Engaged AI Governance: Addressing the Last Mile Cha... | AI Research

Engaged AI Governance: Addressing the Last Mile Challenge Through Internal Expert Collaboration
This paper investigates the "Last Mile" challenge in AI governance: the difficulty of translating high-level legal requirements, such as those found in the EU AI Act, into the daily, practical workflows of software development teams. While many companies have organizational policies, these often fail to reach the developers building the systems. Through insider action research at an AI startup, the authors demonstrate how collaborative workshops can bridge this gap, turning abstract regulatory requirements into shared, actionable development strategies.

The Last Mile Challenge

Governance often fails because it is treated as an external imposition—a set of rules handed down from management or regulators that developers view as administrative "box-ticking." The authors argue that when governance is disconnected from the actual development process, it risks becoming performative. The "Last Mile" challenge refers to this critical transition point where organizational policy must be integrated into the technical decisions made by engineers every day.

A Pipeline for Action

To address this, the researchers developed a "legal-text-to-action" pipeline. Instead of leaving compliance to legal departments, they engaged the development team directly in a collaborative workshop. This process involved five steps: 1. Preparation: Extracting specific, relevant requirements from the EU AI Act. 2. Presentation: Introducing these requirements to the team. 3. Assessment: Having developers evaluate how these rules relate to their current work. 4. Ideation: Brainstorming practical strategies to meet these requirements. 5. Prioritization: Using a collective evaluation method to decide which actions to implement based on impact and effort.

Patterns of Perception

The research identified three distinct ways developers perceive regulatory requirements: * Convergence: The requirement aligns perfectly with existing development priorities, such as improving system quality. * Existing Practice: The team realizes their current work already satisfies the requirement. * Disconnection: The requirement is viewed as unnecessary administrative overhead.
The study found that practitioners are much more likely to genuinely engage with governance when they understand how it serves the end-user or improves their own development needs. When requirements are seen as purely for verification or compliance, they are often treated as superficial tasks.

The Value of Collaboration

The authors conclude that internal expert collaboration is a powerful tool for making governance visible and collective. By involving developers in the translation process, companies can transform compliance from an external burden into a shared ownership model. This approach does not replace the need for external stakeholder engagement or democratic accountability, but it provides a practical mechanism for ensuring that AI systems are built responsibly from the ground up, rather than just documented that way after the fact.