Failures in AI customer service are not just annoyances: poor problem solving and clumsy handovers to human agents cost travel platforms transactions and trust, raise operating costs and carbon footprints, and undermine economic, social and environmental sustainability.

Main Finding

Negative interactions with intelligent customer service (ICS) on travel e‑commerce platforms—captured via 347 user‑reported dissatisfied critical incidents—produce concrete, multi‑dimensional harms to platform sustainability. Failures cluster around six core operational problems (beginning with semantic understanding) and translate into economic losses (transaction loss, higher operating costs), social harms (eroded trust, widened digital divide), and environmental consequences (larger digital carbon footprints and missed opportunities to promote green consumption). The paper argues these microservice failures require a multi‑level collaborative governance response, not just technical fixes.

Key Points

• Core insight: Every failed ICS interaction is a “moment of truth” that can destabilize user–platform relationships and thereby affect long‑run platform sustainability (economic, social, environmental).
• Six recurring problem categories (paper identifies six; the first explicitly listed is):
  • Semantic understanding: poor parsing of complex intents, colloquial or non‑standard inputs.
  • Problem‑solving limitations: inability to resolve non‑routine or multi‑step travel issues.
  • Ineffective human handover: poor escalation or transfer to human agents when needed.
  • Low empathy / impolite human‑computer interaction: perceived coldness or discourteous responses.
  • Poor personalization / localization: inadequate accommodation of diverse user groups and cultural contexts.
  • Interaction inefficiency: prolonged or repetitive exchanges that drive user frustration and resource use.
• Consequences mapped to sustainability pillars:
  • Economic: lost transactions, reduced customer lifetime value, raised operating costs from repeated contacts or forced human intervention.
  • Social: reduced trust, reputational damage, and disproportionate negative effects on less digitally literate users (widening the digital divide).
  • Environmental: longer/inefficient digital interactions increase energy use and carbon footprint; ICS also misses opportunities to nudge users toward greener choices.
• Proposed remedy: a multi‑level collaborative governance framework linking microservice design, operational processes, platform policy, and broader stakeholder coordination.

Data & Methods

• Methodological approach: Critical Incident Technique (CIT) applied to naturally occurring, user‑generated complaints to capture authentic “moments of dissatisfaction.”
• Data collection:
  • Sources: public posts on Weibo, Zhihu, Xiaohongshu (major Chinese social platforms).
  • Period: Aug 24–29, 2025.
  • Search strategy: combined travel platform names (e.g., Trip.com, Fliggy, Booking.com) with keywords such as “intelligent customer service,” “AI customer service,” “not useful,” “hard to reach human agent.”
  • Screening: removed vague venting, complaints focused solely on human agents or unrelated services, promotional or duplicate posts.
  • Final sample: 347 valid dissatisfied critical incidents; all publicly available and anonymized for analysis.
• Analysis:
  • Qualitative coding and thematic consolidation to derive six categories of dissatisfied incidents.
  • Cross‑mapping of incident types to sustainability outcomes (economic, social, environmental).
  • Development of prescriptive recommendations and a multi‑level governance framework (conceptual rather than quantitatively estimated).
• Limitations noted (implicit/derivable):
  • Single short collection window and Chinese social media sources may limit generalizability.
  • Reliance on negative, self‑selected public posts (no random sampling; no quantitative causal estimates).

Implications for AI Economics

• Reassess AI ROI metrics: Economic evaluations of ICS should include downstream costs of failed interactions (churn probability, lost revenue per failed incident, incremental human‑agent costs) not just reduced headcount or response speed gains.
• Incorporate externalities into platform valuation:
  • Social capital (trust) is an economic asset; systematic ICS failures can reduce platform demand and competitive position.
  • Digital inclusion has economic impacts—marginalized user segments may be lost customers if ICS remains unusable for them.
  • Environmental externalities (digital carbon) are becoming economically salient via regulation, carbon reporting, or consumer preferences; inefficient ICS increases these costs.
• Design incentives and governance:
  • Platforms should internalize the cost of escalation and failed resolution in AI deployment decisions (human‑in‑the‑loop thresholds, escalation quality metrics).
  • Market and regulatory incentives could promote ICS transparency (explainability of answers, clear handover processes) and politeness standards, which have measurable economic benefits via trust retention.
  • Encourage investment in capabilities that matter for sustained value: improved intent recognition, robust escalation, personalization for low‑digitally‑literate users, and features to nudge green choices.
• Measurement recommendations for economists and platform managers:
  • Track per‑incident expected cost = P(failure) × [direct cost of resolution + expected lost lifetime value + reputational/marketing cost].
  • Add non‑accuracy ICS KPIs into business dashboards: handover success rate, time‑to‑human when needed, user‑reported politeness/satisfaction, and estimated energy per interaction.
  • Quantify environmental impact: estimate average energy/CO2 per interaction and model reductions from shorter, more effective exchanges and from using ICS to promote greener options.
• Research directions:
  • Quantitative estimation of the economic cost of ICS failures (causal effect on churn, transaction value).
  • Cross‑market studies comparing cultural/localization effects on ICS economics.
  • Experimental tests of governance interventions: improved handover protocols, empathy‑aware dialogue policies, and green‑choice nudges to measure ROI and carbon co‑benefits.

Limitations of this summary: the source paper is qualitative and conceptual; specific numeric estimates of economic and environmental impacts are not provided. The paper’s multi‑level governance framework is proposed at a conceptual level; operationalization and cost‑benefit quantification remain open tasks for AI economics research.