NICE Taps Desktop Mining to Move Enterprise AI Pilots to Production

The Million-Dollar Pilot Trap: Why 89% of AI Agents Never Scale With NiCE — Ai4

Ai4

Enterprise AI pilots fail to deliver ROI because companies deploy costly frontier models on isolated web chat silos instead of automating proven workflows directly from historical interaction data.

AI-assisted summary of The Million-Dollar Pilot Trap: Why 89% of AI Agents Never Scale With NiCE — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 19, 2026, this session features Kevin Lee, Field CTO and Key Pursuits Leader at NICE. According to the video description, the talk addresses why enterprise AI agent initiatives frequently stall in the pilot phase and details the operational foundations, governance frameworks, and scalable architectures necessary to deliver measurable business value in customer experience.

## TL;DR Kevin Lee explains why most enterprise AI agent pilots fail—pointing to excessive token costs, undefined ROI, and isolated chatbot experiments—and presents a proven framework for scaling AI across enterprise contact centers. He outlines NICE's three core pillars for production AI: actively analyzing interaction data, deploying domain-specific models tailored to successful resolutions, and orchestrating seamless omnichannel journeys across digital, voice, inbound, and proactive outbound touchpoints. Drawing on real-world deployments with TripAdvisor, Lufthansa, and Toyota, Lee demonstrates how automated desktop activity mining and historical interaction analysis create resilient, high-ROI silicon agents that effectively collaborate with human teams.

## Atmosphere & style The presentation delivers a pragmatic, fast-paced, and enterprise-hardened tone that cuts through generative AI hype with clear operational realism. Lee commands the stage with confident, contrarian authority, trading buzzwords for hard-nosed industry metrics, relatable pain points, and vivid analogies. It feels like an executive-level briefing crafted for CX leaders who are weary of science experiments and eager for battle-tested engineering blueprints.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) Most enterprise AI initiatives fail to scale: 66% remain stuck in pilots, 95% lack measurable ROI, and 40% are projected for cancellation by 2027 due to steep token economics, unclear value, and governance breakdowns. - [02:26](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=146s) Scaling enterprise AI successfully depends on three pillars: continuous inspection of interaction data, purpose-building agents around proven human resolutions, and orchestrating journeys seamlessly across silicon and human agents. - [02:26](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=146s) Effective artificial intelligence requires historical domain intelligence; specialized CX models focused on customer resolutions outperform generic frontier LLMs answering open-domain queries. - [05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Organizations should avoid deploying costly generative frontier models for simple, repetitive inquiries like order tracking that are better resolved deterministically or via lightweight agentic workflows. - [05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Analyzing 100% of omni-channel voice and digital interactions reveals true customer intent and determines whether automating a dialogue yields justifiable business ROI. - [05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Monitoring agent desktop actions inside systems like Salesforce and SAP enables platforms to automatically synthesize proven prompts and equip virtual agents with precise back-end execution tooling. - [09:30](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=570s) Restricting AI pilots to web chat widgets traps deployments in low-volume silos, leaving high-volume voice and digital messaging channels unaddressed. - [09:30](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=570s) Customer experience orchestration must evolve beyond reactive inbound routing by leveraging telemetry to trigger proactive outbound outreach across voice and digital channels before inbound volume spikes. - [12:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=772s) Agentic platforms act as digital shock absorbers during crisis events, demonstrated by Lufthansa autonomously rebooking travelers at a peak throughput of 12,000 calls per minute during a strike. - [12:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=772s) Connecting AI agents directly to real-time IoT telematics enables fully autonomous, outbound operational workflows, as demonstrated by Toyota scheduling vehicle service from CAN bus alerts. - [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) Before launching AI pilots, enterprises should replace anecdotal whiteboarding with data-backed scrutiny of actual contact volumes, ROI viability, and potential handoff friction with legacy systems.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) The Enterprise Pilot Trap

Kevin Lee introduces himself as the Field CTO and Key Pursuits Leader for NICE, describing the company as the world's preeminent contact center and customer experience (CX) purveyor, facilitating interactions between consumers and the global brands they transact with. Addressing the convention, he immediately surfaces the three most overused yet painful statistics hovering over enterprise AI: 66% of enterprises are currently stuck in pilots, 95% fail to deliver measurable ROI, and 40% of AI projects are projected to be canceled by 2027. Most organizations are desperately seeking ways to be "on the other side of history" regarding these metrics.

The underlying operational symptoms plaguing these stalled pilots are widely felt across the industry: sky-high operational costs driven by steep token economics, vague or non-existent ROI, and critical governance failures. Rather than running into these roadblocks blindly, enterprises must streamline these challenges up front. NICE approaches this with substantial operational credibility, supporting over 27,000 customers across 150 countries, orchestrating more than 20 billion complex voice and digital interactions across channels like SMS, RCS, WhatsApp, LINE, WeChat, and Viber, and processing over two trillion words every single month.

- **Enterprise AI Pilot Reality**: 66% stuck in pilots; 95% fail to show measurable ROI; 40% projected for cancellation by 2027. - **Scale of NICE Operations**: 27,000+ customers globally across 150 countries; 20+ billion interactions orchestrated; 2+ trillion words processed monthly. - **Top Pitfalls**: High run costs/token economics, unclear business ROI, and governance failures.

---

[02:26](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=146s) The Three Pillars of Production AI

To separate failed science experiments from scalable enterprise deployments, successful brands rely on three fundamental pillars. The first pillar is knowing your data ("KYC" or know your customer/data). Successful organizations do not simply dump interaction records into an Enterprise Data Warehouse (EDW) to collect digital dust for audits and compliance; they actively inspect and understand that data at scale. The second pillar is purpose-building AI agents around demonstrably successful real-world interactions rather than guessing or delegating interaction design to a frontier foundation model. The third pillar is seamless orchestration across the full customer journey. As Lee notes, "when all you have is a hammer, everything looks like a nail," but real CX requires meeting consumers wherever they are—whether through an inbound phone call, proactive RCS/SMS, an automated "silicon agent," or a carbon human agent.

```mermaid flowchart LR accTitle: Three Pillars of Scalable Enterprise AI subgraph P1["Pillar 1: KYC / Data"] A["Inspect Interaction Data"] --> B["Identify Best Outcomes"] end subgraph P2["Pillar 2: Purpose-Built AI"] C["Train from Successful Dialogues"] --> D["Deterministic or Agentic Models"] end subgraph P3["Pillar 3: Omnichannel Orchestration"] E["Inbound / Outbound Channels"] --> F["Silicon + Human Collaboration"] end B --> C D --> E ```

Lee offers a contrarian perspective: highly effective artificial intelligence requires actual, historical intelligence first. While NICE heavily leverages LLMs, their stack is enhanced by large language models, not merely built on top of raw third-party endpoints. Backed by 25 years of analyzing interactions for highly complex, strictly regulated global enterprises, NICE has developed over 1,000 domain-specific CX models. Unlike generic market models, these models ignore irrelevant tasks like retrieving stock tickers, reporting weather, or giving driving directions; they are hyper-focused on moving a customer's journey "from PI to buy in the fastest, most effective, and satisfactory manner possible."

- **Pillar 1**: Know your data—move beyond passive EDW storage to active real-time inspection. - **Pillar 2**: Purpose-build AI agents grounded in verified positive business outcomes. - **Pillar 3**: Deliver end-to-end orchestration across inbound, outbound, digital, voice, silicon, and carbon agents. - **CX Domain Focus**: Over 1,000 pre-built CX models engineered specifically for customer resolutions, not generic open-domain trivia.

---

[05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Automated Insights and Auto-Building Prompts

A key hot take shared by Lee is that enterprises do not always need a generative frontier model burning $5 worth of tokens just to answer a routine question like "Where is my order?" Many interactions are better served deterministically or through lightweight agentic workflows. NICE utilizes an engine called Automated Insights, which harnesses roughly a dozen CX models to scan 100% of an enterprise's interaction data—spanning PSTN, WebRTC, SIP voice calls, and digital text across multiple languages—to determine actual interaction outcomes and identify high-value automation levers.

```mermaid flowchart TD accTitle: Automated Insights to Agent Deployment Pipeline subgraph Ingestion["Interaction Ingestion"] A["100% Voice & Digital Data<br>(PSTN, WebRTC, SIP, Text)"] end subgraph Analysis["Automated Insights"] B["Analyze Intents, Actions & Outcomes"] C["Score Automatability & ROI"] end subgraph Generation["Agent Generation"] D["Extract Proven SOPs & Prompts"] E["Observe Desktop Glass Activity<br>(CRM, ERP, Billing Tools)"] end subgraph Deployment["Cognigy Platform"] F["Publish Purpose-Built Agents<br>(Voice, Web Chat, Messaging)"] end

A --> B B --> C C --> D C --> E D --> F E --> F ```

The system evaluates whether an interaction is truly automatable—typically short, definitive dialogues with clearly defined boundaries where a real resolution can be achieved, rather than merely logging a ticket. Furthermore, this analysis answers the critical question: "Is the juice worth the squeeze?" Organizations often embark on building complex API integrations or new Model Context Protocol (MCP) harnesses for use cases that might be more cheaply and effectively handled by business process outsourcing (BPO) partners.

When an opportunity demonstrates clear ROI, NICE's platform auto-builds the prompts and tooling directly from winning historical dialogues. Instead of teams of prompt engineers spending weeks manually tweaking text in isolation, the platform extracts proven conversation paths and continuously updates them as standard operating procedures (SOPs) evolve. Additionally, because human agents work within the NICE platform glass, the software monitors their desktop interactions—observing when they Alt-Tab into Salesforce, SAP, ServiceNow, Sabre, Fiserv, or Jack Henry. By understanding the precise buttons and internal systems human workers touch, the platform automatically equips virtual agents with the exact tools needed to resolve transactions autonomously rather than escalating to a human agent at a $10 penalty.

- **Deterministic vs. Generative**: Avoid using costly frontier models for routine deterministic tasks like basic order tracking. - **Automated Insights**: Ingests 100% of multi-modal, multilingual interactions to identify clear automation targets. - **Desktop Activity Mining**: Observes agent interactions with core back-ends (Salesforce, SAP, ServiceNow, etc.) to expose actionable tooling for AI bots. - **Continuous Prompt Refinement**: Derives prompt templates from successful human conversations rather than manual prompt engineering.

---

[09:30](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=570s) Full-Lifecycle Orchestration: Breaking Beyond the Web Chat Silo

A widespread trap in AI agent pilots is settling for low-hanging fruit—most commonly web chat widgets. Teams deploy a chatbot, consider it an easy starting point, and subsequently discover that web chat accounts for only a sliver of their total contact volume, leaving the core scaling problem unsolved. Through NICE's integration with Cognigy, purpose-built agents escape this trap by deploying across both digital and voice channels without re-architecting the core logic.

Beyond multi-channel delivery, true orchestration must span the entire customer relationship lifecycle: discovery, evaluation, purchasing, and ongoing support. Crucially, it must handle both inbound and outbound directionality. Inbound ACD and IVR routing have long dominated call center thinking, but modern CX requires proactivity. When telemetry and systems indicate an impending issue, virtual agents should proactively outbound a phone call, SMS, or email with full historical context, resolving the issue before an inbound spike ever hits the contact center. Bringing customer data, agentic tooling, human augmentation, and multi-channel orchestration into a single platform like NICE CXone prevents pilots from floundering in isolated architectural silos.

| Pilot Trap Approach | Unified CXone Orchestration | |:--- |:--- | | **Channel Scope** | Web chat only (low-volume island) | Omnichannel (Voice, WebRTC, SMS, WhatsApp, Web) | | **Directionality** | Reactive inbound only | Proactive outbound triggered by contextual telemetry | | **Lifecycle Reach** | Post-purchase care or ticket logging only | Full lifecycle: Discovery, Evaluation, Purchase, and Care | | **System Architecture** | Disconnected LLM experiments & one-off harnesses | Unified platform combining data, silicon bots, and human agents |

---

[12:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=772s) Real-World Production Deployments: TripAdvisor, Lufthansa, and Toyota

NICE operates at massive enterprise scale, supporting organizations that range from small two-person contact centers to operations running hundreds of thousands of concurrent human and silicon agents. This scale demands strict adherence to rigorous compliance standards, regularly navigating multi-syllabic regulatory frameworks like HIPAA, PCI-DSS, and CFPB mandates. Lee details three high-profile production implementations:

- **TripAdvisor**: Deployed enterprise-grade agents into full production in just two and a half months, driving customer satisfaction up to 90% while reducing average handle time by two full minutes per call. - **Lufthansa**: When a pilot union strike abruptly grounded flights and created massive travel disruption, NICE’s Cognigy agentic platform acted as an immediate "digital shock absorber." It intercepted inbound volumes, automatically rebooking and rerouting impacted passengers at a peak throughput of 12,000 calls per minute—averting operational chaos. - **Toyota**: Partnering with Toyota's Connected Care group, the platform monitors real-time vehicle telematics via IoT. When a vehicle's CAN bus signals that scheduled or urgent maintenance is due, the platform proactively initiates outbound messaging, engages the driver with an agentic virtual assistant, and schedules the service transaction seamlessly.

---

[15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) Questions to Ask Before Your Next Pilot

Before launching another AI pilot, Kevin Lee urges enterprise leaders to step away from subjective journey mapping and Post-it note brainstorming sessions, and instead confront hard, data-backed operational questions:

1. **What are customers actually calling about?** Stop guessing top intents via anecdotal whiteboarding exercises. Quantify real contact volumes to ensure proposed automations align with primary customer friction points. 2. **Is the juice worth the squeeze?** Ensure the selected use case will measurably shift business KPIs, MBOs, and ROI metrics before investing months of engineering overhead into complex API integrations. 3. **Where does the handoff break?** Inspect exactly where multi-modal transitions or silicon-to-human agent escalations fail. Determine whether an agent cannot contain an issue because of missing tool integrations, or whether workflows are fundamentally hamstrung by legacy architecture like "an AS/400 in a basement somewhere."

Lee closes by inviting attendees to booth 917 to see NICE Automated Insights in action and watch end-to-end agents built in under ten minutes—demonstrated live on the convention floor by ordering automated coffees and lattes directly via an iPad-based virtual agent.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) 66% of enterprises are currently stuck in pilots, 95% fail to deliver measurable ROI, and 40% of AI projects are projected to be canceled by 2027. — Kevin Lee

> [02:26](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=146s) when all you have is a hammer, everything looks like a nail — Kevin Lee

> [02:26](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=146s) from PI to buy in the fastest, most effective, and satisfactory manner possible. — Kevin Lee

> [05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Where is my order? — Kevin Lee

> [05:19](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=319s) Is the juice worth the squeeze? — Kevin Lee

> [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) an AS/400 in a basement somewhere. — Kevin Lee

> [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) What are customers actually calling about? — Kevin Lee

> [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) Where does the handoff break? — Kevin Lee

## Mentioned Kevin Lee, NICE, Cognigy, NICE CXone, Automated Insights, TripAdvisor, Lufthansa, Toyota, Toyota Connected Care, Salesforce, SAP, ServiceNow, Sabre, Fiserv, Jack Henry, SMS, RCS, WhatsApp, LINE, WeChat, Viber, PSTN, WebRTC, SIP, HIPAA, PCI-DSS, CFPB, AS/400

## Video description Kevin Lee - Field CTO & Key Pursuits Leader, NiCE

AI experimentation is everywhere. Enterprise-scale AI deployment is not. While organizations race to adopt AI agents, many remain trapped in pilots that never deliver meaningful business value. The challenge isn't access to AI. It's building the operational foundation required to deploy AI securely, govern it effectively, and scale it across the enterprise. What can we learn from enterprises already running AI agents in production across millions of customer interactions? Dive into the realities of operationalizing AI, the governance models that build trust, and the measurable outcomes organizations are achieving today.​

[Source](https://www.youtube.com/watch?v=UX0yVExlo_Q)

中文简报

NICE借助桌面挖掘推动企业AI试点走向投产 企业AI试点未能实现投资回报，是因为企业将昂贵的前沿模型部署在孤立的网页聊天孤岛中，而非直接基于历史交互数据对已验证的工作流进行自动化。 Kevin Lee阐释了大多数企业AI智能体试点失败的原因——指出其根源在于高昂的Token成本、不明确的ROI以及孤立的聊天机器人实验——并提出了一个在企业联络中心规模化扩展AI的成熟框架。他概述了NICE生产级AI的三大核心支柱：积极分析交互数据、部署针对成功解决方案定制的领域专用模型，以及在数字、语音、呼入和主动呼出等触点间编排无缝的全渠道旅程。结合TripAdvisor、Lufthansa和Toyota的实际部署案例，Lee展示了自动化桌面活动挖掘与历史交互分析如何打造出具备韧性、高ROI的硅基智能体，并与人工团队展开高效协作。 大多数企业AI项目难以实现规模化：66%仍停留在试点阶段，95%缺乏可衡量的ROI，且因高昂的Token成本、价值不明确及治理崩溃，预计到2027年将有40%的项目被取消。 (0:04) 成功规模化企业AI取决于三大支柱：持续审查交互数据、围绕成熟的人工解决方案针对性构建智能体，以及在硅基智能体与人工之间无缝编排业务旅程。 (2:26) 有效的人工智能需要历史领域智能；专注于解决客户问题的专用CX模型，表现优于回答开放领域查询的通用前沿LLM。 (2:26) 企业应避免将昂贵的生成式前沿模型用于订单查询等简单重复的咨询，这类咨询更适合通过确定性规则或轻量化智能体工作流来解决。 (5:19) 对全渠道语音和数字化交互进行100%全量分析，可以揭示真实的客户意图，并确定对话自动化是否能够产生合理的商业ROI。 (5:19)

Reading summary

NICE Field CTO Kevin Lee argues that enterprise AI pilots overwhelmingly stall due to prohibitive token costs, vague ROI, and siloed deployments, which can be overcome by grounding agents in historical interaction data and end-to-end orchestration. The Enterprise AI Pilot Dilemma (0:04) Enterprise artificial intelligence initiatives frequently stall before reaching production. Lee notes industry figures indicating that 66% of enterprises are stuck in pilots, 95% fail to show measurable ROI, and 40% of AI projects are projected to be canceled by 2027. Stalled deployments typically suffer from high operational costs tied to token economics, unclear returns, and governance vulnerabilities. Addressing these operational barriers up front is essential for organizations attempting to scale. Three Pillars for Scalable Production AI (2:26) Moving beyond experimental pilots requires three foundational principles: inspecting interaction data rather than letting it sit idle in data warehouses, purpose-building agents based on proven historical dialogues, and orchestrating interactions across the full customer journey. Lee asserts that effective AI must rely on historical intelligence rather than delegating interaction logic entirely to third-party foundation models. To support this, NICE relies on more than 1,000 domain-specific customer experience models built from analyzing interactions across 27,000 customers in 150 countries, filtering out general-knowledge tasks to focus strictly on customer resolutions. Data-Driven Prompts and Desktop Mining (5:19) Routine tasks like order tracking often do not warrant the token expense of frontier generative models and are better handled deterministically or with lightweight workflows. NICE's Automated Insights engine scans full multi-modal interaction data across voice and digital channels to score whether an interaction is viable for automation. When automated workflows show clear ROI, the platform builds prompts from high-performing human dialogues and observes human desktop usage—such as navigation across Salesforce, SAP, and ServiceNow—to equip virtual agents with the precise tool integrations required for autonomous resolution. Breaking Out of the Web Chat Silo (9:30) A recurring failure in AI pilots is isolating chatbots on web pages, which often represent only a minor share of total interaction volume. Deploying through platforms like Cognigy allows organizations to use identical agent logic across voice, SMS, and digital messaging channels. Furthermore, effective customer experience requires combining inbound routing with proactive outbound outreach triggered by system telemetry, addressing customer issues across discovery, purchasing, and care before call volumes spike. Production Deployments Across Regulated Enterprises (12:52) Large-scale deployments must satisfy regulatory frameworks like HIPAA, PCI-DSS, and CFPB rules while handling extreme volume swings. Lee highlights TripAdvisor reaching 90% customer satisfaction and cutting average handle time by two minutes within two and a half months of deployment. During a pilot strike, Lufthansa utilized NICE's agentic platform as a digital shock absorber, handling peak rebooking volumes of 12,000 calls per minute. In an IoT application with Toyota Connected Care, telematics data triggers proactive outbound engagement to schedule vehicle maintenance directly with drivers. Before initiating further pilots, enterprise leaders are advised to replace subjective journey mapping with data-backed evaluations: identifying actual contact drivers from operational data, vetting ROI against implementation complexity, and diagnosing architectural bottlenecks in legacy back-ends that disrupt automated handoffs.

精炼阅读摘要

NICE全球现场首席技术官Kevin Lee指出，绝大多数企业生成式AI目前仍受困于高昂成本且缺乏明确回报的试点陷阱；要实现规模化生产级部署，企业必须摆脱通用大模型的盲目调用与孤立的网页聊天插件，转向以专有历史数据为基础、由桌面工作流挖掘驱动并贯穿全渠道的端到端编排体系。 企业试点陷阱与高昂运行成本 (0:04) Lee援引行业统计指出，当前企业AI面临着严峻的落地瓶颈：66%的企业受困于概念验证阶段，95%未能产生可衡量的投资回报率，且预计到2027年将有40%的AI项目因成效或治理问题被取消。阻碍试点走向规模化的主要运营障碍包括陡峭的Token计费模式导致的巨额运营开销、模糊不清的商业回报，以及严峻的合规与治理漏洞。 多数团队在部署AI时往往受困于“低垂果实”陷阱，优先选择搭建网页聊天窗口（Web Chat）。然而在实际业务中，网页聊天通常仅占企业总交互量的一小部分，依赖低容量渠道无法解决企业核心痛点；一旦面对复杂庞大的真实服务需求，缺乏稳健架构支撑的实验性系统极易陷入停滞。 规模化AI三大支柱与专用模型 (2:26) 为了让AI走出实验室并进入生产环境，Lee提出了企业必须依循的三大支柱：首先是“认知你的数据”，将交互数据从被动堆积在数据仓库用于合规审计的状态，转变为大规模的主动实时检查；其次是“目标驱动构建”，直接基于经检验成功的真实业务闭环开发智能体，而非将交互逻辑全权交由通用基础模型猜测；最后是贯穿全客户旅程的无缝全渠道编排，实现数字智能体与人工坐席的无缝协作。 高效的AI必须以深厚的历史真实智能为前提。NICE依托覆盖全球150个国家、2.7万家客户、每月处理逾两万亿词交互的业务底盘，研发了超过1000个预构建的客户体验专用领域模型。这些模型剥离了查询天气、股票等无意义的开放域功能，完全聚焦于将客户从初始意图最快、最有效且满意地引导至最终交易与问题解决。 自动化洞察与坐席桌面行为挖掘 (5:19) 针对成本问题，Lee强调企业并不需要为“查询订单位置”等日常确定性问题调用昂贵的前沿模型而消耗数美元Token，轻量级工作流或确定性方案往往更高效。NICE利用“自动化洞察”引擎扫描涵盖PSTN、WebRTC、SIP语音及多语言数字文本的100%交互记录，精准评估任务的自动化可行性与投资回报率，判断构建复杂API整合是否在商业上划算。 在构建智能体时，系统直接从历史优秀对话中提炼标准作业程序并自动生成提示模板，避免了耗时数周的手工提示词工程。此外，系统通过监测人工坐席在屏幕上对Salesforce、SAP、ServiceNow等后台系统的实际点击与操作路径，为虚拟智能体自动挂载对应的后端工具接口，使AI能够自主解决事务，避免因能力不足升级至人工坐席而产生每次约10美元的转接成本。 全周期编排与生产级实证落地 (9:30) 真正的全渠道编排必须突破单一网页插件的局限，覆盖客户全生命周期的发现、评估、购买与售后环节，并兼顾呼入与主动外呼。借助与Cognigy的集成，企业无需重构核心逻辑即可将智能体部署至语音与数字渠道；当遥测系统捕获异常征兆时，智能体可携带上下文主动外呼或发送短信，在呼入洪峰形成前化解问题。 该技术体系在应对HIPAA、PCI-DSS及CFPB等严苛合规要求的同时，已在超大规模场景中得到验证：TripAdvisor在两个半月内完成智能体投产，客户满意度提升至90%并将平均通话时长缩短两分钟；在汉莎航空因飞行员工会罢工导致大面积停运的危机中，平台充当“数字减震器”，在每分钟1.2万通电话的峰值压力下自动完成旅客改签；丰田互联出行则通过车机物联网遥测数据，在保养临期时主动外呼车主，无缝完成预约服务。 Lee呼吁企业管理者在启动下一个AI试点前，停止依赖主观的便签白板推演，转而基于真实数据回答三个核心问题：客户致电的真实诉求是什么、特定场景的自动化收益是否值得投入工程研发、人机接管断裂究竟源于缺少系统工具还是受制于底层老旧架构。企业唯有直面实际业务痛点与技术债务，才能彻底跨越百万美元的AI试点陷阱。

Detailed notes — zh

## Background 本视频由 Ai4 于 2026 年 8 月 19 日在 YouTube 平台发布。视频记录了 NICE 现场首席技术官兼关键业务拓展负责人 Kevin Lee 的演讲，探讨了为何许多企业在引入 AI Agent 时深陷无法带来实际业务价值的“百万级试点陷阱”，并分享了在生产环境中安全部署、有效治理和规模化扩展 AI Agent 的运营基石与实践经验。 ## TL;DR 演讲揭示了当前企业 AI 落地的残酷现实：66% 的项目卡在试点阶段，95% 无法证明 ROI，其核心症结在于高昂运行成本、商业回报模糊与治理合规缺失。为此，NICE 提出了破局的三大支柱：基于专有 CX 模型深度洞察交互数据并精细评估投入产出比、从优秀交互与坐席屏幕轨迹中自动构建具备工具调用能力的专用 Agent，以及构建覆盖全生命周期与主动触达的双向无缝编排。结合 Tripadvisor、汉莎航空及丰田汽车等标杆案例，Kevin Lee 总结了企业在启动下一个 AI 试点前必须深思的关键自省问题。 ## Atmosphere & style 整体基调务实、犀利且极具工程导向，充满了资深行业老兵对当前企业盲目追逐 AI 概念的清醒审视。演讲者没有停留在泛泛的愿景勾勒上，而是用刺眼的数据和切中要害的现实痛点（如昂贵的 Token 账单、老旧系统的接口断裂）打破行业幻想，并逐步转向结构严密、可落地的实操架构，带给听众兼具警醒感与前瞻性的专业体验。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) 当前企业 AI 项目普遍深陷试点泥潭：行业数据显示 66% 的项目卡在试点阶段，95% 未能产生可衡量的 ROI，预计到 2027 年将有 40% 的项目被正式取消。 - [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) 导致 AI 试点无法规模化落地的三大核心痛点包括：Token 经济学带来的高昂运行成本、模糊不清的投资回报率，以及治理与合规层面的失败。 - [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 要想拥有高效的人工智能（AI），首先必须具备依托领域专业知识增强的“真实智能（Actual Intelligence）”，而非直接简单套用通用大语言模型。 - [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 企业应当利用自动化洞察技术对 100% 的全渠道交互数据进行细粒度分析，精准识别意图与动作，并严格评估自动化场景的投入产出比。 - [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 避免在简单问题（如‘查订单’）上耗费高昂 Token 成本调用前沿生成式大模型，很多时候确定性流程或低成本 BPO 伙伴是更高 ROI 的选择。 - [07:40](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=460s) 摆脱反复调试 Prompt 的传统模式，NICE 通过分析海量真实成功对话自动提取优秀范式并持续优化 Agent 的运行逻辑与提示词。 - [07:40](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=460s) 通过捕获人工坐席在 CRM/ERP 系统的屏幕操作轨迹，自动识别并打通 Agent 所需的底层 API 与工具集成，实现真正端到端解决问题而非简单转接。 - [10:21](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=621s) AI 编排必须覆盖客户发现、评估、购买到持续关怀的全生命周期，并从被动响应转向基于 IoT/遥测触发的主动式双向交互。 - [12:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=772s) Tripadvisor 与汉莎航空等企业通过落地专用 Agent 平台，实现了 CSAT 提升至 90% 以及罢工危机下每分钟承受 12,000 通电话的自动化弹性承接。 - [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) 启动 AI 试点前企业必须进行三项自省：依赖客观数据明确用户真实呼叫原因、精算该意图的自动化 ROI、以及查明人机交接断点与底层遗留系统瓶颈。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) 百万级 AI 试点陷阱：企业规模化落地的现实困境 NICE 现场首席技术官（Field CTO）兼关键业务拓展负责人 Kevin Lee 登台发表演讲。NICE 作为全球领先的联络中心与客户体验（CX）服务商，为全球众多大型企业提供支持，协调消费者与品牌之间的交互。Kevin Lee 抛出了当前 AI 大会中最常被引用却又极为刺眼的三大核心数据：66% 的企业 AI 项目深陷试点泥潭无法推进；95% 的项目未能实现可衡量的投资回报率（ROI）；到 2027 年，将有 40% 的 AI 项目被正式取消。 这些严峻的行业现状背后，反映出当前企业在推进 AI 试点时普遍遭遇的三大核心痛点：昂贵的运行成本（如 Token 经济学带来的开销）、模糊不清的投资回报率，以及治理与合规层面的失败。Kevin Lee 指出，要打破“试点陷阱”并站在成功的历史这一边，企业必须在项目初期就对这些问题进行顶层设计和解决。NICE 凭借在全球 150 多个国家服务超过 27,000 家客户的实践经验，每月处理超过 2 万亿个单词，并成功协调了涵盖 PSTN、WebRTC、SIP 等复杂语音通话及 SMS、RCS、WhatsApp、Line、WeChat 等多渠道数字化交互在内的 200 亿次互动，总结出企业成功实现 AI 规模化落地的三大核心支柱。 ```mermaid flowchart TD accTitle: AI 试点失败的核心痛点与破局三支柱 subgraph P["AI 试点陷阱核心痛点"] P1["高昂运行成本 (Token 经济学)"] P2["不明确的 ROI (无法衡量商业价值)"] P3["治理与合规失败"] end subgraph S["规模化落地三大支柱"] S1["KYC: 深入洞察与分析数据"] S2["基于最佳交互构建专用 AI Agent"] S3["全生命周期与双向无缝编排"] end P --> S ``` - **66%** 的企业 AI 项目卡在试点阶段（Pilot Trap）。 - **95%** 的企业 AI 项目未能提供可衡量的 ROI。 - **40%** 的 AI 项目预计将在 2027 年前被取消。 - NICE 平台每月处理超过 **2 万亿个单词**，累计编排超过 **200 亿次复杂交互**。 --- ### [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 支柱一：KYC（了解你的数据）与“真实智能” Kevin Lee 提出了一个鲜明观点：“要想拥有高效的人工智能（Artificial Intelligence），首先必须具备‘真实智能’（Actual Intelligence）。”在客户体验领域，AI 不能仅仅简单套用通用大语言模型（LLM），而是需要依托领域专业知识进行深度增强。过去 25 年间，NICE 持续分析各行业客户在严苛合规环境下的真实交互，构建了超过 1,000 个针对 CX 场景的专用 AI 模型。这些模型不关注股票代码、天气预报或导航路线，而是专注于以最快、最有效且最令人满意的方式将客户引导完成从“意向（Pi）到购买（Buy）”的全流程。 利用这些专用模型，NICE 推出的 Automated Insights（自动化洞察）技术能够对企业 100% 的全渠道交互数据进行细粒度分析，精准识别对话的意图、动作与业务结果，找出最具自动化潜力的场景。Kevin Lee 强调，企业必须严格评估“投入产出比（Is the juice worth the squeeze?）”。在很多情况下，并非所有问题都需要调用昂贵的前沿生成式大模型：“在回答‘我的订单在哪里’这种简单问题时，根本不需要耗费 5 美元的 Token 成本去调用前沿生成式大模型。”有时确定性流程或低成本 BPO 伙伴反而是更高效的选择。 ```mermaid flowchart LR accTitle: 从全量数据到高 ROI 自动化意图识别 A["100% 全渠道交互数据<br/>(语音/数字文本)"] --> B["1,000+ CX 专用模型<br/>Automated Insights 分析"] B --> C{"高可自动化度评估<br/>(流程明确/ROI 达标)"} C -->|高回报高频场景| D["自动化与 Agent 承接"] C -->|低 ROI/边缘长尾| E["保持确定性流程/BPO 承接"] ``` --- ### [07:40](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=460s) 支柱二：以优秀真实交互自动构建专用 AI Agent 传统 AI 开发模式中，工程师需要花费大量工时反复编写、调试和优化 Prompt。NICE 则采取了截然不同的数据驱动方式：直接从海量成功的真实客户对话中提取优秀范式，自动生成并持续优化 Agent 的 Prompt 与运行逻辑。这种机制能敏捷适应业务流程（SOP）、市场环境及产品信息的动态变化，而非停留在静态的时间快照。 与此同时，针对需要调用外部工具完成复杂业务的情况，NICE 能够直接捕获人工坐席在操作界面上的屏幕交互轨迹。系统可以清晰识别坐席在 Salesforce、SAP、ServiceNow、Sabre、Fiserv 或 Jack Henry 等企业后台系统中的切换与点击动作，进而明确虚拟 Agent 所需集成的具体业务工具与 API 接口。这样构建出的 Agent 能够真正端到端解决用户问题，而不是仅仅“记录一张工单，然后额外花费 10 美元升级转接给人工”。构建完成的 Agent 可以无缝发布至 Cognigy 等平台中，迅速跨越渠道限制，从 Web 聊天窗口快速扩展至全渠道语音等关键场景。 - **Prompt 自动构建与迭代**：利用成功对话数据直接沉淀 Prompt，伴随业务 SOP 变化实现自动化更新。 - **坐席屏幕轨迹挖掘**：精准解析人工坐席对 CRM/ERP 等核心系统的操作，自动推导出 Agent 所需的工具集成与 Action。 - **摆脱单一渠道限制**：避免落入“仅在 Web Widget 试点”的误区，一键将 Agent 部署扩展至承载主力业务量的语音通信网络。 --- ### [10:21](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=621s) 支柱三：全生命周期与主动式交互编排 许多试点项目之所以陷入僵局，是因为其业务范围被死死限定在单一孤岛内。Kevin Lee 强调，真正的 AI 编排必须覆盖消费者全生命周期的各个阶段——从前期的需求发现、产品评估，到交易购买，再到长期的客户关怀。 除了全生命周期的覆盖，交互的方向性（Directionality）与主动性（Proactivity）同样至关重要。传统客服系统往往处于被动响应状态，依赖用户拨打免费热线进入 IVR/ACD 或发起在线咨询。然而，在现代物联网与遥测技术支持下，系统在问题爆发前就能捕获关键触发信号。企业应当充分利用已构建的 Agent 能力，主动发起携带完整上下文的语音外呼、SMS 或邮件通知，在用户遇到故障的第一时间主动解决问题。NICE CXone 平台正是将数据洞察、Agent 构建、人工坐席协同与全流程编排集于一体，彻底解决了企业因多点工具割裂而无法规模化扩展的难题。 ```mermaid flowchart TD accTitle: NICE CXone 全流程与双向编排架构 subgraph LifeCycle["客户全生命周期编排"] L1["需求发现与评估"] --> L2["业务交易与购买"] --> L3["持续关怀与服务"] end subgraph Flow["双向互动模式"] F1["被动接入 (Inbound)<br/>IVR / ACD / 语音 / 数字渠道"] F2["主动触达 (Proactive Outbound)<br/>IoT 遥测触发 / 上下文外呼 / SMS"] end subgraph Platform["NICE CXone 统一平台"] P1["真实智能洞察 (Automated Insights)"] P2["虚拟 Agent (Cognigy)"] P3["人工坐席协同 (Human + Silicon)"] end Platform --> LifeCycle Platform --> Flow ``` --- ### [12:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=772s) 标杆企业实战成果与合规规模化经验 NICE 平台既服务于仅有两名坐席的小型呼叫中心，也支撑着拥有数十万人工与数字坐席的超大型组织，深度适配 HIPAA、PCI、CFPB 等各类严苛的企业级合规与监管标准。Kevin Lee 现场分享了三个典型的企业落地案例： - **Tripadvisor**：在短短 2.5 个月内实现企业级生产环境全面上线，客户满意度（CSAT）提升至 90%，每通电话平均处理时长缩短整整 2 分钟，为大型企业节约了极为可观的运营成本。 - **德国汉莎航空（Lufthansa）**：在面临飞行员工会罢工、导致大批航班延误取消的极端突发危机中，NICE Cognigy 平台充当了关键的“数字减震器”，自动承接了所有退改签与航班重新调度呼叫，峰值处理能力达到每分钟 12,000 通电话，有效避免了客服体系瘫痪。 - **丰田汽车（Toyota）**：与其 Connected Care 团队深度合作，利用车载 CAN 总线与 IoT 遥测数据监控车辆健康状况。当系统检测到维护需求时，主动触发外呼通知并接入虚拟 Agent，协助车主预约保养与车辆维护，将服务从事后补救转变为事前主动关怀。 --- ### [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) 启动下一个 AI 试点前的三项关键自省 演讲最后，Kevin Lee 给计划启动下一个 AI 试点的企业提出了三项必须深思的底层问题，并呼吁企业告别脱离数据的“便签纸式推演”： 1. **用户真正呼叫的原因是什么？** 必须依赖客观的交互数据分析，而不是在战略研讨会上凭空猜测、在白板便签上简单写下“查订单”、“退款”或“补发卡片”。 2. **核心业务意图的 ROI 是否值得投入？** 仔细评估自动化该场景带来的商业回报，确认其是否真正能够达成核心 KPI 与业务目标，算清算细投入产出比。 3. **服务交接（Handoff）在何处断裂？** 明确数字 Agent 向人工坐席升级、或者跨渠道切换时的断点何在，搞清楚是否需要底层 API 集成，以及业务流程是否会被“地下室里的某台老旧 AS400 架构主机”所扼杀。 Kevin Lee 邀请参会者前往 917 号展位，现场体验在 10 分钟内快速构建专属 Agent，并演示了完全由 iPad 虚拟 Agent 驱动的点单制作咖啡与拿铁流程。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=4s) 66% 的企业 AI 项目深陷试点泥潭无法推进；95% 的项目未能实现可衡量的投资回报率（ROI）；到 2027 年，将有 40% 的 AI 项目被正式取消。 — Kevin Lee > [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 要想拥有高效的人工智能（Artificial Intelligence），首先必须具备‘真实智能’（Actual Intelligence）。 — Kevin Lee > [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) 在回答‘我的订单在哪里’这种简单问题时，根本不需要耗费 5 美元的 Token 成本去调用前沿生成式大模型。 — Kevin Lee > [07:40](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=460s) 记录一张工单，然后额外花费 10 美元升级转接给人工。 — Kevin Lee > [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) 用户真正呼叫的原因是什么？必须依赖客观的交互数据分析，而不是在战略研讨会上凭空猜测、在白板便签上简单写下‘查订单’、‘退款’或‘补发卡片’。 — Kevin Lee > [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) 核心业务意图的 ROI 是否值得投入？仔细评估自动化该场景带来的商业回报，确认其是否真正能够达成核心 KPI 与业务目标，算清算细投入产出比。 — Kevin Lee > [15:52](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=952s) 服务交接（Handoff）在何处断裂？明确数字 Agent 向人工坐席升级、或者跨渠道切换时的断点何在，搞清楚是否需要底层 API 集成，以及业务流程是否会被‘地下室里的某台老旧 AS400 架构主机’所扼杀。 — Kevin Lee > [03:42](https://www.youtube.com/watch?v=UX0yVExlo_Q&t=222s) Is the juice worth the squeeze? — Kevin Lee ## Mentioned Kevin Lee, NICE, NICE CXone, Cognigy, Automated Insights, Tripadvisor, 德国汉莎航空 (Lufthansa), 丰田汽车 (Toyota), Salesforce, SAP, ServiceNow, Sabre, Fiserv, Jack Henry, AS400 ## Video description Kevin Lee - Field CTO & Key Pursuits Leader, NiCE AI experimentation is everywhere. Enterprise-scale AI deployment is not. While organizations race to adopt AI agents, many remain trapped in pilots that never deliver meaningful business value. The challenge isn't access to AI. It's building the operational foundation required to deploy AI securely, govern it effectively, and scale it across the enterprise. What can we learn from enterprises already running AI agents in production across millions of customer interactions? Dive into the realities of operationalizing AI, the governance models that build trust, and the measurable outcomes organizations are achieving today.​ [Source](https://www.youtube.com/watch?v=UX0yVExlo_Q)

Source quotes

[0:04] Kevin Lee: 66% 的企业 AI 项目深陷试点泥潭无法推进；95% 的项目未能实现可衡量的投资回报率（ROI）；到 2027 年，将有 40% 的 AI 项目被正式取消。 [3:42] Kevin Lee: 要想拥有高效的人工智能（Artificial Intelligence），首先必须具备‘真实智能’（Actual Intelligence）。 [3:42] Kevin Lee: 在回答‘我的订单在哪里’这种简单问题时，根本不需要耗费 5 美元的 Token 成本去调用前沿生成式大模型。 [7:40] Kevin Lee: 记录一张工单，然后额外花费 10 美元升级转接给人工。 [15:52] Kevin Lee: 用户真正呼叫的原因是什么？必须依赖客观的交互数据分析，而不是在战略研讨会上凭空猜测、在白板便签上简单写下‘查订单’、‘退款’或‘补发卡片’。 [15:52] Kevin Lee: 核心业务意图的 ROI 是否值得投入？仔细评估自动化该场景带来的商业回报，确认其是否真正能够达成核心 KPI 与业务目标，算清算细投入产出比。 [15:52] Kevin Lee: 服务交接（Handoff）在何处断裂？明确数字 Agent 向人工坐席升级、或者跨渠道切换时的断点何在，搞清楚是否需要底层 API 集成，以及业务流程是否会被‘地下室里的某台老旧 AS400 架构主机’所扼杀。 [3:42] Kevin Lee: Is the juice worth the squeeze?

Source quotes

[0:04] Kevin Lee: 66% of enterprises are currently stuck in pilots, 95% fail to deliver measurable ROI, and 40% of AI projects are projected to be canceled by 2027. [2:26] Kevin Lee: when all you have is a hammer, everything looks like a nail [2:26] Kevin Lee: from PI to buy in the fastest, most effective, and satisfactory manner possible. [5:19] Kevin Lee: Where is my order? [5:19] Kevin Lee: Is the juice worth the squeeze? [15:52] Kevin Lee: an AS/400 in a basement somewhere. [15:52] Kevin Lee: What are customers actually calling about? [15:52] Kevin Lee: Where does the handoff break?

Original transcript

[0:04] Thank you very much. Hi everybody. Uh as  you just said, yeah, my name is Kevin Lee. [0:08] I'm the field CTO and key pursuits leader for  NICE. Uh hopefully you know who we are, but by [0:14] way of introduction if you don't. U NICE is the  world's pre-minent contact center and CX purveyor [0:20] uh for some of the world's largest organizations.  we effectively facilitate the interactions [0:26] between your consumers and the brand uh that  those consumers decide to transact with. So [0:33] um to pop up the three most overused metrics  and quotes for this entire uh convention, [0:42] here we go. Uh 66% of uh enterprises are stuck  inside of a pilot. 95% fail to deliver measurable [0:51] ROI and 40% of the AI projects will be cancelled  by 2027. Uh the these are stats that you're [0:58] all probably very familiar with and you want to  figure out how to be on the other side of history [1:04] in these cases. And so we've got some real strong  opinions on this uh backed up by stats and results [1:11] from the brands that partner with us. And so we'd  like to talk a little bit about that. Uh so we've [1:17] all felt the results here, right? High costs to  run, token inconomics were referenced earlier, [1:23] unclear ROI, governance failures. These are three  thematic uh threads that we see plaguing all the [1:31] pilots that uh we come across. And so we endeavor  to streamline these issues and uh think about them [1:38] upfront so that you're on the other side of those  equations. And uh so how do we know this right? Uh [1:45] NICE as an organization has over 27,000 customers  in 150 different countries globally. Uh we have [1:54] orchestrated over 20 billion interactions. That's  the most complex voice interactions that you can [2:00] think of. Uh and digital interactions across all  the different modalities SMS, RCS, WhatsApp, Line, [2:08] WeChat, Vibe, all the different things that are  out there. And uh we process over two trillion [2:14] words every month. Uh and so we're industry  recognized by all the the top analysts out there, [2:21] Gartner, Forester, IDC. Uh and we'd be happy  to share all of those wonderful reports with [2:26] you as well. Uh so three pillars of success,  what sets the brands apart that work with us, [2:33] right? Uh firstly, they KYC or know their  data. They actually inspect this data. They [2:39] get to understand it at scale. It's not just  something that they have sitting in an EDW for [2:45] compliance reasons and for audit reasons. They  actually leverage it, leveraging our technologies [2:51] so that they can understand what's going on.  Uh secondly, they are purposebuilding the AI [2:57] agents based on the most successful interactions.  They're not just guessing. they're not letting a [3:02] frontier model determine what they think is the  best way to resolve the interactions that their [3:09] consumers are having. And then thirdly, the  seamless orchestration of those interactions, [3:14] right? Uh I'm fond of saying when all you have  is a hammer, everything looks like a nail. When [3:20] you have the most comprehensive platform inside of  CX, we will meet your consumers where they want to [3:25] be met. whether that's an inbound voice call or a  proactive RCS SMS uh with a silicon agent or with [3:34] a carbon human agent. So let's talk about those  things and how we do it. And the first thing with [3:42] knowing your data, knowing your customers, uh I'll  start with a kind of a interesting perspective [3:48] maybe. Um, I would contend that to have highly  effective artificial intelligence, you need to [3:54] have actual intelligence. That's something that  we have in spades and it's enhanced with LLMs. [4:00] It is not just built on LLM. And that's a big  distinction I would say inside of our category. [4:07] Uh, we've been analyzing interactions for the  partner brands we work with for over 25 years. [4:12] some of the world's largest highly complex uh  compliance uh ridden organizations unfortunately [4:20] in some cases right so we've partnered with them  we've been able to analyze their interactions and [4:27] from there build over a thousand CX specific AI  models when I say CX specific what do I mean we [4:35] don't care about how to find the ticker price  for your stock we don't care about the weather [4:40] we can't tell you driving directions we don't do  any of those things which lots of other vendors [4:44] uh on the show floor would do. We care about  getting your customers interactions from pi to [4:48] buy in the fastest most effective and satisfactory  manner possible. Right? So we've built models to [4:55] identify those types of interactions the  triumpvirant of success I would say. And when we [5:02] utilize those models, we can then facet down  into the or the interactions and tell you what [5:08] are the the intents, the actions, the outcomes  of those interactions, right? We can understand [5:14] to a high level of granularity what's going  on, what does good look like? And from there, [5:19] we use that actual intelligence to train your  agents. Sometimes they're agentic, sometimes [5:26] they're deterministic. That's another hot sports  opinion I would say here. Uh you don't always have [5:32] to have a generative frontier model answering  questions for $5 worth of tokens especially [5:38] when the question is where is my order right?  Uh so one of our technologies called automated [5:43] insights. It does leverage those uh about a dozen  of those thousand models to help us determine what [5:51] are the outcomes inside of these interactions  and then what business drivers can we affect. [5:57] We processed all of the interactions, 100% of your  organization's interactions, whether they're voice [6:04] interactions over PSTN, WebRTC, SIP, etc., or  digital interactions, asky characters in whatever [6:12] language, right? And we can give you all that  detail, but then you can pull the levers and say, [6:17] "Hey, what are the things I want to do? I want  to in this case want to drive automation." [6:22] Fantastic. We can do that. We can then look at the  interactions that have the highest automatability, [6:28] what are typically automatable, shorter  interactions, very definitive, have a defined [6:34] start and end. We don't want to just log a ticket.  We actually want to resolve the case. And so to [6:40] do that, we have models that help us identify  those types of interactions. And then from there, [6:45] we've identified the the opportunities at hand,  right? And that's another important thing. [6:51] We talked about quantifying the ROI. Is the juice  worth the squeeze? That's a question that we all [6:58] think we ask, but we actually don't do this at  scale across all of the interactions to say, [7:03] is where is my order actually worthwhile? Should I  automate that? Should I go build API integrations, [7:10] a new fangled MCP harness so that I can actually  go find this and talk about it? It may not be [7:16] worth it. In some cases, we have BO partners that  do this for fractions of dollars that are more [7:23] effective at doing it and more cost-effective. So,  finding the right ones that have the right ROI, [7:28] I I would again contend that this is  highly overlooked in a lot of the pilot and [7:34] uh the scoping conversations that are happening  especially in the CX world. So, once we do this, [7:40] when we identify the qualified opportunities, then  we can actually auto build right. Imagine instead [7:46] of saying yes, I found the use case that I want to  build and then having our engineers and our FTEEs [7:53] and all the new terms in the landscape spent time  to build these prompts and iterate and improve [8:00] and iterate again. What if we use those successful  conversations to build the actual prompts? That is [8:06] what we do with the data. That's why we make such  a point about leveraging the data itself. So what [8:13] do we do? We find those successful conversations  and then from there we use those conversations to [8:19] build the actual prompts and improve the prompts  because this is not a snapshot in time. It doesn't [8:24] happen one time. It happens ongoing, right? Things  change, SOPs change, the landscape of our category [8:32] and products change. And this is exactly what  we're doing. We're using that data to then build [8:39] the prompt and the tooling. We actually also have  another very interesting ability because human [8:45] agents are utilizing our platform. We have that  surface, that glass. We see what they're doing. We [8:52] know what they're alt tabbing into. We know that  they're going to Salesforce, SAP, Service Now, [8:57] Saber, Fizerve, Jack, Henry, whatever tools  you may use. And we actually see what they're [9:02] clicking on, right? We can actually leverage that  as well and show the tools that we need to build, [9:09] right? and to give uh access to with our agents  so they can actually complete the transaction, [9:15] not log a ticket and then escalate to a human  agent for 10 more dollars. And so from there, [9:21] we can publish these agents, right? The  agents go directly into the platform and [9:25] then we can actually deliver them and build those  purpose-built agents. Uh and so you're seeing here [9:30] just a screenshot of once those agents are built  inside of our platform of Cogni [snorts] and then [9:37] we can actually deploy them in production and uh  it's not just one place not just one modality. [9:44] This is the other thing I see inside of pilots all  the time. They're like where's the easy place to [9:48] start? Web chat. Fantastic. We'll put a widget on  the website. We'll start there. and then someone [9:53] eventually finds out that web chat only accounts  for a very small percentage of the volume that [9:59] comes into their organization. Going back to  the scale problem and so with us doesn't matter [10:05] what modality it is. We started it inside of web  chat fine we escaped uh you know escape velocity [10:14] and now we can actually deploy this inside of  voice as well. Um, now moving to orchestration. [10:21] The other thing that we see being a challenge  inside of pilots is that we're sometimes mired [10:26] in just a very defined scope and we say, "Hey,  like this is all we care about. If this works, [10:31] we'll continue to pour more resources and dollars  into it." Uh, but the ability to orchestrate it [10:37] across the entire uh life cycle of a consumer,  right? Because they're not always looking for [10:43] just care interactions. Uh sometimes at the very  beginning it's discovery, it's the evaluation of [10:48] your product or your service. Uh then transacting,  the purchase, the ongoing care. You need a [10:54] platform that can address all of the stages of  that life cycle. Also, the actual modalities and [11:04] the directionality matter as well, right? We're  all very familiar with calling a toll-free number [11:09] and having the inbound call come in to an IVR or  ACD uh and then those digital channels. But the [11:16] thing that we're also seeing being left out that  helps address the seat problem, the CX problem is [11:23] proactivity. This in the past has been left out.  We've built these agents. Why are we not using [11:28] them in a proactive manner when we know what  the problem is and it's about to happen and [11:32] we know all the triggers because we have all the  telemetry. Why are we not leveraging that so that [11:37] we outbound that voice call with context? Why are  we not outbounding that SMS or email? Because that [11:44] virtual agent that we built, that agentic agent  that we built can handle it regardless of the [11:49] directionality. So orchestration is the other  piece. And then all of this brought together, [11:54] right? knowing your customers, knowing the  data, uh having all of the agentic capabilities, [12:01] the orchestration, the augmenting of the human  agents and then the orchestration all in one [12:07] platform. Again, this is going back to the  previous slide. If you built a pilot that [12:12] does one thing, you cannot you cannot expand  uh the scope of your pilot quickly. And again, [12:18] this is where pilots end up floundering. And so  all in one platform, that's what we do with NIC [12:24] CX1. Uh and the impacts that we've delivered for  our clients are are plentiful. I won't read all [12:33] the stats here, but you can imagine uh customer  satisfaction, handle time reduction, just the [12:38] amount of interactions that are being inbounded to  human agents uh and the focus they have to exert [12:45] there, quicker reporting are just some of the  things that we have delivered for our clients. [12:52] And so this is just a very small sampling of some  of the world's largest and complex organizations [12:57] that we partner with today. Uh the point I'll  make here is that we have everybody from you [13:03] know two person contact centers all the way up to  hundreds of thousands of human and silicon agents [13:09] on our platform. Uh so that enterprise scale rigor  compliance standards that you have to adhere to [13:16] those are things that are not unique to us. all  the multi- salobic acronyms HIPPA PCI the CFPB [13:26] uh type organizations that have to look at all  these uh compliance issues we are very familiar [13:32] with them and so I'll highlight just a couple of  uh very interesting results we've had starting [13:38] with Trip Adviser here so uh Trip Advisor was  able to increase their satisfaction up to 90 uh [13:46] 90% right And we were able to do this inside of  two and a half months to production. Not just, [13:54] hey, let's try something out in a very small uh  sample size. This is actual enterprisegrade uh [14:01] production. And we were able to save 2 minutes a  call. You can imagine how much that type of impact [14:10] uh would be felt inside of an enterprise scale  organization. And then moving to Lufansza. So, [14:16] Lufansza, uh, German airline, if you're familiar  with them, so something that Germans like to do is [14:22] strike. So, earlier this year, uh, their  pilot union actually had a strike. Uh, [14:30] and when that happened, our platform,  Cognagi, answered all of the calls. [14:37] All the flights that got cancelled and delayed  uh were then via our agentic platform rerouted, [14:45] rebooked and handled with that digital shock  absorber of our agentic agents. And had that [14:51] happened without this and they were just waiting  for those voice calls, you can imagine there would [14:56] have been a uh quite the shock wave that got  sent through the organization. And so at the [15:02] peak we were handling up to 12,000 calls uh a  minute inside of that organization. So okay and [15:11] uh lastly Toyota right we talked about proactivity  something that is underutilized all too often and [15:18] instead of just waiting for those interactions  we worked with Toyota and their AIE care group. [15:23] We outbounded the interactions when the CAN bus  in their vehicles uh alerted that there was some [15:30] maintenance that needed to happen. Well, they're  using IoT of course and so that triggered our [15:35] outbound messaging notification to then start the  case, bring in the right virtual agent and help [15:42] those drivers facilitate the uh maintenance  that they needed inside of their vehicles. [15:52] Okay. And so just to bring it all together here,  uh before you move on to your next pilot, there's [15:58] a couple of questions that we would encourage your  your organizations to ask, right? Firstly, uh what [16:04] are the people actually calling about? Right?  Let's actually know. Let's not guess. All too [16:08] often, I get into these sessions where we journey  board and story map and all the other terms that [16:14] we use and then we end up putting post-it notes  that say, "Hey, where is my order? Where is my [16:18] refund? Send me a new ID card." And then I would  ask you, well, is it worth the squeeze? Are you [16:24] going to get the ROI? Is this going to move the  needle? Will it hit KPIs and MBOs's? And I would [16:30] start with that. Then uh what are the top intents,  right? We just talked about that ROI and if it's [16:36] worth it or not. Uh that's just kind of dovetailed  into the first one. And then where does the [16:40] handoff break, right? Moving from one modality to  another or the escalation from that silicon agent [16:47] to the human agent. uh where does it no longer get  contained and how can we address that because that [16:53] will answer the question of do we need to build  integrations? Do we have the ability to integrate [16:59] this process? Is it uh going to be hamstrung  by an AS400 in a basement somewhere? And so [17:06] those are the types of things that uh we would  encourage you to take a look at and uh start [17:13] conversations with us as well. We're happy to help  facilitate that. Um, we have something we talked [17:17] about earlier called automated insights. We can  run these analysis efforts on your interactions [17:23] to help you understand what's going on there.  Uh, so if you'd like to see any more of this, [17:27] we are at booth 917 on the show floor. Uh, we  can actually show you how to build a demo uh, [17:34] in less than 10 or build a demo build an actual  agent in under 10 minutes. In fact, we are doing [17:39] coffees and lattes there and it's all being  done through our virtual agent on the iPad. So, [17:44] if you'd like to take a look at that, please come  visit us at 9:17. I'll be around for a few minutes [17:49] if there's any other questions that you may  have. Uh, and with that I yield the balance of my