What are the six pillars of trustworthy AI in industry?
Trustworthy industrial AI rests on six pillars defined by the European Commission and the OECD: robustness, transparency, data governance, human oversight, fairness and accountability. These principles apply to any AI solution deployed in an industrial setting, from failure diagnosis to predictive maintenance.
In an industrial context, where the stakes are high (safety, reliability, cost, compliance), AI can only be genuinely useful if it is trustworthy. This is the approach taken by industrial intelligence platforms such as Mimorian, which models equipment, structures failure diagnosis and captures the know-how of field maintenance teams through a multi-agent AI architecture. But what do we mean by "trustworthy AI"? Several international frameworks (notably the Ethics Guidelines for Trustworthy AI from the European Commission, updated by the AI Act, and the OECD principles) give a clear view of what this involves. Here are the six essential pillars, what they mean in practice, and why they matter for businesses.
1. Robustness & technical safety
Robustness means that AI withstands not only normal use, but also edge cases, failures, attacks or manipulation. This covers accuracy, reliability, redundancy, the ability to "fail safe" (fallback), and the reproducibility of results. Stratégie numérique de l'UE+2OECD+2 According to Siemens, unplanned downtime costs the world's 500 largest companies 1.4 trillion dollars a year, underlining how critical robust AI systems are for failure prevention.
For a factory or a workshop, this means that a diagnosis or prediction system must not "run away" because conditions change (dust, temperature, wear, and so on). Contingency plans or a safe shutdown must therefore be in place. Otherwise, the risk of breakdown or accident is real, with its financial and human consequences.
2. Transparency & explainability
This pillar requires that the system's decisions, processes and limits are understandable to the people concerned: engineers, operators, maintenance teams. This means being able to trace where the data comes from, explain how the AI reaches its conclusions, and flag any uncertainties. Stratégie numérique de l'UE+2arXiv+2
In practice, in a maintenance context, this means that technicians know why the AI flags an anomaly, on what basis, with what level of confidence, rather than a "black box" warning with no explanation.
Explainability also means owning uncertainty. Trustworthy AI is not perfect AI, it communicates its grey areas rather than hiding them.
3. Privacy & data governance
Data is the fuel of AI, but its collection, storage, processing and sharing must respect the rules, not only legal ones (GDPR, sector-specific standards), but also those of quality, integrity and traceability. Data governance covers the origin of the data, its freshness, its representativeness, and legitimate access to it. Stratégie numérique de l'UE+2Centrum für europäische Politik: cep.eu+2
In industry, this translates into: reliable machine data, calibrated sensors, cleaning and correction processes, transparent data histories, documented transformations. Otherwise, the model can drift, produce false alerts, or worse, make poor decisions.
4. Human agency & oversight
No AI should operate in a fully automatic mode without the possibility of human intervention. "Human agency" means that the user must be able to understand, correct and stop the AI if needed. Oversight ("human oversight") guarantees human control mechanisms: during development (testing), at the point of use (monitoring), and after use (audit). Stratégie numérique de l'UE+1
In a workshop, this shows in clearly defined roles: who validates alerts, which operator can cancel an automatic action, how field feedback feeds improvement.
5. Fairness, non-discrimination & inclusion
Trustworthy AI must avoid biases (in the data, in the uses) that could lead to unfair or discriminatory outcomes. It is also about designing for all users, managing differences (abilities, context, culture) so that the system is accessible and relevant. Stratégie numérique de l'UE+2Centrum für europäische Politik: cep.eu+2
In practice: training data with representativeness, cross-testing, feedback from field operators in different contexts, equal access to tools, an interface usable by everyone.
6. Accountability & societal / environmental wellbeing
This pillar covers several dimensions: the legal and moral accountability of the actors who develop and deploy AI, auditability, the ability to correct errors, and redress mechanisms for those affected. And of course, accounting for the direct and indirect effects on society (employment, safety, civil rights) and on the environment (carbon footprint, energy consumption, sustainability). Stratégie numérique de l'UE+1
For the industrial company, this means: anticipating the social impacts when choosing technologies (for example, automation), carrying out technical as well as societal assessments, documenting responsibilities, ensuring that safeguards are in place in the event of an incident, and that the choices made respect sustainability.
How can these pillars be built into an operational strategy?
These six pillars are not abstract concepts, but actionable levers. According to McKinsey, 78% of organisations now use AI in at least one business function (compared with 55% a year earlier) [Source: McKinsey, State of AI 2024], a figure that rose to 88% in 2025 [Source: McKinsey, State of AI 2025]. The urgency of adopting a trustworthy approach from deployment onwards is no longer up for debate. Here is how businesses can build them in:
- From the design phase, define the requirements for each pillar: robustness, transparency, and so on.
- Put in place indicators and metrics (robustness testing, audits, bias scoring, performance monitoring with varying conditions).
- Develop governance processes around the data: who collects it, who validates it, who shares it, how versions are tracked.
- Train the field teams (maintenance, operators) to understand what the AI does, its limits, how to respond, and how to report alerts or anomalies.
- Plan for audit & field feedback mechanisms to continuously improve the system, adjust parameters, and correct biases.
- Account for legal obligations and social expectations: regulatory compliance, transparency towards stakeholders, environmental impact.
This is the approach Mimorian follows: transparency translates into assumptions that are traceable back to their source (documentation, history, relational graph). Human oversight is built into the process: the technician validates, invalidates or sets aside each lead at any time. Data governance improves naturally, since every intervention produces a structured intervention report, generated automatically from the voice exchange.
Why do these pillars really matter?
For an industrial company, failing to respect one or more of these pillars can translate into:
- High risks of failures, operational errors or costly poor decisions.
- Loss of trust among technical teams, operators, and even customers, holding back adoption.
- Risks of legal or regulatory non-compliance, which can lead to penalties or recall obligations.
- Negative impacts on reputation or over the long term (unsustainable uses, adverse social effects).
- Sub-optimal performance: if the AI works well only in ideal conditions or if users reject it, the ROI fails to materialise.
For an overview of the topic, see our complete guide: What is trustworthy AI in industry? A complete guide for maintenance. For the detailed regulatory dimension (AI Act timeline 2026-2027 after the Digital Omnibus, sector-specific obligations), see our complete guide to the AI Act and industrial maintenance.
📚 Sources :