Impact pathway: beyond automation and augmentation – a relational ontology of AI in operations and supply chain management

Artificial intelligence (AI) is rapidly moving from a specialized optimization tool to a general-purpose technology embedded in everyday operations and supply chain management (OSCM). Contemporary AI systems can generate novel outputs, engage in open-ended problem-solving and operate with increasing autonomy across loosely defined tasks (Jackson et al., 2024; Krakowski, 2025). BMW, for example, uses agentic AI systems to streamline corporate fleet ordering. A dedicated agent autonomously transfers data to internal applications and initiates the necessary steps, replacing around 90% of previously manual tasks (BMW Group, 2026). Industry surveys document accelerating deployment of agentic AI across core OSCM processes, even as many organisations struggle to capture its value (BCG, 2026; PwC, 2026). These developments illustrate how OSCM decision-making increasingly takes place within configurations in which humans and AI are enmeshed.

Recent OSCM scholarship acknowledges these developments. Roehrich et al. (2025) diagnose a structural shift beyond incremental optimisation: AI “is beginning to reshape how operational systems sense, decide and adapt – often in ways that displace traditional human judgment and organizational routines”, affecting “what decisions are made as well as how they are justified and by whom”. First contributions have begun to unpack this shift, revealing the paradoxes AI adoption creates in sales and operations planning (Jazairy et al., 2025) and explorative learning (Dai et al., 2025). Cohen et al.'s (2026) vision statement positions AI as an “adaptive intelligence layer” that, when integrated with OSCM domain knowledge, can yield “faster, smarter, and more resilient supply chains” (Cohen et al., 2026, p. 692). The central challenge, they argue, is combining “what AI does well” (e.g. pattern recognition) with “what operations management provides” (a structural understanding of how supply chains work) (Cohen et al., 2026, p. 3).

The depiction of AI as a technological layer added to existing operations reflects a specific ontology: AI is a discrete tool with stable, pre-determined properties whose potential is unlocked through integration into existing workflows. The same ontology underlies the automation-augmentation dichotomy that dominates both research and practice (Krakowski, 2025). Both framings rest on three assumptions: humans and AI are pre-constituted entities entering the interaction with specific roles and capabilities; the key question is how to split tasks across these capability sets (which actor should do which portion of the work, and with what degree of oversight) and performance is evaluated against stable, predefined criteria.

This discrete-tool ontology, however, fails to account for the emergent and unpredictable outcomes of AI deployment that often go beyond the user's intentions (Retkowsky et al., 2026). As an alternative, we propose a relational ontology of AI that treats roles, capabilities and performance not as predetermined but as constituted through ongoing human–AI interactions (Huysman, 2026). The central challenge for OSCM scholars and practitioners is no longer how the right division of labour between humans and AI looks like, but how AI reconfigures existing roles, processes and performance, and with what effects.

The dominant approach in OSCM research treats AI as a standalone tool with predefined properties deployed to unlock efficiency gains. Jackson et al. (2024) conceptualize AI's capabilities as an “internal resource” which can “substitute or complement” human cognitive capabilities to achieve competitive advantages (p. 6122). Cohen et al. (2026) similarly frame AI as a set of capabilities (speed, predictive insight, consistency) to be leveraged effectively. In both cases, human–AI collaboration is reduced to the challenge of dividing labour between human and machine, with humans interpreting AI outputs, managing autonomous systems and contributing oversight in “co-pilot” mode.

What this perspective obscures is the fundamentally relational and emergent character of human–AI collaboration. As humans and AI systems interact, they do not simply divide preexisting work; they reshape the work itself (Huysman, 2026). Decision processes are reconfigured, informational flows reinterpreted and boundaries of expertise renegotiated. The worker is no longer merely a decision maker supported by a tool, nor is the AI merely an artifact executing a bounded function; both are continually constituted through use, feedback, adaptation and organizational context.

Some recent work acknowledges this. Dai et al. (2025) show how AI shapes explorative organizational learning, which in turn refines the AI system's algorithms – a reciprocal dynamic that moves beyond linear accounts of technological impact. Spring et al. (2022) come closest to a relational ontology, rejecting the idea that AI has fixed effects determined by its designers and arguing that the same artefact can produce different outcomes depending on user and setting. Yet, the underlying premises remain largely unchanged: Capabilities and identities are properties that humans and AI bring to the interaction, rather than something continuously constituted through it.

This article therefore adopts a relational ontology that treats human–AI interactions not as exchanges between stable, pre-constituted entities, but as processes through which OSCM capabilities, roles and performance criteria continuously form. The relevant unit of analysis is neither the human nor the AI in isolation, but the human–AI configuration.

To analyze how AI is framed in current OSCM practitioner discourse, we sampled 63 public practitioner PDFs published between 2023 and 2026, comprising 52 industry/consultancy reports and 11 company/use-case documents (details on the corpus and the full list of sources are available upon request). The unit of analysis was the AI-context sentence: a sentence containing automation or augmentation language where an AI anchor appeared in the same or adjacent sentence; this procedure yielded 1,244 sentences. We classified each sentence's framing of AI as a discrete-tool, hybrid (relational language layered on a discrete-tool framing), genuinely relational or insufficiently informative. 975 sentences (78.4%) anchored AI as a discrete tool – (54.4%) exclusively and (24.0%) in hybrid form – whereas a genuinely relational framing appeared in only five (0.4%); the remainder were insufficiently informative. The discrete-tool paradigm thus strongly dominates practitioner discourse. Table 1 connects the coded report language about roles, processes, capabilities and performance criteria to the three impact pathways.

To ground our impact pathways, we complement the corpus analysis with four exemplary cases, each rendering visible a reconfiguration of roles, processes or performance criteria the discrete-tool framing cannot account for: Walmart's autonomous negotiation agent Pactum, Starbucks' AI-based inventory management system and Amazon's robotised fulfilment centres, reconstructed from public company documents and press coverage (all sources are available upon request) as well as an intervention-based field study of a GenAI platform in healthcare conducted by one of the authors (Song et al., 2025).

A major challenge for practice is that AI will create new OSCM roles and change existing ones in ways leaders struggle to anticipate. A 2026-Gartner report estimates that by 2030, 20% of procurement professionals will hold new “AI-driven roles [that] leaders aren't ready for” (O'Doherty, 2026). The 2026 Stanford AI Index confirms that organizations are creating AI governance roles as accountability structures, expertise requirements and career paths are restructured.

Walmart's deployment of Pactum, an AI system that autonomously negotiates contracts with thousands of tail-end and mid-tier suppliers “24/7/365” within parameters set by procurement teams illustrates how the discrete-tool paradigm frames this reconfiguration (Procurement Magazine, 2026). From this view, procurement work comprises two stable layers – a transactional layer absorbed by AI and a strategic layer that remains the domain of human experts. Yet this layered view obscures how both are reshaped. Procurement professionals must now interpret aggregate negotiation patterns the AI surface, intervene on edge cases and manage supplier relationships mediated – and sometimes strained – by agent behaviour rather than direct human contact. When suppliers interact with an AI buyer-agent, their trust, satisfaction and willingness to continue the relationship are shaped by the chatbot's negotiation style, making the technology an active participant in the exchange rather than a neutral support tool (Herold et al., 2025). Role reconfiguration is thus not an ex ante design choice but an ongoing, emergent phenomenon: roles, decision rights and accountability are continually negotiated through human–AI interactions.

The value of a relational lens is further demonstrated by a recent study by one of the authors on a GenAI platform providing emotional and medical feedback to patients after hospital discharge (Song et al., 2025). Rather than treating AI as a static tool, the authors show how it functions as a non-human actor embedded in a dynamic network of clinicians, developers, patients and family members. The authors observed an expansion of all roles through AI deployment: patients shifted from passive recipients of healthcare services to proactive managers of their own health. Physicians took on dual roles as clinicians and AI evaluators; as a cardiologist stated: “At first, I thought my role would be to provide clinical data, but during the evaluation meetings, we gradually realized that we needed to be directly involved in the training and evaluation of M-AI …. This dual role actually made me more enthusiastic, because I could see for myself how the AI was being adapted based on our comments.” The study shows that when these actors collaborated adaptively, patients experienced less anxiety and greater self-confidence after discharge. From a relational perspective, the challenge is not to protect existing social structures from AI disruption, but to understand how roles – with their embedded expertise, judgment and trust – evolve through ongoing human–AI interaction.

A second challenge is that AI deployment reshapes processes and tasks well beyond the ones it is initially deployed to improve. The Stanford AI Index (2026) documents this pattern: gains of 14–26% in structured tasks come alongside slowdowns in adjacent judgment-heavy work and longer-term skill development penalties (p. 219). These results are difficult to explain under the discrete-tool paradigm, since AI is not simply slotted into a stable process while surrounding elements remain fixed. Yet, supply chain executives hesitate to reorganise processes around AI; this may explain why 89% of operations leaders report less-than-expected results from technology investments (PwC, 2026), and why only 7% of supply chain managers report having found value in agentic or generative AI at all (BCG, 2026, p. 20).

Starbucks' 2025 AI deployment illustrates this. The company rolled out an AI-powered inventory management tool, replacing manual inventory counts with camera and LIDAR scanning (Cunningham, 2026). Despite vendor-advertised 99% accuracy, Starbucks experienced chaotic swings between excess deliveries and supply shortages. The same technology performs adequately at other U.S. food chains, and Starbucks benefits from a comparable system in its China stores. The U.S. failure reflects a blind spot the discrete-tool framing cannot see: Starbucks' fragmented U.S. supplier base – 1,500 cup-and-lid pairings from different vendors, numerous small regional suppliers with limited surge capacity – accumulated through procurement decisions made long before any AI deployment. AI's effects are thus determined less by the technology than by the material and operational infrastructure into which it is embedded.

The healthcare case illustrates the generative side of this dynamic. When the GenAI platform was deployed, the patient care process changed: hospital staff continued initial in-hospital care while AI ensured continuity after discharge (Song et al., 2025). A hospital leader noted: “Leadership was initially only considering providing more in-hospital psychological care. But after many discussions, we agreed that a sustainable solution must go beyond the hospital walls.” The AI did not enable a faster or more efficient version of the existing care process but a different one: post-discharge support moved from outside to inside the process boundary.

The relational reframing redirects where managerial investment should flow. Rather than asking how processes should be redesigned to accommodate AI, a relational ontology asks how processes – and AI's capability within them – are co-constituted through its embedding. Thus, instead of scoping AI deployment as a discrete automation project with a fixed before-and-after, managers would do better to treat the supplier base, the parameter-setting and oversight practices and the agent's capabilities as a single configuration, since it is their interaction, not the model alone, that determines what the procurement function can do. The relevant question is no longer “which tasks can the agent take over?,” but “what does procurement become once continuous, large-scale negotiation is possible?”.

Conventionally, performance management assumes that tasks, roles and competencies are stable over time. As human–AI interactions become more deeply integrated into work processes, identifying individual actors responsible for specific outputs may become difficult, if not conceptually meaningless. Performance management must therefore move beyond individual-level assessment to evaluate human–machine configurations through which work is accomplished.

Amazon fulfilment centres illustrate the consequences of failing to make this shift. Amazon's productivity “rate”, that is the output target each warehouse associate must meet, is continuously recalibrated based on what human-robot ensembles achieve. The rate is constitutively relational: it reflects what the configuration of workers and robots (as embodied AI) can produce, not what any individual could achieve unaided. Workers report that targets jumped from “70 per hour” in manual fulfilment facilities to “at least 180” in robotic ones; same worker, different configuration, different performance (Fairwork, 2024). Yet, the new rate is applied back to the individual, with workers in the “bottom five percent” relative to peers facing disciplinary actions. Any technical problems (robot downtime, dispatch errors) are absorbed by workers, driving persistent injury rates and turnover. These are not implementation problems to be optimised away but structural consequences of making individuals accountable for the performance of human-robot ensembles.

A relational ontology points toward a different approach: tracking interaction quality, that is how effectively humans and AI systems communicate, challenge each other and co-produce outputs. The most successful firms will be those that achieve the most effective human–AI configuration, not those that employ the most capable individuals or implement the most advanced technology. Some companies are already moving in this direction. A data science manager at a large payments company that we interviewed described how AI use is now formally embedded in year-end goals: “depending on the role, it might be like using [the technology] versus contributing [towards its development] versus building [e.g. training models].” This illustrates how performance could be measured and managed in human–AI systems, in contrast to approaches that preserve individual-level attribution while simply adding AI as a new performance variable (e.g. Accenture recently linking promotions to AI usage intensity).

Beyond the unit of analysis, AI also changes what counts as good performance. As AI becomes more agentic, performance criteria cannot be fully defined in advance. Pactum's procurement agent, could opt more for a competitive negotiation strategy – optimizing for price discounts, payment terms and speed – or a collaborative one, optimizing for supplier trust and relationship continuity. Thus, the same AI-enabled process can enact different, potentially conflicting performance logics and the choice between them is not made once at deployment but continually, through the parameters buyers set and reset as they observe what the agent produces.

This renegotiation becomes consequential when the logic AI enacts diverges from human norms. AI may introduce alternative performance logics, such as prioritizing supplier-diversification over per-contract savings, that human actors alone could not have initiated. Or it may reinforce existing patterns and reduce an organization's ability to pursue novel paths, when models are based on historically conservative or biased data. In both cases, the relevant question is not what the AI advocates, but how the ongoing interaction between the agent's optimization, buyers' parameter adjustments and suppliers' responses continuously produces and transforms what counts as good performance.

Table 2 summarizes the three impact pathways and their exemplary research questions.

The three impact pathways call for a reframing of the central questions in OSCM in the age of AI: not who does what, but how roles are continuously reconstituted through interaction; not how processes should be redesigned to accommodate AI, but how process configurations emerge and stabilize through relational embedding and not which metrics best capture performance, but how performance criteria are continuously produced and transformed through ongoing human–AI interaction. For OSCM scholars, this reorientation opens new questions about emergence, stabilization and unintended transformation. Practitioners need frameworks to govern the evolution of human–AI configurations, and not just to optimize the initial design.