Summary of "MSC 2026 | 사례 세션: The Secret Agent - AI 서비스 구축 사례 (Channel Corp., Wrtn AX, LBOX, Config)"
Overview
Four customer case studies demonstrating how to design, operate, and scale AI services, with emphasis on agentic AI, operations, data and model strategy, and infrastructure. Common theme: move beyond one-off POCs to robust, production-grade AI that acts (calls tools, executes tasks) and is operable at scale.
Key focus: building AI services that judge and act (tool-calling, orchestration, long context) rather than relying on brittle static workflows.
Channel Corporation — Channel Talk / ALP (ALF)
Product
- Channel Talk: SaaS conversational/CRM platform.
- ALP (ALF): counseling/agent product that answers knowledge queries and automates operational tasks via a “Task” engine that sequences actions and decisions.
Features & Architecture
- Knowledge-based QA combined with task automation (agents execute multi-step operational flows).
- Branch-aware answers (store/branch-specific knowledge management).
- Voice agents: custom TTS model trained from scratch to improve counselor-like tone and accurate numeric pronunciation to reduce hallucinations.
- Heavy LLM usage with many LLM calls per conversation; moving toward agentic search to better navigate fragmented/unstructured client knowledge.
Metrics & Operations
- Consultation resolution rate: ~70% for ALP.
- ALP MRR: approximately 250M KRW (Feb 2026).
- Volume: ~160,000 AI-handled consultations/week; ~840M tokens processed; tens of thousands of requests — stability is critical.
Platform & Roadmap
- Platform choice: Azure OpenAI (model availability, TTFT = time-to-first-token, SLA guarantees, data security/retention, responsive Microsoft support).
- Roadmap: agentic search, longer context inputs, improved LRM (LLM + retrieval/model orchestration), and leveraging consultation data as a moat.
Lytten AX (Wrtn AX)
Company & Goals
- Consumer AI app scaled to ~7M MA; expanding into B2B.
- Key problems: keep AI costs sustainable while offering free access, and increase internal productivity with limited AI talent.
Technical Approach
- Model orchestration / multi-model strategy: route simple/fast queries to cheap models and use higher-quality models for complex requests to balance cost and quality.
- Built internal agent classes: CX agents, financial agents, front-end/back-end development agents, and data agents (natural language → SQL + visualization).
Lessons for AI Transformation (AX)
- Success factors: identify true productivity bottlenecks; clearly separate internal vs. external technical responsibilities; secure management buy-in.
- Failure risks: mismatched quality expectations, cost misestimation, missing data/infrastructure, poor UX, and low data literacy.
- Process advice: diagnose problems first, build agents with clear ROI, integrate human-in-the-loop and feedback loops, and design solutions to scale.
Platform
- Heavy Azure usage for scaling, rapid model experimentation, and cost/performance tuning.
L-Box (legal tech) — Box AI
Product & Rationale
- Products: legal judgment search and Box AI.
- Moved from static workflow-based features to an agentic architecture because legal work requires high accuracy, deep reasoning, and traceability.
Why Agentic Design
- Static workflows limit depth and scalability and do not fully benefit from improved LLM capabilities.
- Agentic architecture uses a main LLM agent to plan and orchestrate sub-agents/tools, enabling:
- Better utilization of improved LLM reasoning.
- Easier addition of tools/sub-agents.
- Improved handling of complex legal intents and multi-step tasks (document search, drafts, file processing).
Evaluation & QA
- Emphasis on domain-specific benchmarks (e.g., bar exam questions) to measure legal knowledge and reasoning.
- Need for a virtuous cycle: product → user feedback → model improvement → A/B testing and metrics.
Operational Concerns & Roadmap
- Stability is paramount for enterprise legal users; token usage is high so platform reliability (Azure OpenAI cited) and integrated model management are critical.
- Roadmap: behavior-driven legal workflows that propose solutions while leaving final judgment to lawyers.
Config (robotics / physical AI)
Business & Core Problem
- Provides data + model services enabling robot automation.
- Core problem: shortage of robot teleoperation data; large amounts of human demonstration data are abundant but not directly usable by robots.
Technical Solution
- Human-to-robot data conversion model: uses a small set of robot data plus large human datasets to synthesize robot-format training data with high precision.
- Train a foundation model (video encoder + language encoder + transpo-decoder) on 100k+ hours of converted data; fine-tune with ~10 hours of teleop data for specific tasks (target >80% success).
- CI/CD-like automated pipeline: data capture → conversion → foundation training → fine-tune.
Data Ops & Infrastructure
- Data factories in Vietnam (human demonstrations) and Seoul (robot teleop) producing large volumes (claimed totals: 100k hours, 3k tasks, 4M episodes).
- Compute-heavy infra using Azure services (CosmosDB/SQL DB, Blob Storage, AKS) to store/serve data and run computation.
- Emphasis on elevating physical-world events to software to enable agent orchestration over physical models.
Vision
- Connect low-level physical intelligence models to agents for long-horizon, multi-step automated tasks in factories and homes.
Cross-cutting Technical Themes, Recommendations, and How-to Guidance
- Operation-as-product: design for production from the start — consider latency, throughput, governance, and observability; POCs must evolve into operational services.
- Model strategy: employ multi-model orchestration to balance latency, cost, and quality; avoid reliance on a single best model.
- Data-first design: focus on data flows, structure, and feedback loops; consultation/user interaction data is often the moat.
- Agentic AI: prefer agents that judge and act (tool-calling, orchestration, long context) rather than brittle static workflows — this enables scalability and continuous improvement as models improve.
- Benchmarks & evaluation: use domain-specific evaluation sets and internal benchmarks for iterative development.
- Stability & infra: enterprise services must ensure SLA-backed, low-latency serving (TTFT) and robust data security/retention — Azure OpenAI was frequently cited.
- Organizational/process advice: diagnose productivity bottlenecks first, align management, decide internal vs. external responsibilities, and build human-in-the-loop feedback and governance.
Products / Platforms Mentioned
- Channel Talk (Channel Corp) — ALP/ALF agent product + custom TTS
- Lytten AX — agent & data agents (NL→SQL, visualization)
- Box AI (L-Box) — agentic legal assistant
- Config — robotics foundation model pipeline
- Azure OpenAI, Microsoft Foundry (integrated platform for models, data, infra, ops)
- Additional components: custom TTS training, ReRank modules, document intelligence models
Main Speakers / Sources
- Dongjun Baek (Microsoft) — session host / opening & summary
- Jeong-hee Hwang — AI Team Lead, Channel Corporation (Channel Talk / ALP)
- Kim Tae-woo — Lytten AX (Wrtn AX)
- Choi Jin-hwan — L-Box / Box (legal tech, Box AI)
- Seo Jun-hoe — Config (robotics / foundation models for physical tasks)
- Microsoft Azure / Microsoft support team referenced heavily as platform/support partner
Category
Technology
Share this summary
Is the summary off?
If you think the summary is inaccurate, you can reprocess it with the latest model.
Preparing reprocess...