Author Name: Kamal Naghili
Department, Nakhchivan State University, Nakhchivan, Azerbaijan
Purpose: Knowledge about AI in organisations is based mainly on large companies. Small and medium-sized enterprises are applying AI as well, but under quite different circumstances, where strategic decisions are usually made by one owner-manager. The aim of this literature review is to investigate the factors that influence the implementation of AI in SMEs and its effect on the strategic decision-making process.
Design/methodology/approach: Structured integrative review of peer-reviewed research and established working papers, centred on 2015 to 2025 but drawing on earlier theories of technology adoption. A set of anchor reviews was expanded through keyword searches and citation tracking. Eleven empirical studies form the core evidence and are read against the adoption, decision-making and capability literatures.
Findings: For SMEs, adoption is more about the person who runs the firm (the owner-manager), the data and competencies within the firm, and pressures from the customer base and competition. On decision-making, the evidence favours augmentation over replacement. AI appears to speed up decisions and widen the information behind them, and it helps less experienced workers most. It can also mislead when applied to problems outside its scope. Gains depend heavily on whether the firm can judge when to trust the system.
Originality/value: The research paper links two separate literatures that hardly overlap, namely those on SME technology adoption and AI-based decision-making. Six propositions for empirical testing are provided, along with guidelines for owner-managers and policymakers.
Keywords: artificial intelligence; SMEs; technology adoption; strategic decision-making; TOE framework; human-AI collaboration
1. Introduction
A forty-person logistics firm and a multinational bank now have much the same access to AI technology. With cloud services and generative models, the cost of access has become extremely low. But they have almost nothing else in common: the bank has data scientists, lawyers, clean datasets and the freedom to fail in experiments. The logistics company has an owner who reads about AI at night and wonders whether it is worth the trouble.
That contrast sits at the centre of this paper. SMEs make up the vast bulk of businesses within many economies, as well as accounting for a considerable part of total employment, but they lag behind large businesses in the use of new digital technologies (OECD, 2021). AI research into organisations has tended to track the funding and the data, meaning that it has been dominated by large corporations. Ransbotham et al. (2017), surveying executives around the world, noted a considerable gap between AI aspiration and adoption. For SMEs that gap is likely to be wider still.
A second reason for attention is less obvious. In a big company, AI implementation normally happens function by function, and decisions regarding strategy go through committees, analysts and layers of review. In an SME, the decision about strategy normally comes down to the owner-manager alone, who often acts on intuition and limited formal analysis (Thong, 1999). If AI changes how decisions are made anywhere, it may change them most visibly here.
The literature has begun to respond. Schwaeke et al. (2025) reviewed 106 articles on AI adoption in SMEs and sorted the findings into eight clusters using the technology, organisation and environment model. Their review is valuable, but like most adoption research it stops at the point of adoption. What happens next, how AI changes the way an SME actually decides, belongs to a separate body of work on human-AI decision-making that rarely considers small businesses (Jarrahi, 2018; Shrestha et al., 2019; Raisch & Krakowski, 2021).
This paper tries to bring the two together. It is guided by three questions:
RQ1. Which technological, organisational and environmental factors drive or block AI adoption in SMEs?
RQ2. How does AI change the way strategic decisions are made in SMEs?
RQ3. What outcomes, positive and negative, are linked to AI use in SME decision-making?
In answering them, I develop a framework that links adoption drivers to changes in decision-making and to outcomes, and I set out six propositions for empirical testing. Section 2 reviews the theory behind this. Section 3 explains the method. Findings related to adoption and decision-making are presented in Section 4. The model building and propositions are provided in Section 5, while Sections 6 through 8 provide implications, limitations and conclusions.
2. Theoretical Background
2.1 What is meant by AI here
““Artificial intelligence” can be defined in such a broad way that it encompasses anything from a spam detector to an autonomous vehicle. For the purpose of management, a more practical definition would be needed as compared to a philosophical one. Agrawal et al. (2018) provide an effective definition for the current AI: essentially, AI is a technology that makes predictions cost-effective. From predicting demand to warning about a risky client to generating an email response or recommending a product, all are predictions from data. Their argument has a corollary that matters for this paper. The cheaper that prediction becomes, the more valuable the human activity of judgment, knowing what to do with a prediction, becomes.
Davenport and Ronanki (2018) provide an operational classification, separating process automation, cognitive insights (patterns within data), and cognitive engagements (customer or employee interactions using chatbots). Since 2022, a fourth category has become impossible to ignore: generative AI, which creates text, images and code. Generative AI matters for SMEs because it needs little data from the firm itself and can be used directly by non-specialists.
2.2 Explaining adoption: the TOE framework
The most common framework that can be applied in the analysis of technology adoption by organizations is the technology, organization and environment (TOE) framework by Tornatzky and Fleischer (1990). This theory postulates that the adoption of technology by organizations is influenced by factors relating to the technology itself, organization and the environment. Its strength is flexibility. Baker (2012) notes that researchers have filled each of its three boxes with different variables depending on the technology studied, and Oliveira and Martins (2011) found it among the most robust adoption models at firm level.
Small business research added an important refinement long before AI. According to Iacovou et al. (1995), when it comes to small firms, the factors that explain the adoption of electronic data interchange include perceived advantages, organisational readiness, and pressure from trading partners. Thong (1999) went even further and revealed that the characteristics of the chief executive, notably innovativeness and knowledge of information systems, carry great weight. The owner is not just one factor among many; in a small firm, the owner is often where the organisational box begins and ends.
2.3 Individual acceptance: TAM and UTAUT
It is not organisations that make use of technology; it is individuals. According to the Technology Acceptance Model (TAM) by Davis (1989), perceived usefulness and perceived ease of use lead to intention to use the technology. The Unified Theory of Acceptance and Use of Technology (UTAUT) adds social influence and facilitating conditions to this model (Venkatesh et al., 2003). In SMEs, these individual-level models and the firm-level TOE framework overlap more than they do in large organisations, because the person whose acceptance matters most, the owner-manager, is also the one deciding on adoption for the whole firm. Chatterjee et al. (2021) combined TAM and TOE for exactly this reason when studying AI adoption in manufacturing firms.
2.4 AI and decision-making
Managers were seen by Simon (1955) as having bounded rationality, where their decision-making involves limitations on the time they have, the information that is available to them, and even their capabilities, which results in them settling for satisfactory rather than optimal decisions. Mintzberg et al. (1976) showed how strategic decisions take place through several complex phases of recognising problems, considering alternative solutions, and then choosing from among them. AI could, in principle, loosen some of these limits by widening the information available, scanning more options and testing them faster.
How this plays out depends on how humans and machines divide the work. Shrestha et al. (2019) have identified different types of decision structures: full delegation of a decision to an algorithm; hybrid sequences in which algorithms filter options from which humans choose (or the other way round); and aggregation of human and AI judgments. Different structures suit different decisions, depending on how clearly the decision is defined, how far the reasoning must be explained, and how fast the decision must be made. Jarrahi (2018) advocates a relationship that allows AI to handle complexity and large volumes of data, whereas humans deal with the uncertainty, ambiguity and politics involved in the decision-making process. Raisch and Krakowski (2021) point out that automation and augmentation are not simple alternatives to each other. Augmentation can become automation over time, and companies that automate too early may lose the human understanding they later need.
Behind all this is trust. As Glikson and Woolley (2020) conclude from an examination of empirical studies on human trust in AI, it is dependent on the nature of the AI, its transparency and its reliability. People can also err in both directions. As Dietvorst et al. (2015) report, people quickly abandon algorithms after observing one of their mistakes, even if an algorithm is better than humans in general performance. On the other hand, as Logg et al. (2019) reported, lay people often prefer algorithmic advice to human advice. Neither aversion nor appreciation is the same as calibrated trust.
2.5 Capabilities
Finally, adoption does not necessarily equal value. The dynamic capabilities approach suggests that firms gain advantage by sensing opportunities, seizing them and reconfiguring their resources (Teece, 2007). Absorptive capacity, or the capacity to absorb external knowledge and put it into practice, is contingent upon a company’s internal knowledge (Cohen & Levinthal, 1990). Mikalef and Gupta (2021) extend this argument to the realm of AI, suggesting that only when AI capability is created through a combination of data, technology, skills and organisational culture can value be generated.
3. Methodology
The review is done in the form of an integrative review in line with what was suggested by Tranfield et al. (2003) to be used in management studies, that is, with the use of a formulated question, criteria for selection and a transparent synthesis. This type of review is appropriate in this case since the available evidence comes from diverse disciplines and uses different methods, which rules out a statistical meta-analysis.
The point of departure is based on a number of anchor reviews: Oliveira and Martins (2011) concerning firm-level adoption models; Dwivedi et al. (2021) concerning AI in various fields of study; Glikson and Woolley (2020) regarding trust in AI; and Schwaeke et al. (2025) regarding AI adoption by SMEs. I traced their references backwards and later citing work forwards. Keyword searches in Scopus and Web of Science completed the set, using the string (“artificial intelligence” OR “machine learning” OR “generative AI”) AND (SME* OR “small business*” OR “small and medium”) AND (adoption OR “decision making” OR “decision-making”).
Studies that were included had to be peer-reviewed journal articles, scholarly books or working papers from established research institutions; written in English; and offer evidence or theory on AI use in companies, on AI-based decision-making, or on technology adoption in small companies. Most studies dealing with AI issues are recent (after 2015), while foundational theory of adoption and decision-making was considered irrespective of the date of publication. Studies purely technical in nature (only discussing algorithms) were excluded.
Of these, eleven studies have been selected as the empirical basis for the research based on their examination of the adoption of AI in small businesses or the robustness of their examination of the impact of AI on decision-making and working practices. They are summarized in Table 1. Some of these are from outside the SME domain, and I highlight where results from large firms or experiments may not be transferable. For each study I recorded the setting, design, focus and main finding, then grouped the material by research question.
The limitation of this approach should be acknowledged at the start. There are still few rigorous empirical studies of AI-assisted decision-making specifically inside SMEs. Part of the argument therefore rests on extending evidence from other settings, and the propositions in Section 5 are offered as hypotheses to be tested, not as conclusions.
Table 1. Core empirical studies
| Study | Setting | Design | Focus | Main finding |
| Iacovou et al. (1995) | Small organisations | Multiple case study | EDI adoption | Perceived benefits, organisational readiness and external pressure jointly explained adoption. |
| Thong (1999) | Small businesses, Singapore | Survey | IS adoption | Chief executive characteristics and organisational features such as size and staff knowledge shaped adoption. |
| Dietvorst et al. (2015) | Laboratory | Experiments | Trust in algorithms | People abandoned algorithms after seeing them err, even when the algorithm beat human forecasts. |
| Ransbotham et al. (2017) | Firms worldwide | Executive survey | AI strategy | A wide gap separated firms’ AI ambitions from actual deployment. |
| Davenport & Ronanki (2018) | Large firms | Project analysis and survey | AI implementation | Modest, well-scoped projects tended to succeed more often than ambitious transformation efforts. |
| Logg et al. (2019) | Laboratory | Experiments | Trust in algorithms | Lay participants often preferred algorithmic advice to human advice; experts were less receptive. |
| Chatterjee et al. (2021) | Manufacturing firms, India | Survey, integrated TAM and TOE | AI adoption | Technological, organisational and environmental factors, together with perceived usefulness and ease of use, shaped intention to adopt. |
| Mikalef & Gupta (2021) | Firms | Survey, scale development | AI capability | An AI capability built from tangible, human and intangible resources was associated with creativity and performance. |
| Dell’Acqua et al. (2023) | Consulting firm | Field experiment | Generative AI and knowledge work | AI improved speed and quality on suitable tasks but made errors more likely on tasks beyond its capability. |
| Noy & Zhang (2023) | Professionals, online | Experiment | Generative AI and writing tasks | Access to a chatbot cut task time and raised rated quality, with weaker writers gaining most. |
| Brynjolfsson et al. (2025) | Customer support, 5,172 agents | Staggered rollout | Generative AI and productivity | Productivity rose about 15% on average; less experienced agents gained most, while top performers saw little gain. |
Source: compiled by the author from the studies listed.
4. Findings
4.1 What drives and blocks adoption
Organising the evidence by the three TOE categories, a consistent picture emerges. The technology itself is rarely the main barrier. The organisation, and above all its owner, usually is.
Technological factors. Perceived value and compatibility with existing systems are relevant issues, just as they are in any other case of technology adoption (Iacovou et al., 1995; Chatterjee et al., 2021). However, the specific challenge related to AI adoption is data readiness. Most applications of machine learning algorithms require huge amounts of clean and structured data, but such data is unavailable to most small companies. Generative AI has changed the picture here, because general-purpose models arrive already trained and can be useful with little or no proprietary data. Costs have fallen too, though hidden costs such as integration, staff time and growing subscriptions are easy to underestimate.
Organisational factors. This is where the SME story differs most from the large-firm story. Thong’s (1999) finding that the chief executive’s attitudes and knowledge drive adoption in small businesses applies with particular force to AI, a technology surrounded by hype and confusion. An owner who is curious and reasonably informed can push adoption through quickly, with none of the committee approvals a large firm requires. An owner who is sceptical, or who feels out of depth, can block it just as quickly. Skills are the second organisational constraint. Schwaeke et al. (2025) identify knowledge, resources and culture among the main clusters in the SME literature, and the shortage of in-house expertise comes up repeatedly. Davenport and Ronanki’s (2018) observation that modest, well-defined projects succeed more often than grand ones is useful here, since SMEs rarely have the slack to survive an ambitious failure.
Environmental factors. The demand by consumers, suppliers, and competing firms forces smaller firms to utilise technologies they would otherwise ignore (Iacovou et al., 1995). In the case of artificial intelligence (AI), this occurs due to the rise of platforms that offer accounting software, electronic commerce websites, and customer relationship systems that include AI features, so some SMEs adopt AI almost without deciding to. Regulation is a newer environmental factor. The EU Artificial Intelligence Act (Regulation (EU) 2024/1689) sets forth different requirements depending on the risk category of the application. According to Schwaeke et al. (2025), legal compliance issues have not received much attention from SME research scholars, even though the cost burden is higher for smaller firms.
The first thing that links all three categories is adoption. Adoption is not a yes or no decision. Many SMEs play around with free or low-cost products without making an explicit decision, whereas others have embedded AI in their processes. This difference is crucial and will be discussed further in my framework development.
4.2 How AI changes strategic decision-making
Direct studies of AI in SME strategic decisions are scarce. The broader evidence, though, allows several reasonable inferences.
The first concerns information. Bounded rationality as developed by Simon (1955) applies most in the case of small enterprises, where the owner often lacks both time and analysts. AI techniques can help broaden the informational base of the decision-making process by summarising market reports, analysing sales data or generating scenarios, at a cost previously considered prohibitively high. Agrawal et al. (2018) would put it this way: the prediction part of a decision has become cheap, which leaves the judgment part as the scarce resource.
The second concerns who benefits. Perhaps the most striking result in recent research is that AI helps less experienced people most. Brynjolfsson et al. (2025), studying 5,172 customer support agents, found an average productivity gain of about 15%, concentrated among newer and less skilled agents; the most experienced saw little improvement and, in some respects, a slight decline in quality. Noy and Zhang (2023) found a similar compression in writing tasks. For SMEs, which often cannot afford experienced specialists in every function, this is good news. A small firm may use AI to bring a junior employee closer to the performance of a seasoned one.
The third finding is a warning. Dell’Acqua et al. (2023) describe AI capability as a “jagged frontier”. On tasks inside the frontier, consultants using AI worked faster and produced better work. On a task just beyond it, they were noticeably more likely to reach a wrong answer than colleagues working without AI, because the AI’s output looked convincing. Strategic decision-making in smaller organizations could be one example of that sort of ambiguous decision-making situation where the seemingly natural fluency of AI might not correspond to its reliability. Similarly, Jarrahi (2018), looking at the problem conceptually, notes that uncertainty and ambiguity still represent a challenge for machines.
The fourth concerns trust. An owner who has been burned once by a bad AI recommendation may stop using it altogether, the algorithm aversion described by Dietvorst et al. (2015). An owner impressed by early results may lean on it too heavily, closer to the appreciation found by Logg et al. (2019). Both are dangerous. What makes the difference is calibrated trust: knowing which kinds of questions the tool answers well (Glikson & Woolley, 2020). In large companies, this is achieved by testing and documentation processes. In SMEs, it depends largely on the owner’s own experience and judgment.
All things considered, the case for augmentation seems stronger than that of replacement in terms of SME strategy. According to the work of Shrestha et al. (2019), the most suitable structure is likely to be a hybrid one, whereby AI gathers, filters and drafts while the owner makes the decision. It is easier to justify delegating decision-making authority where tasks are repetitive and narrowly focused on operations (e.g., stock replenishment and minor price changes) rather than strategy. According to Raisch and Krakowski (2021), there is a potential danger inherent in delegating increasingly more preparatory work to AI: the owner may gradually lose the detailed understanding of the business that made good judgment possible in the first place.
4.3 Outcomes and risks
The positive outcomes reported in the literature include faster work, better quality on suitable tasks, broader information for decisions and, at firm level, links between AI capability and creativity and performance (Mikalef & Gupta, 2021). Mikalef and Gupta’s point, though, is that the gains come from capability, not from the tools alone. A firm that buys a subscription but lacks the data, skills and culture to use it well should not expect much.
There is also a timing issue. According to Brynjolfsson et al. (2019), the improvements in productivity that stem from general-purpose technologies like AI are only realised after a delay, since companies have to make complementary investments first. Companies operating on thin profit margins can either give up on using AI technology even before they see any benefits or fail to make complementary investments altogether.
The risks are just as real. Too much dependence on applications that work well on average yet fail unpredictably can lead to overconfident errors (Dell’Acqua et al., 2023). Data protection and regulatory obligations add compliance burdens. Biases in training data can spill over into decisions about customers or staff. And dependence on a few large AI providers leaves small firms vulnerable to pricing changes and shifting terms of service.
5. Conceptual Framework and Propositions
5.1 The framework
Figure 1 brings the findings together. The left-hand side shows the TOE drivers of adoption, with the owner-manager placed inside the organisational box because of the weight the evidence gives to that role. These drivers influence the extent of AI adoption, which varies from experimenting to fully embedding the technology in core processes. Extent of adoption, in turn, influences the decision-making process: the information base, speed, and the division of labour between the owner and the machine. These changes produce outcomes, both gains and risks. Absorptive capacity and calibrated trust serve as moderating variables.
Figure 1. Framework linking AI adoption, decision-making and outcomes in SMEs
Source: developed by the author from the reviewed literature.
5.2 Propositions
The framework suggests six propositions. Each follows from the evidence above and can be tested in future research.
P1. In SMEs, the owner-manager’s knowledge of and attitude toward AI is a stronger predictor of adoption depth than firm size or financial resources. The basis for this is Thong’s (1999) evidence on chief executive influence in small firms, combined with the low cost of entry for current AI tools, which reduces the weight of financial constraints.
P2. Platform-embedded AI raises the share of SMEs using AI but has a weaker effect on decision-making than deliberate adoption. Adoption that happens by default through software updates is unlikely to be accompanied by the process changes and learning that capture value (Brynjolfsson et al., 2019; Mikalef & Gupta, 2021).
P3. AI-assisted decision-making yields larger performance gains in SMEs that lack specialised staff than in SMEs that have them. This extends the finding that less experienced workers benefit most (Brynjolfsson et al., 2025; Noy & Zhang, 2023) from the individual to the firm level.
P4. The effect of AI use on strategic decision quality is moderated by calibrated trust: it is positive where owners can distinguish tasks AI handles well from tasks it handles poorly, and may turn negative where they cannot. This follows from the jagged frontier evidence (Dell’Acqua et al., 2023) and the literature on trust (Dietvorst et al., 2015; Glikson & Woolley, 2020).
P5. Hybrid decision structures, in which AI prepares and humans decide, are associated with better strategic outcomes in SMEs than either full delegation or minimal use. The argument draws on Shrestha et al. (2019) and Jarrahi (2018).
P6. Absorptive capacity strengthens the link between adoption depth and firm performance. Firms with more prior technical knowledge will extract more value from the same tools (Cohen & Levinthal, 1990; Mikalef & Gupta, 2021).
6. Discussion and Implications
The main theoretical contribution of this review is to join the adoption question to the decision question. Adoption research usually takes the use of AI technology as the endpoint, while decision-making research usually assumes the existence of AI technology and focuses on large organisations or labs. In SMEs, adoption and decision-making are intertwined in the person of the owner-manager, who decides whether to adopt and then personally lives with the effects of that decision on how decisions are made. This means that the owner becomes both the major proponent of adoption and the major moderator of its impact.
A second contribution concerns the meaning of adoption. Defining it as a simple yes or no ignores the distinction between an organisation in which employees only sometimes interact with a chatbot and an organisation in which forecasting has been transformed by the application of machine learning. Studying the depth of adoption, rather than its mere presence, should make future findings far easier to interpret.
There are three key issues which must be considered by owner-managers in practice. Firstly, focus on specific and narrowly defined problems where the outcomes can be verified, in line with Davenport and Ronanki (2018). Secondly, use AI for decision preparation, not strategic decision-making. Thirdly, take time to learn where the tools fail, because that knowledge is what turns AI from a risk into an asset.
For policymakers, the evidence suggests that simply subsidising AI tools will achieve little. Support for skills, advice on data practices and help with regulatory compliance are likely to matter more, especially as regulations like the EU AI Act are implemented. Programmes that build owner-managers’ understanding of AI, its strengths as well as its limits, address the factor that the evidence identifies as decisive.
7. Limitations and Research Agenda
The main limitation has already been flagged: research on AI-enabled decision-making in SMEs is still thin, and some of the findings presented here come from large companies or laboratory experiments. Whether they will apply to small companies or not remains to be seen, and that is why such findings are presented in the form of propositions. The review is also limited to English-language sources, and much of the empirical work comes from high-income countries.
Future research could take several directions. Longitudinal studies following SMEs over several years would show whether the productivity lag described by Brynjolfsson et al. (2019) applies, and whether some firms give up before the benefits appear. Measures of adoption depth need to be developed and validated. In addition, field studies of how owner-managers actually use AI in making strategic decisions, perhaps through diaries or observation, would provide important information not covered by surveys. Comparative work across countries at different stages of development would be especially valuable.
8. Conclusion
AI technology has become sufficiently affordable that it can be adopted by small businesses. However, affordability does not equal value. As per this review, adoption in SMEs depends mostly on the owner-manager, the firm’s skills and data, and pressure from its environment. Once adopted, AI needs to be perceived as a support for judgment, not a substitute for it. AI increases the scope of information used in decisions and boosts the performance of less experienced staff, but it can lead people astray if applied to issues it cannot help solve. The firms most likely to benefit are those that learn where to trust their tools and where not to. For a small business, that learning may turn out to be the real investment.
