Does automation eliminate manual work or repurpose it?

Grønsund and Aanestad's two-year qualitative case study (September 2017 to October 2019) of a maritime shipbroking firm ("ShipCo") that introduced algorithmic support for geospatial satellite data analysis and prediction finds that the firm's response to automation was not replacement but a "human-in-the-loop pattern" built from two new roles of "augmentation work": auditing — "the generation of a ground truth and assessment of the algorithmic output against this" — and altering — "the work of altering the algorithm and the data acquisition architecture." As the conclusion states, "some tasks, namely auditing and altering the algorithm, were new and constituted the core of augmentation work." The firm's pre-existing role of manual data classification "was not replaced, but repurposed to provide the auditing process with a ground truth against which the algorithmic outcome was compared."

The two roles are, in the authors' terms, "mutually dependent on each another, eventually forming a feedback loop" (their Fig. 6): auditing generates the ground truth that lets algorithmic output be assessed; that assessment surfaces the gaps that altering then acts on, by adjusting the algorithm itself or the data-acquisition architecture feeding it; the altered system produces new output for auditing to assess again. The paper frames this as a response to Markus's (2017) call to "focus on the organizational and environmental conditions that promote or inhibit the effective performance of humans in the control loop," and argues the human-in-the-loop configuration becomes a "strategic capability" precisely because the feedback loop supports learning and keeps algorithmic performance aligned to "the organization's requirements and changes in the environment."

This gives an empirical, pre-LLM precedent for framings the library already holds for generative-AI collaboration. It reads as a concrete instance of How should humans and AI systems divide decision-making tasks?: ShipCo's repurposed classification role is a Type B-leaning retention of human judgment (ground truth generation needs a human), parallel to how How should AI agents and humans divide research tasks? keeps final decisions with humans even as proposal work is delegated — except here the human role is pulled toward auditing and altering rather than toward decisions per se. It also anticipates the vocabulary in How do humans and technology share control over organizational routines?: altering is protocol development that this case keeps human-controlled, which that later taxonomy would classify as an augmenting-technology quadrant, even though Grønsund and Aanestad's fieldwork concluded the year before that taxonomy was published and neither paper cites the other.

The study is a single case at a single firm in one domain — geospatial satellite data analysis and prediction for maritime trade — using interpretive methods (participant observation across 39 meetings, informal interviews, no audio recording, retrospective field notes), so the "strategic capability" claim is this case's interpretation, not a cross-organization finding. The "algorithm" here is a classification and prediction system, not an autonomous agent taking goal-directed action, and the fieldwork ended in October 2019, before generative AI or LLM-based agents existed. The paper does not establish that auditing-and-altering is the pattern organizations will adopt for more autonomous agents; the implication it supports is narrower: when organizations introduce algorithmic automation into analysis work, expect new human roles to form around auditing and altering the system rather than assuming the automation simply removes labor — whether that pattern holds for today's more agentic systems is a separate, open question.