AI Agents That Hunt Factory Problems Before They Become Losses

Every manufacturing plant has problems nobody owns. A failing bearing quietly disrupts production schedules, degrades product quality and wastes energy, yet each symptom is observed by a different department that has no way of seeing the full picture. A new perspective article published in Discover Artificial Intelligence argues that this blind spot, one of the most stubborn features of industrial operations, could be attacked with an unlikely tool: teams of large language model agents arranged to mirror the expertise structure of the factory itself. The paper, by Mohammed Amin Adoul of Université de Mostaganem and Abdelhamid Ghoul of Luleå University of Technology, is deliberately framed as a hypothesis rather than a demonstrated result, but its specification is detailed enough to be built, tested and refuted.

The starting point is a well-documented misalignment between what researchers study and what factories need. The most consequential industrial problems, from equipment failures and quality defects to supply chain disruptions, safety incidents and energy waste, are often distributed across departments, so that the true cost of any one of them is invisible to any single unit. Established identification methods struggle with exactly this property. Failure Mode and Effects Analysis, the dominant structured technique in manufacturing, relies on multi-day workshops with expensive cross-functional participation. Lean audits and consultancy engagements share the same weaknesses: they are episodic, dependent on the knowledge of whoever happens to be in the room, and organised along departmental lines. The authors distil four recurring limitations of these approaches, describing them as episodic, expertise-dependent, siloed and qualitatively biased.

The proposed alternative is a conceptual framework in which six domain-specialised agents, covering production, quality, maintenance, supply chain, safety and energy, collaborate through a five-phase protocol. The architecture has three layers: a human interface layer where workers interact with the agents, a collaboration layer where agents communicate through structured protocols on a publish-subscribe message bus, and an output layer producing prioritised problem rankings, a financial impact report and a register of safety-critical items. Each agent would combine a domain-specific system prompt encoding standards such as ISO 55000 for maintenance or ISO 50001 for energy, structured interview templates, short-term working memory, a persistent knowledge base and an interface to the message bus. The authors describe the overall concept as a cognitive organisational twin, a software mirror of the firm’s knowledge structure updated by conversation rather than by telemetry, in deliberate contrast to the digital twin that mirrors physical assets.

The five-phase protocol begins with worker-driven discovery, in which each agent conducts structured interviews within its own department and converts unstructured testimony into standardised problem records containing severity, frequency, estimated financial impact, affected performance indicators, suspected root causes and current workarounds. In the second phase, every discovered problem is broadcast to all agents, each of which scores its relevance to its own domain, assigns an independent severity assessment and flags financial impacts invisible to the originating department. The third phase searches for relationships between problems, distinguishing candidate common causes, candidate causal influences and reported co-occurrences, with the authors insisting that every such link remains a hypothesis for human verification rather than a detected causal relationship. The fourth phase produces consensus-based financial estimates, and the fifth aggregates the agents’ ranked preferences through debate.

What distinguishes the paper from many AI framework proposals is the mathematical specificity of its prioritisation procedure. Agent financial estimates, normalised to expected annualised cost over a twelve-month horizon, are combined by taking the median across the six agents, with an interquartile dispersion ratio used to attach confidence labels: values of 0.25 or below are labelled consistent, up to 0.75 dispersed, and above 0.75 unreliable. Criteria weights are elicited from plant management through the analytic hierarchy process, and the resulting decision matrix is scored by TOPSIS, a standard multi-criteria method, while the agents’ holistic rankings are aggregated by a normalised Borda count. The two components are blended in a combined score whose weighting parameter must be explicitly reported along with a sensitivity analysis. Notably, the framework refuses to compute the familiar FMEA Risk Priority Number, because multiplying ordinal severity, occurrence and detection scores as though they were ratio quantities is precisely the defect the authors identify in the traditional method.

Two safeguards anchor the design. The first is a non-compensatory safety gate: if any of the six severity assessments assigns a score of eight or more, corresponding to potential fatality, permanent injury or major loss of containment, the problem is placed on a mandatory review register and referred for expert safety assessment regardless of where it ranks. The need for this gate emerges starkly from the authors’ worked calculation. In their hypothetical example, a combustible dust hazard with the highest possible severity is ranked only third by the compensatory scoring system, because its rarity drives down its expected annual cost and allows recurring losses to offset it. The financial figure responsible for pushing the hazard down is also the one the procedure itself labels unreliable. The second safeguard is a mandatory human verification gate: no output of the system is a decision, every financial figure must be re-estimated by a human from company data, and any causal claim must be checked against process or maintenance records before action follows.

Because the entire input to the framework comes from conversations between workers and agents, the authors devote unusual attention to the ethics of that interaction. They specify six design requirements: voluntary participation with an unconditional opt-out, no individual attribution in any record, non-punitive use of the data, operation in the languages actually spoken on the shopfloor, access to a human facilitator who is not part of line management, and provenance labels distinguishing what a worker reported from what a model inferred or what was retrieved from company documents. In several European jurisdictions, consultation with employee representation before deployment is a legal precondition. Whether workers will actually speak candidly to such a system, given concerns about job security and prior experience of monitoring technology, is treated as an open empirical question rather than an assumption.

The authors are equally candid about the limits of their own proposal. The most consequential weakness is that six agents built on one base model do not constitute six independent estimators: their errors are correlated by construction, so agreement among them should be read as consistency, never as corroboration. Financial estimates produced by language models are uncalibrated, and no study identified in the review grounds such estimates against realised industrial costs. Hallucination poses a specific hazard here, since a fabricated problem record would be indistinguishable in style from a genuine one, and long-term knowledge bases can propagate and entrench errors. A systematic review cited in the paper, which screened 2341 publications on foundation-model agents in industrial automation, found most systems at technology readiness levels four to six, with genuine deployments in only about nine per cent of cases, and recurring barriers including generalisation failure, hallucination and output instability.

Perhaps the most striking feature of the article is its own worked example turning against the concept it illustrates. In a hypothetical vignette of a Nordic steel plant with 450 employees, the agents would surface a recurring crusher bearing failure, unrecovered furnace heat and an accumulating combustible dust hazard, and would hypothesise a common cause in deferred capital investment, exactly the kind of cross-departmental insight a single-department assessment would miss. Yet the final ranking places the catastrophic hazard third, below routine losses, and only the safety gate and the dispersion label catch it. The authors present this outcome as a property of compensatory aggregation rather than of the chosen figures, and they conclude that the framework’s most honest claim is a negative one: the first question on their fourteen-question research agenda is whether six agents achieve anything a single agent, or a well-run workshop, does not. Until that is answered, the proposal stands as a precisely stated hypothesis about organising AI at the industry-academia interface, one explicit enough to be refuted.

Subject of Research: Large language model multiagent systems for discovering and prioritising industrial manufacturing problems

Article Title: Large language model multiagent systems for industrial problem discovery and prioritization

Article References: Adoul, M. A., & Ghoul, A. (2026). Large language model multiagent systems for industrial problem discovery and prioritization. Discover Artificial Intelligence, 6(1), Article 1456. https://doi.org/10.1007/s44163-026-02468-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02468-6

Keywords: large language models, multiagent systems, manufacturing, FMEA, industrial problem discovery, prioritisation, TOPSIS, Borda count, Industry 5.0, human-AI collaboration, safety gate, cognitive twin