This talk examines how human mediation is critical to sustaining algorithmic management in employee-based, on-demand work. Through a case study of a public transportation service, we show that algorithmic systems designed for optimization and control often generate tensions with human-centered work realities. We identify four domains of conflict: work allocation and pacing, information and communication, work organization, and performance assessment. Managers and operators actively mediate these tensions by buffering drivers from algorithmic pressure, interpreting opaque directives, fostering professional communities, and contextualizing evaluations. Our findings shift the focus from gig work to employee settings, revealing that human mediation is not a peripheral fix but a necessary infrastructure for effective governance of algorithms. This has important implications for IS research and practice, emphasizing the need to invest in algorithms and the organizational capacities that sustain them.