EXCELYA: G7 Hiroshima AI Process (HAIP) Transparency Report

Excelya does not use one fixed numerical threshold to determine when a severe AI risk becomes unreasonable.

We use a risk-based governance framework. This framework includes AI governance, privacy, quality, security, and human oversight requirements.

We assess each AI system based on:

• Its intended use.
• The type of data that it processes.
• Its potential effect on people and the organization.
• Applicable legal and regulatory requirements.
• The effectiveness of available risk controls and mitigations.

The AI Lab performs technical, privacy, quality, and security reviews.

In addition, all significant AI use cases receive an independent review from the Digital Committee before deployment.

The Digital Committee operates separately from the AI Lab.

The Committee provides an additional level of governance and oversight.

During the review, the Committee evaluates:

• The business need for the AI use case.
• The risk profile of the AI system.
• Regulatory and ethical considerations.
• Data governance requirements.
• Potential effects on users.
• The adequacy of proposed controls.

The Committee must complete its review before approval for production use.

If the Committee identifies concerns, it can require:

• Additional risk mitigations.
• Additional testing.
• Additional documentation.
• Additional monitoring measures.
• Restrictions on the deployment scope.

The organization must implement these actions before authorization is granted.

Governance responsibilities are divided as follows:

1. The AI Lab designs and validates the solution.
2. Business, Data Protection, and Quality Assurance stakeholders review risks in their areas of responsibility.
3. The Digital Committee independently approves, rejects, or challenges deployment decisions.

Our organization deploys AI agents exclusively in the Microsoft ecosystem. This ecosystem includes Microsoft 365 Copilot, Copilot Studio, Azure AI services, and related orchestration services.

We assess the risks of interactions between AI agents. This assessment is part of our AI governance process, architecture review, security review, and production readiness review.

We do not permit unrestricted autonomous interactions between AI agents.

AI agents must operate:

• Within approved workflows.
• Within defined authorization limits.
• Within documented business processes.

We define monitoring requirements during solution design. These requirements remain active during the complete operational life of the solution.

Monitoring activities can include:

• Review of execution logs.
• Review of workflow history records.
• Review of error records.
• Review of conversation logs, when available and enabled.
• Review of test evidence.
• Monitoring of system performance.
• Review of incident management records.

We use this monitoring information to detect issues, verify compliance with requirements, and support continuous improvement of AI solutions.

Our organization uses a defined governance process before deployment of an AI model or system.

The AI Lab designs, develops, tests, and validates the solution.

Business stakeholders, Data Protection, Quality Assurance, and Security representatives review risks in their areas of responsibility.

The Digital Committee performs an independent review before deployment. The Digital Committee has authority to approve, reject, or request additional controls for significant AI use cases.

Before deployment, the project team must provide evidence for:

• Functional performance.
• Security controls.
• Privacy controls.
• System robustness and reliability.
• Compliance with applicable requirements.
• Risk assessment results.
• Test, evaluation, validation, and verification activities.
• Known limitations and residual risks.

Risk evaluation results must be interpreted in the context of the intended use, operating environment, data quality, available controls, and identified limitations. Risk metrics alone do not determine deployment approval.

The following approvals can be required before deployment:

• Business Owner approval.
• Data Protection approval, when applicable.
• Quality Assurance approval, when applicable.
• Digital Committee approval for significant AI use cases.

Before deployment, the organization must implement appropriate safeguards. These safeguards can include:

• Technical and procedural risk mitigations.
• Access controls and authorization limits.
• Human oversight measures.
• Monitoring and logging controls.
• Incident management procedures.
• Usage restrictions and operational guardrails.
• Additional testing and validation activities.

The organization can restrict deployment scope or delay deployment if the identified controls are not sufficient to reduce risk to an acceptable level.

After deployment, the organization continues monitoring activities and reviews the solution regularly to verify compliance, performance, and effectiveness of controls.

a { text-decoration: none; color: #464feb; } tr th, tr td { border: 1px solid #e6e6e6; } tr th { background-color: #f5f5f5; }

After deployment, our organization reviews incident reports from internal and external sources.

We use these reports to identify new risks, recurring issues, and changes in the risk profile of an AI system.

The review team assesses the relevance of each incident to the deployed system and its intended use.

Each identified issue receives a risk score. The risk score determines the required response time and corrective actions.

The organization uses defined remediation timelines that range from 2 days to 30 days, based on the assigned risk score.

If the review identifies a risk, the organization updates the risk assessment and determines the required actions.

Corrective actions can include:

• Additional monitoring.
• Additional testing and validation.
• Updates to documentation.
• Changes to controls or mitigations.
• Restrictions on system use.
• Model replacement or retirement when required.

For example, if a model reaches its expiry date or no longer meets governance requirements, the organization reviews the impact and takes the appropriate action before continued use.

The organization records identified issues and tracks corrective actions through its monitoring and incident management processes.

The organization verifies that corrective actions are completed within the required timeline for the assigned risk score.

Monitoring and incident management activities remain active throughout the operational life of the AI system.

Our organization makes reporting mechanisms available according to applicable legal requirements, internal policies, and the risk level of the reported issue.

Employees are trained to report AI-related incidents via email to an internal AI inbox. Suspected data breaches are reported to IT.

• Reporting to regulatory authorities as relevant

Incident report for serious incidents under the AI Act (High-risk AI systems)

• NIST AI Risk Management Framework (AI RMF)
• Industry-led best practices or codes of conduct

We partner with industry groups such as Drug Information Association (DIA) and MAPS (Medical Affairs Professional Society) to share benchmarks, experience and best practices. We are a member of IASEAI, and collaborate with academic centers (The Open University, KCL...) on specific topics (specifically automation bias and HiTL in medical affair processes). 

Our organization applies risk management during the complete AI system lifecycle.

The AI Lab designs, tests, validates, documents, and monitors AI systems. Business, Quality, Security, and Data Protection stakeholders review risks in their areas of responsibility. The Digital Committee performs an independent review of significant AI use cases before deployment. Significant risks can be escalated to the Digital Committee for additional review and decision.

Our AI systems use Microsoft technologies, including Microsoft 365 Copilot, Copilot Studio, Azure AI services, Power Automate, and Microsoft Graph connectors. We use Microsoft evaluation and observability capabilities to assess quality, grounding, hallucination rates, prompt injection resistance, indirect prompt injection resistance, output safety, content filtering effectiveness, agent tool usage, performance, and reliability. We also use workflow run history, error logs, conversation logs (when enabled), audit records, and operational monitoring data to support governance, traceability, and operational oversight.

Before deployment, each AI system undergoes risk assessment, testing, evaluation, validation, security review, privacy review, and documentation review. Higher-risk systems receive additional controls and oversight. Agentic AI systems receive additional review of autonomy, orchestration logic, authorization limits, tool access, multi-agent interactions, and human oversight mechanisms.

After deployment, the organization monitors evaluation results, system performance, execution logs, workflow histories, error records, conversation logs, and incident reports. Each issue receives a risk score. The risk score determines the escalation path and remediation timeline. Corrective actions must be completed within defined timelines ranging from 2 to 30 days. 

The organization reviews its AI governance practices regularly and whenever significant events occur, which in practice has meant every quarter! Risk assessments, evaluation results, and controls are updated as required

Our organization does not develop or train foundation AI models.

We use third-party AI models and apply governance controls to the data, knowledge sources, testing datasets, and evaluation processes used in our AI solutions.

Before use, datasets and knowledge sources undergo review for quality, relevance, accuracy, completeness, and compliance with applicable privacy requirements.

Personal information is removed or anonymized where required. Data used for testing and evaluation is curated and approved before use.

The organization uses a combination of internal datasets and publicly available reference datasets where appropriate. In the clinical research and clinical trial domain, publicly available scientific, regulatory, and industry-standard data sources can be used to support data quality assessment, benchmark development, and evaluation activities.

Domain experts review datasets and evaluation results to verify accuracy, consistency, and suitability for the intended use case.

Data can be labelled manually by subject matter experts or through controlled data-processing pipelines. The organization performs quality checks to verify the correctness of labels and metadata.

The organization uses Microsoft evaluation capabilities in Azure AI services and Copilot Studio to assess output quality, grounding, hallucination rates, safety, fairness, reliability, and performance. Test datasets are used to evaluate system behavior under representative use conditions.

The organization reviews known limitations of datasets, evaluation datasets, and knowledge sources. Risk assessments consider data coverage, potential bias, domain limitations, and the ability of evaluation results to generalize to real-world use cases.

If testing identifies quality, fairness, safety, privacy, or performance concerns, the organization implements corrective actions before deployment. These actions can include dataset updates, additional review by subject matter experts, changes to prompts or guardrails, additional testing, or restrictions on system use.

The organization maintains monitoring and periodic review activities to verify that data quality, system performance, and risk controls remain effective throughout the operational life of the AI system.

This is shared across IT, the AI_Lab, and selected internal Champions.

• Not applicable

• Not applicable

Our organization applies privacy and data-protection controls throughout the AI system lifecycle.

We use a combination of technical, organizational, and procedural controls to protect data and reduce privacy risks.

Personal information is removed, anonymized, or minimized when required by the intended use case and applicable regulations.

We use role-based access controls to limit access to data, systems, tools, and AI capabilities. Users receive only the access required for their job responsibilities.

We use encryption and secure data-handling practices to protect data during storage and transmission.

We implement guardrails, access restrictions, content controls, and approved skills or tools to prevent unauthorized actions, unauthorized data access, and unintended system behavior.

Before deployment, AI systems undergo privacy, security, and compliance reviews. These reviews verify that data sources, processing activities, access permissions, and controls are appropriate for the intended use case.

• Conduct regular cybersecurity risk and vulnerability assessments tailored to AI models or systems
• Perform threat modeling that includes AI-specific risks (e.g., data poisoning, prompt injection)
• Maintain or update security policies aligned with recognized standards (e.g., NIST CSF, ISO/IEC 27001, SOC 2, CIS Critical Security Controls), and review them regularly in response to emerging risks or regulatory guidance
• Train staff on cybersecurity best practices and AI-specific threats, including how to recognize and respond to misuse or unintended outcomes of advanced AI systems
• Implement incident response, disaster recovery, or business continuity plans, including escalation procedures for high-severity incidents
• Maintain vulnerability management processes (scanning, channels for external reporting, patching, remediation)
• Integrate cybersecurity AI risks into broader cybersecurity risk management activities
• Regularly review security measures and update them in response to emerging risks, incidents, or major releases
• These responsibilities are delegated contractually to model or infrastructure providers

• Apply multi-factor authentication (MFA) and role-based access controls for advanced AI systems
• Encrypt AI assets (datasets, models, configurations) in storage and in transit as appropriate
• Restrict physical access to servers, GPUs, or devices hosting AI models, systems, and/or applications
• Monitor system activity and log access to detect anomalies or misuse
• Conduct secure AI development and/or deployment practices
• Conduct security testing focused on AI-specific risks (e.g., red-teaming, adversarial testing, stress tests)

• These responsibilities are delegated contractually to model or infrastructure providers

• Conduct pre-employment screening or role-sensitive vetting
• Monitor for unauthorized internal access or suspicious activity
• Provide staff training on cybersecurity best practices and AI-specific threats
• Implement insider-risk escalation or response procedures

Cannot provide response for security reasons!

• Human oversight or interruptibility mechanisms (approval requirements for sensitive actions, escalation paths, interruption interfaces)
• Monitoring (anomaly detection, multi-agent interaction monitoring)
• Logging of agent actions
• Constraints on the agent's action space (e.g., tool-use permissions, environment sandboxing, allowlists/blocklists of resources, scope limits per task)

Evaluation results are shared with stakeholders according to the purpose, risk level, and applicable legal or regulatory requirements.

Internally, evaluation results are reviewed by the AI Lab, business owners, Quality Assurance, Information Security, Data Protection, and the Digital Committee when required.

In France, information related to AI deployments is also shared with employee representative bodies when required by applicable legal obligations, including consultation and information processes involving the Comité Social et Économique (CSE).

Evaluation results support deployment decisions, risk assessments, corrective actions, monitoring activities, and governance reviews.

The organization provides both technical and non-technical summaries. Technical reports are used for validation, governance, and audit purposes. Plain-language summaries are used when communicating with business stakeholders, employees, management, and other non-technical audiences.

Selected results, lessons learned, and governance practices can also be shared with customers, partners, professional networks, industry forums, academic collaborators, conference audiences, or regulatory stakeholders when appropriate.

The organization shares information in a manner that protects confidential information, personal data, intellectual property, and applicable regulatory obligations.

Not applicable

Our organization integrates AI risk management into its overall governance framework.

The AI Lab provides technical, operational, and governance expertise. The AI Lab performs risk assessments, testing, validation, documentation, monitoring, and recommendations for AI systems.

Business stakeholders, Data Protection representatives, Quality Assurance personnel, Information Security personnel, and subject matter experts provide advisory review within their areas of responsibility.

The Digital Committee performs an independent governance review of significant AI use cases. The Digital Committee has decision-making authority for approval, rejection, escalation, or the application of additional controls before deployment.

Final accountability for AI risk-related decisions remains with the designated business owner and the organization's governance structure. Deployment cannot proceed without the required reviews and approvals.

AI governance policies, risk assessments, controls, and operating procedures are reviewed on a regular basis and whenever significant events occur, in practice this has been every quarter.

Review and update triggers can include:

• Security incidents.
• Privacy incidents.
• Compliance findings.
• Regulatory or legal changes.
• Significant model or system performance issues.
• Material changes to AI system functionality.
• New AI capabilities or agentic features.
• Newly identified risks.
• Model expiry or retirement events.

When a review identifies new or increased risk, the organization updates the risk assessment and determines the required corrective actions. Corrective actions can include additional testing, additional monitoring, updated controls, deployment restrictions, model replacement, or model retirement.

For higher-risk AI systems, including agentic AI systems, the organization applies additional governance, review, monitoring, and human oversight requirements before and after deployment.

Our organization provides AI literacy and AI governance training to employees and relevant stakeholders. The objective is to ensure safe, responsible, compliant, and effective use of AI systems. Training programs are aligned with AI governance requirements and AI literacy obligations. 

Training methods include:

• Mandatory onboarding and foundational AI training.
• Role-based training modules tailored to business responsibilities.
• Regular awareness and upskilling sessions.
• AI workshops, webinars, and practical learning activities.
• Guidance documents, prompt libraries, and approved-use resources.
• Assessments and certification activities to verify understanding of AI risks and responsibilities. 

Training topics include:

• AI governance and legal responsibilities.
• AI risk awareness and escalation procedures.
• Safe prompting practices.
• AI hallucinations and output verification.
• Data sensitivity, confidentiality, and privacy protection.
• Human oversight and accountability.
• Tool scope, limitations, and approved use.

Employees can raise questions, request guidance, or report concerns through established governance, quality, security, data-protection, and incident-management channels. Issues can be escalated to the AI Lab, relevant governance stakeholders, or the Digital Committee, depending on the nature and severity of the concern.

Training content and governance guidance are reviewed and updated when significant regulatory changes, new AI capabilities, policy updates, or new risk scenarios are identified.

AI-related incidents are managed through the organization's Quality Management System (QMS) and follow the same incident-management process used for other quality, compliance, security, and operational incidents.

Incident records are maintained in the QMS and include the incident description, impact assessment, risk score, investigation records, corrective actions, preventive actions, approvals, and closure status.

The organization performs investigations and root-cause analysis when required. The organization uses incident records to identify trends, evaluate recurring issues, verify the effectiveness of corrective actions, and improve risk controls.

Each incident receives a risk score. The risk score determines the response priority, escalation requirements, and remediation timeline. Defined remediation timelines range from 2 days to 30 days, depending on the severity of the issue.

Significant incidents are escalated through the organization's governance structure. Escalation can involve business owners, Quality Assurance, Information Security, Data Protection representatives, the AI Lab, the Digital Committee, or senior leadership, depending on the nature and severity of the issue.

Where required by applicable legal, regulatory, contractual, or organizational obligations, incidents can also be reported to external authorities or other relevant parties.

The organization tracks all corrective and preventive actions until completion and verifies their effectiveness before incident closure.

• ISO/IEC standards (e.g., ISO/IEC 27001, ISO/IEC 42001)
• NIST AI Risk Management Framework (AI RMF)
• Industry-led best practices or codes of conduct
• OECD AI Principles
• Other international or regional standards (please specify): EU AI ACT

We mostly follow NIST, and ISO, within overarching EU AI Act considerations.

Excelya informs users when they interact with AI systems.

All chatbot solutions include a clear disclaimer that identifies the system as an AI-based system and explains that outputs require appropriate human review.

The organization applies transparency measures during solution design and deployment. Users receive information about the purpose, scope, capabilities, and limitations of the AI system when appropriate.

The organization made an early design decision not to use human avatars or AI-generated human images in situations where users could reasonably mistake an AI system for a human.

Instead, the organization uses a clearly identifiable non-human character, the robot teddy bear BEX, in interfaces, communications, and training materials where there is a risk of confusion between human and AI-generated content.

This approach helps users distinguish AI systems from human users and supports the organization's transparency, trust, and responsible AI objectives.

The organization reviews transparency measures as part of its AI governance process and updates them when new capabilities, new use cases, or new regulatory requirements are introduced.

Excelya invests in research and practical initiatives that support responsible use of AI in regulated environments.

A major area of focus is Human-in-the-Loop (HITL), professional judgement, and human oversight in Medical Affairs, Medical Information, pharmacovigilance, regulatory, and compliance workflows. Our work examines what professionals need in order to exercise meaningful judgement when reviewing AI-assisted outputs, including access to evidence, decision authority, escalation pathways, and the ability to challenge or reject AI-generated content.

We are particularly interested in situations where individuals remain accountable for decisions supported by AI. Current work explores how organizations can prevent oversight from becoming a simple approval step and instead support genuine review, critical thinking, and professional responsibility.

Research topics include professional judgement, cognitive workload, automation bias, trust in AI systems, human-AI collaboration, auditability, and governance. The objective is to develop practical approaches that organizations can use to evaluate whether Human-in-the-Loop controls are effective in real-world regulated workflows.

We intend to share selected findings through scientific and professional channels. Planned outputs include a potential manuscript for Therapeutic Innovation & Regulatory Science (TIRS) and conference presentations or posters at events such as AISEAI. The goal is to contribute practical evidence and governance guidance for organizations that deploy AI in regulated environments.

Excelya supports AI literacy, education, and awareness initiatives for employees and external audiences.

Internally, the organization provides AI training programs, AI literacy modules, workshops, awareness sessions, governance guidance, and practical learning resources. These activities help employees understand the capabilities, limitations, risks, and appropriate use of AI systems. Training topics include AI governance, AI risk awareness, data protection, confidentiality, human oversight, output verification, and responsible use of AI. 

The organization also supports AI education through internal communities and knowledge-sharing initiatives, including AI Champions and incentivized employee engagement programs. These activities encourage responsible AI adoption, practical learning, and discussion of emerging AI risks and opportunities.

In addition, Exelyates participate in community education activities outside the workplace. These activities include discussions and awareness sessions with parents, teachers, and students in preschool and primary school settings. The objective is to help children and caregivers understand the nature, capabilities, limitations, and responsible use of AI technologies.

These initiatives place particular emphasis on AI literacy, critical thinking, human oversight, privacy awareness, and (in)appropriate reliance on AI-generated content. The objective is to improve understanding of AI among both professional users and the wider community.