Designing a methodology: Responsible AI by Design (2019) 

Implementing AI principles in an organization requires a methodology. We have designed the “Responsible use of AI by Design” methodology to facilitate this process. The main ingredients of this methodology are: 

The AI principles An awareness and training program A questionnaire to assess the different AI applications Technical tools such as for explainability, fairness and environmental impact. A governance model  

The first step to bring the methodology to life was to create an online course on AI Ethics, explaining the basics of AI along with basic ethical questions that may arise when using the technology. So far, around 4000 employees have finished the course.  

Experimenting with how to implement the methodology (2020-2021) 

The second step was creating a questionnaire to facilitate performing ethical assessments of the products and services of the company. The first version consisted of a small set of question for each ethical principle. In a second version, we added a risk assessment phase where we asked product owners to assess the impact in case the AI system made mistakes in terms of severity, scale, and probability of occurrence. In collaboration with the Environment, Society, and Governance (ESG) department, this version was trialed in one of our main business operations in Spain by including it in the formal product development process. In a third version, we also included the future requirements of the European Union’s AI regulation, the AI Act.  

Running pilots to confirm and adjust the methodology (2022) 

Once we confirmed the questionnaire from a content and usability perspective, we designed an AI governance model that we wanted to test in several business units. Our governance model introduced three new organizational roles: 

Responsible AI (RAI) Champions: a diverse and multidisciplinary group in charge of fostering the responsible use of AI in the business areas they represent, providing ethical guidance to the product owners, proposing measures to mitigate possible risks, and ensuring training and awareness raising on the topic.  An AI Ethics “Committee”: a multidisciplinary group of experts in AI where ethical risks can be escalated for guidance and/or decision-making.  An AI “Office”: a small team responsible for leading the change management process needed for the responsible use of AI in the business units where AI systems are developed, used, acquired, or marketed. The Office coordinates and/or executes the necessary risk assessments, and monitors the implementation of the requirements. The team also coordinates the activity of the RAI Champions and the AI Ethics Committee.  

The governance model also introduced a workflow, and assigned responsibilities in the workflow to different roles and/or departments, such as the product owner, the RAI champion, and the ethics committee. Moreover, we designed a problem-based escalation process. AI use cases that were assessed as high-risk, or had disagreements in the assessments, were escalated to the ethics committee for debate and recommendations. If the debate was not conclusive, final decisions were supposed to be taken by the Office of Responsible Business. From this point, we strengthened our collaboration with the Compliance department via weekly meetings.  

After running pilots in four different global business units (Innovation, Human Resources, Enterprises, CDO), we extracted many learnings, some of which include: 

The importance of having a clear and well-known definition of AI. The need to identify new essential roles to drive the cultural change implied in the process, such as the RAI Champions, the Ethical AI Committee, and an “AI Office” that drives the entire program. The AI value chain is more complex than simply providers and users, and a thorough review of the different types of roles a company can have and the responsibility it entails is necessary. The roles include AI providers, deployers, partners, buyers, etc.  Foster a culture of responsible design for instance by creating guides and training programs for the different profiles in the AI value chain. This culture has motivated employees working with AI to develop a sensibility for ethical thinking; to develop technical tools to mitigate specific risks (bias, black box algorithms); and bring the consumer voice into our AI product development processes.  

Approving an internal AI governance regulation (2023) 

Equipped with the practical experience of the governance model obtained in the pilots, and in close collaboration with the Compliance area, we created the official AI governance model of Telefónica which was approved in December 2023. This final governance model also includes the requirements stemming from the European Union’s AI Act, UNESCO’s Recommendation on the Ethics of AI, and also from some other relevant initiatives such as the US National Institute of Standards and Technology (NIST) and the OECD.