Schwaeke, Peters, Kanbach, Kraus & Jones — The new normal: the status quo of AI adoption in SMEs
TL;DR
The wiki’s first source built on a classical adoption theory. A systematic literature review of 106 peer-reviewed articles on AI adoption in small and medium-sized enterprises, organised with the technology–organization–environment (TOE) framework of Tornatzky & Fleischer (1990). It is the most-cited recent TOE review of organisational AI adoption (≈200 citations by September 2026), and the reason it matters here is structural: until this ingest the wiki explained enterprise AI adoption entirely through practitioner maturity models and the dynamic-capabilities spine, with no adoption theory in the IS-research sense. See technology-adoption-theories for how the four classical models relate.
The finding is a map, not an effect size. The 106 studies sort into eight clusters under TOE’s three contexts:
| TOE context | Cluster | Studies | What it covers |
|---|---|---|---|
| Technology | Compatibility | 19 | Fit of AI with existing IT, strategy, work practices and values |
| Technology | Infrastructure | 16 | IT readiness; AI as integrated system, not isolated tool |
| Organization | Culture | 19 | Learning culture, communication, leadership style and commitment |
| Organization | Resources | 17 | Money, people, AI specialists; the SME constraint |
| Organization | Knowledge | 10 | Employee expertise, digital proficiency, willingness to change |
| Environment | Ecosystem | 9 | Collaborative R&D with partners and universities |
| Environment | Competition | 8 | Competitive pressure as driver |
| Environment | Regulation | 8 | Policy, government support, training curricula |
The named gaps. The authors flag two blind spots in the literature: trend identification as a driver of adoption is overlooked, and legal requirements for AI implementation are neglected. Environment is the thinnest TOE context in the literature.
Key claims
Compatibility is the first condition. The largest cluster, and the authors’ own summary: successful adoption “depends on aligning existing practices and cultures and addressing customer needs.” For manufacturing SMEs running machinery over long periods, avoiding compatibility issues is itself the adoption decision. Perceived usefulness and ease of use — TAM’s two constructs — appear here as sub-factors of compatibility rather than as a separate theory.
Resources are the most-cited barrier, and management commitment is the named compensator. Limited funds, lack of AI specialists and poor technical expertise recur across the resource cluster. The discussion argues that senior-management commitment “can compensate” for resource scarcity by strengthening communication between experts; smaller firms are credited with streamlined communication and direct leadership involvement in R&D.
Culture has three components: a learning culture (shared learning spaces, continuous skills development), communication (open dialogue with employees through constant change), and leadership style — “leadership plays a strong role in the dynamic capabilities of firms through the ongoing development of new practices” (citing Quansah & Hartz 2022). This is the review’s one explicit bridge to the dynamic-capabilities vocabulary.
A positive ROI is decisive. SMEs “face a challenge in embracing digital transformation without a positive return on investment”, which the authors position as the constraint policy should address.
Dynamic-capabilities reading
The review is TOE, not Warner & Wäger, so the cells below are the wiki’s reading of its clusters rather than the paper’s own vocabulary.
contextual/internal-barriers— the resources cluster: limited funds, AI-specialist shortage, resistance, compatibility and legal complexity listed together as the challenges SMEs must overcome.contextual/internal-enablers— the culture cluster’s leadership component: senior-management commitment as the factor that facilitates action-plan implementation and compensates for scarce resources.contextual/external-triggers— the competition cluster: “if competing organizations perform more effectively, other organizations are pressured to compete as well.”digital-transforming/navigating-innovation-ecosystems— the ecosystem cluster is this cell almost verbatim: SMEs partner with external organisations and universities for collaborative R&D to bypass the internal investment required to build technological capability.digital-transforming/improving-digital-maturity— the knowledge cluster: employee expertise, digital proficiency and continuous skills development as what makes a technology transformation succeed.
Neighbour sources
Cimino et al. (2025) studies the same population with a different theory — Teece’s dynamic capabilities rather than TOE — and explains all of its null results by limited financial and technical resources. Schwaeke’s resource cluster is the literature-level name for that constraint. The two are complementary lenses on SME adoption, not competing findings.
Carrier (MIT, 2026) argues from operations practice that adoption speed and system-fit beat access to the technology; the review’s largest cluster, compatibility, is the same first-order condition reached from the literature.
Linked entities and concepts
- Concepts: technology-adoption-theories, enterprise-ai-adoption, dynamic-capabilities
- Dangling (single-source mention, deferred per the second-source promotion rule): Julia Schwaeke, Anna Peters, Dominik K. Kanbach (HHL Leipzig), Sascha Kraus (Free University of Bozen-Bolzano), Paul Jones (Swansea University, corresponding author).
Scope and reliability
What was read. The complete article including Appendices A–C (database table, 36 search strings, list of the 106 contributions). Figures 1 and 4 — the TOE model after Nguyen et al. (2022) and the eight clusters crossed with the three TOE contexts — are images and did not survive conversion.
A frequency map, not a meta-analysis. The clusters count research attention. The authors say so in their limitations: the capabilities reported most often “may not necessarily possess the strongest correlation with performance outcomes”, and they call for meta-analyses per category. Read the cluster sizes as what researchers have studied, not as what matters most.
The methods section contradicts itself twice. The review window is stated as “before and including August 2023” and, a paragraph later, as “2003 until June 2024”; Figure 2 supports the later date. The screening funnel is non-monotone (305 papers after abstract reading, then 317 after cross-referencing, then 106), and a drop from 5,768 to 768 is labelled “screening for duplicates”. The raw file records both.
“AI” is defined broadly. The definitional section counts blockchain, IoT and social media among “AI-related technologies”, so the sample is not restricted to machine-learning systems — and almost none of it concerns generative AI, which postdates most of the reviewed work.
Peer-reviewed journal articles only, English and German, which the authors flag as a publication-bias risk.