Large language models can speed multilingual coordination in global projects but often mirror dominant cultural norms and lack robust evaluation of culturally appropriate communication, risking pragmatic misalignment in cross‑cultural teams.

Main Finding

LLMs can improve selected efficiency and coordination outcomes in cross-cultural project teams (e.g., drafting, translation, routine coordination), but they also risk reproducing dominant (often Western-centred) cultural norms and producing pragmatically inappropriate outputs when deployed without human contextualisation. Existing evidence is concentrated on semantic/task-oriented metrics, leaving major gaps in understanding LLMs’ effects on cultural‑pragmatic alignment, relational trust, and governance in global projects.

Key Points

• Four thematic domains synthesised:
  • A: Human–AI collaboration in project management — LLMs as assistants for drafting, summarisation, translation, and routine coordination tasks.
  • B: Cross-cultural communication & cultural intelligence (CQ) — project outcomes depend on culturally sensitive timing, tone, role boundaries and trust; CQ is a useful conceptual lens for what “culturally adaptive” should mean.
  • C: LLM cultural/multilingual optimisation strategies — approaches include multilingual pretraining, fine‑tuning, prompting strategies, localized data augmentation, and human‑in‑the‑loop adaptation; many models still biased toward English/Western corpora.
  • D: Gaps in measuring cultural–pragmatic performance — most technical evaluations use semantic similarity, BLEU/ROUGE-type metrics, or task accuracy; few measure pragmatic appropriateness, relational effects, or culturally contingent misalignment.
• Risk profile:
  • Semantic adequacy ≠ pragmatic appropriateness: outputs can be fluent yet culturally inappropriate (e.g., tone misfit, misinterpreted politeness or status cues).
  • Reproduction of dominant norms: training data imbalance leads to systematic cultural bias in outputs.
  • Institutional risks when teams rely on LLM outputs without professional mediation or accountability.
• Methodological and scope notes:
  • Evidence is interdisciplinary but dispersed; technical papers, PM literature, translation/interpreting studies and CQ research have developed largely in parallel.
  • The review highlights the need to integrate translation/interpreting expertise and CQ considerations into LLM deployment in projects.

Data & Methods

• Review type: Scoping review following Arksey & O’Malley framework, PRISMA-ScR and JBI recommendations; thematic analysis with four a priori domains.
• Corpus: 69 interdisciplinary studies synthesised.
• Sources searched: Web of Science, Scopus, ABI/INFORM, EBSCO Business Source, plus targeted screening of ACL/NeurIPS proceedings and relevant arXiv preprints.
• Timeframe of search: October 2025 – January 2026; no publication-year restriction.
• Inclusion criteria: English-language publications with sufficient methodological detail; focus on cross-cultural project communication, LLMs, CQ and evaluation.
• Coding/synthesis: Deductive thematic scaffold (A–D) applied; Zotero for reference management and Excel to document screening and coding.
• Limitations acknowledged by authors:
  • English‑only search may bias thematic distribution and underrepresent locally published/non‑English evidence.
  • Predefined domains narrowed scope and possibly excluded adjacent, relevant work.
  • The heterogeneous nature of studies precluded meta‑analytic effect estimates.

Implications for AI Economics

• Productivity vs. relational costs: Economic evaluations of LLM adoption in international projects must include not only task-time savings (drafting, translation) but also potential negative externalities — increased misunderstandings, reduced trust, coordination breakdowns — that can impose transaction and rework costs.
• Labour-market impacts:
  • Demand shift rather than simple displacement: greater reliance on LLMs for first‑pass work may reduce certain routine translation/drafting tasks but increase demand for culturally specialised reviewers, localisers, and human-in-the-loop QA.
  • Valuation of cultural expertise: translation/interpreting professionals and cross-cultural project managers perform economic value that LLMs do not fully substitute; firms should price and budget for that expertise when deploying LLMs.
• Investment priorities:
  • Data diversity & localisation: investing in culturally representative training/fine‑tuning data and regionally targeted evaluation yields higher expected returns by reducing cultural-misalignment risks.
  • Human-in-the-loop governance: allocating resources to review, escalation paths, and role definitions (who accepts LLM suggestions) can materially change realized benefits.
• Measurement & metrics for cost–benefit analysis:
  • Current LLM metrics (BLEU, ROUGE, semantic similarity) are inadequate for economic modelling of cross-cultural deployments. New metrics should quantify pragmatic alignment, impact on trust, frequency/severity of miscommunication, and downstream rework costs.
  • Firms and researchers should run RCTs or quasi‑experimental evaluations comparing project outcomes (time-to-complete, budget variance, stakeholder satisfaction, rework incidence) with and without LLM mediation.
• Governance & market regulation:
  • Standards and auditing for cultural/pragmatic performance (including documentation of cultural coverage, provenance of localized data) will shape market trust and adoption rates — potential regulatory or procurement requirements could affect costs of adopting LLMs in global projects.
• Research directions relevant to AI economics:
  • Develop empirical estimates of productivity gains net of cultural-friction costs across sectors and languages.
  • Model complementary labor reallocation (how much human cultural-review capacity is needed per unit of LLM-assisted output).
  • Design pricing models and procurement contracts that internalise cultural‑misalignment risk (e.g., warranties, liability allocations, mandated human sign‑off for high‑stakes communications).
  • Create pragmatic evaluation metrics that can be used to predict economic impacts and incorporate them into ROI frameworks.

Suggested short actionable steps for economic stakeholders: - When piloting LLMs in cross-border projects, measure both efficiency (time/cost savings) and relational outcomes (stakeholder satisfaction, rework). - Budget explicitly for human cultural review and include CQ expertise in project teams. - Prioritise investments in localized training data and evaluation protocols before full deployment. - Collaborate with translators/interpreters to define acceptable pragmatic thresholds; incorporate those thresholds in procurement and auditing.

If you want, I can draft a short economic model outline (inputs, parameters, and equations) to estimate net benefits of deploying LLMs in a cross-cultural project setting that includes cultural‑misalignment risk.