This case translates a literature review into practical interaction principles for enterprise IM AI assistants, with a focus on the boundary conditions of anthropomorphic cues.
本案例将文献综述转译为企业 IM AI 助手可执行的交互原则，重点讨论拟人化线索的效用边界。
In high AI-literacy teams, does stronger anthropomorphic design always improve human-AI collaboration, or can it reduce efficiency, autonomy, and judgement quality in complex work? (Liu et al., 2025; Carrasco-Farré & Hakobjanyan, 2026)
在高 AI 素养团队中，更强的拟人化设计是否一定提升人机协作？还是会在复杂任务中损害效率、自主感与判断质量？（Liu 等, 2025；Carrasco-Farré & Hakobjanyan, 2026）
In enterprise IM, interaction cadence is high and decision cycles are short. If anthropomorphic framing is overused, users may experience hidden social pressure, slower verification, and reduced confidence in high-stakes contexts (Reeves & Nass, 1996; Carrasco-Farré & Hakobjanyan, 2026).
企业 IM 的交互频率高、决策节奏快。若拟人化表达过强，用户在高风险任务中可能出现隐性社交压力、核验变慢与决策信心下降（Reeves & Nass, 1996；Carrasco-Farré & Hakobjanyan, 2026）。
This is especially salient in teams where AI is already a routine production instrument. In such environments, interaction quality is evaluated less by social warmth and more by controllability, consistency, and traceable reasoning.
这在“AI 已成为日常生产工具”的团队里尤其明显。此时交互质量更少由社交温度决定，而更多取决于可控性、一致性与推理可追溯性。
Interpretation: in early-stage human-AI collaboration, anthropomorphic cues can function as scaffolding, reducing initial uncertainty and lowering entry barriers to task engagement.
理论解释：在人机协作早期阶段，适度拟人化可作为“脚手架”机制，降低不确定感与进入任务的心理门槛。
Evidence base: this aligns with findings from social response paradigms and onboarding-oriented HCI observations (Reeves & Nass, 1996; Liu et al., 2025). The effect is supportive rather than universally optimal.
文献支持：该结论与社会反应范式及上手阶段 HCI 观察结果一致（Reeves & Nass, 1996；Liu 等, 2025）。其作用是“支持性”而非“普适最优”。
Interpretation: as user expertise rises, interaction bandwidth is increasingly allocated to analytical verification and decision synthesis; surplus social framing competes for attentional resources.
理论解释：随着用户专业能力提升，交互带宽将更多用于分析核验与决策整合；多余社交表达会竞争注意资源。
Evidence base: this pattern is consistent with Cognitive Load Theory and Social Facilitation/Inhibition in complex tasks (Sweller, 1988; Bond & Titus, 1983), and with team-level interdependence findings in AI-supported work (Carrasco-Farré & Hakobjanyan, 2026).
文献支持：该模式与认知负荷理论及复杂任务中的社会促进/抑制效应一致（Sweller, 1988；Bond & Titus, 1983），也与 AI 支持工作中的团队互依研究结论一致（Carrasco-Farré & Hakobjanyan, 2026）。
Interpretation: high-literacy users prioritize autonomy, traceability, and parameter-level control. Interaction quality is therefore judged by controllability and reliability over social fluency.
理论解释：高素养用户更重视自主权、可追溯性与参数级控制，因此交互质量更取决于可控与可靠，而非社交化流畅度。
Evidence base: this aligns with Self-Determination Theory and organizational learning perspectives, where stable shared knowledge and explicit artifacts support team performance (Deci & Ryan, 2000; Nonaka & Takeuchi, 1995; Edmondson, 1999).
文献支持：该结论与自我决定理论及组织学习视角一致，即稳定的共同知识与显式产物有助于团队绩效（Deci & Ryan, 2000；Nonaka & Takeuchi, 1995；Edmondson, 1999）。
“High AI literacy” in this project is operationalized using a workplace-oriented Generative AI Literacy (GAIL) framework (Liu et al., 2025), covering five dimensions: foundational technical understanding, prompt optimization, output quality evaluation, innovation-oriented application, and ethical/compliance awareness.
本项目中的“高 AI 素养”采用面向职场的生成式 AI 素养（GAIL）框架进行操作化定义（Liu 等, 2025），包含五个维度：基础技术理解、提示优化能力、输出质量评估、创新应用能力、伦理与合规意识。
At the sample level (n = 78), computed GAIL scores indicate an upper-middle literacy profile (M = 3.84/5, SD = 0.57). Since the original GAIL publication does not provide official proficiency cutoffs, this project uses descriptive operational bands for reporting only (high-operationalized ≥ 4.0: 42.3%; very-high-operationalized ≥ 4.5: 15.4%). These boundaries are anchored to Likert semantics and sample interpretability, not standardized GAIL norms.
在样本层面（n = 78），GAIL 得分呈现中上水平（M = 3.84/5，SD = 0.57）。由于 GAIL 原始文献未提供官方能力分层 cut-off，本项目仅采用描述性“操作化分组”进行汇报（高素养-操作化 ≥ 4.0：42.3%；极高素养-操作化 ≥ 4.5：15.4%）。该分界依据为 Likert 语义锚点与样本可解释性，不代表 GAIL 官方常模。
Design interpretation: retain onboarding-oriented social cues at early stages, then progressively shift toward instrument-like interaction as user competence stabilizes.
设计解释：在上手阶段保留支持性社交线索；随着用户能力稳定提升，逐步收敛为工具化交互。
Evidence support: staged adaptation aligns with literacy-differentiated interaction findings and avoids one-size-fits-all anthropomorphic design (Liu et al., 2025; Carrasco-Farré & Hakobjanyan, 2026).
文献支持：分阶段适配符合素养分层交互结论，可避免“一刀切”的拟人化策略（Liu 等, 2025；Carrasco-Farré & Hakobjanyan, 2026）。
Design interpretation: expose decisive variables, confidence signals, and reversible actions so users can verify and intervene without relying on rhetorical interpretation.
设计解释：显式暴露关键变量、置信信号与可回退操作，使用户可直接核验与干预，而不依赖修辞理解。
Evidence support: explicit control improves perceived competence and reduces cognitive ambiguity in high-complexity workflows (Sweller, 1988; Deci & Ryan, 2000).
文献支持：显式控制在高复杂流程中有助于提升胜任感并减少认知歧义（Sweller, 1988；Deci & Ryan, 2000）。
Design interpretation: use social framing only where coaching or emotional buffering is necessary; for expert tasks, preserve clear role hierarchy and operational accountability.
设计解释：仅在辅导或情绪缓冲场景使用社交框架；在专家任务中，维持清晰角色层级与操作责任边界。
Evidence support: clear boundary management supports team learning efficiency and reduces pseudo-social dependency risks (Nonaka & Takeuchi, 1995; Edmondson, 1999).
文献支持：清晰边界管理有助于团队学习效率，并降低伪社交依赖风险（Nonaka & Takeuchi, 1995；Edmondson, 1999）。
Interaction layer: reduce affective fillers and keep response structure explicit. Control layer: expose key parameters and confidence cues. System layer: tune anthropomorphic intensity based on user literacy and task criticality (Liu et al., 2025; Carrasco-Farré & Hakobjanyan, 2026).
交互层：减少情绪性填充语，保持响应结构显式。控制层：暴露关键参数与置信线索。系统层：根据用户素养与任务关键性动态调节拟人化强度（Liu 等, 2025；Carrasco-Farré & Hakobjanyan, 2026）。
Governance layer: define explicit switching rules for “social mode” (e.g., onboarding, coaching) and “instrument mode” (e.g., auditing, approval, risk triage), and record switching rationale for accountability and later review.
治理层：明确“社交模式”（如上手、辅导）与“工具模式”（如审核、审批、风险分诊）的切换规则，并记录切换依据，保证可问责与可复盘。
Primary method: structured questionnaire based on the five-dimension GAIL construct, with semi-structured interviews for triangulation. Participants were recruited from interaction design, engineering, product, QA, and localization functions in AI-intensive workplace settings (n = 78).
主方法：基于 GAIL 五维构念的结构化问卷，辅以半结构访谈进行三角互证。样本来自 AI 使用密集的职场团队，覆盖交互设计、研发、产品、测试与国际化翻译岗位（n = 78）。
AI literacy distribution: M = 3.84/5, SD = 0.57; high-operationalized band (≥ 4.0): 42.3%; very-high-operationalized band (≥ 4.5): 15.4%. (Project-defined descriptive bands; rationale stated in Section 02.) Follow-up UX validation metrics (task time, error rate, trust/autonomy) are reserved for the next iteration.
AI 素养分布：M = 3.84/5，SD = 0.57；高素养-操作化分组（≥ 4.0）占 42.3%；极高素养-操作化分组（≥ 4.5）占 15.4%。（项目自定义描述性分组，界定依据见第 02 节。）后续 UX 验证指标（任务时长、错误率、信任/自主感）将在下一迭代补充。
Validation plan: run 5-8 participants from high-AI-literacy roles through three core tasks (information retrieval, cross-role handoff, and uncertainty handling); track completion time, clarification turns, and confidence-correctness alignment; accept release when median time decreases by >= 15%, clarification turns drop by >= 20%, and no critical trust mismatch appears in post-task review.
验证计划：招募 5-8 名高 AI 素养岗位用户，完成 3 个核心任务（信息检索、跨角色交接、不确定性处理）；记录任务时长、二次澄清轮次、以及“置信度-正确性”一致性；发布门槛为中位完成时长下降 >=15%、澄清轮次下降 >=20%，且复盘中无重大信任错配问题。
Citation note: Liu et al. (2025); Carrasco-Farré & Hakobjanyan (2026); plus CLT / SDT / Social Facilitation / Organizational Learning frameworks.
引用说明：Liu 等（2025）、Carrasco-Farré & Hakobjanyan（2026），及认知负荷、自我决定、社会促进、组织学习等理论框架。
Reference list for key claims and design translations used in this case study.
用于支撑本案例核心主张与设计转译的参考文献列表。
To increase academic traceability, this section maps core theories to concrete design implications and expected UX effects in enterprise IM AI assistant scenarios.
为增强学术可追溯性，本节将核心理论与企业 IM AI 助手场景中的具体设计含义、预期 UX 效果进行映射。
| Claim研究主张 | Evidence Source证据来源 | Design Translation设计转译 | Validation Indicator验证指标 |
|---|---|---|---|
| In high-literacy expert teams, excessive anthropomorphic wording can add cognitive overhead. 在高素养专家团队中，过度拟人化措辞会增加认知负荷。 | Cognitive Load Theory (Sweller, 1988); GAIL workplace profile (Liu et al., 2025). 认知负荷理论（Sweller, 1988）；职场 GAIL 画像（Liu 等, 2025）。 | Default to compact, structured, tool-like response cards with explicit fields and reduced social fillers. 默认采用结构化、工具化响应卡片，字段显式，减少社交填充语。 | Task completion time, secondary clarification turns, NASA-TLX mental demand. 任务完成时长、二次澄清轮次、NASA-TLX 心理负荷。 |
| Expert users need autonomy and override capability more than affective companionship. 专家用户优先需要自主控制与可覆写能力，而非情感陪伴。 | Self-Determination Theory (Deci & Ryan, 2000); interview observations from AI-intensive roles. 自我决定理论（Deci & Ryan, 2000）；AI 高强度岗位访谈观察。 | Expose parameters, confidence, and rationale; provide one-click edit-and-rerun paths. 暴露参数、置信度与依据；提供一键编辑重跑路径。 | Perceived autonomy/competence scores, manual override rate, rerun success rate. 自主/胜任感评分、手动覆写率、重跑成功率。 |
| Instrumental interaction improves reliability in high-stakes, multi-step collaboration. 在高风险、多步骤协作任务中，工具化交互更有利于稳定性。 | Social Facilitation/Inhibition (Bond & Titus, 1983); team interdependence findings (Carrasco-Farré & Hakobjanyan, 2026). 社会促进/抑制（Bond & Titus, 1983）；团队互赖研究（Carrasco-Farré & Hakobjanyan, 2026）。 | Use role-aware response templates and uncertainty labeling in planning/review phases. 在规划/评审阶段使用角色感知模板与不确定性标注。 | Cross-role handoff errors, review rework ratio, trust calibration (confidence vs. correctness). 跨角色交接错误率、评审返工率、信任校准度（置信度-正确性一致性）。 |
| Scalable team performance depends on explicit and reusable knowledge artifacts. 可扩展团队绩效依赖显式、可复用的知识产物。 | Organizational Learning (Nonaka & Takeuchi, 1995; Edmondson, 1999). 组织学习理论（Nonaka & Takeuchi, 1995；Edmondson, 1999）。 | Standardize prompt conventions, trace logs, and review checklists as shared team assets. 将提示规范、追溯日志与评审清单标准化为团队共享资产。 | Knowledge reuse rate, onboarding time-to-first-correct-output, audit trace completeness. 知识复用率、新人首个正确输出时间、审计追溯完整度。 |