Artificial intelligence can turn fragmented project information into structured working material, but machinery safety is not a text-generation exercise. Under ISO 12100, conclusions must reflect the actual machine, its users, operating modes, life phases and foreseeable interventions. Plant managers and engineering teams therefore need clear controls over where AI contributes, where its output must be challenged and who makes the final decision.
The useful role: preparing material for engineering review
A machine risk assessment often begins with information dispersed across specifications, layouts, control narratives, operating instructions and discussions with production staff. AI can help organise this material into proposed descriptions of intended use, user groups, operating modes and machine limits. It can also expose gaps by asking whether the available information covers installation, commissioning, operation, adjustment, cleaning, fault finding, maintenance and decommissioning. This is valuable preparation, but it is not evidence that every relevant condition has been captured.
The practical purpose of an is to produce an editable draft rather than an approved conclusion. Used carefully, such a tool can convert notes into consistent language, suggest questions for the team and reduce the time spent staring at an empty form. The engineer must still compare every proposal with drawings, software functions, installation conditions, tooling, materials, energy sources and observed working practices. A polished sentence that does not describe the real machine remains technically wrong.
ISO 12100 requires scenarios, not generic hazard labels
ISO 12100 calls for a structured process: determine machine limits, identify hazards, estimate and evaluate risk, apply risk-reduction measures and reconsider the remaining risk. Entries such as “moving parts”, “electricity” or “hot surface” identify possible hazard sources, but they do not explain exposure. A useful scenario states who is exposed, what task is being performed, in which life phase and zone, what hazardous situation can arise, which event may lead to harm and what that harm could be.
An AI system may draft scenarios by combining known tasks, zones and hazard sources. The AI-supported machine risk assessment module can therefore help a review team test whether descriptions are complete and use terminology consistently. It cannot see an unexpected access route, hear an unstable bearing, observe an operator bypassing a sensor or know that maintenance personnel routinely enter a space omitted from the layout. Those facts require inspection, consultation and competent engineering judgement.
Defining limits without allowing assumptions to harden into facts
A well-controlled assistant for machinery risk assessment can propose categories for use, space, time and environmental limits. It may prompt the team to consider expected competence, production materials, operating modes, access points, interfaces with adjacent equipment, service life, component wear, dust, temperature or utility conditions. Its output should be treated as a hypothesis to verify against the machine and its planned environment.
Reviewers should distinguish supplied facts from generated assumptions. If a prompt does not contain the maximum workpiece mass, an AI-generated value must not be accepted simply because it appears plausible. If maintenance intervals are not established, the draft should identify the missing decision rather than invent one. The same discipline applies to foreseeable misuse: suggestions may help start a workshop, but designers, operators, maintenance staff and safety specialists must decide which behaviours are reasonably foreseeable in the actual setting.
Risk estimation and protective measures remain engineering decisions
AI may help check whether the chosen risk-estimation method has been applied consistently, but it should not turn a score, colour or category into a declaration that the machine is safe. Severity, exposure, probability and avoidance assumptions depend on the defined scenario and available evidence. Small changes to access frequency, speed, stopping performance or the possibility of escape can materially alter the analysis. A competent reviewer must examine those assumptions rather than accept a generated rating.
The same boundary applies to risk reduction. ISO 12100 gives priority to inherently safe design measures, followed by safeguarding and complementary protective measures, then information for use. AI can prompt that sequence and identify where a proposed measure is described too vaguely. It cannot validate a safety function, determine the achieved performance level under ISO 13849 or confirm that a guard works in every relevant mode. Calculations, circuit review, testing and validation remain separate technical activities.
Review controls that make AI assistance defensible
A controlled workflow should require a named reviewer to accept, edit or reject each proposal. Generated material should not move directly into an approved assessment. The reviewer needs access to the source information behind the entry and should record unresolved questions, assumptions and required verification. For complex lines, discipline-specific review is particularly important: mechanical, electrical, controls, process and maintenance specialists may each identify conditions invisible to the others.
Change control matters as much as initial approval. A useful record preserves the previous and revised state, the person making the change, the date and the technical reason. Maintaining a technical decision history with the full context of changes helps teams understand why a scenario was closed, reopened or returned for further risk reduction. This trace is especially important after software modifications, retrofit work, tooling changes or altered production materials, when an earlier conclusion may no longer apply.
Residual risk must be described, communicated and reconsidered
After protective measures are applied, the team must reassess the scenario and identify what remains. AI can help keep the description connected to the original hazard, exposed person and task, avoiding entries that reduce residual risk to a colour or isolated label. It can also flag cases where a measure appears in the documentation without a clear relationship to the risk it is intended to reduce.
The final judgement, however, belongs to the responsible engineering team and ultimately does not transfer the manufacturer’s responsibility to software. Reviewers must decide whether further design improvement is reasonably practicable, whether safeguards have been verified and whether remaining risks require information, training or other measures. AI can improve structure and consistency; it cannot approve residual risk or provide legal certainty.
A practical governance rule for manufacturing teams
Manufacturers should use AI where the work is linguistic, organisational and diagnostic: structuring project data, drafting descriptions, proposing questions and finding missing scenario elements. They should stop automation at the point where a decision depends on physical reality, specialist calculation, test results, applicable type-B or type-C standards, or acceptance of remaining risk.
The operating rule is straightforward: AI prepares material; competent people verify facts, challenge assumptions and decide. With that division of work, the technology can shorten the route from scattered information to a reviewable assessment without skipping the reasoning demanded by ISO 12100. Without it, faster documentation may merely conceal weak analysis behind more fluent text.











