Artificial intelligence is rapidly becoming central to business operations. It automates processes, assists employees, accelerates research, optimises infrastructure and contributes to increasingly important decisions. Yet, as its capabilities grow, one question becomes impossible to ignore: Can we truly trust the AI systems we deploy?
The answer cannot be based on performance alone. A system may deliver excellent results during a demonstration, then underperform when exposed to real-world data, change its behaviour after an update or generate an answer that appears credible but is incorrect. Trust must therefore be designed, measured and documented.
This is precisely the ambition of Trustworthy AI methodologies & technics.
High-performing AI is not enough
Trustworthy AI refers to artificial intelligence systems that are designed and operated in a lawful, ethical, and technically robust manner. According to OECD AI Principles, trustworthy AI should respect human rights and democratic values, be transparent and accountable, remain robust, secure and safe throughout its lifecycle, and contribute to inclusive growth, sustainable development, and human well-being.
Trustworthy AI is therefore neither a single technology nor a universal certification. It is a comprehensive approach designed to ensure that an AI system operates reliably, transparently, securely and under appropriate control throughout its entire lifecycle.
This approach applies to the complete system, not just the model. It includes the data, the model itself, orchestration mechanisms, user interfaces, infrastructure, tools accessible to the AI and the human procedures surrounding its use. The same model integrated into two different architectures can therefore present very different levels of risk.
Trust can be structured around six key dimensions.
-
Explainability makes it possible to understand why a decision or recommendation was produced. This does not necessarily mean exposing every internal mechanism of a model. Instead, it means giving users enough information to interpret a result, identify its limitations and make an informed decision.
-
Transparency and traceability make it possible to reconstruct the path that led to a decision, including the origin of the data, the model version, the parameters used, intermediate processing steps and the final action. This audit trail is essential for investigating an incident, responding to a customer or demonstrating that the system remains under control.
-
Robustness measures the AI system’s ability to remain reliable when conditions change. Data evolves, usage patterns drift and unexpected situations arise. Organisations must therefore define the operational domain in which the system has been validated, monitor changes over time and provide fallback mechanisms when the system operates outside that domain.
-
Fairness and bias management seek to prevent AI systems from reproducing or amplifying imbalances contained in the data. The objective is not simply to monitor generic indicators, but to understand which forms of fairness matter in each specific context and what trade-offs they may involve.
-
Data protection and cybersecurity cover data, models, infrastructure and, increasingly, AI agents themselves. They require control over identities and access rights, as well as protection against prompt injection, data poisoning, model extraction and unauthorised tool use.
-
Finally, human oversight ensures that people retain a genuine ability to understand, intervene and arbitrate. However, simply adding “human in the loop” to an architecture diagram is not enough. Operators must have the time, information and interface required to detect an error and respond effectively.
Another dimension - Sustainability is key in Bull strategy, but we intentionally do not cover it here, as it is covered as a stream on its own as part of Bull’s Corporate and Social Responsibility policy.
Why should every company care
Trustworthy AI is often seen as a topic primarily relevant to regulated industries such as healthcare, financial services, transportation, defence or the public sector. These industries are naturally highly exposed because an AI failure can have significant human, financial or regulatory consequences.
However, companies operating outside regulated industries are also directly concerned.
A customer service agent that invents a refund policy creates a commercial and contractual risk. An assistant processing confidential documents may expose trade secrets. A poorly evaluated recruitment tool may introduce discrimination. A model silently updated by a supplier may change the behaviour of a product without the company being informed.
Customers and partners are also passing their own requirements down to suppliers. Questions about data provenance, processing location, reversibility, version control and auditability are increasingly appearing in requests for proposals, procurement processes and security questionnaires.
Furthermore, building trust into AI systems from the outset is far less costly than reconstructing it afterwards. Data lineage, evaluation results and architecture decisions are relatively easy to document while the work is taking place. Several months later, when teams have changed and intermediate artefacts have disappeared, rebuilding this evidence becomes expensive and sometimes impossible.
The challenge therefore extends well beyond regulatory compliance. It concerns user adoption, reputation, operational resilience, intellectual property protection and the ability to bring AI-powered solutions to market sustainably.
From general principles to practical engineering
For several years, Trustworthy AI was primarily described through broad principles such as transparency, fairness, robustness, accountability and human oversight. These principles were necessary, but they offered limited practical guidance to project teams trying to determine which tests to perform, which metrics to use and which evidence to retain.
This is now changing. Just in Europe, regulations like GDPR, the EU AI Act, NIS2/DORA or CRA (and similar regulations exist in other regions such as Asia and Latin America), standards like ISO/IEC 42001 (AI management systems), ISO/IEC 23894 (AI risk management) , emerging AI Act standards created by CEN-CENELEC JTC21, and sector-specific frameworks and regulations (e.g for aerospace, healthcare, automotive or financial services) are gradually transforming broad principles into concrete requirements for engineering, governance and evaluation.
It is no longer sufficient to claim that an AI system is responsible. Organisations must be able to provide evidence supporting that claim.
The response must nevertheless remain proportionate and tuned for the use-case in addition to the industry sector specifics. An internal AI assistant does not require the same assurance framework as a medical system or an industrial control function. As well, since regulations such as the AI Act apply by use case rather than by industry: a recruitment tool or a system assessing access to services falls under the same obligations whether it is deployed by a hospital or a retailer.
The right approach is to identify actual risks, select the relevant controls and explicitly document the decisions that have been made.
Bull’s experience: connecting research, engineering and real-world use cases
Bull has been actively involved in this transformation for several years. Its contribution to the French Confiance.ai programme, launched in 2021, helped shape its approach to Trustworthy AI. This ecosystem brought together industrial organisations, research institutes and technology companies to bridge the gap between scientific research and operational constraints.
Since then, Bull’s R&D teams and customer-facing teams have pursued four complementary objectives:
-
Defining the trust dimensions applicable to Bull’s AI solutions
-
Continuously assessing relevant techniques and technologies
-
Integrating these methods into products and projects when they create tangible value
-
Translating advances in Trustworthy AI into practical responses to business challenges
This experience is supported by a diverse portfolio of projects, including medical decision support, the use of AI to combat human trafficking, maritime surveillance, regulatory compliance, predictive maintenance for computing infrastructure and intelligent assistance for industrial technicians.
These initiatives highlight an essential reality: priorities are never exactly the same from one sector or use case to another. In some situations, explainability is critical. In others, robustness, confidentiality, cybersecurity or model version control become the primary concern.
Let‘s define the priorities for trust together
There is no universal Trustworthy AI roadmap. Each organisation must define its own trajectory according to its use cases, data, customers and the potential consequences of an AI system producing an incorrect result.
This is why Bull wants to engage with companies that develop, integrate or use AI systems.
Which use cases do you consider the most sensitive? Where are you currently encountering the greatest difficulties: evaluation, explainability, security, governance, architecture choices, supplier control or human oversight? What evidence are your customers and partners beginning to request
The objective is not to impose yet another framework. It is to jointly develop an operational understanding of AI risks, identify shared priorities and determine which methods, tools and architectures can enable AI to be deployed with confidence.
Because the next stage of artificial intelligence increasing systems autonomy will not be defined solely by more powerful systems. It will be defined by our ability to demonstrate their value, control their behaviour and ensure that their use remains sustainable, responsible and widely accepted. With great power comes great responsibility.
