Introduction

As generative AI becomes increasingly embedded in managerial work, one of the most important questions is no longer whether managers use these systems, but how they decide when to rely on them. This is especially significant in small and medium-sized enterprises (SMEs), where strategic decisions are often made under conditions of resource constraint, incomplete information, and limited formal decision support.

Current discussions of AI in organisations often frame trust too simplistically. Managers are assumed either to trust AI or to distrust it. Yet this binary framing is analytically weak. In practice, reliance on AI is rarely stable, uniform, or unconditional. It shifts across tasks, over time, and in response to experience. A manager may use generative AI confidently for drafting, cautiously for analysis, and sceptically for strategic recommendations. The key issue, therefore, is not trust in the abstract, but trust calibration: the ongoing process through which managers adjust reliance on AI relative to its perceived capability, the task at hand, and the organisational consequences of error.

This article argues that trust calibration is emerging as a central issue in AI-enabled decision-making in SMEs. Generative AI does not simply provide information; it reshapes the conditions under which managers evaluate information, form judgements, and decide whether to act. Understanding how trust is calibrated under uncertainty is therefore essential to understanding the real organisational impact of generative AI.

Why trust calibration matters more than “trust”

The concept of trust is often invoked in AI adoption research, but it is frequently used as a broad attitude measure rather than a process variable. This is a problem because a manager can express general trust in AI while still rejecting a specific recommendation in a high-stakes decision. Conversely, a manager may claim scepticism toward AI but still rely heavily on it under time pressure.

Trust calibration offers a more precise lens. It shifts the focus from whether managers trust AI in general to whether they rely on it appropriately for a given task and context. A well-calibrated decision-maker does not trust or distrust AI categorically. Instead, they adjust reliance in proportion to the system’s actual usefulness, the quality of available evidence, and the risks associated with the decision.

This distinction matters because the organisational costs of miscalibration can be substantial. Over-reliance may lead managers to accept plausible but weak recommendations without adequate scrutiny. Under-reliance may lead them to ignore potentially valuable analytical support, especially in areas where generative AI can improve speed, synthesis, or scenario exploration. The challenge is therefore not to maximise trust, but to calibrate it.

Research in human–AI decision support increasingly treats calibration as central to effective collaboration. Recent work in human–AI decision-making shows that trust behaviour is shaped not only by system performance but also by task difficulty, confidence in one’s own judgement, and the evolving interaction between user and system over time. In other words, trust is dynamic rather than fixed, and appropriate reliance depends on more than technical accuracy alone.

Why SMEs are a particularly important setting

Trust calibration is especially important in SMEs because decision structures are often compressed. Unlike large organisations, SMEs may lack dedicated analytics teams, formal AI governance mechanisms, or layered review structures. Strategic decisions may be concentrated in a founder-manager, director, or small senior team, and the distance between AI-generated output and managerial action may therefore be relatively short.

This organisational structure has two implications.

First, generative AI may be unusually attractive in SMEs because it appears to compensate for missing resources. A manager without access to specialist analysts can ask a system to summarise market conditions, compare competitors, generate customer segmentation ideas, or outline strategic options. The apparent gain in speed and cognitive support is considerable.

Second, the absence of formal checks may make miscalibration more consequential. If a large organisation uses AI poorly, the effects may be buffered by review routines, internal challenge, or specialist oversight. In an SME, a poorly calibrated manager may move directly from AI output to strategic action with much less friction. That does not make SMEs inherently more vulnerable, but it does mean that the quality of managerial judgement becomes even more important.

Recent OECD evidence reinforces why this matters. Generative AI is already being used in a sizeable share of SMEs, and many firms report gains in performance and support for skill shortages. At the same time, the same evidence suggests that use remains concentrated in relatively peripheral tasks and that many SMEs are not yet putting in place structured training, internal guidelines, or governance routines for responsible use. This creates a tension: adoption is moving faster than the development of the organisational capabilities needed to use AI well.

Trust calibration as a managerial process

A more realistic account of trust calibration in SMEs begins with the recognition that managers do not encounter generative AI as a neutral machine output. They encounter it as a recommendation, synthesis, interpretation, or draft scenario that must be judged in context. The decision to rely on that output is shaped by several interrelated considerations.

1. Perceived task risk

Managers are likely to calibrate trust differently depending on what is at stake. Low-stakes drafting tasks invite a different level of reliance than hiring decisions, pricing changes, supplier selection, or market-entry decisions. As the consequences of error increase, the threshold for acceptance should rise.

2. Perceived AI competence

Trust is partly shaped by beliefs about what the system is good at. Managers may see generative AI as useful for summarising and brainstorming but weak at context-sensitive strategic judgement. These beliefs need not be accurate, but they strongly influence reliance.

3. Confidence in one’s own expertise

Managers with strong domain knowledge may use AI as a comparative tool rather than as an authority. Managers with lower confidence or less experience may be more vulnerable either to over-reliance or to complete rejection. Trust calibration is therefore partly relational: it depends on confidence in the self as well as confidence in the system.

4. Prior interaction history

Trust is not reset with each prompt. It accumulates through experience. Repeated useful outputs may encourage greater reliance; visible errors may trigger scepticism or defensive checking. Research on trust dynamics in automation shows that trust evolves through moment-to-moment interaction rather than existing as a static attitude.

5. Organisational norms and routines

Managers do not calibrate trust in isolation. If an organisation encourages challenge, verification, and evidence checking, AI outputs are more likely to be treated as provisional inputs. If speed and decisiveness are prioritised without equivalent review routines, over-reliance becomes more likely.

These factors suggest that trust calibration is best understood not as a single managerial trait, but as a situated organisational process.

The three main forms of trust miscalibration

To understand trust calibration analytically, it is helpful to identify the main ways it can go wrong. In SME settings, three forms of miscalibration are especially important.

1. Over-reliance

Over-reliance occurs when managers accept AI outputs too readily relative to their actual reliability or relevance. This may happen because the output is fluent, well-structured, and delivered with apparent confidence. Generative AI is particularly susceptible to this problem because rhetorical quality can easily be mistaken for analytical quality.

In an SME context, over-reliance may appear in decisions such as:

• accepting AI-generated market analysis without validating assumptions
• adopting pricing suggestions without testing local competitive conditions
• using AI-generated hiring criteria without examining bias or fit
• relying on AI summaries in place of primary evidence

Over-reliance is not simply a technical failure. It is a judgement failure in which the manager treats plausibility as proof.

2. Under-reliance

Under-reliance occurs when managers dismiss or avoid AI outputs even when those outputs could improve efficiency, broaden option generation, or sharpen analysis. This may arise from negative prior experiences, low technical confidence, or a general belief that “real” strategic thinking must remain entirely human.

In SMEs, under-reliance can matter because managerial time and analytical capacity are often scarce. A manager who refuses to use AI even for bounded exploratory tasks may lose a potentially valuable cognitive aid. Under-reliance therefore deserves attention alongside over-reliance; both represent forms of poor calibration.

3. Context-blind reliance

A third and less discussed form of miscalibration is context-blind reliance. Here the problem is not simply too much or too little trust overall, but the failure to vary trust across tasks. A manager may apply the same reliance pattern to all uses of AI, even though different tasks require different standards of verification and different tolerances for error.

This matters because generative AI is not used for a single organisational purpose. It may support writing, synthesis, ideation, customer communication, forecasting, strategic scanning, and decision framing. Treating all of these as equivalent uses is analytically careless and managerially risky.

Why uncertainty complicates calibration

Trust calibration becomes more difficult under uncertainty because uncertainty weakens the very cues managers normally use to evaluate advice. In stable environments, outcomes provide feedback. A manager can compare predictions to results, identify patterns, and refine reliance accordingly. Under uncertainty, feedback is often delayed, ambiguous, or confounded by external conditions.

This is precisely why SMEs provide such an important setting for research. Many strategic decisions in SMEs involve uncertainty that is difficult to reduce quickly: changing customer demand, volatile costs, regulatory shifts, competitive entry, financing constraints, or unstable supply conditions. In these environments, it may be difficult for managers to know whether an AI recommendation is insightful, generic, or simply wrong.

Generative AI also complicates uncertainty because it often masks uncertainty rhetorically. It can produce clear, coherent, and persuasive text even when the underlying evidence base is weak. Recent research on LLM-assisted decision-making suggests that interface cues and confidence signals can affect perceived reliability in ways that do not necessarily improve actual decisions. In some cases, uncertainty cues may even amplify behavioural over-reliance rather than correct it.

This creates a difficult managerial problem. The manager is not only trying to solve a strategic issue; they are also trying to infer whether the AI system is behaving as a useful collaborator, a superficial synthesiser, or a source of false confidence.

A sensemaking perspective on trust calibration

One reason trust calibration is so difficult is that it is not just a technical judgement. It is also a sensemaking process.

Managers do not merely ask whether an output is “correct.” They ask:

• What is this output telling me about the situation?
• Does it fit what I already know?
• Is it exposing a possibility I had not considered?
• Is it simplifying a complex issue too aggressively?
• What kind of evidence would make this recommendation actionable?

These are questions of interpretation, not just verification.

From a sensemaking perspective, trust calibration involves the construction of plausible meaning around AI-generated output. Managers extract cues from the output, compare them to prior knowledge, test them against organisational constraints, and decide whether the recommendation is credible enough to influence action. This means that trust calibration is partly cognitive, partly social, and partly organisational.

It also means that calibration cannot be solved by better technical accuracy alone. Even highly capable systems can be poorly used if managers lack routines for interrogation, challenge, and contextual evaluation.

What better calibration might look like in SMEs

If trust calibration is a managerial and organisational process, then the practical question becomes: what would better calibration actually look like?

A useful starting point is to distinguish between AI as answer and AI as input. In SMEs, better calibration is likely to involve treating generative AI as a provisional analytical input rather than as a final strategic authority. That does not mean using it timidly. It means using it with explicit boundaries.

A better-calibrated SME decision process might include several practices:

1. Task differentiation

Managers should distinguish clearly between low-stakes support tasks and high-stakes strategic tasks. Reliance can be higher in the former and should be more conditional in the latter.

2. Verification thresholds

The level of checking should rise with the consequences of error. A competitor summary may need quick plausibility checks; a market-entry recommendation may require independent evidence, financial scrutiny, and internal discussion.

3. Comparative prompting

Rather than asking AI for a single answer, managers can ask for multiple scenarios, assumptions, or counterarguments. This makes the system more useful as a thinking aid and less likely to be treated as a single source of truth.

4. Challenge routines

Where possible, SMEs should create simple internal routines for challenging AI-assisted recommendations. Even lightweight practices—such as asking what evidence would disconfirm the recommendation—can reduce over-reliance.

5. Reflective learning from errors

Managers need opportunities to examine where AI outputs helped, where they misled, and why. Without this feedback loop, trust calibration remains impressionistic rather than cumulative.

These are not full governance systems. They are pragmatic calibration mechanisms suited to resource-constrained organisations.

Implications for research

The literature on AI in organisations still has a tendency to treat trust as a secondary variable attached to adoption. That is increasingly insufficient. If generative AI is becoming part of managerial cognition and decision support, then trust calibration should be treated as a central explanatory mechanism rather than a peripheral concern.

Three research directions appear especially important.

1. From adoption to reliance patterns

We need to know not only whether SMEs adopt generative AI, but how reliance varies across tasks, decision phases, and organisational roles.

2. From attitudes to process

Survey measures of “trust in AI” are useful but limited. We need richer evidence on how managers actually interpret, challenge, and act on AI outputs in live organisational contexts.

3. From individual trust to organisational calibration

Trust is often measured at the level of the individual user. But in organisations, calibration is also shaped by routines, culture, and decision structures. SME research should therefore examine how trust becomes embedded in practice, not just in attitudes.

Conclusion

As generative AI becomes more common in SMEs, the critical managerial question is not simply whether these systems are used, but how managers decide when and how far to rely on them. That is a trust calibration problem.

Trust calibration matters because generative AI does not merely accelerate information processing. It enters the decision environment as a persuasive, adaptive, and often rhetorically confident source of input. In SMEs—where resources are constrained, decision authority is concentrated, and formal review structures may be limited—the consequences of miscalibration can be significant.

Over-reliance can produce weak decisions disguised as efficient ones. Under-reliance can waste potentially valuable analytical support. Context-blind reliance can flatten important distinctions between exploratory tasks and high-stakes strategic choices.

For these reasons, trust calibration should be treated as a central concern in the study of AI-enabled decision-making. The future of generative AI in SMEs will depend not only on what these systems can do, but on whether managers can learn to rely on them selectively, critically, and proportionately under uncertainty.

Author Note

Saba Shahzadi is an emerging researcher working at the intersection of generative AI, organisational decision-making, and SME digital transformation. Her doctoral research focuses on how SME managers use generative AI under uncertainty, with particular attention to sensemaking, trust calibration, and human–AI interaction.

References and Further Reading

• OECD (2025). Generative AI and the SME Workforce: New Survey Evidence. OECD Publishing. DOI: 10.1787/2d08b99d-en.
• Yang, X. J., Schemanske, C., & Searle, C. (2021). Toward Quantifying Trust Dynamics: How People Adjust Their Trust After Moment-to-Moment Interaction With Automation. Human Factors. DOI: 10.1177/00187208211034716.
• Naiseh, M., Al-Thani, D., Jiang, N., & Ali, R. (2023). How the different explanation classes impact trust calibration: The case of clinical decision support systems. International Journal of Human-Computer Studies. DOI: 10.1016/j.ijhcs.2022.102941.
• Carter, O. B. J., Loft, S., & Visser, T. A. W. (2024). Meaningful Communication but not Superficial Anthropomorphism Facilitates Human-Automation Trust Calibration. Human Factors. DOI: 10.1177/00187208231218156.
• Zeng, J. et al. (2025). ContractMind: Trust-calibration interaction design for AI contract review tools. International Journal of Human-Computer Studies. DOI: 10.1016/j.ijhcs.2024.103411.