AI Law, Risk & Governance

Responsible AI: Principles That Work in Day-to-Day Operations

Marin Niehues17 February 202511 min read

Handshake over a table with documents

Responsible AI works in day-to-day operations when every principle has an owner and a fixed place in the workflow. The term stands for developing and operating AI according to verifiable principles. Closely related are trustworthy AI, the guiding term of the EU, and AI ethics, which asks about the values behind it; AI governance is the control system that enforces all of this within the company. Whether guidelines take effect is decided at the point where a business unit wants to introduce an application and someone has to say whether it is approved. This article describes how executive and divisional leadership turn principles into a delivery logic. This article provides an overview and does not constitute legal advice.

Why principles alone have no effect

A sentence such as “We use AI fairly and transparently” describes an attitude. It answers none of the questions that actually arise in a project. Who checks which data a provider used to develop its model? Above what error rate does an assistant not go into operation? Who may override a result of the system, and who is informed? Where these answers are missing, every team decides at its own discretion, and the company ends up with as many interpretations of the guideline as it has projects.

From a delivery perspective, Responsible AI is therefore a task for the organization. It requires decision rights and review criteria, together with a rhythm in which both are applied. The content of the principles is largely settled, and the major frameworks agree at their core. The difficulty lies in translating them into operations, and that translation is management work.

Which Responsible AI frameworks exist?

Five frameworks matter for companies in the EU, and only the EU AI Act among them is binding. The others provide orientation or allow voluntary certification. All share a small core of principles; they differ in purpose and in how binding they are.

  • The OECD AI Principles, adopted in 2019, are the first intergovernmental standard on AI. Among other things they call for transparency and robustness, together with accountability for outcomes, and serve as a reference point for many national strategies.
  • The Ethics Guidelines for Trustworthy AI by the European Commission's High-Level Expert Group, published in 2019, coined the term trustworthy AI and shaped the logic of the later AI Act.
  • The AI Risk Management Framework of the US National Institute of Standards and Technology (NIST) organizes the work into four functions (Govern, Map, Measure, Manage). It is voluntary and works well as a structure for a company's own risk management.
  • The standard ISO/IEC 42001 describes a management system for AI on the pattern of ISO 27001 and is certifiable.
  • The EU AI Act sets the binding framework with obligations by risk class, ranging from transparency obligations to comprehensive requirements for high-risk systems.

Many companies also formulate their own frameworks that tailor these sources to their business. A company framework makes sense as long as it is aligned with the AI Act and does not become a parallel world beside it. The article on the EU AI Act sets out the obligations in detail.

What is trustworthy AI?

Trustworthy AI is the EU's guiding term for AI that is lawful and ethical and technically robust. The expert group's Ethics Guidelines translate it into seven requirements:

  • human agency and oversight
  • technical robustness and safety
  • privacy and data governance
  • transparency
  • diversity, non-discrimination and fairness
  • societal and environmental well-being
  • accountability

Responsible AI describes the same goal from the perspective of the organization that implements it. In practice, the seven requirements can be condensed into five testable principles.

Which principles belong to Responsible AI?

Responsible AI comprises five principles: transparency and fairness, data protection with information security, human oversight, as well as reliability. Each receives a review question that can be answered in a project, and an owner.

  • Transparency: employees and affected persons know when they are interacting with AI; for chatbots and voice assistants as well as for AI-generated content, the AI Act makes this an obligation. Review question: is the application labeled, and is there a short record of purpose and data basis together with known limitations? Owner: the product team of the application.
  • Fairness: results do not systematically disadvantage any group. Review question: are there test cases for the relevant groups, and were they repeated after the last change to the model or data? Owner: the business unit together with the data team.
  • Data protection and information security: sensitive information does not leave its intended scope. Review question: which data class may enter this system under the data classification? Owner: data protection and information security.
  • Human oversight: for decisions with consequences for people, a person remains responsible, as the GDPR and the AI Act require for automated individual decisions and high-risk systems. Review question: who checks results, and how do they override them? Owner: the head of the business unit.
  • Reliability: in operation the application delivers the quality the business case assumes. Review question: which quality threshold applies before launch, and who monitors it afterwards? Owner: the operations team of the application.

Oversight has to be described before anyone can exercise it, and a label has to be built in by a team before it exists. How the committees for this are set up is covered in the article on building AI governance.

The delivery path: from principle to approval

The most effective lever is an approval path that works in proportion to risk. An assistant that summarizes internal documents needs a different depth of review from a system that pre-sorts job applications or proposes terms to customers. Treating everything the same slows down uncritical applications and reviews critical ones too superficially. A path with four stations has proven effective.

The first station is classification. When the idea arises, the business unit answers a few questions: does the application affect decisions about people? Does it process confidential data? Does it have an external effect? Does it fall into a category of the AI Act? The answers determine the risk class, and the class determines the depth of review.

The second station is review. Depending on the class, data protection and information security assess the application together with the business unit and legal, using fixed review questions. The review runs in parallel with the development of the prototype so that it creates no waiting time of its own. Where the review touches legal questions, it belongs with the responsible legal advisers.

The third station is approval. A named owner in the business unit decides, the decision is documented, and conditions receive due dates. Decision rights follow risk: divisional leadership approves applications in the low class, while executive management decides on high classes.

The fourth station is operation. Quality and usage are reported at a fixed cadence, incidents immediately. The risk assessment thus stays current. The article on AI risk management describes the framework for this.

The application file also serves as evidence for auditors and supervisory authorities. A company that maintains it consistently already holds most of the documentation that certification under ISO/IEC 42001 requires.

Metrics that show executive management the effect

Responsible AI becomes measurable as soon as the approval path exists. Four metrics show whether the principles are taking hold in daily work. The share of applications with completed classification shows whether the path is used. The lead time for approval shows whether it leaves the business units their pace. The number of incidents and the time to resolve them describe quality in operation. Finally, the share of applications with named and actually exercised oversight shows whether human control is more than an entry on a form.

These metrics belong in the regular AI reporting to executive management. A company that knows the lead time for approval reduces it deliberately. A company that does not know it experiences approval as an obstacle and works around it. The link to the employee level is provided by the AI policy for employees, which turns the principles into understandable rules of conduct for daily work. How Responsible AI fits into the overall governance of AI is shown in the article AI management: what executive leadership must steer.

Niehues Consulting as a partner for implementation

Niehues Consulting supports companies in building Responsible AI as part of their delivery capability. In AI management this includes the selection and sequencing of applications and the steering of the AI portfolio with metrics that incorporate the criteria of responsible AI from the outset. In AI prototypes, operationalisation and rollout, the approval path and responsibilities are developed together with the business units, along with user enablement, until the applications are accepted in operation. In delivery management and organizational design, roles and decision paths are shaped so that the organization delivers reliably even under time pressure.

The approach follows the Niehues Delivery Model with the phases Understand, Build and Roll out. For Responsible AI, three of its review questions matter most. In Understand: who decides, who delivers, who uses? In Build: does the prototype prove its value in the real workflow with real data? In Roll out: who is accountable for operation from day one? Answering these questions for every application brings the principles into the workflow. The entry point is a free initial consultation of 30 minutes via the contact page.

Frequently asked questions

What does Responsible AI mean?

Responsible AI means developing and operating AI according to verifiable principles such as transparency or human oversight. Within a company, the term is put into practice through fixed procedures with named responsibilities. It is closely related to the EU's concept of trustworthy AI.

What is the difference between Responsible AI, AI ethics and AI governance?

AI ethics asks about values, Responsible AI translates them into principles for development and operation, and AI governance is the control system of committees and roles that enforces these principles. The three terms build on one another. For executive management, governance is the lever, because it defines decision rights.

Who is responsible for Responsible AI in a company?

Overall responsibility lies with executive management. For each application, a named owner in the relevant business unit is accountable, supported by data protection, information security and legal advisers. A central team coordinates standards and metrics.

How do you implement Responsible AI in practice?

The starting point is classifying all applications by risk. This is followed by review questions per risk class and a documented approval path with regular reporting. The applications with the highest risk are addressed first, the rest afterwards.

What is ISO/IEC 42001, and is certification worth it?

ISO/IEC 42001 is the international standard for an AI management system, and certification pays off when customers or tenders require verifiable evidence. This applies particularly to companies that build AI into products for others or are already certified to ISO 27001, because the two management systems can be combined. Companies that use AI only internally use the standard as a structure and forgo the certificate.

Does a company need its own AI policy?

Yes. There is no explicit obligation to have a policy, yet the AI Act requires every company that uses AI to ensure sufficient AI literacy among its staff. A policy combined with training is the most direct way to demonstrate this.

How do you test AI for fairness and bias?

Fairness is tested with test cases for the relevant groups that run before approval and after every change to the model or data. Error rates and outcomes are compared between the groups. The business unit defines which groups are relevant and which deviation is unacceptable.

Let’s talk about your initiative.

In 30 minutes we clarify where your organization stands and which step will pay off most now.

Request a consultation

30 minutes · free · no obligation