Lime, a global micromobility company, implemented Forethought's AI solutions to scale their customer support operations. They faced challenges with manual ticket handling, language barriers, and lack of prioritization for critical cases. By implementing AI-powered automation tools including Solve for automated responses and Triage for intelligent routing, they achieved 27% case automation, 98% automatic ticket tagging, and reduced response times by 77%, while supporting multiple languages and handling 1.7 million tickets annually.
Lime is a global leader in micromobility, providing electric bikes and scooters in over 200 cities across five continents. With a mission to build sustainable, affordable transportation alternatives, the company has powered more than 250 million rides. As with any high-volume consumer service, customer support is a critical operational function, and Lime faced significant challenges in scaling their support operations to match their exponential business growth.
This case study examines how Lime partnered with Forethought, an AI-powered customer experience platform, to automate and optimize their customer support operations. The implementation demonstrates practical applications of AI and LLM-based technologies in production environments for ticket classification, automated response generation, and workflow automation.
Before implementing AI-powered solutions, Lime’s support center operated in a traditional, manual fashion that presented several critical challenges:
Lack of Intelligent Routing: Every support agent handled every type of ticket without any system to prioritize or route specific tickets to specialized agents. This created inefficiencies as agents with specific expertise were not matched to relevant tickets.
Multilingual Complexity: As a global company, Lime receives support tickets in many different languages. Agents frequently had to toggle between Google Translate and their ticketing system to handle common inquiries, creating friction and slowing response times.
No Priority-Based Processing: Despite having strong compliance partnerships with cities, Lime did not prioritize support tickets based on criticality. Critical issues like rider accidents or city official compliance inquiries were handled in the same order as routine questions about charges from weeks ago. This approach posed significant risks to both customer safety and business partnerships.
Limited Self-Service Channels: The company lacked a chat widget or comprehensive self-service options. All support inquiries came through email, phone, the Lime app, or web forms, requiring agent intervention for every case regardless of complexity.
With over 1.7 million tickets annually and exponential business growth, manual processing was unsustainable.
Lime implemented Forethought’s multi-agentic AI platform, deploying several interconnected components to address their support automation needs.
The Triage Agent serves as the initial processing layer for incoming support tickets. It performs two key functions:
Language Detection and Routing: The system automatically identifies the language of incoming tickets and routes them to dedicated language-speaking agents. This eliminates the manual translation workflow and ensures customers receive native-language support.
Category Classification and Priority Assignment: Triage layers language classification with category predictions to determine service levels based on severity. The system uses custom tags and triggers so that when Forethought predicts a specific language and category combination, the case is automatically routed to the appropriate queue. Critical inquiries (such as accident reports or city compliance issues) are now prioritized over routine questions.
The classification system predicts over 2.5 million tags annually for language and category, achieving 98% automatic tagging accuracy. This automated classification directly contributed to a 77% reduction in time to first response.
The Solve Agent handles direct customer interactions through chat widgets and email channels. Key capabilities include:
Intent Interpretation: The system interprets the intent behind customer requests, moving beyond simple keyword matching to understand what customers actually need.
Knowledge Base Search: Solve searches the entire database of knowledge articles and previously resolved tickets to find the most accurate response. This represents a form of retrieval-augmented generation (RAG) where the AI draws from the company’s institutional knowledge to generate responses.
Multilingual Support: The Solve Agent currently supports 4 languages globally, with plans to expand. This enables consistent automated support across Lime’s international operations.
Intelligent Escalation: When customer questions are too complex for automated resolution, the system automatically routes customers to the help center web form, which is then passed to a human support agent.
The Solve Agent deflects approximately 27% of cases that come through email and web channels, representing significant agent time savings at Lime’s scale.
Lime uses Forethought’s Workflow Builder to create Robotic Process Automation (RPA) guided workflows. These automated workflows handle end-to-end processes through pre-set rules and integrate with internal tools to complete specific automation tasks:
These workflows represent deeper integration between the AI layer and Lime’s operational systems, enabling full automation of routine processes without human intervention.
Several aspects of this implementation highlight important LLMOps considerations:
The solution integrates across multiple touchpoints: email, chat widget, web forms, and the Lime mobile app. This multi-channel deployment requires consistent AI behavior and response quality across different input modalities. The system also integrates with internal tools for workflow automation, suggesting API-based connectivity to Lime’s backend systems.
Supporting 4 languages globally with plans for expansion represents a significant production challenge. The AI models must accurately detect language, correctly classify intent across languages, and generate appropriate responses in each supported language. This requires either multilingual models or language-specific model deployments.
Processing 2.5 million tags annually with 98% accuracy demonstrates reliable production performance. However, it’s worth noting that 2% error rate still represents approximately 50,000 incorrectly tagged tickets annually. The system’s integration with human agent queues provides a safety net for misclassified or complex cases.
The architecture maintains clear escalation paths from automated systems to human agents. This hybrid approach ensures that AI limitations don’t result in customer frustration while still capturing efficiency gains on automatable inquiries.
The implementation delivered measurable results:
The case study claims these improvements led to “significant cost savings while improving customer satisfaction,” though specific cost figures are not provided. The efficiency gains are logical given the volume of tickets and the high percentage of automation achieved.
While the case study presents impressive results, several points merit consideration:
Vendor-Provided Data: This case study originates from Forethought’s marketing materials, meaning the metrics presented are self-reported and selected to present the solution favorably. Independent verification of results would strengthen confidence in the claims.
Deflection vs. Resolution: The 27% case automation rate measures deflection, but it’s unclear what percentage of these deflected cases actually resolved the customer’s issue versus merely providing some response. True resolution rates would be a more meaningful metric.
Customer Satisfaction Impact: While improved response times generally correlate with satisfaction, the case study doesn’t provide specific CSAT or NPS metrics to demonstrate actual customer experience improvements.
Complexity of “Simple Questions”: The case study notes that Solve handles deflection of “simple questions,” but the boundary between simple and complex is not defined. Understanding this threshold would help assess the solution’s true capabilities.
Lime plans to expand their use of Workflow Builder to create additional automated processes, which they predict will significantly scale their global support footprint. The company also plans to add more language support through the Solve Agent.
The partnership approach highlighted by Dawn Phelps (Senior Customer Support Content and AI Manager) suggests ongoing collaboration with Forethought for continuous improvement and expansion of AI capabilities, which is a healthy approach to managing production AI systems that require ongoing tuning and enhancement.
This case study demonstrates a practical, production deployment of AI-powered customer support automation at significant scale. The multi-agent architecture—with separate systems for classification/routing, automated response, and workflow automation—represents a modular approach that allows incremental implementation and optimization. The results, while vendor-reported, suggest meaningful operational improvements through AI automation, particularly in multilingual, high-volume customer support environments.
Snorkel developed a specialized benchmark dataset for evaluating AI agents in insurance underwriting, leveraging their expert network of Chartered Property and Casualty Underwriters (CPCUs). The benchmark simulates an AI copilot that assists junior underwriters by reasoning over proprietary knowledge, using multiple tools including databases and underwriting guidelines, and engaging in multi-turn conversations. The evaluation revealed significant performance variations across frontier models (single digits to ~80% accuracy), with notable error modes including tool use failures (36% of conversations) and hallucinations from pretrained domain knowledge, particularly from OpenAI models which hallucinated non-existent insurance products 15-45% of the time.
Coinbase developed CB-GPT, an enterprise GenAI platform, to address the challenges of deploying LLMs at scale across their organization. Initially focused on optimizing cost versus accuracy, they discovered that enterprise-grade LLM deployment requires solving for latency, availability, trust and safety, and adaptability to the rapidly evolving LLM landscape. Their solution was a multi-cloud, multi-LLM platform that provides unified access to models across AWS Bedrock, GCP VertexAI, and Azure, with built-in RAG capabilities, guardrails, semantic caching, and both API and no-code interfaces. The platform now serves dozens of internal use cases and powers customer-facing applications including a conversational chatbot launched in June 2024 serving all US consumers.
A comprehensive overview of ML infrastructure evolution and LLMOps practices at major tech companies, focusing on Doordash's approach to integrating LLMs alongside traditional ML systems. The discussion covers how ML infrastructure needs to adapt for LLMs, the importance of maintaining guard rails, and strategies for managing errors and hallucinations in production systems, while balancing the trade-offs between traditional ML models and LLMs in production environments.
