AI as a cognitive multiplier: explaining performance variance in organizational decision-making

Artificial intelligence (AI) is being used more and more in business applications. As AI has developed, these applications have moved from initial applications as mere customer support chatbots into tools used to support managerial judgment and decision-making. AI is often presented as a tool that increases consistency, speed and decision quality. These same apparent qualities that make AI attractive as a tool also make it persuasive as an authority. Research shows that the speed and the apparent certainty of AI-generated answers lead their users to trust the quality of the outputs and to delegate decision-making.

However, studies that observe the actual results of AI implementation on decision-making paint a contradictory picture. On one hand, some empirical studies report improved decision-making efficiency, greater speed of decision-making and improved predictive accuracy, especially in data-intensive contexts (Edwards et al., 2000; Shrestha et al., 2019). On the other hand, empirical studies also report the risks of excessive reliance, reduced scrutiny and limited skill development in decision-making environments supported by AI (Sutton et al., 2018; Duan et al., 2019; Taddeo et al., 2019). This branch of research warns that over-trusting AI systems can create safety and governance vulnerabilities.

Evidently, the empirical record shows that identical AI systems do not produce uniform effects. To go further, experimental evidence that's been conducted in both professional and entrepreneurial settings seems to show substantial divergence in AI/supported decision-making outcomes, with some users showing substantial improvement, while others stagnate or decline relative to baseline (Yu et al., 2024; Otis et al., 2024). The findings of these experiments are difficult to reconcile when viewed through the lens of typical AI and decision-making research, which typically emphasizes average performance effects.

It is also worth noting that these experiments were conducted in discrete decision-making episodes. In organizational settings, however, decisions are often made in sequence, with the outputs of one decision becoming the inputs for another. Intuitively, one can imagine that the divergence would compound over time.

The central question that this paper is looking to address is not whether AI improves average decision quality; it is instead why exposure to the same AI system produces divergence in decision-making outcomes.

Decision-making, defined here as the selection of a preferred path among its alternatives to reach a preferred outcome (Simon, 1987), is generally assumed to benefit from the consistency and replicability of algorithms. Following from this definition, problem solving, here understood as the application of knowledge and strategies to resolve novel or complex situations (Martinez, 1998), might also be expected to improve under AI support.

Under these assumptions, differences in outcomes would feasibly stem from technological limitations, poor implementation of technology, or variation across AI tools used. However, emerging empirical evidence seems to suggest that exposure to the same AI systems does not produce uniform effects (Yu et al., 2024; Otis et al., 2024). Moreover, much of the recent literature relies on commonly used AI systems, such as ChatGPT, which serves to reduce technological differences as a root cause of the divergent outcomes.

Given that variation across AI technologies does not explain the discrepancy, naturally attention should shift to the remaining source of variation, which is the decision-makers themselves.

Much of the current body of research evaluates whether AI produces net gains or net losses relative to baseline human judgment. While informative, this approach emphasizes aggregate average effects and overlooks the distributional consequences of AI use. Differences in how individuals frame problems, how willing they are to accept outputs at face value and how exactly they integrate AI-generated material into their reasoning are only partially captured by these prevalent modes of research.

In other words, there is a gap in understanding why identical AI systems cause divergent outcomes across decision-makers. Existing frameworks highlight complementarity between human and AI systems but they typically treat human capability as stable within that interaction.

This paper seeks to develop a conceptual framework explaining why exposure to the same AI system can cause dispersion in decision outcomes. It focuses on the cognitive abilities of the human users as the deciding factor. In so doing, this paper integrates insights from cognitive psychology with management theory and information sciences.

To clarify how AI reshapes the distribution of decision performance across users, the paper proposes testable research questions:

With the recent advances in machine learning as well as the increased integration of AI systems into everyday environments, research on AI has also expanded rapidly, with AI-engaged publications increasing roughly thirteen-fold across scientific fields in the past decade (Duede et al., 2024).

Initially, papers in this area focused on AI as an automation tool, highlighting its potential to improve efficiency, consistency and predictive accuracy in decision-making environments (Duan et al., 2019). Here, AI was viewed as a rational decision-making aid whose algorithms would serve to increase human predictability. As such, AI was believed to be consistently capable of outperforming human judgment, especially in data-intensive tasks (Kmen et al., 2026). Subsequent research then shifted from AI as a replacement for human decision-makers, and viewed it as a decision-support system. These papers are focused on human-to-AI collaboration, claiming that optimal outcomes emerge when humans retain the final authority, while AI systems provide analytical input (Lai et al., 2021). Concepts such as centaur decision-making were introduced to emphasize complementarity in decision-making between humans and AI systems (Saghafian and Idan, 2024).

Much of the research in both camps is framed around whether AI systems enhance or undermine average decision quality. Some studies reported visible gains in decision-making accuracy and speed under AI support (Ouanes and Farhah, 2024), while others warned of excessive reliance on AI recommendations and reductions in critical scrutiny of the users (Buçinca et al., 2021).

Diverging from this mean effect framing, controlled studies have documented heterogeneity in AI-supported performance under the same technological exposure. In a study of radiologists using diagnostic AI, Yu et al. (2024) show that, relative to their baseline, some users improve significantly with AI assistance, others show minimal change and a subset performs worse. Similarly, Otis et al. (2024) find uneven effects of generative AI on entrepreneurial performance, with users similarly showing patterns of improvement, stagnation or decline, all while using the same system. Research on trust in automation offers some insight into this heterogeneity of outcomes by looking at trust dynamics. It posits that reliance depends on whether trust is appropriately calibrated to the system, task and context (Lee and See, 2004; Hoff and Bashir, 2014). Further research shows that individuals may discount algorithmic advice after observing errors (Dietvorst et al., 2014), adding further weight to human-AI trust dynamics.

A gap remains in explaining how these differences in user outcomes evolve over repeated interaction with AI systems. The existing explanations do not specify whether AI interaction reinforces existing engagement patterns or whether dispersion widens, stabilizes or compresses over time. The distributional consequences of AI use therefore remain theoretically underspecified.

A substantial body of research within the field of cognitive psychology shows that humans routinely rely on heuristics and cognitive effort-reducing strategies, especially when making decisions under time pressure or high cognitive load (Tversky and Kahneman, 1974; Shah and Oppenheimer, 2008). Such strategies, which often operate automatically (Kahneman, 2011), allow individuals to conserve cognitive resources while still performing the task at hand.

It can be theorized that in AI-supported decision-making, the pull towards cognitive shortcuts becomes stronger. Rapid, seemingly confident outputs of AI systems reduce the felt cost of engaging deeply with a problem, making cognitive offloading more tempting than it would be with slower or less accessible sources. In this sense, AI can be understood as a supernormal stimulus, given that it amplifies existing opportunities for cognitive offloading beyond levels previously encountered in everyday life (Tinbergen, 1948). As a result of these supernormal qualities, individuals may reduce internal analytical effort and overly trust or even rely on AI systems as cognitive substitutes (Sparrow et al., 2011).

Work in cognitive psychology also indicates that individuals differ in their baseline tendency to scrutinize information. When AI systems are introduced into a decision-making workflow, these baseline cognitive characteristics are not neutralized; they are stabilized and, over repeated use, may be adjusted. In other words, users who approach AI-generated outputs with sustained scrutiny may strengthen that pattern, while users who rely on outputs without critically examining them reduce occasions for independent problem formulation and skill development (Barr et al., 2015; Buçinca et al., 2021).

Research on decision-making and problem solving emphasizes that the quality of decisions depends on access to information as well as the individual's capacity to evaluate, structure and reflect on that information. In novel or ill-defined problem situations, effective problem solving and decision-making require problem formulation, the evaluation of alternative courses of action and the monitoring of one's own reasoning process (Newell and Simon, 1972; Dunbar, 1998).

These cognitive capabilities are commonly grouped together under the umbrella of higher-order thinking skills (HOTS), which are typically understood to include critical thinking, problem solving and metacognition (Bloom, 1956; Anderson and Krathwohl, 2001). Metacognition, understood here as the ability to monitor and regulate one's own thinking, including the calibration of confidence against available evidence, plays a particularly important role as it supports both critical evaluation and adaptive problem solving (Martinez, 2006; Moore and Healy, 2008). These skills are understood to be unevenly distributed across individuals and shaped by prior experience and learning opportunities.

Recently, studies have been conducted that examine how the interaction with AI systems affects higher order thinking skills. As is the case in broader AI research, the results vary. Some experiments, conducted in educational settings, suggest that AI can serve to support the structuring of complex problems and to expose students to alternative perspectives when used as a collaborative tool (Du et al., 2025). Other research, meanwhile, cautions that overreliance on AI-generated solutions may reduce opportunities for independent thought and reflective engagement, especially among users with weaker metacognitive habits (Rastogi et al., 2020).

As previously stated, much of the existing research evaluates AI-supported decision-making in terms of average changes in accuracy, speed or efficiency. Where heterogeneous effects on decision-making outcomes are documented, variation across the users is typically treated as a function of baseline capability and not a structural outcome in its own right. As such, consequences of AI use on the distribution of decision-making outcomes remain conceptually underdeveloped.

Additionally, current research often emphasizes system performance technical specifications or task characteristics, while offering limited explanation of the cognitive mechanisms through which AI systems interact with human judgment. Where human capability is considered, it is often assumed to be a stable component within collaboration. As such, there is a gap in accounting for how repeated interaction with AI systems may amplify differences between the users working in the same decision-making environment. This omission is especially relevant for business settings where decisions are made in sequence and institutionalized. Prior research in organizational learning shows that repeated patterns of action can store, reproduce and reinforce organizational practices, behavior and strategic responses over time (Levitt and March, 1988; Feldman and Pentland, 2003; Van Lieshout et al., 2021). In the case of AI-supported decision-making, small differences in engagement levels and decision quality can compound across decision-making cycles, causing cascading effects.

Addressing these gaps requires a framework that links AI-supported decision-making to the underlying cognitive processes and explains how baseline cognitive engagement strategies get reinforced under repeated use.

This study is conceptual and aimed at theory-building. As such, its objective is to explain the dispersion of decision-making quality in AI-assisted decision-making at a cognitive mechanism level. It does not collect primary data. The literature review and conceptual synthesis were conducted over seven months, between September 2025 and March 2026. The analysis is consistent with broadly accepted conceptual design guidance in management research (Meredith, 1993; Jaakkola, 2020). The framework presented in this paper was developed through an iterative process:

First, attention was given to identifying how existing papers operationalize performance change in AI-supported environments, as well as to identifying where variation across users is acknowledged but not sufficiently explained; or to instances where existing constructs (e.g. automation bias, over-reliance) describe the symptom but leave the divergence across users insufficiently explained.

Second, streams of literature were selected based on their relevance. Three major research streams were established:

1. AI-supported decision-making in organisational settings, to establish the form of divergence;

2. Cognitive effort and heuristic processing, to explain effort substitution and offloading dynamics; and

3. Higher-order thinking and metacognition, to explain why evaluative engagement varies across individuals.

These streams were further supplemented by works on algorithmic reliance, trust calibration and organizational routines in order to develop a comprehensively exhaustive construct of cognitive engagement, feedback responsiveness and repeated decision episodes.

The data analysis followed a qualitative and comparative logic and did not engage in statistical aggregation. Studies from these three research streams were read for how decision quality and cognitive engagement are implicitly defined; what mechanisms are proposed or assumed to link AI use to outcomes and whether dispersion across users is treated as explanatory or incidental. Construct derivation followed a traceability rule: each construct retained in the framework had to address a specific explanatory gap, as well as connect to an established line of theory discussed in the literature review (e.g. bounded rationality accounts of decision effort, augmentation perspectives, higher-order thinking and metacognitive regulation).

Third, these streams of research were connected into a unified theory, using abductive reasoning. From there, the cognitive multiplier construct was developed as it integrates and explains contradictory results of prior research (“AI is beneficial” and “AI is harmful”) in a single cognitive mechanism.

The resulting framework is presented as a set of testable propositions (see: 4.4) which serve to support future empirical examination.

This section introduces the central contribution of this paper – when AI systems interact with human users, they function as a cognitive multiplier. This theory in turn addresses the research questions (RQ1–3) posed in section 1.3.

The term cognitive multiplier is here defined as a cognitive mechanism that produces increased sensitivity of the quality of decision-making outcomes to the baseline cognitive engagement during the human-to-AI interaction.

Cognitive engagement in turn refers to the levels of critical scrutiny, metacognitive monitoring and confidence calibration of a user in an AI-supported decision-making episode. This differs from a user's overall cognitive capability – a user with generally high cognitive capabilities might still exhibit low cognitive engagement in a discrete decision-making episode, due to various externalities, such as distractions, time pressure, etc (Kahneman, 2011).

We can represent the baseline cognitive engagement as E₀. In this theory, decision quality – represented by Q - is always a function of cognitive engagement. With AI acting as a cognitive multiplier, it increases the responsiveness of Q to variation in E₀. In other words, the claim is that:

In AI-supported decision-making, ∂Q/∂E₀ is larger than in unsupported decision-making. Small differences in baseline cognitive engagement produce disproportionately larger differences in decision quality under AI-supported decision-making.

The cognitive multiplier increases the sensitivity of decision quality to baseline cognitive engagement by altering the perceived cost of effort associated with the decision-making episode. AI, by way of its ease of use, speed of response and apparent confidence when generating replies, reduces the cognitive load placed on the decision-makers. This shift in the perceived cost of effort multiplies the initial differences in user cognitive engagement, making them much more consequential:

Users with high initial cognitive engagement, who might approach outputs as provisional material for further reasoning, augment the quality of their decisions by deepening their analytical exploration.

Users with low initial cognitive engagement, who might treat outputs as substitutes for independent reasoning, reduce the quality of their decisions by lowering their analytical exploration.

Figure 1 illustrates this effect as contained in a discrete decision-making episode. As can be seen, initial small differences in baseline cognitive engagement are, under AI-supported conditions, multiplied into larger differences, ranging from cognitive augmentation to cognitive atrophy (section 4.3).

However, it is also important to note that engagement patterns stabilize across repeated decision-making episodes (Ouellette and Wood, 1998). When decisions are made in sequence (as is often the case within organizations), each discrete episode uses as input the cognitive engagement shaped by the prior episode. The impact of the cognitive multiplier therefore compounds across decision-making episodes. We can represent this over-time effect in the following way:

Engagement at episode t is E_t. Then, E_t+1 = f(E_t, interaction outcomes).

Figure 2 depicts this dynamic updating process:

As shown, across repeated decision-making episodes, initial differences in baseline cognitive engagement widen. Simply put, users who consistently interrogate outputs throughout repeated interactions strengthen their evaluative discipline, while those users who rely on minimally examined outputs reduce their independent formulation capabilities. Across repeated decision-making episodes, this results in increased dispersion in E and given that the cognitive multiplier effect increases ∂Q/∂E, dispersion in Q expands accordingly.

The cognitive multiplier can thus be observed at two distinct levels:

1. Within discrete decision-making episodes: AI increases the sensitivity of outcomes to baseline engagement (AI increases ∂Q/∂E).

2. Across repeated decision-making episodes: Reinforcement updates baseline cognitive engagement, expanding initial differences over time (reinforcement updates E, expanding dispersion in Q).

As previously mentioned, the cognitive multiplier framework suggests that cognitive engagement does not remain fixed across AI-supported decision-making episodes but that it disperses across a continuum ranging from cognitive augmentation to cognitive atrophy.

Cognitive augmentation is here understood as a positive shift in sustained cognitive engagement, supported by using AI systems as metacognitive scaffolding and reflected in improved problem formulation, active revision of AI-generated content and ultimately better quality of decisions. When used as a scaffold, AI systems may support exploration of outcomes without replacing human judgment. The algorithmic speed of analysis enables users to consider a broader range of alternatives or simulate complex scenarios that would otherwise be impractical (Dellermann et al., 2019). Cognitive augmentation is therefore associated with appropriate reliance, as users treat AI outputs as useful inputs while retaining responsibility for evaluation (Vasconcelos et al., 2023).

Cognitive atrophy is understood as a negative shift in sustained cognitive engagement, supported by the use of AI as an offloading tool and reflected in reduced problem formulation, passive acceptance of AI-generated material and lower overall quality of decisions. Cognitive atrophy is therefore associated with overreliance, as users treat AI outputs as substitutes for judgment rather than inputs into judgment (Vasconcelos et al., 2023). This negative trajectory can be further explained by recent experimental work which describes cognitive debt under sustained AI reliance; evidenced by reduced indicators of deep engagement and weaker recall of a user's own arguments (Kosmyna et al., 2025). It further fits into documented concerns about automation bias and over-reliance (Buçinca et al., 2021). Importantly, cognitive atrophy may occur while acceptable short-term decision outcomes are still produced, given that the negative shift in cognitive engagement compounds and becomes much more prominent across repeated decision-making episodes, as illustrated in Figure 2. In organizational settings where the same or similar decisions are made in codified sequence, these seemingly small shifts may quickly accumulate.

While this theory presents the cognitive multiplier as a ubiquitous mechanism found in human-to-AI interaction, it also posits that the rate and magnitude of decision quality performance dispersion depend on moderators, which are understood here as individual and situational conditions that shape a user's cognitive engagement. These moderators determine both how strongly AI systems amplify differences in baseline cognitive engagement within a discrete decision episode (how ∂Q/∂E is at the point of interaction) and how quickly baseline cognitive engagement shifts across repeated decision-making episodes (the slope of E_t+1 = f(E_t, outcomes)).

Discrete episode moderators primarily consist of problem framing and perceived ownership of the task at hand. These factors shape how much of an individual's overall cognitive capability is being deployed within a discrete episode. It stands to reason that when individuals perceive themselves as being responsible and/or accountable for the quality of a decision-making outcome, they would display higher levels of cognitive engagement when evaluating the AI-generated outputs. On the other hand, in decision-making episodes which are perceived to be routine, not important or for which the decision-maker does not feel personally responsible and/or accountable, it stands to reason that the decision-makers would default to a pattern of lower cognitive engagement.

Across-episode moderators include feedback responsiveness, understood as the disposition to act on detected feedback signals. As such, it builds upon the feedback signal detection governed by metacognitive capabilities, and adds a behavioral component – noticing AI-generated errors doesn't mean much unless the decision-maker chooses to act on this observation. Decision-makers who both observe errors in AI-generated outputs and adjust their patterns of interaction accordingly should be more likely to interrupt the drift of cognitive engagement across repeated decision-making episodes than those who do not.

This framework also proposes one cross-level moderator, which affects the impact of the cognitive multiplier on both discrete and repeated decision-making episodes. This proposed moderator is the need for cognition (Cohen et al., 1955), defined as the individual's tendency to engage in effortful cognitive activities and to enjoy thinking; or as the presence of the motivation for effortful cognitive activities. Within a discrete decision-making episode, individuals with a high need for cognition should be more likely to treat AI-generated outputs as a jumping-off point for subsequent reasoning, thus sustaining the levels of cognitive engagement across the interaction. In environments with repeated decision-making episodes, the motivation for effortful cognitive activities should help in maintaining high levels of cognitive engagement even where low-effort patterns would otherwise solidify, thus acting against cognitive atrophy.

To enable the empirical examination of the cognitive multiplier framework, the following testable propositions are clearly laid out:

As shown in the previous section, the cognitive multiplier framework proposes a new way of interpreting AI-supported decision-making. Its central result aims to show that AI-supported decision-making increases the consequences of differences in users' cognitive engagement. This is especially relevant in business settings, due to the large number of decisions that are made in series. Simply put, the cognitive multiplier framework shifts attention towards how AI systems redistribute cognitive engagement of users over time. When viewed through this lens, divergence in decision-making outcomes is interpreted as a structural form of human-AI interaction and not statistical noise. This follows the established cognitive psychology theory that variability in human–technology engagement reflects systematic differences in cognitive strategy (Salomon, 1990).

The framing offered by the cognitive multiplier view of human-AI interaction also serves to reinterpret the observed phenomena of over-reliance and automation bias. Typically, these are treated as deviations from optimal use or as cognitive errors to be corrected (Parasuraman and Riley, 1997). When observed through the cognitive multiplier lens, they are instead treated as points along a broader distribution of cognitive engagement. In other words, over-reliance and automation bias are not isolated and uniform but they are consequential as indicators of cognitive engagement patterns trending towards cognitive atrophy.

As mentioned in the literature review, the current debate regarding AI-supported decision-making often presents a dichotomy between AI systems enhancing human capabilities and AI systems substituting for human capabilities (Nguyen and Elbanna, 2025). The cognitive multiplier framework reimagines these two positions as directions of the same underlying cognitive mechanism. As such, it builds a bridge between both sides of the debate by specifying how identical AI systems produce significant divergence in outcomes, and can be responsible for both cognitive augmentation and cognitive atrophy.

The framework also serves to extend experimental accounts which emphasize the positive outcomes of AI-assisted decision-making (Dellermann et al., 2019). These experimental accounts propose that gains in performance are limited to users with stronger pre-existing capabilities. The cognitive multiplier framework adds a temporal dimension to this perspective. In so doing, and despite lexical similarities, the cognitive multiplier framework differs from accounts of a cognitive amplifier (An, 2025) by specifying a temporal and distributional effect. In other words, instead of suggesting that AI systems scale performance (suggesting uniformity), it posits that AI systems increase sensitivity to baseline differences in cognitive engagement. The framework also differs from the concept of a statistical moderator (Baron and Kenny, 1986), as it does not alter relationships between variables but serves to multiply the influence of existing inputs.

The framework also connects to debates on algorithm aversion, confidence calibration and trust in automation. These streams aim to explain why users may accept, reject, overuse or underuse AI-generated outputs (Dietvorst et al., 2014; Lee and See, 2004; Logg et al., 2019; Vasconcelos et al., 2023). The cognitive multiplier framework treats these explanations as pathways through which cognitive engagement is shaped. In this view, appropriate reliance supports cognitive augmentation when AI outputs are treated as inputs for judgment, while inappropriate reliance supports cognitive atrophy when AI outputs are treated as substitutes for judgment.

In the field of behavioral strategy, the cognitive multiplier framework builds upon well-documented theories of bounded rationality (Simon, 1955) and systematic variation in cognitive processing across decision-makers (Powell et al., 2011) by stating that AI systems may serve to scale the limitations of human users, instead of alleviating them. This positioning is further related to research describing tools as means of extending human capabilities (Engelbart, 1962) and redistributing cognitive operations (Clark and Chalmers, 1998). Here, the multiplier model challenges the notion that external tools serve as an extension of capabilities, instead proposing that cognitive augmentation is only one end of the possible spectrum of outcomes.

Finally, the temporal component of the cognitive multiplier framework links it to organizational learning and routines literature, since divergent outcomes across repeated AI-supported decision-making episodes may stabilize into organizational behaviors, practices and routines (Levitt and March, 1988; Feldman and Pentland, 2003; Van Lieshout et al., 2021). This positions the framework within calls for further research on the use and effects of AI systems in organizational settings (Loureiro et al., 2020; Heimberger et al., 2024; Nguyen and Elbanna, 2025) while also specifying the mechanism through which AI-assisted decision quality may diverge over time.

This study suggests that variance in output quality in AI-supported decision-making episodes is a core issue of human-AI interfaces, as well as that human cognitive capability is not a stable input. Instead, repeated interaction with AI systems may reshape the cognitive engagement of the user. In other words, AI-supported decision-making should not be viewed as a discrete episode but as a process through which future engagement patterns are reinforced. The framework therefore suggests not only that some users benefit more from AI than others but that AI systems may widen the distance between users over time by stabilizing different trajectories of cognitive augmentation and cognitive atrophy.

Viewed through the lens provided by the cognitive multiplier framework, no matter how well-executed an AI-system deployment within organizations is, it cannot be expected to uniformly reduce performance dispersion across employees. This is, of course, due to the individual differences in both cognitive capacity and cognitive engagement of the employees. As such, the role of management can be viewed as ensuring positive moderation towards cognitive augmentation among the employees or at least preventing them from sliding into cognitive atrophy. This has consequences for hiring practices, employee training, workflow design and structures of accountability.

When hiring new employees that will interface with AI systems, the framework implies that organizations should look past pure technical skills or supposed AI literacy, focusing instead on the individual cognitive profile, looking to filter candidates more likely to approach decision-making with a high baseline level of cognitive engagement. One of the key characteristics to look for might be the individual need for cognition, as it moderates the levels of cognitive engagement both in discrete decision-making episodes and across repeated decision-making episodes.

When planning and conducting employee training, organizations should focus on enhancing employees’ higher-order thinking skills, thereby building their overall cognitive capacity. Organizations should also look to train their employees to be able to assess outputs and be responsive to feedback, thus moderating the potential negative drift towards cognitive atrophy across repeated decision-making episodes.

In workflow design, where possible organizations should build routines that require users to perform unassisted cognitive work before interacting with AI systems, so that AI-assisted outputs may be reviewed against prior assumptions. This could aid in maintaining the perception of ownership in individuals, thus positively moderating their levels of cognitive engagement in a decision-making episode.

The implications of the framework on structures of accountability focus on the need for clear attribution of responsibility and/or accountability for AI-assisted decision-making on the individual decision-maker. The quality of decisions should be subject to a periodic evaluation, thus implementing a kind of cognitive oversight into existing review structures. These evaluations would track variance in decision-making quality and frequency of the user overriding AI-generated outputs in order to detect negative drifts in the levels of cognitive engagement.

In the context of business education, the cognitive multiplier framework carries important implications, especially given the growing presence of generative AI systems in educational settings. With research showing that educational environments in part shape the cognitive strategies and evaluative habits of individuals (Zimmerman, 2002), it naturally follows that educational environments play a part in shaping the distribution of baseline cognitive engagement levels that AI systems later on amplify.

Of special relevance is the development of higher-order thinking skills, namely problem formulation, critical thinking and metacognition. These all influence the overall cognitive capacity of an individual, as well as enable the individual to moderate their drift towards cognitive augmentation or cognitive atrophy. The framework thus encourages educational practices that place emphasis on cultivating these skills. On a fundamental level, students must first be made aware of the existence and importance of these skills. On a more concrete level, one educational theory that might prove to be especially relevant is problem-based learning (PBL) (Barrows, 1986).

In PBL methodology, students work in groups to solve an unstructured problem, with the instructor serving to encourage, support, advise and monitor the learning activities of the students. The goals of PBL are stated as helping students develop flexible knowledge, effective problem-solving skills, self-directive learning skills, effective collaboration skills and intrinsic motivation (Barrows and Kelson, 1995). As can be observed, these goals closely match the moderating factors proposed by the cognitive multiplier framework. In particular, self-directive learning is closely related to metacognitive awareness (Hmelo-Silver, 2004) and feedback responsiveness, while intrinsic motivation might be linked to the need for cognition. In the context of business education, PBL may be linked to the established practice of learning via business case studies, which seems to show positive results in improving the students' problem-solving skills and performance on case-study-based educational materials (He, 2015).

If a learning institution wishes to integrate AI systems in curricula, the timing of such an integration should also be taken into consideration. If AI systems are introduced too early, meaning before the students have the opportunity to develop their own cognitive capabilities, the stabilization of their cognitive engagement patterns toward atrophy may occur faster.

The cognitive multiplier framework as presented in this study addresses the observed divergence in decision-making outcomes both between empirical studies, and between individuals tested in these studies. It proposes that in human/AI interfaces, AI systems increase the sensitivity of decision outcomes to the baseline levels of users’ cognitive engagement. AI systems act as multipliers here due to their capabilities of reducing perceived cognitive effort by offering structured outputs readily available at a pace and volume not previously encountered.

Across repeated decision-making episodes and as the variance between the user's baseline levels of cognitive engagement and their decision-making outcomes compounds, the cognitive multiplier effect serves to stabilize engagement patterns towards cognitive augmentation or cognitive atrophy. The framing of augmentation and atrophy as poles across a continuum addresses current debates on the net benefits of the integration of AI systems by claiming that the effects of integration are not binary, or static.

Specifying the mechanism-level explanation of how identical AI systems may cause diverging outcomes across users allows for other researchers to evaluate the consequences of AI integration in terms other than aggregate performance, while allowing organizations to moderate their AI integration efforts in such ways to encourage upward drift in decision-making quality.

This conceptual study presents a theoretical framework. It specifies the framework through a set of testable propositions but does not empirically test them. As such, the cognitive multiplier as a framework requires subsequent systematic validation.

The framework is by design limited to the cognitive mechanism-level within individuals, and the implications this has for organizations. It does not address broader drivers that may shape the human/AI interface, such as technical features of AI systems, or employee incentives. Additionally, the framework is only applicable to decision-making systems where there is a human/AI interface – in other words, fully automated decision-making systems, with minimal human oversight and/or intervention, the cognitive multiplier mechanism is not applicable.

The cognitive multiplier framework would be significantly challenged by empirical findings that explicitly prove compression and not dispersion, of decision-making outcomes in AI-supported decision-making. Therefore, this paper invites future empirical research in decision-making outcome dispersion and shifts in cognitive engagement across repeated AI-supported decision-making episodes.

Future research may empirically test if differences in baseline cognitive engagement predict divergence in decision outcome quality in discrete decision-making episodes. Baseline cognitive engagement may be quantified through behavioural indicators (such as independent problem formulation), or more directly using neuroimaging (observing activation patterns in brain regions associated with higher-order thinking skills).

In order to empirically determine if repeated AI-supported decision-making episodes produce a substantial drift in cognitive engagement, there is a strong need for longitudinal studies. The same quantifying tests specified for testing cognitive engagement may be deployed at stable intervals in order to detect drift.

The proposed moderating role of perceived task ownership on the levels of cognitive engagement within a discrete decision-making episode could be empirically examined by experiments manipulating task framing, while measuring for levels of cognitive engagement specified above. Additionally, field studies could examine if employees with clearly attributed roles and responsibilities display higher levels of cognitive engagement in AI-supported decision-making.

The proposed moderating role of feedback responsiveness on the levels of cognitive engagement across repeated decision-making episodes could be further examined by longitudinal experiments detecting both the individual's detection of feedback signals and their behavioural response to them. Specifically, experiments may track the detection of deliberately introduced AI-generated errors, as well as the subsequent changes in the user's patterns of verification.

The correlation between high levels of need for cognition and the levels of cognitive engagement within a discrete decision-making episode may be examined using the 18-item need for cognition scale (Cacioppo et al., 1984) and the quantifying tests specified for testing cognitive engagement levels. Additionally, it would be useful to test the willingness of individuals with a high need for cognition to engage with AI systems, as it could be possible that these individuals only engage with AI systems after a period of independent cognitive work, thus raising their levels of cognitive engagement before an AI-supported decision-making episode.

Longitudinal studies could build on the potential research on the correlation between high need for cognition and high baseline cognitive engagement by observing if a high need for cognition acts as a buffer against stabilizing engagement patterns towards cognitive atrophy. This could be tested by evaluating the participants' need for cognition, grouping the participants by their scores (low-medium-high) and tracking the changes in their cognitive engagement levels across repeated AI-supported decision-making episodes.

As mentioned before, the cognitive multiplier framework would be significantly challenged by consistent empirical evidence showing that variance in the quality of AI-supported decisions compresses, instead of expands. It would also be challenged and require revision, by findings that show that perceived ownership, feedback responsiveness and need for cognition do not influence the levels of the user's cognitive engagement.

Luka Lučić: Conceptualization, methodology, formal analysis, writing – original draft, writing – review and editing.

The author declares that the final content of this manuscript represents the sole intellectual efforts and original contributions of the author. The author used ChatGPT-4o (OpenAI, accessed 2025) for grammar checking purposes only. All AI-generated content was thoroughly reviewed, fact-checked, and substantially revised by the author to ensure accuracy, originality, and adherence to academic standards.