This critical integrative reappraisal makes three contributions. First, it integrates three previously separate strands of the author’s work—the effects of English use on negotiation and marketing, the exporting barriers facing non-Anglophone business owners, and the evolution of business-English learning—and reinterprets their implications under generative-AI mediation. Second, it develops a two-layer model that distinguishes surface competence in linguistic production and literal comprehension from a deeper form of communicative competence termed cultural fluency. Third, it translates this framework into five testable propositions and a human-AI decision framework for determining what to automate, what to undertake collaboratively, and what to retain under human control. These contributions arise from the present reappraisal and do not appear in the three focal publications.

The model advances beyond the general claim that AI augments rather than replaces human communicators by proposing that AI’s effects are layer-dependent and may alter the economic and strategic value attached to different forms of communicative competence. The propositions developed in Section 7 are theoretically grounded predictions offered for empirical testing rather than established findings.

Three terminological distinctions are important. The term non-native English speaker is retained in its conventional sense while recognizing that the native-non-native distinction simplifies a continuum of proficiency and multilingual experience ([2]). The term fluency refers narrowly to the real-time production and comprehension of grammatically correct and appropriately worded English. It is distinguished from cultural fluency, which comprises the cultural, pragmatic, and relational capabilities that jointly produce counterpart trust, and which AI may support but does not at present reliably supply. The third distinction separates performance from competence. Communicative performance refers to the quality of the language that reaches an interlocutor on a given occasion; communicative competence refers to the capability the communicator personally holds. AI mediation raises the former without necessarily raising the latter, and the gap between them is widest in the surface layer, where a system can supply pronunciation or render rapid speech intelligible in real time while the user’s own articulatory and listening capability remains unchanged. Statements below that AI “handles” or “supplies” a surface component should accordingly be read as claims about performance on a given occasion, not about the acquisition of the underlying capability.

Accordingly, the reappraisal addresses three research questions:

RQ1: What common linguistic, cultural, psychological, and situational challenges facing non-native business communicators emerge from the author’s three prior publications?

RQ2: Which of these challenges can generative AI substantially reduce, and which will continue to require human cultural, pragmatic, and relational competence?

RQ3: How does generative AI reshape the value of English fluency and the strategic position of non-native communicators in international business?

To address these questions, the study adopts a critical integrative reappraisal of three interconnected publications by the author, supplemented and critically evaluated through independent scholarship on business English, cross-cultural communication, and AI-mediated translation. This method is suited to synthesizing heterogeneous sources and to generating conceptual frameworks rather than aggregating homogeneous findings ([35]; [39]). It draws on two bodies of material. The primary corpus is the author’s three prior publications: [24], a research article on the effects of English use in negotiation and marketing; [25], a scholarly book on the evolution of business-English learning; and [26], a research article on the exporting challenges of non-Anglophone owners. Each was re-read and analytically examined across six themes: linguistic, cultural, psychological, and situational barriers; business outcomes; and the implications of generative-AI mediation. Because no formal coding protocol with predefined units and reliability testing was applied, the term coding is avoided. The contextual literature is a body of independent scholarship used not merely to support but to test, qualify, and where necessary contradict the reinterpretation drawn from the primary corpus.

The contextual literature was purposively identified through structured, iterative searches of Scopus, Web of Science, and Google Scholar through January 2026, run over title and keyword fields. Terms were entered both singly and in pairwise combinations across the three domains, rather than as a single fixed Boolean string, and the combinations were varied iteratively according to what each search returned. The language domain comprised “English as a lingua franca,” “business English,” “non-native speakers,” and “communicative competence.” The business domain comprised “international business,” “negotiation,” “marketing,” “exporting,” and “cross-cultural communication.” The technology domain comprised “machine translation,” “generative artificial intelligence,” and “real-time translation.” The results of each search were screened before the next term or pair was chosen, so that lines returning little of relevance were set aside, and productive ones pursued further. Reference lists of key sources were examined for further material. The aim was analytical saturation across the six analytical themes rather than exhaustive enumeration; the process was therefore purposive and iterative rather than a systematic count-based review, and no completeness is claimed for the contextual literature. Sources were included when they comprised peer-reviewed scholarship, authoritative scholarly books, or relevant primary research bearing directly on one of the six analytical themes or on the capabilities and limitations of language technology, and were excluded when non-scholarly, off-topic, or superseded. For very recent technological developments not yet covered by peer-reviewed work, a small number of reputable primary and industry sources were used and are labeled as such rather than treated as scholarly evidence.

Screening proceeded in four stages. At identification, records returned by the database searches and by reference-list checking were compiled. At title-and-keyword screening, a record was retained when it bore on at least one of the six analytical themes or on the capabilities and limitations of language technology. At full-text assessment, the retained records were read against the inclusion and exclusion criteria stated above. At the saturation stage, searching within a theme stopped once further records returned claims already represented in the retained set. Because the process was iterative rather than enumerative, cumulative counts of records identified and screened at the first two stages were not maintained; the retained set is therefore reported in full in their place. Thirty-nine contextual sources were retained alongside the three focal publications, and all appear in the reference list. Their distribution across the analytical themes is as follows: linguistic barriers, 6; cultural barriers, 7; psychological barriers, 3; situational barriers, 3; business outcomes, 8; generative-AI mediation and language technology, 9; and competence theory and review methodology, 5. Several sources bear on more than one theme, so the theme-level counts sum to more than thirty-nine. Table 1 summarizes the search and selection approach.

Table 1. Search-and-selection approach to the contextual literature.

Because the present author authored the three focal publications, this reappraisal is at risk of confirmation bias and interpretive bias. To mitigate it, the reinterpretation is critically compared with independent scholarship, contradictory evidence is actively sought and reported, and no central claim is treated as established solely on the basis of the author’s prior work; the author’s own findings are treated as the object of reappraisal rather than as privileged evidence. Table 2 sets out the three publications, their original findings, and their reinterpretation under generative-AI mediation, and it anchors the analysis that follows.

Table 2. The three focal publications and their reinterpretation in the age of generative AI.

Because corpus-level conclusions are drawn from these three works, their designs and acknowledged limitations require statement at the outset. Table 3 profiles the design, evidence base, setting, analysis, and principal limitations of each publication as reported in the original work. Two features of the profile bear directly on the weight the reappraisal can carry. First, the evidence base is heterogeneous: one quasi-experimental mixed-methods study, one qualitative interview study, and one expository monograph resting on documentary and secondary sources without primary data. Second, each work carries acknowledged limitations of sample, scope, or evidentiary basis, and [24] reports neither a final sample size nor effect sizes or group-level statistics, which limits the precision of any quantitative claim drawn from it. Corpus-level statements in the sections that follow are accordingly framed as convergent patterns across heterogeneous sources rather than as pooled findings, and each is checked against independent scholarship.

Table 3. Design, evidence base, setting, and reported limitations of the three focal publications.

As Table 4 shows, no single publication addresses every theme with equal weight; the barriers are identified collectively across the corpus rather than each being documented in full by each work. Generative-AI mediation is not substantively addressed in prior work, which is precisely why the present reappraisal offers a systematic treatment rather than a summary of them.

Table 4. Thematic coverage of the three focal publications (based on analytical examination of each work).

Note.“Yes” indicates that the theme is substantively addressed; “Limited” indicates that the theme is acknowledged but is not a primary focus or systematically examined; “—” indicates that the theme is not substantively addressed.

A lingua franca is a language adopted as a common medium between speakers whose first languages differ. English occupies this role in international business more completely than any language has in modern history, functioning as the default medium in multinational corporations, cross-border negotiations, international finance, science, and technology ([29]). Its ascent reflects a compounding of advantages: the historical reach of the British Empire, the twentieth-century economic and cultural weight of the United States, the early dominance of English in computing and the internet, and the network effects that accrue to any standard once a critical mass of users adopts it. Each additional speaker who learns English increases the value of English to everyone else, producing a self-reinforcing dynamic that has proved remarkably durable ([30]).

Importantly, the majority of business conducted in English now takes place between non-native speakers rather than between native speakers or between native and non-native speakers. This empirical reality gave rise to the study of English as a business lingua franca, which reframes English not as the property of its native speakers but as a shared resource continually adapted by its users ([20]). In BELF interactions, participants tolerate a wide range of accents, simplify syntax, negotiate meaning collaboratively, and prioritize clarity and goal attainment over native-like correctness ([38]). This scholarship matters for the present argument because it established, well before the AI era, that effective business communication in English was never primarily about grammatical perfection; it was about the pragmatic accomplishment of shared goals. That insight becomes central once machines assume responsibility for surface correctness.

The architecture of the digital economy further consolidated the dominance of English. English was the language of early computing, of the foundational protocols and documentation of the internet, and of the software and platforms through which global commerce is increasingly conducted ([30]). This digital entrenchment is not a historical curiosity; it is directly relevant to the AI turn, because the same asymmetry that made English the default language of the internet has shaped the training data and default behavior of the large language models that now mediate communication. English is overrepresented in the corpora on which these systems learn, with consequences examined below. The genealogy of English dominance and the genealogy of AI capability are thus intertwined, which is one reason to doubt the simplest predictions that AI will straightforwardly dissolve the advantages of English.

The English imperative is not merely an individual concern; it is an organizational strategy. Multinational firms increasingly adopt English as an official corporate language to standardize internal communication, integrate geographically dispersed teams, and facilitate exchange with international customers and partners regardless of national origin. [23] documented both the rationale for such mandates and their disruptive human consequences, showing that the imposition of a common language redistributes status and anxiety within an organization, advantaging fluent speakers and marginalizing others irrespective of their substantive expertise. Subsequent research on the reception of corporate language policies confirmed that employees experience them ambivalently, embracing the coordination benefits while resisting the loss of voice and the implicit hierarchy that a single official language creates ([19]).

The experience of firms that adopted English-only mandates illustrates the stakes concretely. When a large non-Anglophone corporation requires its entire global workforce to operate in English, the policy simultaneously advances integration. It imposes a high transitional cost on employees who must now conduct their professional lives in a second language, sometimes under explicit proficiency requirements tied to advancement. [23] analysis showed that such mandates produce anxiety, status anxiety in particular, as employees who were authoritative in their first language find themselves diminished in the mandated one, while the meaning and burden of the policy are experienced very differently by native speakers, high-proficiency non-native speakers, and low-proficiency non-native speakers. These distributional effects are precisely what AI mediation stands to alter, since a technology that lets employees contribute in their own language, or that raises the floor of everyone’s English output, changes who bears the burden of the common-language requirement and how heavily.

These dynamics reveal a crucial point: language competence in business is entangled with power. The capacity to speak fluently in the dominant language confers not only communicative efficiency but also perceived competence, authority, and legitimacy ([38]). A non-native speaker who commands the substance of a discussion but cannot deploy the language with ease is frequently underestimated and may be excluded from the informal exchanges where influence is actually distributed. Any technology that alters who can produce fluent-sounding English therefore has the potential to redistribute organizational power, a theme to which the analysis returns.

The sheer scale of English use underlines why its dominance is unlikely to be dislodged, even by a technology that reduces the need to learn it. English serves as a first or additional language for well over a billion people and functions as the predominant language of the internet, of international science, and of global commerce, a position established over centuries through the compounding of imperial, economic, and technological advantage ([29]; [30]). Dominance at this scale is self-sustaining: the more English is used, the greater the incentive to use it, and the deeper it embeds itself in the institutions, standards, and technologies through which global activity is organized. This self-reinforcing dynamic is why analyses predicting the imminent irrelevance of English tend to overreach. A technology that lowers the cost of participation without requiring personal mastery of English does not diminish English’s centrality; if anything, by making cross-lingual access easier, it may extend the reach of English-language content, standards, and norms still further.

At the individual and national levels, English proficiency has functioned as a form of human capital that commands a measurable premium. Research across Asian economies has associated English proficiency with stronger economic performance and with access to higher-value forms of participation in the global economy ([18]). At the individual level, English-speaking skills have been linked to career progression, broader professional opportunities, and enhanced bargaining power ([27]). For firms, communicative competence in English is associated with export performance and with the ability to build durable relationships in foreign markets ([26]).

This premium is the economic engine that has driven the enormous global investment in English-language education. It also frames the disruptive potential of AI in stark terms. If a substantial part of the English premium derived from the scarcity of the ability to produce and comprehend English, and if AI now makes a serviceable version of that ability abundant and cheap, then the premium should, in principle, erode at the margin. Whether it does so, and for which components of communicative competence, is an empirical and conceptual question that the later sections address. What matters here is recognizing that the stakes of the AI turn are not merely technical or pedagogical; they are distributive, bearing on who captures the returns from global communication.

Early approaches to teaching English for business, inherited from the broader tradition of foreign-language instruction, emphasized grammar, translation, and the memorization of vocabulary and rules. This orientation reflected an implicit theory that language competence consisted primarily in the mastery of linguistic form, and it produced learners who could parse sentences and conjugate verbs but who often struggled to communicate effectively in real interaction. Over the second half of the twentieth century, the field shifted decisively toward communicative approaches, which reconceived the goal of instruction as the ability to accomplish real communicative tasks rather than the mastery of form for its own sake. Business English, as a branch of English for specific purposes, developed alongside this shift, tailoring instruction to the concrete communicative needs of professional contexts—negotiation, presentation, correspondence, and meetings—rather than to the abstract structure of the language ([34]).

This historical shift is directly relevant to the AI turn, and in a way that is easy to miss. The movement from grammar-centered to communication-centered instruction was itself a recognition that surface correctness is not the ultimate object of language competence—that the point is to do things with words in social contexts. Generative AI, by automating surface correctness, does not overturn this recognition; it extends it, forcing the field to confront the question of what remains once form is handled even more directly. In this sense, the discipline has been preparing for the AI turn for half a century, gradually relocating the center of value from form toward communication, and the AI turn accelerates a movement already long underway.

The late twentieth and early twenty-first centuries brought successive waves of technological advances to language learning. Computer-assisted language learning, and later mobile and app-based learning, expanded access and enabled new modes of practice, allowing learners to study anywhere and to receive immediate feedback ([16]). Gamification introduced motivational structures drawn from game design, and evidence indicated that well-designed gamified environments could support engagement and learning in second-language acquisition ([6]). Informal digital learning—the incidental acquisition of English through media, social platforms, and online interaction—became an increasingly significant complement to formal instruction, particularly for motivated learners with access to English-language content.

These developments established a pattern that frames the AI turn. Each new technology expanded access, lowered cost, and automated some portion of the learning process. Each generated both enthusiasm about democratization and concern about superficiality and the displacement of deeper learning. In each case, the technology proved to be a powerful complement to, rather than a replacement for, the human dimensions of instruction and practice—the feedback, modeling, and social interaction through which competence and confidence actually develop. Generative AI intensifies this pattern to an unprecedented degree, because it does not merely support the learning of English but can, for many practical purposes, substitute for having learned it. This capacity to substitute for learning, rather than only to support it, is what makes the current moment genuinely novel, and it is why the question at issue is not simply pedagogical but strategic and even existential for the field: when the machine can supply the output, what is the human learner learning for?

The most visible obstacles are linguistic. Business English is saturated with specialized vocabulary, technical terminology, and field-specific jargon that even proficient general speakers may not command, and it is further complicated by idiomatic expressions whose meaning cannot be derived from their constituent words ([41]). Grammar presents a persistent difficulty, given the language’s irregular structures, its system of articles and prepositions, and the subtleties of tense and aspect that non-native writers and speakers frequently mishandle ([28]). Pronunciation and intonation pose distinct problems because a speaker’s first language may lack certain English sounds and because English relies heavily on stress and rhythm to convey meaning; mispronunciation can impede comprehension and, more subtly, lead listeners to underestimate a speaker’s competence ([34]).

Writing constitutes a separate and often underappreciated barrier. Professional business writing demands not only accurate grammar and vocabulary but also command of genre conventions—the structure of an effective email, the register of a formal report, the norms of a proposal—and non-native writers who have not internalized these conventions may produce documents that mislead readers or diminish their own credibility ([41]). Fluency in speech, understood as the capacity to produce language smoothly and without excessive hesitation, is a further challenge; disfluency, filler, and visible effort can undermine the persuasive force of an otherwise sound argument. It is precisely this cluster of surface-level linguistic barriers—vocabulary, grammar, spelling, mechanical writing quality, and literal comprehension—that contemporary AI systems are most capable of neutralizing, a point developed below.

Language never travels alone; it carries culture. Effective business communication depends on norms that vary across societies and that are largely invisible to those who have not been socialized into them. Directness that reads as honesty in one culture reads as rudeness in another; silence signals disagreement in one setting and respect in another; the pacing of a negotiation, the sequencing of relationship-building and deal-making, and the acceptable expression of disagreement all differ systematically across cultural contexts ([37]). A non-native speaker may command the vocabulary and grammar of English yet still cause offense or misread intentions because the pragmatic and cultural rules governing its use differ from those of the interlocutor’s culture.

These cultural challenges are qualitatively different from linguistic ones. They are not solved by producing correct English sentences, because the difficulty lies not in the sentences but in the appropriateness, timing, and interpretation of communicative acts within a relationship. Qualitative research on exporting found that non-Anglophone entrepreneurs consistently identified cultural adaptation—rather than vocabulary or grammar alone—as decisive for success in Anglophone markets, and that overcoming linguistic barriers required a holistic approach encompassing cultural understanding and sustained relational effort rather than mere translation ([26]). This finding is pivotal for the AI argument: the dimension of communication that non-native speakers themselves identified as most consequential is the dimension that automated translation addresses least well.

The third domain is psychological. Foreign-language anxiety—the apprehension associated with using a non-native language—operates as a catalyst that converts latent linguistic difficulty into active communicative breakdown ([1]). Non-native speakers frequently report self-doubt, fear of making mistakes, and heightened self-consciousness, and these states impair performance independently of actual proficiency, producing hesitation, withdrawal from discussion, and a reluctance to assert views in high-stakes settings. In multinational corporations, such anxiety can lead capable professionals to disengage from precisely the conversations in which their expertise is most needed, with consequences for both the individual and the organization ([1]).

Confidence and proficiency are mutually reinforcing: low proficiency erodes confidence, and low confidence suppresses the practice and participation through which proficiency grows. This psychological dimension complicates any simple prediction about AI. On one hand, tools that guarantee correct output may relieve anxiety and encourage participation. On the other hand, dependence on such tools may deepen the underlying insecurity and displace the effortful practice through which durable competence and genuine confidence are built. This ambivalence is revisited when human agency is considered later.

A fourth challenge is situational rather than linguistic per se. Business communication frequently unfolds at speed, requiring rapid comprehension, quick decision-making, and immediate responses amid the overlapping talk of meetings and negotiations. Non-native speakers must perform additional real-time cognitive work—parsing unfamiliar input, formulating output in a non-native code, and monitoring for errors—while native speakers devote those same cognitive resources to the substance of the exchange ([2]). This asymmetry can cause non-native participants to fall behind the flow of a fast-moving discussion even when their underlying knowledge is superior. Real-time AI translation is specifically aimed at this asymmetry, and its capacity to reduce the burden of real-time processing is among its most significant potential contributions.

The four domains of challenge—linguistic, cultural, psychological, and situational—are rarely encountered in isolation; they compound one another, making the lived difficulty greater than the sum of its parts. A linguistic gap, such as an unfamiliar idiom, imposes cognitive load; the load slows comprehension; the slowed comprehension causes the speaker to fall behind a fast-moving discussion; falling behind triggers anxiety; and the anxiety further degrades both comprehension and production, in a self-reinforcing spiral ([1]; [2]). A cultural misreading can lead to a linguistic error, and a linguistic error can be interpreted through a cultural lens as rudeness or incompetence, damaging the relationship and heightening the speaker’s anxiety in subsequent interactions. The barriers are, in this sense, a system rather than a list, and interventions that address one component while ignoring the others may achieve less than expected.

An account focused exclusively on barriers would misrepresent the situation of non-native speakers, for the literature also documents substantial benefits that accompany competence in business English. Command of the global lingua franca opens access to international partnerships, to cooperation with foreign customers and colleagues, and to markets and opportunities that would otherwise be closed ([30]). At the individual level, English competence is associated with career advancement and expanded professional horizons ([27]); at the firm and economy levels, it is associated with export performance and fuller participation in the global economy ([18]; [26]). Non-native speakers who acquire strong business English thus convert a barrier into an advantage, and the multilingual competence they possess—command of their own language and culture in addition to English—can itself be a distinctive asset in cross-cultural business that monolingual native speakers lack.

Framing the situation as a ledger of barriers and benefits clarifies what AI changes. To a significant degree, AI reduces the costs recorded on the barrier side of the ledger—the effort, anxiety, and disadvantage associated with surface-level linguistic demands. It does not, however, straightforwardly deliver the benefits recorded on the other side, because those benefits flow from the deeper competence that AI does not confer. Indeed, by making surface fluency abundant, AI may erode the specific advantage that fluency once conferred, even as it leaves untouched—or enhances—the advantage available to those who possess the multilingual and multicultural competence to operate with judgment across markets. The benefits side of the ledger, in other words, migrates along with the value, toward the deep layer.

Discussions of the non-native speaker’s difficulty often privilege production—speaking and writing—over reception, yet listening comprehension is an equally consequential and frequently underestimated barrier. Active listening, the capacity to comprehend and respond appropriately to a speaker in real time, is fundamental to effective business interaction, and non-native speakers must perform it under conditions that native speakers do not face: unfamiliar accents, rapid and overlapping speech, colloquialism, and the absence of the redundancy that shared linguistic and cultural background provides. In fast-moving meetings and negotiations, a lapse in comprehension can cascade, as the listener falls behind while attempting to decode one utterance and consequently misses the next. The reception-side barrier is important to the AI argument because real-time captioning and interpretation address it directly, converting speech that a non-native listener would have struggled to parse into intelligible input, thereby relieving one of the least visible but most disabling of the situational burdens.

Comprehension is not merely lexical; it is inferential and cultural. Understanding what a business interlocutor means, as opposed to what they literally say, requires grasping implicature, register, and the pragmatic force of indirect speech—knowing, for instance, that a polite hedge may signal a firm refusal or that an enthusiastic affirmation may be a courtesy rather than a commitment. These inferential dimensions of comprehension lie in the deep layer, and it is precisely there that literal translation, however accurate, can mislead, because a faithful rendering of the words may strip away the pragmatic and cultural cues that carry the actual meaning. The reception-side barrier therefore has both a surface component that AI addresses well and a deep component that it addresses poorly, mirroring the structure of the argument as a whole.

Business communication is not exhausted by words. Research on the success of non-native speakers’ business pitches has shown that nonverbal and paraverbal cues—eye contact, facial expression, gesture, posture, and intonation—exert a substantial influence on persuasive effectiveness, independent of linguistic competence in the narrow sense ([14]). A pitch delivered in flawless English but with incongruent or culturally inappropriate nonverbal behavior may fail, while a linguistically imperfect pitch delivered with compelling presence may succeed. This finding is doubly relevant to the AI turn. It confirms that a significant portion of communicative effectiveness lies outside the text that translation systems operate on, and it locates that portion squarely in the embodied, relational, culturally variable domain that constitutes the deep layer. As AI increasingly handles the verbal surface, the nonverbal and paraverbal dimensions—and the cultural literacy required to deploy them appropriately—become relatively more decisive, not less.

A quasi-experimental study examined how English use affects non-native speakers in negotiation and marketing, focusing on two functions in which communicative competence directly affects economic outcomes ([24]). Negotiation is a domain in which fluency, vocabulary, and confidence translate into bargaining power: a negotiator who cannot rapidly articulate positions, respond to counteroffers, or deploy persuasive framing operates at a structural disadvantage regardless of the merits of their case. Marketing likewise depends on linguistic precision and cultural resonance, since messaging that is grammatically correct but culturally flat, or that mistranslates the connotations of a brand, fails to move audiences ([17]). The study’s comparison of groups, including the effect of a targeted English communication training program, supported the general proposition that improvements in language and communication skills produce measurable gains in these business-critical functions. These findings do not stand alone: independent research converges on the same pattern, associating linguistic competence with bargaining power and status in lingua-franca interaction ([38]), documenting how foreign-language anxiety degrades participation and performance irrespective of underlying ability ([1]), and linking English-speaking skills to professional advancement and opportunity ([27]). The convergence of experimental, qualitative, and survey evidence strengthens confidence that the effect is real rather than an artifact of any single study or method.

Read in the light of the present argument, these findings invite a refined interpretation. The components of negotiation and marketing performance that depend on surface fluency and literal accuracy are candidates for AI augmentation. In contrast, the components that depend on reading a counterpart’s intentions, calibrating persuasive appeals to cultural values, and building the trust that underwrites agreement are not. The value of the earlier findings is thus not diminished by the AI turn; rather, they help identify precisely which elements of communicative competence remain scarce.

A qualitative investigation of non-Anglophone business owners exporting to Anglophone markets extended the analysis from individual performance to firm-level internationalization ([26]). Using semi-structured interviews and thematic analysis grounded in an interpretive paradigm, the study found that linguistic barriers manifest across the entire export value chain—negotiation, marketing and advertising, customer service, documentation, and supply-chain coordination—and that inadequate language capability can create an impression of unprofessionalism, erode customer confidence, and forfeit business opportunities. Participants overcame these barriers through a repertoire of pragmatic strategies: engaging professional translators and interpreters, employing bilingual liaisons, investing in staff language training, localizing materials, and using audiovisual aids.

Two features of these findings are salient for the AI argument. First, several of the strategies that exporters historically relied upon—professional translation, interpreting, and localization of routine materials—are precisely the services that generative AI now automates at a fraction of the former cost and delay, which implies a genuine lowering of entry barriers for smaller firms. Second, the study emphasized that accurate translation of critical information, though necessary, was not sufficient: success also required cultural intelligence, relationship-building, and the adaptation of communicative strategy to local expectations ([26]). These conclusions align with an independent body of work associating language capability and communicative competence with export performance and with fuller participation in the global economy ([18]; [30]), which lends the firm-level claims support beyond the single qualitative study. The export evidence thus prefigures the central thesis at the level of the firm: AI addresses the translational layer that exporters had to purchase or build, while the relational and cultural layer that distinguished successful exporters remains a human accomplishment.

Cognitive load theory holds that working memory has a limited capacity and that learning and performance suffer when that capacity is exceeded ([15]; [40]). Operating in a non-native language imposes an additional, extraneous load: cognitive resources that a native speaker devotes to the substance of a task are, for the non-native speaker, partly consumed by the mechanics of decoding and encoding language ([2]). This framework explains the situational barrier described above and predicts that any technology which offloads the mechanical processing of language should free cognitive capacity for higher-order reasoning. Real-time AI translation can be understood precisely as a mechanism for reducing extraneous linguistic load. The theory also issues a caution, however: the offloading of cognitive work to external tools can, over time, attenuate the internal capacities that effortful processing would otherwise develop, so that reduced load in the moment may entail reduced competence in the long run.

Social learning theory emphasizes that people acquire capabilities by observing and interacting with others, and that a sense of agency and self-efficacy is central to sustained performance ([3]; [10]). Language learning is quintessentially social, advancing through participation, feedback, and modeling. This theory addresses the psychological barriers non-native speakers face and the ambivalent role of AI. If AI tools scaffold participation, allowing users to observe well-formed models and to engage in exchanges they would otherwise avoid, they may enhance self-efficacy and accelerate learning. If, instead, users delegate communication entirely to the tool, they forgo the participatory practice through which self-efficacy and competence are built. They may experience a diminished sense of agency—a risk that follows directly from social learning theory’s emphasis on participation and self-efficacy ([3]).

Communication accommodation theory describes how speakers adjust their communicative behavior toward or away from their interlocutors to manage social distance, gain approval, or assert identity ([42]). In BELF interactions, accommodation is pervasive: participants converge toward simplified, mutually intelligible forms and collaboratively repair misunderstandings ([38]). This theory foregrounds a dimension of communication that is inherently relational and dynamic—the moment-to-moment calibration of one’s speech to another person’s needs and reactions. It is difficult to automate because it depends on continuous social sensing and shared goals that go beyond the literal content of an utterance. Accommodation is therefore a strong candidate for the durable, human layer of competence, and its prominence in successful BELF interaction suggests that AI mediation may alter the mechanics of communication without displacing the relational work that accommodation performs.

Cultural dimensions theory provides a systematic vocabulary for the cultural variation that shapes business communication, characterizing societies along dimensions such as the relative emphasis on directness, hierarchy, individual versus collective orientation, and tolerance for ambiguity ([11]; [32]). It explains why communicative acts that are appropriate in one cultural setting misfire in another and why cultural adaptation cannot be reduced to translation. The dimensional tradition associated with Hofstede has been influential. However, it is also heavily criticized because it treats nations as culturally homogeneous, reifies dynamic, contested cultures into static scores, and relies on data of limited contemporary validity ([22]). The present argument does not depend on specific dimensions or on any claim that national cultures are uniform; it requires only the weaker, widely accepted premise that communicative appropriateness is culturally variable and cannot be read off linguistic form alone. Understood in that deflationary way—as a reminder of systematic cultural variation rather than as a taxonomy of fixed national types—the theory identifies a zone of competence that automated translation, operating primarily on form, addresses only partially. As the linguistic layer is automated, this culturally variable layer becomes the locus of competitive differentiation, whether or not one accepts any particular dimensional scheme.

The resource-based view of the firm holds that sustainable competitive advantage arises from valuable, rare, difficult-to-imitate, and non-substitutable resources and capabilities. Applied to communication, this view distinguishes sharply between capabilities that are becoming commoditized and those that are not. The mechanical production of correct English, once a scarce and valuable capability, is being rendered abundant and imitable by widely available AI tools, and therefore ceases to satisfy the criteria for sustainable advantage. By contrast, culturally intelligent, relationship-embedded communicative capability—built through experience, difficult to codify, and specific to particular markets and relationships—retains the properties of a strategic resource. The resource-based view thus predicts, at the level of the firm, exactly the migration of value that arises at the level of the individual: away from the fluency that machines supply and toward the cultural fluency that they do not.

Bridging the cultural dimensions and communication accommodation perspectives is the construct of cultural intelligence, which integrates rather than adds to the five theories above and denotes the capability to function and communicate effectively across culturally diverse settings ([7]). Cultural intelligence integrates a cognitive component (knowledge of cultural norms and conventions), a metacognitive component (awareness and monitoring of cultural assumptions during interaction), a motivational component (the drive to engage across difference), and a behavioral component (the capacity to adapt one’s verbal and nonverbal conduct accordingly). Research links cultural intelligence to the performance of global virtual teams and to the effective leveraging of language proficiency, indicating that proficiency and cultural intelligence are complements rather than substitutes: language competence realizes its value only when paired with the cultural judgment to deploy it appropriately ([37]). For the present argument, cultural intelligence names the core of the deep layer with unusual precision. It is learned through experience, difficult to codify, specific to relationships and markets, and exactly the capability that AI-generated fluency neither contains nor confers. As the surface layer is automated, cultural intelligence becomes the pivotal human capability, and the theories assembled here converge on it as the locus of enduring value.

It is important to note that these theories do not speak with one voice about AI; they generate a differentiated set of predictions, which is a strength rather than a weakness of the framework. Cognitive load theory predicts short-term performance gains from offloading but long-term risks to competence; social learning theory predicts benefits when AI scaffolds participation but harms when it displaces it; communication accommodation and cultural dimensions theory predict that the relational and cultural layer will resist automation; and the resource-based view predicts that value will migrate toward whatever remains rare and hard to imitate. The convergence is not that AI is good or bad, but that its effects are layer-dependent: broadly enabling at the surface, largely inert or even corrosive at the depth, unless human competence is deliberately preserved and cultivated alongside it.

Before drawing these perspectives together, it is worth noting what they share, since their convergence is what justifies treating them as a single framework. The five theories originate in different disciplines—instructional design, psychology, sociolinguistics, cross-cultural management, and strategic management—and none was developed to explain artificial intelligence or even business English. Nevertheless, each, in its own vocabulary, locates the substance of communicative competence somewhere other than the correct production of words: in the processing burden beneath fluent speech, in the relational adjustment surrounding it, in the cultural meaning underneath it, in the tacit capability behind it, and in the effortful practice that produces it. Each is, in this sense, a theory about context and relationship rather than about the isolated speaker, and each carries an implicit claim that the valuable part of communication is contextual, relational, tacit, and hard to acquire or imitate. That shared premise—reached independently by five unrelated frameworks—is what makes the two-layer model more than an artifact of any one of them.

Taken together, the five theories—cognitive load theory, social learning theory, communication accommodation theory, cultural dimensions theory, and the resource-based view—support a two-layer model of communicative competence in business (Figure 1). The surface layer comprises the mechanical production and literal comprehension of language—vocabulary, grammar, spelling, pronunciation, and basic translation—and imposes cognitive load, generates anxiety, and historically conferred a scarce advantage. The deep layer comprises culturally attuned pragmatics, relational accommodation, and persuasive and rhetorical judgment; it is relational, tacit, context-specific, and difficult to codify, and its exercise produces the trust and credibility on which cooperation depends. The theories converge on the prediction that generative AI is highly effective at the surface layer and substantially less reliable at the deep layer. Two qualifications travel with that prediction. It concerns performance rather than competence at the surface, since a system that supplies correct form does not thereby confer it on the user. And it concerns the present state of the technology rather than a permanent boundary, the conditions governing its durability being set out in Section 7.3. The consequences of that asymmetry are the subject of the sections that follow.

Figure 1. The two-layer model of communicative competence.

The surface layer is increasingly automatable in performance terms; the deep layer is at present handled unreliably by automated systems and is retained under human control; competitive value is expected to migrate from the former to the latter. The boundary between the layers is treated as contingent rather than fixed. The two-layer model deliberately continues a long tradition of theorizing communicative competence, and its contribution lies not in inventing new components but in reorganizing familiar ones around the fault line that AI introduces. Since [12] distinguished communicative competence from grammatical knowledge, applied linguistics has consistently held that knowing a language exceeds knowing its rules. [5] decomposed communicative competence into grammatical, sociolinguistic, and strategic components, later adding a discourse component, and [36] distinguished organizational competence, encompassing grammatical and textual knowledge, from pragmatic competence, encompassing the functional and sociolinguistic use of language. The surface and deep layers proposed here map recognizably onto these prior distinctions: the surface layer corresponds broadly to grammatical and organizational competence, the deep layer to the sociolinguistic, pragmatic, and strategic competences, enriched by the intercultural construct of cultural intelligence ([7]) and by the relational dynamics of accommodation. Therefore, the model does not claim to discover a new anatomy of competence. It claims that a technological development has, for the first time, made one part of this long-recognized anatomy—the grammatical and organizational surface—cheaply reproducible by machines, while leaving the pragmatic and cultural depth substantially less well served, and that this asymmetry has consequences the earlier frameworks had no occasion to consider. The binary of surface and deep is offered as an analytical simplification of these finer-grained schemes, justified by its aptness to the specific fault line that AI introduces; the term cultural fluency is used, in preference to the more technical vocabulary of pragmatic and sociolinguistic competence, to name the composite of deep-layer capabilities as they bear on business outcomes, and to mark that this composite, not grammatical fluency, is what now differentiates. Where precision is required, the underlying constructs from the established frameworks remain available and are intended to be used.

Figure 2 summarizes how the six theoretical perspectives assembled here map onto the two layers, together with the differentiated prediction each contributes: cognitive load theory bears primarily on the surface layer; social learning theory

Figure 2. The integrated theoretical framework mapped onto the two layers.

and the resource-based view cut across both; and communication accommodation theory, cultural dimensions theory, and cultural intelligence anchor the deep layer that resists automation.

Each perspective is positioned by the layer it most illuminates, with its central prediction; together they imply that AI’s effects are layer-dependent.

The immediate driver of the disruption is a step change in the quality, speed, modality, and accessibility of machine translation. Earlier rule-based and statistical systems produced translations that were often awkward and context-blind; neural and generative systems, trained on vast multilingual corpora, produce output that is markedly more fluent and context-aware. The scope of that improvement requires specification rather than general description. Benchmark evaluations report that generative systems translate written text competitively with, and on some pairs better than, dedicated commercial engines when the pair is high-resource and includes English, while performance degrades substantially for low-resource languages and for translation between two non-English languages ([4]; [9]; [13]). Evaluations of this kind assess sentence- and document-level written translation in general and news-like registers; they do not establish equivalent quality for speech, for specialized or high-consequence registers such as legal and contractual text, or for output released without human review. Claims made below about AI-supplied quality are accordingly confined to written and spoken text in routine business registers, in high-resource pairs involving English, and at a level adequate for tasks whose errors are cheap to detect and correct; where any of those conditions fails, the empirical basis for describing the output as high-quality or native-like is presently weak. Equally significant is the move beyond text. Contemporary systems operate across speech, images, and video, enabling live spoken conversation between people who share no common language, and unified multilingual and multimodal models can handle both speech and text across up to roughly one hundred languages within a single system ([33]).

Three features of the current moment distinguish it from earlier waves of translation technology. First, real-time capability: translation now occurs with low enough latency to support live meetings and conversations rather than only the asynchronous rendering of documents. Second, embeddedness: these capabilities are integrated directly into the platforms where business is already conducted, including major videoconferencing and productivity tools, and increasingly run on personal devices, so that using them requires no specialist workflow. Third, breadth of adoption: generative AI has diffused through the economy with unusual speed, with a majority of surveyed organizations reporting regular use across multiple functions by 2025 ([21]). The combination of quality, immediacy, ubiquity, and low cost is what converts a longstanding technology into a genuine structural change in how cross-lingual business communication can be conducted.

The trajectory of the technology matters as much as its current state, because it indicates the direction in which the analysis must anticipate change. Machine translation began as a tool for rendering discrete texts from one language to another, an essentially transactional function performed on demand. The current generation of systems is moving beyond this transactional model toward something more agentic: systems that ingest multiple sources of context, infer communicative intent, and participate in extended, multimodal interaction rather than translating isolated utterances ([13]). Multimodal models that handle speech, text, images, and video within a single architecture and operate across scores of languages exemplify this shift from tool to agent ([33]).

This trajectory has a double-edged effect on the central thesis. On the one hand, more agentic systems encroach further into the surface layer and begin to touch the lower reaches of the deep layer, handling some contextual and pragmatic features that earlier systems ignored, suggesting that the boundary between the automatable and the human will continue to move. On the other hand, the more communication is delegated to autonomous agents, the more acute the risks of agency loss, over-reliance, and the erosion of authentic human relationships become, and the higher the stakes of preserving the human competencies that the deep layer comprises. The trajectory thus does not refute the thesis; it raises the stakes of getting the human-machine division of labor right, and it makes the cultivation of deep-layer competence more urgent rather than less.

The capabilities most directly relevant to the barriers cataloged earlier fall into several categories. In written communication, AI writing assistants and translation tools can correct grammar, refine vocabulary, adjust register, and render text between languages, allowing a non-native professional to produce polished, genre-appropriate documents that would previously have required either advanced proficiency or the services of an editor ([13]). In spoken communication, real-time interpretation reduces the situational and cognitive-load barriers by handling decoding and encoding on the speaker’s behalf, so that a non-native participant can follow and contribute to a fast-moving discussion with less extraneous processing burden. In routine translation and localization, tasks that exporters once outsourced to professional translators—product descriptions, standard correspondence, marketing collateral—can be substantially automated, lowering the cost and delay that previously constrained smaller firms ([26]).

These capabilities map precisely to the surface layer of the two-layer model, addressing the vocabulary, grammar, spelling, comprehension, and real-time processing demands that dominated the earlier literature and led to non-native speakers being underestimated despite their expertise. The mapping is to surface performance rather than to surface competence, and the two can diverge durably. A system that supplies pronunciation, or that renders rapid speech intelligible through live captioning, improves what an interlocutor receives on that occasion without developing the user’s own articulatory or listening capability; the barrier is bypassed rather than removed, and the relief lasts only as long as access to the tool does. In principle, this relief should reduce the anxiety associated with surface errors, broaden participation, and lower the linguistic entry barrier to international business, particularly for individuals and firms that previously lacked the resources to acquire advanced proficiency or to purchase professional language services.

These capabilities manifest differently across business functions. In internal communication, real-time translation and captioning let employees follow and contribute to meetings in a language they do not master, mitigating the marginalization that English-only policies produce ([23]). In external correspondence, AI raises the floor of written quality so that non-native professionals can meet genre expectations without the errors that once undermined credibility ([13]; [41]), and in documentation and routine transactional communication automation is most complete and the gains for smaller firms most direct ([26]). Marketing is more instructive: AI can localize content at scale, but literal localization often renders messages that are correct yet flat or carry unintended associations ([2]); customer service shows the same split between automatable volume and human empathy. In negotiation, where the stakes are highest, AI can support comprehension and drafting but cannot read counterparts, calibrate concessions, or build the trust on which agreement depends.

The capabilities are real, but so are the limits, and they cluster in the deep layer. Current systems struggle with the very features that distinguish high-stakes and culturally sensitive communication: idiom, humor, tone, connotation, and the pragmatic force of an utterance within a specific relationship and cultural setting ([2]). In legal, contractual, marketing, and other high-consequence contexts, the cost of subtle error is high, and human oversight remains necessary to ensure that a translation conveys not merely the words but the intended meaning and effect. The prevailing model in professional practice is accordingly not full automation but human-in-the-loop collaboration, in which AI performs the mechanical work and human experts supply cultural judgment, verify intent, and preserve voice ([2]).

Beyond accuracy, three risks bear on the argument. The first is the erosion of human agency. Delegating the generation of communication wholesale to AI, rather than using it for targeted assistance, may diminish users’ sense of ownership and agency, with implications for motivation and for the relationships built through such communication. This risk follows from social learning theory’s account of agency and self-efficacy ([3]). The second is over-reliance and skill atrophy. If non-native professionals cede language production to machines, the effortful practice through which durable competence and confidence develop may be displaced, consistent with the cautions issued by cognitive load and social learning theory. The third is confidentiality. Cloud-based translation of sensitive commercial information raises data-security concerns, a consideration driving interest in on-device processing that keeps information local. Each risk underscores that AI reconfigures rather than removes the demands on human communicators.

One qualification is essential, because the limits of AI are a moving target: contemporary systems already handle some pragmatic and cultural features—adjusting register, recognizing certain culturally marked expressions, localizing content—so the claim is not that these are permanently inaccessible to machines but that current systems handle them unreliably, especially in the high-stakes, relationship-specific, and low-resource cases where error is most costly ([9]). Whether a core of the deep layer remains resistant to automation even as capability advances—and why accountability, authenticity, and relationship ownership may be structurally rather than merely technically human—is taken up in the discussion of boundary conditions below.

Beyond the cultural and relational limits already discussed, two reliability characteristics reinforce the need for human oversight. First, generative models can produce output that is fluent and confident yet subtly or grossly inaccurate. That very fluency masks the errors—a particular hazard for the non-native user who relies on the tool precisely because they are least equipped to catch them. The competence required to supervise the tool is thus continuous with the competence it ostensibly replaces, which argues against treating AI as a wholesale substitute for language learning rather than an instrument to be wielded with judgment ([9]). Second, systems encode the biases of their training data, which overrepresents some languages and cultural perspectives—English prominent among them—so that output may be fluent yet culturally misaligned with the intended audience. Because the tool’s defaults are not culturally neutral, only a human with the relevant cultural competence can recognize when polished output is wrong for its audience. Reliability and bias thus locate the indispensable human contribution in the same place as the cultural and relational limits: the deep layer of judgment, oversight, and attunement.

A further consequence complicates any expectation that AI will simply level the linguistic playing field. Because these systems are trained disproportionately on English-language data and often perform best in English, and because English frequently serves as an intermediary or pivot language within translation pipelines, generative AI may reinforce rather than dilute the global dominance of English even as it lowers the barrier to participating through other languages. The infrastructure of automated communication is, in an important sense, English-centric, so that English retains a privileged position in the very technology that ostensibly renders it optional. The English imperative, on this view, is not abolished by AI; it is relocated—from the individual speaker, who no longer must personally master the language, to the technical substrate, which encodes English dominance into the machinery of global communication. This paradox is central to a balanced assessment: AI both loosens and tightens the grip of English, and the net effect on any given actor depends on how the technology is used.

This paradox deserves fuller development than it usually receives, because it cuts against the intuitive reading of AI as a straightforward equalizer. The English-centricity of the technology may operate at several levels. At the level of performance, comprehensive evaluations find that systems translate most accurately into and out of English and other high-resource languages, while speakers of low-resource languages—often those in the least advantaged economies—receive a markedly inferior service ([4]; [9]), so that the benefits of AI mediation appear to be distributed in rough proportion to the existing hierarchy of languages rather than in a way that flattens it. At the architectural level, English may function as a pivot in some multilingual translation systems, through which other language pairs are routed, potentially privileging English norms and framings in the output. Furthermore, to the extent that training corpora and the most capable systems are English-dominant, the communicative conventions that AI treats as default may be disproportionately Anglophone, thereby propagating them into the communication of non-Anglophone users. The consequence, on this reasoning, is that AI may lower the barrier to participating in the English-dominated global economy while simultaneously deepening that economy’s English character, extending the reach of English-language norms and standards even to interactions in which no participant is producing English directly. For linguistic diversity, the long-run implications are ambiguous and potentially troubling: a technology that reduces the individual incentive to learn English may nonetheless strengthen the systemic dominance of English, with uncertain effects on the vitality and status of other languages. The paradox thus reframes the equity question, so that the pertinent worry is not only whether individuals can access the tools but whether a communication infrastructure that may encode the advantage of one language can be a neutral equalizer—a question the optimistic narrative of democratization tends to elide. These architectural and normative mechanisms are advanced as reasoned conjecture; establishing them would require direct evaluation of multilingual systems and their training data, which the closing section flags as a research priority.

If widely available tools are commoditizing the surface layer of communicative competence, then the scarce and valuable competence is expected to migrate to the deep layer. Fluency—understood as the capacity to personally produce grammatically correct, appropriately worded English in real time—was historically scarce, costly to acquire, and therefore a source of advantage. It is becoming more abundant because a serviceable version can now be summoned on demand. By the logic of the resource-based view, a capability that is no longer rare or difficult to imitate ceases to confer a sustainable advantage. The differentiating competence is shifting toward cultural fluency: the capacity to communicate appropriately, persuasively, and trustworthily across cultural boundaries; to read and accommodate an interlocutor; to calibrate persuasive appeals to cultural values; and to build and sustain the relationships and reputation on which business ultimately depends.

This migration does not render language competence worthless; rather, its center of gravity moves: the correctness, literal accuracy, and mechanical polish that AI supplies decline in relative value, while the judgment, appropriateness, relational skill, and credibility that it does not supply rise. For the non-native professional, the strategic imperative shifts from eliminating surface errors toward cultivating the cultural and relational capabilities that machines cannot furnish and that AI-generated output, however fluent, cannot guarantee by itself.

Because cultural fluency is the central construct proposed here, it warrants an explicit specification rather than an intuitive one, and that specification must separate the capabilities a communicator holds from the outcome those capabilities produce. Cultural fluency is defined as the capability to communicate appropriately, persuasively, and trustworthily across cultural boundaries, and it comprises three antecedent capabilities. The first is cultural intelligence: knowledge of, and the capacity to adapt to, the norms, values, and expectations of a given cultural setting ([7]). The second is pragmatic competence: the ability to use language to accomplish intended effects in context—managing implicature, register, politeness, and the illocutionary force of an utterance—rather than merely to produce correct forms. The third is relational competence: the capacity to build, accommodate, and sustain relationships through communication, including the continuous adjustment to an interlocutor that accommodation theory describes ([42]). Counterpart-rated trust is positioned as the outcome that the three jointly produce rather than as a fourth capability alongside them.

Separating the antecedents from the outcome resolves a circularity that an earlier formulation of the construct carried. Treating trust-building as a dimension of competence placed trust on both sides of the model, as an element of the capability and as the result that capability produces, and rendered the construct unfalsifiable in a specific way: evidence that trust had been earned would also have counted as evidence that the capability was present. The revised specification removes the difficulty by locating each element where it belongs. Cultural intelligence, pragmatic competence, and relational competence are properties of the communicator, measurable independently of any particular counterpart. Trust is a property of the relationship, measurable only by asking the counterpart, and it accrues over time through the repeated exercise of the three capabilities under conditions of consistency and authentic engagement. The model accordingly predicts that the three antecedents jointly raise counterpart-rated trust, and it is testable in that form: if measured cultural, pragmatic, and relational competence failed to predict counterpart-rated trust in culturally distant business relationships, the specification would be disconfirmed. Table 5 defines the three antecedents and the outcome, distinguishes each from the others, and indicates how each could be measured.

Table 5. Cultural fluency specified as three antecedent capabilities and one relational outcome: definitions, distinguishing focus, and candidate indicators.

Two of the antecedents warrant a clarification the construct table makes visible. Cultural intelligence and relational competence are related but distinct: the former concerns knowledge of and adaptation to a culture in general, the latter the moment-to-moment adjustment to a particular interlocutor, so that a person may possess broad cultural knowledge yet accommodate a specific counterpart poorly, or the reverse. What the earlier inclusion of trust-building was intended to capture is preserved by the revised specification and stated more precisely. Fluent automated output does not by itself earn trust, because the capabilities that generate trust operate on the relationship rather than on the text; trust is therefore the outcome that AI-supplied surface performance leaves unaddressed, and it is reached only through the sustained exercise of the three antecedent capabilities.

The claim that the deep layer endures as human work is contingent rather than categorical, and the conditions carrying it can be stated. The first is reliability: appropriateness, persuasive effect, and relational adjustment are defined relative to a context and a relationship that automated systems, operating on linguistic form, capture only partially and at present unreliably, most of all where the relationship is specific and the register unusual ([2]; [9]; [32]; [37]; [42]). Reliability is an empirical condition and may be met by future systems. The second is accountability: responsibility for what is communicated attaches to a person or a firm, and a system bearing no responsibility cannot discharge it, so a human remains in the position of warranting the communication whatever the system’s fluency. The third is relationship context: trust accrues between counterparties rather than to the tool that mediates them, and in relationship-dependent settings the perception of genuine human engagement is part of what is exchanged. The first condition is expected to weaken as capability advances; the second and third are properties of the social relation within which communication occurs rather than of the text produced, and are correspondingly more durable. The proposition advanced here is therefore that the deep layer remains human-led for as long as reliability is unmet and accountability and relationship ownership stay with human parties—a claim open to disconfirmation rather than a permanent boundary. On the same reasoning, the model predicts that perceived automation may reduce counterpart-rated trust in relationship-sensitive contexts, a prediction that direct empirical work on AI-mediated communication and perceived authenticity would need to test (see Section 11).

This human core was anticipated by the evidence reappraised in Table 2: the exporting study found that successful non-Anglophone entrepreneurs distinguished themselves not by eliminating linguistic error but by cultural adaptation and relationship-building ([26]), and the negotiation and marketing findings decompose into an automatable component and a residual human one of persuasion, cultural resonance, and the reading of counterparts ([24]). The reframing is thus continuous with the evidence, naming the dimension it already implied was decisive.

For non-native speakers, the reframing carries a mix of relief and challenge. The relief is substantial: the surface barriers that generated anxiety, consumed cognitive resources, and led capable professionals to be underestimated can, with judicious use of AI, be materially reduced. A non-native entrepreneur can now enter Anglophone markets, correspond, market, and transact with a level of surface polish that once required either advanced proficiency or costly professional services. To the extent that the earlier literature described a playing field tilted against non-native speakers by surface-level demands, AI tilts it back toward a level playing field.

There is, moreover, a distinctive advantage the reframing brings into focus and that Proposition P4 formalizes: among communicators with comparable AI-supported surface performance, those who have developed bilingual and bicultural competence—disproportionately, though not exclusively, non-native professionals who have long operated across languages and cultures—are expected to hold a relative advantage in culturally and relationally demanding tasks. Bilingualism is not a deficit; not every non-native speaker is bicultural, and some native speakers are highly interculturally competent. The claim is comparative, and it turns on who has cultivated the deep-layer capabilities that now differentiate.

The challenge is that the terms of competition shift to a plane where surface fluency no longer masks or substitutes for deeper competence. When everyone can produce polished English, polished English ceases to differentiate, and the differentiators become cultural intelligence, judgment, and relational skill—capabilities that are not evenly distributed and that AI does not confer. Non-native speakers who invest only in AI-mediated correctness while neglecting the deeper layer may find they have solved yesterday’s problem. Those who use AI to offload the surface burden and redirect their effort toward cultural and relational mastery are best positioned to convert the technology into durable advantage. The strategic counsel that follows from the analysis is thus not to abandon English learning but to reorient it.

Because language competence is entangled with power, a change in who can produce fluent communication is also a change in the distribution of power. The earlier literature showed that fluent speakers enjoyed status, authority, and access that non-fluent speakers were denied, often independently of substantive merit ([23]; [38]). If AI raises the floor of everyone’s surface fluency, it compresses the status differential that surface fluency conferred, potentially enfranchising capable professionals whose contributions were previously discounted because of the way they spoke or wrote. In this respect AI has a genuinely democratizing potential, redistributing communicative power away from the accident of first-language advantage.

A dimension easily overlooked in an efficiency-focused analysis is the relationship between communication, identity, and authenticity. Communication presents a self and builds a relationship. The perception of engaging with an authentic human interlocutor is part of what makes business relationships hold—a limit on delegation qualitatively different from the accuracy limits discussed earlier. Even where an AI system could generate a message indistinguishable in quality from a human-authored one, the knowledge or suspicion that communication has been fully automated may alter how it is received in high-trust, relationship-dependent contexts; the value of human authorship there lies in the signal of attention, effort, and genuine engagement it conveys.

For non-native speakers, this carries a particular poignancy: the accent and the effortful reach for the right word, which the earlier literature treated as deficits, are also markers of genuine cross-cultural effort that can, in the right relational context, build rather than undermine trust. The counsel is therefore to use AI with judgment about when polish serves the relationship and when unmediated human presence serves it better—the scarce competence being not the production of flawless English but the judgment to know what a given relationship and moment require.

The distributive implications follow directly and are formalized in Proposition P3. Insofar as part of the English premium rewarded the scarce ability to produce and comprehend English, that part should compress as AI makes the ability abundant; yet the premium is unlikely to disappear, because the deep layer AI does not supply continues to command a return, and the English-centric character of the AI infrastructure preserves a structural advantage for English and, indirectly, for those fluent in it. The premium is thus restructured rather than eliminated: it decreases rewards for mechanical fluency and increasingly rewards the cultural and relational competence that fluency once signaled but no longer guarantees.

To make the argument falsifiable and to distinguish what the two-layer model predicts from what established competence frameworks already hold, the central claims are stated as five propositions. Each is offered as a prediction for empirical scrutiny rather than as an established finding, and each concerns not the existence of surface and deep competence—long recognized—but the way generative AI is expected to change the relationship between the two layers and the economic value attached to each.

P1 (surface commoditization). As generative AI and machine translation diffuse, the surface layer of communicative competence—vocabulary, grammar, spelling, pronunciation, and literal comprehension—is expected to become increasingly abundant and low-cost, and its scarcity value to decline.

P2 (migration of value). As surface competence is commoditized, the share of variance in business communication outcomes explained by deep-layer competence (cultural fluency) is expected to rise relative to the share explained by surface competence.

P3 (restructured English premium). The labor-market premium associated with English is expected to be restructured rather than eliminated: compressing on its mechanical-fluency component while persisting or rising on its cultural and relational component.

P4 (relative advantage of bicultural competence). Among communicators with comparable AI-supported surface performance, individuals possessing developed bilingual and bicultural competence are expected to demonstrate a relative advantage in culturally and relationally demanding business tasks. The proposition concerns a comparative advantage under matched surface performance, not a claim that bilingualism is a deficit, that all non-native speakers are bicultural, or that all native speakers lack cross-cultural competence.

P5 (entrenchment paradox). AI systems whose training data, evaluation standards, or translation architectures disproportionately privilege English are expected to reinforce its structural dominance even as they lower the barrier to participation without requiring personal mastery of it. Consistent with this conditional form, comprehensive evaluations find that current systems perform markedly better on high-resource languages—English foremost—than on low-resource ones, so that the benefits of automation accrue unevenly across language communities ([4]; [9]).

These propositions are interlocking rather than independent, and the closing section specifies how each could be operationalized and tested. What distinguishes them from prior competence theory is not the surface-deep distinction itself but the directional, AI-driven predictions about commoditization, value migration, and distributive consequence that follow from it—predictions the earlier frameworks had no occasion to make.

The guiding principle is augmentation rather than substitution: deploy AI to handle the mechanical layer—drafting, correction, translation, and real-time comprehension support—while continuing to develop the capabilities that remain human. Augmentation preserves the professional as the author and agent of communication; substitution cedes authorship to the tool and, with it, the agency and authenticity on which competence and trust depend. In practice, this suggests several priorities:

Treat AI as a tool for offloading surface-level linguistic work, freeing cognitive and temporal resources for substantive reasoning, relationship-building, and cultural adaptation rather than as a replacement for communicative competence.Invest deliberately in cultural intelligence: study the business norms, communicative conventions, and values of target markets, since these determine the appropriateness that automated translation cannot reliably supply.Preserve and cultivate genuine language competence sufficient to exercise oversight, because effective human-in-the-loop use of AI requires the ability to judge whether its output conveys the intended meaning and effect.Guard against over-reliance and skill atrophy by maintaining active practice and participation, consistent with the requirements of durable self-efficacy and competence.Retain human agency in relationship-critical communication, using AI for targeted assistance rather than wholesale generation where trust and authenticity are at stake.

The rise of capable AI translation invites firms to revisit the language strategies they adopted in an earlier era. Where a rigid English-only mandate once served to standardize communication, firms now have the option of a more flexible, technology-enabled multilingualism that may reduce the marginalization of non-fluent employees documented in the corporate-language-policy literature ([19]; [23]). Recommended priorities include:

Reassess corporate language policy in light of AI-enabled translation, considering whether flexible multilingual arrangements can capture coordination benefits while more equitably distributing the communicative burden.Invest in cross-cultural and communicative capabilities as a strategic resource, recognizing that, under the resource-based view, culturally intelligent communication is now a principal source of sustainable advantage, whereas mechanical fluency is not.Establish governance for AI-mediated communication, including standards for human oversight in high-stakes contexts and safeguards for the confidentiality of sensitive commercial information.Attend to the human dynamics of AI adoption, since tools that alter who can produce fluent communication also redistribute status and voice within the organization.

A firm that acts on these priorities treats communicative capability as a portfolio to be actively managed rather than a fixed attribute of its workforce: it uses AI to raise the floor of routine communication across the organization—capturing coordination benefits and reducing the marginalization of non-fluent employees—while investing selectively in deep-layer competence where returns are highest, in customer-facing, negotiation-intensive, and culturally distant roles. Managed this way, a potentially disruptive technology becomes a source of both efficiency and inclusion, drawing more fully on the expertise of employees whose contributions were previously muted by surface-level linguistic demands.

The reframing has particularly direct consequences for education, whose traditional emphasis on the surface layer is precisely the terrain that AI now covers. The pedagogical center of gravity should shift toward the deep layer, and toward the meta-skill of using AI well. Priorities include:

Rebalance curricula from a predominant focus on grammatical and lexical correctness toward pragmatic competence, intercultural communication, negotiation, persuasion, and relationship management.Teach AI literacy for communication: how to use translation and writing tools critically, how to verify and edit their output, and how to preserve voice and agency while doing so.Preserve foundational language learning sufficient for oversight and judgment, resisting the assumption that AI availability makes underlying competence unnecessary, and drawing on evidence that unreflective reliance can undermine motivation and learning ([3]).Use AI as a scaffold for participation and practice, harnessing its capacity to lower anxiety and model well-formed language in ways consistent with social learning theory, while ensuring that learners remain active producers rather than passive delegators.

For smaller firms, the analysis is cautiously optimistic. Several of the language-related barriers that historically constrained internationalization—the cost and delay of professional translation and localization, and the difficulty of routine cross-lingual correspondence—are materially reduced by affordable AI tools, lowering the threshold at which a firm can credibly pursue export markets ([26]). The strategic caution mirrors the individual case:

Exploit AI to automate routine translation, localization, and correspondence, redirecting the resources thereby freed toward market-specific cultural adaptation and relationship-building.Retain human expertise for high-stakes communication, including contracts, regulatory documentation, and brand-critical marketing, where subtle error carries disproportionate cost.Recognize that lowered linguistic barriers intensify competition, so that durable success in export markets increasingly depends on the cultural and relational capabilities that distinguished successful exporters even before the AI era.

The professionals whose livelihoods are most directly affected are translators, interpreters, and language-service providers, for whom the automation of routine translation is simultaneously a threat and an opportunity. Survey evidence indicates that a majority of translators expect generative AI to reduce their future incomes, and some question the enduring value of foreign-language skills altogether ([8]). Nevertheless, the same analysis that predicts the commoditization of routine translation also predicts a rising premium on the deep-layer competencies that these professionals uniquely possess. The strategic response mirrors the pattern seen throughout the analysis:

Move up the value chain from routine translation toward high-stakes, culturally sensitive, and creative work—legal and contractual translation, transcreation of marketing materials, and the cultural consulting that surrounds them—where human judgment remains indispensable.Adopt the human-in-the-loop model, using AI to handle volume and drafts while supplying the verification, cultural judgment, and quality assurance that clients increasingly need precisely because AI output is fluent enough to be trusted uncritically.Reframe the professional offering from language conversion to cross-cultural communication consulting, positioning deep-layer expertise—not the mechanical rendering AI now performs—as the core service.

Policymakers concerned with trade, education, and inclusive economic development have a stake in how these benefits and risks are distributed. Access to capable AI tools, and to the complementary skills required to use them well, is uneven, and that unevenness could translate into new inequalities even as old ones diminish. Relevant priorities include:

Support equitable access to AI communication tools and to the digital infrastructure and skills required to use them, so that the lowering of linguistic barriers benefits smaller and less-resourced actors rather than entrenching existing advantage.Reorient language-education and workforce-development policy toward the durable, deep-layer competencies, updating support for English instruction to reflect the changed value of its various components.Attend to the confidentiality, security, and dependency risks of AI-mediated communication in sensitive sectors, and to the implications of an English-centric AI infrastructure for linguistic diversity and equity.

Before turning to boundary conditions, the three research questions posed in Section 1.1 can now be answered directly from the reappraisal.

RQ1asked what common challenges emerge across the three prior publications. Read collectively, the three publications identify a recurring configuration of linguistic, cultural, psychological, and situational barriers that frequently interact and compound one another ([24], [25], [26]). The barrier facing non-native business communicators is thus not primarily lexical or grammatical but a layered combination of surface and deep demands.

RQ2 asked which of these challenges generative AI can substantially reduce, and which continue to require human competence. The reappraisal, interpreted alongside the contextual AI literature, suggests that AI can substantially reduce the linguistic barriers and part of the situational burden—producing fluent text, translating in real time, and lowering the load of operating in a non-native language—while it cannot guarantee the pragmatic, relational, and trust-building competence on which culturally and relationally demanding communication depends. The relief is expected to be real but confined largely to the surface layer.

RQ3 asked how AI reshapes the value of English fluency and the strategic position of non-native communicators. Because AI makes mechanical fluency abundant, it reduces the scarcity value of that fluency and, correspondingly, raises the comparative importance of cultural fluency; the strategic position of non-native communicators shifts from compensating for a surface deficit to cultivating the deep-layer capabilities that now differentiate. The English premium is, on this account, restructured rather than eliminated, as set out in Propositions P2 through P4.

The argument advanced here is deliberately general, and its generality entails boundary conditions and limitations that an honest account must state. Doing so also sharpens the thesis by specifying where it applies most and least forcefully.

The most concrete boundary conditions concern the type of communicative task, and specifying them shows where the two-layer model bites hardest. At one end, routine, low-stakes, transactional communication—standard correspondence, product descriptions, routine customer emails, and the translation of documentation—loads almost entirely on the surface layer, and here AI can dominate with only light human verification. At the other end, communication that is high in stakes and dense in relational and cultural content—strategic negotiations, cross-border mergers and acquisitions, legal disputes, and executive relationship management—loads heavily on the deep layer, and here human judgment must remain in control, with AI confined to a supporting role. Between these poles lie mixed cases such as customer service and marketing, in which AI handles volume and surface rendering while humans provide empathy, cultural resonance, and exception handling. The model therefore predicts not a uniform outcome across business communication but a task-dependent gradient—from near-full automation at the transactional end to reserved human control at the relational end—corresponding to the decision framework set out in Figure 3.

A natural objection is that these boundary conditions are temporary: that a sufficiently advanced AI will eventually acquire cultural intelligence and pragmatic judgment, collapsing the deep layer into the automatable one. The objection has force for the knowledge-based components of cultural fluency, which future systems may increasingly approximate, and the account does not deny that the automatable frontier will continue to advance. However, a core of the deep layer is resistant to automation for structural rather than merely technical reasons, and would survive even a culturally intelligent machine. Accountability—that a person or firm can be held responsible for what is communicated—cannot be delegated to a system that bears no responsibility. Authenticity—the value attached to genuine human engagement—is diminished rather than satisfied by perfect automated imitation, precisely because its worth derives from a human actually standing behind the message. Furthermore, relationship ownership—the fact that trust accrues between the parties, not to the tool that mediates them—remains irreducibly human. These are properties of the social relation within which communication occurs, not features of the text that a more capable generator could produce, which is why the reframing is robust to further advances in AI capability rather than contingent on the limitations of today’s systems.

A further boundary condition concerns the maturity and reliability of the technology. The claim that AI dissolves the surface layer holds most strongly for routine registers and widely resourced language pairs, and least strongly for specialized, high-stakes, or low-resource contexts, where current systems remain error-prone. In legal, medical, and other high-consequence domains, the residual error rate is consequential enough that the surface layer is not fully dissolved but rather rendered cheaper to address under human oversight ([9]). The analysis should therefore be read as describing a strong and continuing tendency rather than a completed transformation, and the balance between automatable and human work will continue to shift as the technology matures.

The second limitation concerns the unevenness of access and the risk of deskilling, which together temper the optimistic reading most sharply. The claim that AI levels linguistic barriers presupposes access to capable tools, adequate connectivity, and the complementary skills to use them critically; where these are lacking—disproportionately in less advantaged economies and among speakers of low-resource languages, whom current systems serve least well—the benefits accrue unequally and may widen rather than narrow existing gaps, an effect compounded by the entrenchment paradox discussed earlier. Access is therefore a policy variable, not a given, and the democratizing potential of the technology is conditional on deliberate choices about its distribution. A distinct but related risk is deskilling. If reliance on AI displaces the effortful practice through which language competence and communicative confidence develop, a generation of communicators may become dependent on tools they cannot adequately supervise, eroding precisely the oversight capacity that responsible use requires and leaving both individuals and firms exposed when the tools fail, mislead, or become unavailable. The cautions of cognitive load and social learning theory apply here with force, and they qualify the more optimistic readings above: the gains from offloading the surface layer are real, but they are contingent on preserving enough underlying competence to keep the human genuinely in the loop rather than nominally so. These considerations do not overturn the thesis, but they mark the conditions under which its benefits are realized and the populations for whom they may not be.

A third limitation is that this is a conceptual synthesis of secondary evidence rather than a presentation of new data. Its central thesis about the migration of value from fluency to cultural fluency is a theoretically grounded prediction. While it is consistent with prior empirical findings and recent evidence on AI capabilities and limits, it has not been tested by a study designed for this purpose. The claims about what AI cannot do rest partly on the current state of the technology, which is changing rapidly, and future systems may encroach further on the deep layer than present ones do. However, the relational and trust-based components of that layer appear more resistant to automation than the cultural-knowledge components. Its value lies in organizing the field and generating testable propositions, and its claims should be held with the provisionality that a fast-moving domain demands.

A further nuance concerns native English speakers, treated here as a comparison class: if the surface advantage of fluency compresses, those whose position rested on native fluency rather than deep-layer competence stand to lose a relative advantage, and native speakers who never had occasion to develop cross-cultural competence may be least prepared for an environment in which it is decisive. The reframing thus runs in opposite directions for native and non-native speakers, and the framework applies to all participants in cross-cultural business communication whatever their first language; the focus on non-native speakers reflects the particular salience of the AI turn for those on whom the burden of the English imperative historically fell.

A fourth consideration is that the two-layer model, like any model, simplifies. The surface and deep layers are not cleanly separable in practice; producing appropriate language and exercising cultural judgment are interwoven in the flow of real communication, and the boundary between them is porous and context-dependent. The model earns its keep as an analytical device that clarifies where value is migrating and how effort should be allocated, not as a literal description of two separable modules. Recognizing its status as a simplification guards against the error of treating the surface layer as unimportant—it remains a necessary condition of communication, now more cheaply satisfied—or the deep layer as wholly independent of linguistic competence, which in fact underwrites the judgment required to use AI well.