs through the whole program. Verify every AI-touched procedure against the real equipment before it governs a human's hands.

Mining incident and near-miss data. Most plants sit on years of incident reports, near-miss logs, and first-aid records that nobody has the time to analyze. AI is good at exactly this: reading a thousand free-text near-miss reports and surfacing that slips on the same dock plate in winter account for a quarter of them, or that a particular line generates a cluster of pinch-point near-misses on the same shift. That is a Pareto the safety lead never had time to build, and it points a finite prevention budget at the tallest bar. This is advisory in the cleanest sense: it informs where to put a guard, a sign, a training, or a redesigned task, and a human decides and implements. The leverage is real because near-miss data is the richest leading indicator of safety risk a plant owns, and it has historically been the least mined.

Training and onboarding a green crew. The talent cliff that drives this whole program, roughly two million manufacturing workers needing reskilling against about five hundred thousand unfilled roles, lands hard on safety. The crew is greener, and a green crew is a more-injured crew, because experience is itself a safety control. AI-assisted training content, plain-language and multilingual work instructions grounded in the plant's actual procedures, and a grounded question-answer assistant a new operator can ask "what is the lockout for this press" all help close the experience gap faster. Again advisory: the assistant retrieves and explains the verified procedure, it does not authorize the work, and a competent person still signs the permit.

Monitoring and surfacing, not deciding. An AI can watch a video feed and flag that someone entered a restricted zone, or read environmental sensors and alert that a solvent vapor concentration is trending toward the action level. That is useful as an alert to a human who investigates and acts. It crosses the line the moment the same system is wired to shut the line down by itself as the primary protection, because then it has become the safety function, and a safety function has to meet functional-safety standards a statistical model cannot meet. The pattern to hold is simple: AI may raise its hand and say "look here." A validated, deterministic system, or a qualified human, decides what happens next.

The Accountability Rule on the Floor

The cardinal rule of this program, that the customer audits you and not the vendor, has a safety twin that is even more unforgiving: OSHA cites you, not the model. When an inspector arrives after an incident, there is no line on the citation form for "the AI was confident." The general duty clause obligates the employer to provide a workplace free of recognized hazards likely to cause death or serious harm, and that obligation does not transfer to a software vendor or an integrator no matter what the purchase order said. If an AI-influenced decision was in the chain that led to an injury, the questions an investigator asks are aimed squarely at the plant and the named people who run it: Who decided this function could rely on the model? What was the validated failure rate? What hard safeguard backed it up? Who verified the procedure? Where is the documentation? "We bought it from a reputable vendor" is the start of that conversation and never the end of it.

This is why the human-accountable principle is not soft language in a safety context. Every safety-relevant role on the floor carries a named accountability that AI cannot absorb. The authorized employee who applies the lock owns the energy isolation. The competent or qualified person who signs the permit owns the judgment that the work is safe to proceed. The engineer who specifies a safeguard owns its adequacy. The EHS lead owns the program. An AI tool can hand each of those people a better draft, a faster analysis, a clearer instruction, and it can make a thin crew far more effective at the paperwork and the pattern-finding that safety runs on. It cannot be the authorized employee. It cannot sign the permit. It cannot be the competent person. The instant a deployment is structured so that the model effectively occupies one of those roles, the deployment has crossed the line, regardless of how the org chart describes it.

There is a documentation discipline that follows directly, and it is the same audit-grade habit the governance lessons teach, pointed at safety. Any time AI touches a safety-relevant artifact, a procedure, a hazard analysis, an incident root cause, an environmental report, the record has to show that a qualified human reviewed and approved it, by name and date, against the real equipment or the real data. That record is not bureaucracy. It is the evidence that the accountability stayed where the law requires it, and it is the difference between an inspection that finds a verified procedure with a human signature and one that finds an AI-generated lockout sheet nobody checked, which is itself a willful-category problem waiting to be cited.

Environmental Compliance and the Confidently Wrong Number

The E and the second part of EHS, environmental and the broader health exposure side, deserve their own treatment because the failure mode there is quieter than a crushed hand and just as serious in front of a regulator. Plants live under permits: a Title V air permit with emission limits, a wastewater discharge permit with concentration caps, hazardous-waste manifests, stormwater plans, reporting thresholds that trigger filings to the Environmental Protection Agency or the state agency. Much of the work is data handling and report drafting, which is precisely the kind of task a generative model will do quickly and, when wrong, will do confidently.

The danger is the invented or mis-stated number. Ask a model to summarize a quarter of continuous-emission-monitor data and draft the deviation report, and it can hand you a clean, professional document with an emission figure that is plausible and wrong, a unit silently converted the wrong way, an exceedance rounded out of existence, or a reporting threshold misremembered. A wrong number in an environmental report is not a quality escape that a customer catches. It is a false certification to a regulator, signed by a responsible official under penalty, and a knowing false statement carries personal and criminal exposure, not just a corporate fine. The exact same verification discipline applies: the AI may assemble and draft, the responsible human reads every number back to the source data, the monitor log, the lab result, the permit limit, before it goes out under a signature. The model is a faster typist and a useful first reader of a large dataset. It is never the source of truth and never the signatory.

There is a real upside on the environmental side too, and it is worth naming so the lesson is balanced. AI is genuinely good at the continuous, tedious watching that environmental compliance requires and that a one-person EHS function cannot sustain by hand: trending a discharge parameter and flagging that it is drifting toward the permit cap two weeks before it would breach, cross-checking a waste manifest for a missing field before it ships, reading utility and process data to find the energy and emissions waste a sustainability program is chartered to cut. All of that is advisory and high-value. It surfaces, a human verifies and acts, and the plant gets ahead of a violation instead of explaining one after the fact. The pattern holds across the whole function: keep AI on the side of seeing more and drafting faster, keep the human on the side of certifying and deciding, and the worst day stays bounded.

A Floor Test Before Any Safety-Touching Deployment

You do not need to be a functional-safety engineer to apply the boundary correctly. You need a short, blunt test you can run in the morning production meeting before anyone signs off on an AI deployment that touches the floor. Run these questions in order, and the first hard no stops the deployment.

One: on its worst day, can this AI's error directly cause a hazardous motion or release a hazardous energy? If yes, it is in the loop of a safety-critical function and the answer is no. Stop here. The deployment does not proceed in that form. If no, continue, because you are in advisory territory and the rest of the test makes it safe rather than deciding whether it is allowed at all.

Two: what hard, deterministic safeguard stands between any AI error and a human body? Name it. A light curtain, an interlock, a two-hand control, a physical guard, a permit-and-verify step with a human signature. If the honest answer is "the AI is the safeguard," you are back at question one with a disguised in-the-loop deployment. There must be something that is not the model holding the safe state.

Three: who is the named, qualified human who verifies the AI's safety-relevant output, and is that verification documented? A drafted procedure, a hazard analysis, an environmental number, a root cause that feeds a corrective action all need a person who owns the check, by name, against the real equipment or data, on the record. No name, no deployment.

Four: was the model validated on the conditions it will actually face, including the worst ones? The opening story failed here: trained on day-shift lighting, deployed against a dark winter morning and a resin change. For any advisory use this question sets your trust calibration, because drift across shifts, materials, and seasons is the rule on a real floor, not the exception, and a model validated only on easy conditions will fail exactly when conditions are hard.

Five: if this AI vanished tomorrow, would the line still be safe? The answer must be an immediate yes. A safe line does not depend on a statistical model for its safety; it depends on engineered safeguards and competent humans, with AI making those humans faster and better informed. If the honest answer is no, the AI has quietly become load-bearing for safety, and that is the condition this entire lesson exists to prevent.

Five questions, two minutes, and they will catch the dangerous deployment before it reaches the line. They encode the whole lesson: keep AI advisory, keep a deterministic safeguard or a competent human between every model error and every human body, document the verification, calibrate trust to the conditions the model was actually tested on, and never let the line's safety come to depend on a guess.

Key Takeaways

• The single hard boundary of floor AI: a human stays accountable for anything that can hurt someone, and a safety-critical control loop, the sense-decide-act chain whose failure can injure a person, is not where AI goes. This is a category difference, not a maturity gap that another training run closes.
• Functional-safety standards (ISO 13849, IEC 62061, IEC 61508) and OSHA require deterministic behavior, validated failure modes, and a quantified probability of dangerous failure. A statistical model offers none of the three natively. "Right ninety-nine percent of the time" is a good inspector and a fatal safety device, because the failures cluster on the unusual conditions you most need handled.
• Sort every proposed use into advisory or in the loop, and ask the load-bearing question: on the AI's worst day, what is the most that can happen to a person? Advisory uses keep a deterministic safeguard or a human between the error and any body, so the worst case is bounded. In-the-loop safety uses are a no, because the worst case is an ambulance.
• The economics are not close. One saved machine cycle is worth cents; OSHA penalties run past sixteen thousand dollars for serious and one hundred sixty-five thousand for willful or repeat, and a single serious recordable injury commonly costs forty thousand to well over a hundred thousand fully loaded, with a disabling injury far beyond that.
• AI helps EHS enormously when kept advisory: drafting lockout-tagout and procedures for a human to verify, mining years of incident and near-miss data into a prevention Pareto, accelerating training for a greener crew, and surfacing alerts a human investigates. It may raise its hand and say "look here." It never decides what happens next.
• OSHA cites you, not the model, and the general duty clause does not transfer to a vendor. Every safety role (the authorized employee who locks out, the competent person who signs the permit, the engineer who specifies the safeguard, the EHS lead) carries a named accountability AI cannot absorb. Document that a named, qualified human reviewed every AI-touched safety artifact against the real equipment.
• On the environmental side the failure mode is the confidently wrong number: an invented emission figure or a mis-converted unit in a report is a false certification to a regulator under penalty, with personal and criminal exposure. The human reads every number back to the source data before signing; the model never becomes the source of truth.
• Run the five-question floor test before any safety-touching deployment: can its error cause hazardous motion or release; what hard safeguard backs it up; who verifies and is it documented; was it validated on the worst real conditions; and would the line still be safe if the AI vanished tomorrow. The first hard no stops the deployment.