Table 2 Impact pathways, illustrative... | Emerald Publishing

Table 2

Impact pathways, illustrative case evidence and exemplary research questions

Impact pathway	Illustrative case evidence	Exemplary research questions
Rethinking role design in human–AI systems	Walmart–Pactum: procurement roles are reconfigured in use rather than by design. Buyers come to interpret aggregate negotiation patterns, intervene on edge cases and manage relationships the agent now mediates	Designing for role drift: As AI agents assume coordination responsibilities, roles shift in ways leaders did not plan for, how can firms detect and manage role drift by building checkpoints that reassign decision rights and accountability as the human–AI configuration evolves, rather than fixing them at deployment?
	Walmart–Pactum: suppliers' trust, satisfaction and willingness to continue the relationship are shaped by the agent's negotiation style (Herold et al., 2025)	Managing AI-mediated supplier relationships: When an AI agent negotiates on a firm's behalf, its style shapes supplier trust and willingness to continue the relationship, how should procurement teams set and adjust agent parameters and redesign their own relationship-management roles, to protect long-term supplier relationships that the agent now mediates?
	Healthcare GenAI (Song et al., 2025): all roles expand through deployment: clinicians take on AI training and evaluation alongside clinical duties, patients shift from passive recipients to proactive managers of their own health	Role expansion across the network: As AI deployment generates new tasks for actors across the network, how can organizations anticipate and resource these expanded roles – including the training and incentives they require – rather than assuming AI simply removes work?
Rethinking process boundaries in human–AI systems	Starbucks–NomadGo: The same AI technology performs adequately in China and at other U.S. chains but fails in Starbucks' fragmented U.S. supplier base	Assessing process readiness beyond the technology: As AI capability is co-constituted with the operational infrastructure into which it is embedded, how can firms assess whether their existing configuration of processes, supplier base and data infrastructure can actually support an AI deployment (before procuring a tool on the basis of vendor performance claims)?
	Healthcare GenAI: the boundary of the care process expands beyond the hospital walls to encompass post-discharge support. AI thus enables a different process, not a faster version of the existing one	Redesigning the process, not just layering AI on it: As AI deployment changes the boundaries of what a process covers (rather than simply speeding up the existing process), how can organizations recognize when a deployment calls for rethinking process scope and boundaries, and manage that broader transformation rather than layering AI onto unchanged routines?
	Corpus evidence: reports confine agents to “automate specific decisions within clear guardrails” (BCG, 2026) and route exceptions back to human experts, yet in practice these dividing lines drift, as human overrides and workarounds gradually shift which decisions the agent actually handles	Supporting process stabilization: As persistent human override can prevent Human–AI processes to become reliable, how can firms support stabilization – and how should they deal with persistent human override: as evidence of a failed deployment, or as an improvement worth formalizing into the process?
Rethinking performance measurement in human–AI systems	Amazon fulfillment: the productivity “rate” reflects what the human-robot configuration produces, yet is applied back to individual workers which can drive injury, turnover and inequity as structural consequences (Fairwork, 2024)	Measuring performance at the right level of analysis: When output cannot be meaningfully attributed to individuals separately from the AI they work with, how should firms redesign performance metrics to evaluate the human–AI configuration – and what would fair, workable measures of configuration-level performance actually look like?
	Walmart–Pactum: what counts as good performance cannot be fully specified at deployment: the same agent can optimise for price and speed (a competitive logic) or for supplier trust and relationship continuity (a collaborative one), and the operative logic is set and reset through the parameters buyers adjust as they observe what the agent produces	Managing performance criteria that cannot be fixed in advance: As AI becomes more agentic, how should firms manage this ongoing parameter-setting – and make visible which performance logic is actually being enacted, given that it emerges from the agent's training data and inferences as much as from deliberate choices?
	Walmart–Pactum and corpus evidence: agents may surface alternative performance logics that humans alone might not have pursued (e.g. supplier-base diversification over per-contract savings) or reinforce historically conservative patterns embedded in training data	When AI expands or constrains performance possibilities: As the logic an AI agent enacts can expand or narrow performance, how can firms tell whether their AI systems are opening up or constraining the options they consider, and intervene before a narrowed logic quietly becomes the organization's definition of good performance?

Impact pathway	Illustrative case evidence	Exemplary research questions
Rethinking role design in human–AI systems	Walmart–Pactum: procurement roles are reconfigured in use rather than by design. Buyers come to interpret aggregate negotiation patterns, intervene on edge cases and manage relationships the agent now mediates	Designing for role drift: As AI agents assume coordination responsibilities, roles shift in ways leaders did not plan for, how can firms detect and manage role drift by building checkpoints that reassign decision rights and accountability as the human–AI configuration evolves, rather than fixing them at deployment?
	Walmart–Pactum: suppliers' trust, satisfaction and willingness to continue the relationship are shaped by the agent's negotiation style (Herold et al., 2025)	Managing AI-mediated supplier relationships: When an AI agent negotiates on a firm's behalf, its style shapes supplier trust and willingness to continue the relationship, how should procurement teams set and adjust agent parameters and redesign their own relationship-management roles, to protect long-term supplier relationships that the agent now mediates?
	Healthcare GenAI (Song et al., 2025): all roles expand through deployment: clinicians take on AI training and evaluation alongside clinical duties, patients shift from passive recipients to proactive managers of their own health	Role expansion across the network: As AI deployment generates new tasks for actors across the network, how can organizations anticipate and resource these expanded roles – including the training and incentives they require – rather than assuming AI simply removes work?
Rethinking process boundaries in human–AI systems	Starbucks–NomadGo: The same AI technology performs adequately in China and at other U.S. chains but fails in Starbucks' fragmented U.S. supplier base	Assessing process readiness beyond the technology: As AI capability is co-constituted with the operational infrastructure into which it is embedded, how can firms assess whether their existing configuration of processes, supplier base and data infrastructure can actually support an AI deployment (before procuring a tool on the basis of vendor performance claims)?
	Healthcare GenAI: the boundary of the care process expands beyond the hospital walls to encompass post-discharge support. AI thus enables a different process, not a faster version of the existing one	Redesigning the process, not just layering AI on it: As AI deployment changes the boundaries of what a process covers (rather than simply speeding up the existing process), how can organizations recognize when a deployment calls for rethinking process scope and boundaries, and manage that broader transformation rather than layering AI onto unchanged routines?
	Corpus evidence: reports confine agents to “automate specific decisions within clear guardrails” (BCG, 2026) and route exceptions back to human experts, yet in practice these dividing lines drift, as human overrides and workarounds gradually shift which decisions the agent actually handles	Supporting process stabilization: As persistent human override can prevent Human–AI processes to become reliable, how can firms support stabilization – and how should they deal with persistent human override: as evidence of a failed deployment, or as an improvement worth formalizing into the process?
Rethinking performance measurement in human–AI systems	Amazon fulfillment: the productivity “rate” reflects what the human-robot configuration produces, yet is applied back to individual workers which can drive injury, turnover and inequity as structural consequences (Fairwork, 2024)	Measuring performance at the right level of analysis: When output cannot be meaningfully attributed to individuals separately from the AI they work with, how should firms redesign performance metrics to evaluate the human–AI configuration – and what would fair, workable measures of configuration-level performance actually look like?
	Walmart–Pactum: what counts as good performance cannot be fully specified at deployment: the same agent can optimise for price and speed (a competitive logic) or for supplier trust and relationship continuity (a collaborative one), and the operative logic is set and reset through the parameters buyers adjust as they observe what the agent produces	Managing performance criteria that cannot be fixed in advance: As AI becomes more agentic, how should firms manage this ongoing parameter-setting – and make visible which performance logic is actually being enacted, given that it emerges from the agent's training data and inferences as much as from deliberate choices?
	Walmart–Pactum and corpus evidence: agents may surface alternative performance logics that humans alone might not have pursued (e.g. supplier-base diversification over per-contract savings) or reinforce historically conservative patterns embedded in training data	When AI expands or constrains performance possibilities: As the logic an AI agent enacts can expand or narrow performance, how can firms tell whether their AI systems are opening up or constraining the options they consider, and intervene before a narrowed logic quietly becomes the organization's definition of good performance?