How can splitting AI accountability across roles keep models answerable?

Asatiani, Malo, Nagbøl, Penttinen, Rinta-Kahila, and Salovaara (2021) report that Denmark's Business Authority (DBA), a roughly 700-employee government entity that maintains the VIRK business register and checks "about 292,000 annual statements per year" for errors and fraud, governs its machine-learning projects by splitting accountability across four distinct roles rather than concentrating it in one team. The in-house Machine Learning Lab, established in 2017, builds prototypes; "domain experts (hereafter 'case workers')" supply domain knowledge during development and are separately "charged with evaluating each model and with its retraining" once deployed; a DBA steering committee "decides which models will enter production use and when"; and an external implementation unit, brought in because DBA culture expects "vendors take responsibility for the support and maintenance functions related to their code," is "accountable for implementing the model and overseeing its technical maintenance."

The authors present this as a deliberate choice, not an accident of org-chart history: DBA chose to hire its own data engineers and data scientists rather than rely on consultants throughout, for "cost-management concerns and a desire to retain relevant knowledge within the agency," but then offloaded production deployment to outside vendors once a model cleared the proof-of-concept stage. The ML Lab's "role is restricted largely to experimentation and development," and a model only reaches the steering committee's decision point once "the proof of concept and documentation such as the evaluation plan" exist. The result is that no single role both builds a model and decides whether it goes live, and the role that answers for a model's day-to-day "operational correctness" (case workers) is organizationally distinct from the role that answers for its upkeep (the implementation unit).

This gives a second, contrasting instance to set against How should AI agents and humans divide research tasks?: that paper describes an agentic model proposing methods and implementing revisions while humans keep "most final decisions," a two-way split between an acting AI and a deciding human. DBA's split instead distributes decision rights among multiple humans around a non-agentic classifier, with no step where the model itself proposes or decides anything — the accountability structure is entirely among people, and the AI is the object being evaluated, approved, and maintained. It also complicates the mechanism in Does labeling AI as an employee change how managers oversee it?, which finds accountability drifting toward an "AI employee" once org charts list it as such; DBA's structure keeps the AI off the accountability chain entirely by naming the case workers, not the model, as answerable for operational correctness.

The excerpt describes one public-sector agency's arrangement, reached through "purposive sampling" of a single case precisely because it was expected to be "an empirically rich setting," and the authors flag that they "could not examine... long-term implications" of the structure. Whether this four-way split generalizes to private firms "driven by differently weighted objectives," as the authors themselves wonder, or to models with more autonomy than DBA's document-checking classifiers, is left open. What the case does show is a concrete, rather than merely argued-for, division of labor in which accountability for an inscrutable model is distributed by stage — develop, evaluate, approve, maintain — among named human roles.