Dialpad CTO Says Standalone AI Agents Face Rapid Commoditization
Ai4
Because raw machine intelligence is becoming a cheap commodity, enterprise competitive advantage will shift from foundation models to closed-loop systems trained on proprietary human interaction data.
## Background
Published on YouTube by the channel Ai4 on August 20, 2026, this keynote features Dialpad co-founder and CTO Brian Peterson. According to the description, the session explores how technological revolutions commoditize expensive assets and argues that the future of enterprise AI belongs not to standalone agents, but to holistic systems centered on context, orchestration, governance, and continuous learning.
## TL;DR
Brian Peterson explains that because artificial intelligence is commoditizing raw cognitive intelligence, standalone AI agents and static prompt wrappers are rapidly becoming commodities rather than defensible advantages. He argues that foundational large language models represent only 5% to 10% of a production agentic system, with the remaining 90% consisting of complex context integration, governance, and safety infrastructure. To avoid brittle point automations, enterprises must build compounding learning flywheels powered by four architectural pillars: skills mining from human interactions, an assisted Agent Studio, rigorous historical simulation in a Proving Ground, and strict phased rollouts. Ultimately, long-term enterprise differentiation depends on proprietary customer data and self-improving organizational intelligence rather than the underlying commodity models.
## Atmosphere & style
Peterson delivers a pragmatic, clear-eyed reality check against prevailing industry hype, speaking with the grounded credibility of an engineering leader who previously built and scaled Google Voice. The talk unfolds like an executive architectural briefing, blending macroeconomic history lessons on technological commoditization with sharp, relatable critiques of fragile customer support chatbots and sidebar copilots. The narrative arc balances strong cautionary guidance against building DIY agent infrastructure with an optimistic, actionable blueprint for continuous human-in-the-loop intelligence.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) Standalone AI agents are merely a stepping stone; long-term enterprise value lies in the overarching intelligence layer and self-improving system harness built around the models.
- [02:12](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=132s) Major technology waves shift competitive advantages by making expensive assets cheap; as raw intelligence becomes commoditized, competitive moats move to context and proprietary learning systems.
- [04:45](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=285s) First-generation AI rollouts like sidebar copilots and isolated chatbots create brittle point automations that lose context across handoffs and require constant manual engineering maintenance.
- [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) Enterprise AI requires a compounding learning flywheel where omnichannel customer interactions continuously feed structured context and refine future system decisions.
- [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) Specialized Small Language Models (SLMs) trained on proprietary enterprise data will increasingly replace slow, monolithic LLMs to deliver deterministic execution and lower latency.
- [12:01](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=721s) Proprietary corporate interactions form an irreplaceable data moat because major foundation model providers cannot access internal organizational data.
- [12:01](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=721s) Human-in-the-loop interactions serve as the primary training fuel for automated workflows by capturing the exact language and resolution steps used in successful customer resolutions.
- [16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) Foundation LLMs represent only 5% to 10% of a production agentic platform, while the remaining 90% consists of integration, validation, memory, and governance infrastructure.
- [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) Enterprises should avoid building agentic platforms from scratch and instead leverage architectures structured around skills mining, guided agent creation, historical simulation, and governed rollouts.
- [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) Simulating proposed agent modifications against historical conversation logs is essential for benchmarking safety and accuracy before deploying changes to live customer environments.
- [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) AI initiatives must target concrete operational bottlenecks and be measured against pre-AI baseline metrics to demonstrate demonstrable return on investment.
- [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) Because every competitor has access to the same baseline AI models, lasting market differentiation depends on unique organizational intelligence and continuous feedback loops.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) From Google Voice to Dialpad: The Shift Beyond Agentic AI
Brian Peterson, co-founder and CTO of Dialpad, introduces himself and the engineering heritage behind the customer experience AI platform. The founding team originally built Google Voice, scaling it to roughly 10 million consumer users before realizing how broken enterprise communication tools were. They initially pitched building a business-focused communication platform inside Google, but Google opted to focus corporate resources on Google+, prompting the team to spin out and found Dialpad. Today, Dialpad operates in more than 50 countries across over 70,000 corporate customers, serving enterprise clients such as Netflix, Uber, Motorola, and the Indianapolis Colts.
Walking the expo hall at AI conferences, Peterson observes that nearly every company's marketing follows the exact same script: after LLMs and ChatGPT, the entire industry shifted to claiming "we finally figured out agents and agents are going to take over the world" with promises to "scale agents." Countering this industry groupthink, Peterson argues that AI agents are merely a stepping stone rather than the final destination. The true long-term value will reside in the overarching intelligence layer and self-improving system—sometimes referred to in the industry as harnesses—that sits above raw models and agents.
- **Founding Heritage:** Built by the former Google Voice team (which had 10M users) after Google chose to prioritize Google+.
- **Enterprise Footprint:** 70,000+ business customers across 50+ countries, including Netflix, Uber, Motorola, and professional sports teams like the Indianapolis Colts.
- **Industry Trendline:** Progression moves from raw Large Language Models (LLMs) to Agentic workflows, and ultimately to continuous Reinforcement Learning intelligence harnesses.
---
[02:12](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=132s) The Economics of Abundance: When Intelligence Becomes a Commodity
Every technological revolution is fundamentally defined by making something historically expensive suddenly abundant and cheap, which inevitably relocates where economic moats and competitive advantages reside. The internet made publishing essentially free with the click of a button, shifting value from content distribution to audiences and aggregator platforms. Cloud computing made infrastructure abundant, commoditizing server management and allowing a 10-person startup at Dialpad's inception to build what previously required 100 engineers, turning software and application design into the new differentiator. The explosion of open-source software subsequently democratized complex codebases, pushing the competitive moat toward superior product design and user experience (UX).
In the current cycle, artificial intelligence is rendering raw intelligence abundant and cheap. Because baseline cognitive and generation capabilities are commoditized across models from major providers, standalone AI agents built strictly for point automation will rapidly become commodities themselves. True sustainable advantage cannot be achieved by merely wrapping an LLM in a static agent prompt; it requires an integrated intelligence layer capable of dynamic, continuous reinforcement learning.
[04:45](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=285s) The Limits of Point Automations and Fragile AI Agents
The sudden mainstream arrival of ChatGPT led every SaaS vendor to tack on a "sidebar co-pilot that like you probably never use" or spin up a customer support chatbot that claims to sound human. However, these first-generation AI deployments have largely introduced fragmented point automations that frustrate users. Chatbots remain limited to basic keyword retrieval, forcing escalations to human agents who inevitably demand, "What are you calling us about? And who are you?" requiring full re-authentication and repeated explanations.
These disconnected automations do not communicate with one another, share context, or maintain shared memory. Because they rely on rigid workflows, the moment an underlying variable or edge case shifts, the point automation breaks. Deploying these early agentic setups has become so cumbersome that companies have had to invent entire roles like "forward deployed engineers" just to babysit brittle systems. True artificial intelligence must have native mechanisms to autonomously improve over time rather than relying on endless manual tuning and engineering interventions.
- **Sidebar Copilots:** Superficial SaaS add-ons that fail to solve structural workflow issues.
- **Context Loss:** Broken handoffs between chatbots and live representatives create repetitive user loops.
- **Brittle Workflows:** Fragile point automations degrade easily, creating an ongoing need for forward deployed engineers to manually adjust system parameters.
---
[08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) The Compounding Learning Loop and the AI Intelligence Harness
To escape the trap of static agents, organizations must construct holistic AI systems centered around context, governance, shared memory, and continuous feedback loops. Standard LLMs hallucinate, change their answers from one prompt to the next, and generate confident falsehoods based on unstructured web data. While this unpredictability may be acceptable for a conversational sounding board or personal therapist, enterprises require deterministic execution, compliance guardrails, and consistent customer memory across voice, chat, email, and meeting channels.
```mermaid
flowchart LR
accTitle: The Compounding Learning Flywheel
subgraph Capture ["Data Intake"]
A["Omnichannel Interactions<br/>(Voice, Chat, Email)"]
end
subgraph Intelligence ["Learning & Execution"]
B["Enterprise Context &<br/>Governance Layer"]
C["Understanding &<br/>Better Decisions"]
end
subgraph Outcome ["Continuous Feedback"]
D["Optimized Customer<br/>Experience"]
end
A --> B
B --> C
C --> D
D --> A
```
A winning architecture forms a closed-loop compounding learning flywheel: omnichannel interactions feed structured understanding, understanding drives superior business decisions, better decisions produce refined customer experiences, and those elevated interactions generate new contextual data to continually retrain and optimize the system. Rather than relying solely on gigantic, monolithic general models, the architecture will increasingly incorporate specialized Small Language Models (SLMs) tuned precisely to proprietary enterprise and customer-level data.
- **Governance and Memory:** Enforcing guardrails and historical continuity across multiple interaction channels.
- **Small Language Models (SLMs):** Specialized, domain-specific models tailored to proprietary data to replace slow, over-contextualized general LLMs.
- **Compounding Flywheel:** Continuous interaction analysis generates better decisions, which recursively improve system intelligence without manual code changes.
---
[12:01](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=721s) Outcomes Over Tech: Human-AI Coordination and Private Data Moats
Peterson emphasizes that enterprise leaders must stop treating AI as an end in itself simply to satisfy boardroom hype and start focusing on the specific business problem being solved. Debunking apocalyptic predictions that AI will wipe out 100% of jobs within a decade, Peterson points out that human interaction remains the fundamental fuel for training enterprise AI systems. Major foundation model providers like OpenAI cannot access proprietary corporate data, meaning an enterprise's private data interactions represent its ultimate defensive moat.
In Dialpad’s infrastructure, human agent conversations are ingested alongside automated interactions to create a live coordination layer. By analyzing the exact language, resolution steps, and empathetic behaviors that human agents use when customers report high satisfaction, the platform uses real-world human success to train and fine-tune its automated agents. Ultimately, passing millions of context tokens to a giant LLM remains too slow and cumbersome; future customer interactions will be powered by customer-specific models that instantly recall an individual's flight history, personal preferences, and past support tickets before the user even speaks.
- **Focusing on Business Outcomes:** Starting with the end goal (e.g., eliminating re-authentication friction) rather than deploying AI for vanity metrics.
- **The Private Data Moat:** Proprietary interactions and organizational data remain inaccessible to foundation model vendors, providing unique defensibility.
- **Human-in-the-Loop Training:** High-performing human agent resolutions directly serve as the training ground for customer service automation.
- **Hyper-Personalized Micro-Models:** Transitioning to customer-tailored models capable of zero-friction recognition and historical personalization.
[16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) Why LLMs Are Only 5–10% of a Production Agentic System
Great AI cannot remain static, yet most current implementations are fundamentally rigid. While swapping out an underlying large language model might yield slight performance bumps, model improvements quickly plateau because models only possess their static pre-training data. This limitation explains the stark divide between flashy demos and enterprise realities: demos look magical because they rely on tightly scripted, hard-coded workflows, whereas real-world production environments demand a massive surrounding apparatus.
```mermaid
flowchart TD
accTitle: Dialpad Agentic System Lifecycle
accDescription: Architectural pipeline showing how human conversation data is mined, developed into agents, simulated against historical data, and deployed under governance.
subgraph Validation ["2. Pre-Deployment Testing"]
C --> D["Proving Ground Simulator"]
E["Historical Conversation Logs"] --> D
D --> F{"Passes Simulation?"}
end
subgraph Deployment ["3. Governed Rollout"]
F -- Yes --> G["Canary / Percentage Rollout"]
G --> H["Live Production System"]
I["Governance, Security & Guardrails"] -. Oversees.-> H
H --> A
end
```
The recent boom in agentic AI stems from the realization that building robust automation workflows requires vastly more than just prompt engineering an LLM. While the foundation model serves as the crucial catalyst that made agentic workflows possible in the first place, it constitutes merely 5% to 10% of the complete system. The remaining 90% represents the hardest engineering challenges: contextual data integration, governance, testing, and continuous organizational learning.
- **The Demo Trap:** Hard-coded workflows create an illusion of intelligence in staged environments but fail when exposed to unpredictable edge cases.
- **The 90/10 Rule:** The LLM represents only 5% to 10% of an agentic platform; the remaining 90% is the infrastructure required to scale, validate, and govern autonomous actions.
---
[17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) The Four Pillars of Agentic Architecture: Don't Build It at Home
Organizations should resist the urge to build custom agentic architectures from scratch internally due to the immense engineering complexity: "Please don't try to build agentic yourself. It is so complicated... Don't try it at home yourself." Any enterprise purchasing or building AI systems must demand four foundational architectural layers to ensure agents learn, scale, and stay safe:
1. **Skills Mining (Human Learning Layer):** The system continuously listens to and transcribes live calls, analyzes what human agents are doing repetitively, and proactively suggests new automation workflows to the agent platform.
2. **Agent Studio:** Building multi-step agents manually is intensely difficult; AI-assisted development tools must guide users through constructing and tuning complex agent logic.
3. **Proving Ground (Historical Simulation Engine):** The single biggest blocker to deploying AI agents is trust. Companies cannot simply flip a switch and risk an agent insulting customers, corrupting CRM databases, or honoring absurd hallucinated promises—such as the widely publicized incident where an automotive chatbot agreed to sell a customer a truck for one dollar. Proving Ground solves this by running proposed agent modifications against every historical customer interaction ever recorded (both human- and AI-led) to benchmark accuracy and safety before code touches production.
4. **Governance, Observability, and Phased Rollouts:** Production agent deployment must follow strict operational discipline. Teams must never execute an all-at-once cutover; systems require progressive percentage rollouts, strict versioning, real-time observability, and automated kill-switches to sever agent authority if boundaries are breached.
---
[19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) Organizational Intelligence and the Shift in Value
When executed correctly, an AI deployment creates a continuous feedback loop where human workers and AI agents learn from one another rather than operating in silos. Business leaders must stop approaching strategy with the mentality of "I need to apply AI," and instead identify concrete operational bottlenecks: "What do I do repetitively or could be better?" Deployments must be rigorously measured against pre-AI baseline metrics; too many companies deploy AI tools and are subsequently unable to prove whether the initiative delivered meaningful ROI.
As baseline intelligence becomes cheap, abundant, and commoditized across the industry, companies must confront a critical question: "If every competitor had access to the exact same AI tomorrow, what would make your customer experience different?" Sustainable competitive advantage no longer resides in the models themselves, but in proprietary organizational intelligence—the compounding flywheels of unique company data, historical context, domain workflows, and continuous self-reinforcing learning systems.
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) we finally figured out agents and agents are going to take over the world — Brian Peterson
> [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) scale agents — Brian Peterson
> [04:45](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=285s) sidebar co-pilot that like you probably never use — Brian Peterson
> [04:45](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=285s) What are you calling us about? And who are you? — Brian Peterson
> [04:45](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=285s) forward deployed engineers — Brian Peterson
> [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) Please don't try to build agentic yourself. It is so complicated... Don't try it at home yourself. — Brian Peterson
> [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) sell a customer a truck for one dollar — Brian Peterson
> [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) I need to apply AI — Brian Peterson
> [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) What do I do repetitively or could be better? — Brian Peterson
> [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) If every competitor had access to the exact same AI tomorrow, what would make your customer experience different? — Brian Peterson
## Mentioned
Ai4, Brian Peterson, ChatGPT, Dialpad, Google, Google Voice, Google+, Indianapolis Colts, Motorola, Netflix, OpenAI, Uber
## Video description
Brian Peterson - CTO, Founder, Dialpad
Every major technology shift makes something expensive cheap—and when that happens, competitive advantage moves. AI is making intelligence abundant, but most organizations are still treating AI as a feature instead of redesigning the systems that create value. In this session, Dialpad CTO Brian Peterson explores why the future belongs not to standalone AI agents, but to organizations that build AI systems around context, orchestration, governance, and continuous learning.
Brian Peterson, co-founder and CTO of Dialpad, argues that as raw machine intelligence becomes an abundant commodity, enterprise advantage will shift from standalone AI agents to integrated, self-improving systems powered by proprietary data.
The Commoditization of Raw Intelligence (2:12)
Drawing parallels to past technological shifts—where the internet commoditized distribution and cloud computing commoditized infrastructure—Peterson asserts that artificial intelligence is turning baseline cognitive generation into an inexpensive commodity. In prior waves, economic value relocated to platforms and product design; similarly, commoditization across major foundation models means that standalone agents built around simple prompts will rapidly lose defensibility.
Peterson argues that sustainable enterprise value cannot be achieved through static wrappers around third-party language models. Instead, long-term competitive moats will belong to organizations that operate self-improving intelligence harnesses capable of continuous reinforcement learning and context integration.
Limitations of Fragile Point Automations (4:45)
Early enterprise deployments of generative AI have largely introduced disconnected point automations, such as customer support chatbots and superficial sidebar copilots. Peterson points out that these tools frequently fail during live operations because they lack shared memory and unified context. When escalated to human representatives, customers are forced to restart authentication and repeat their issues.
Because these initial workflows rely on rigid, hard-coded logic, edge cases cause them to degrade easily. Peterson notes that enterprises have had to rely on forward deployed engineers to manually adjust and maintain these fragile agent setups, underscoring the need for architectures that improve autonomously rather than requiring constant human intervention.
Infrastructure Beyond the Foundation Model (16:35)
While foundation models made autonomous workflows feasible, Peterson estimates that large language models represent only 5% to 10% of a production-grade enterprise system. The remaining 90% consists of the testing, integration, and governance infrastructure required to make autonomous operations reliable and secure.
Peterson cautions companies against building custom agent frameworks internally, highlighting four foundational requirements: continuous skills mining from human conversations, guided agent design studios, simulation engines that benchmark proposed changes against historical customer logs before deployment, and phased canary rollouts equipped with automated kill-switches.
Compounding Flywheels and Private Data Moats (8:40)
To build resilient systems, Peterson advocates for closed-loop learning flywheels that feed omnichannel interaction data—spanning voice, chat, and email—into structured decision-making layers and specialized small language models (SLMs). Rather than routing millions of context tokens through slow, monolithic models, architectures will increasingly use domain-specific models trained on proprietary organizational data.
In this framework, high-performing human agent resolutions serve as continuous training data for automated workflows. Because third-party foundation model vendors cannot access an enterprise's private interaction history, these proprietary feedback loops serve as the primary defensive barrier against competitors with access to the same underlying AI models.
Peterson concludes that enterprise leaders must evaluate AI deployments against concrete operational bottlenecks and pre-implementation baselines rather than adopting technology to fulfill boardroom hype. As access to baseline intelligence standardizes across the market, lasting differentiation will depend on how effectively an organization's proprietary data and feedback loops inform its customer experience.
## Background
本内容由 Ai4 于 2026 年 8 月 20 日发布在 YouTube 平台上。视频由 Dialpad 联合创始人兼首席技术官 Brian Peterson 主讲，深入探讨了在 AI 让智能变得丰裕廉价的时代背景下，企业的核心竞争优势为何不再属于单体 AI Agent，而是属于围绕上下文、编排、治理与持续学习构建的系统级智能架构。
## TL;DR
Brian Peterson 指出，技术演进规律表明当核心要素廉价化后竞争优势必然发生转移，AI 时代的基础智能正迅速沦为通用大宗商品。在真实的 Agentic AI 生产系统中，底层大语言模型仅占 5% 到 10% 的比重，剩余 90% 以上是上下文管理、安全治理与工程编排等复杂难题。企业应当打破脆弱的单点自动化，依托真实人机协作（Human-in-the-loop）沉淀私有数据，并通过技能挖掘、智能体工坊、仿真试验场与治理发布四大支柱构建自我进化的强化学习飞轮。最终能够拉开企业差距的不是公开模型，而是深度融合专有业务知识与持续学习能力的组织智能。
## Atmosphere & style
演讲整体风格务实深邃、逻辑严密，透露出主讲人深耕企业通信与大规模系统工程多年的从容与清醒。Brian Peterson 以平静克制但极具穿透力的语调，刺破了当下行业对 Agent 概念的浮躁炒作与玩具级 Demo 泡沫，精准切入生产落地的残酷现实。整场分享兼具科技史周期的宏观洞察与架构设计的微观工程底气，带给听众一种拨开迷雾、回归业务本质的理性思考体验。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) AI 的核心本质是让“智能”本身变得丰裕且廉价；当一种要素由稀缺变为泛滥，单体 AI Agent 终将同质化为通用大宗商品，不再是企业壁垒。
- [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) 回顾技术演进规律：互联网让发布变廉价并推动平台崛起，云计算让基础设施廉价并催生系统软件，AI 则将竞争优势推向系统级智能层与持续学习机制。
- [00:06](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=6s) 侧边栏 Copilot 与传统 Chatbot 导致上下文割裂，而脆弱的“单点自动化”缺乏共享内存与自演进机制，稍有流程变动就会崩溃。
- [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) 企业在 AI 时代的商业胜出核心在于搭建强化学习飞轮（RL Flywheel），通过统一智能框架串联长短期记忆、多模态通道与安全治理。
- [13:48](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=828s) 人类不会被 AI 完全替代（Human-in-the-loop），真实的人机交互行为和人工客服解决复杂难题的经验，是训练与微调 AI 模型最核心的养料。
- [13:48](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=828s) 公有大模型无法触及企业内部私有的客户历史与业务偏好，跨渠道全量沉淀私有交互数据是构建技术护城河的关键。
- [16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) 在真实的 Agentic AI 生产系统中，底层 LLM 仅占 5% 到 10%，剩下 90% 到 95% 是上下文管理、编排调度、仿真评估与安全治理等复杂工程架构。
- [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) 生产级 Agentic 系统依赖四大支柱：技能挖掘（Skills Mining）、智能体工坊（Agent Studio）、仿真试验场（Proving Ground）与安全治理（Governance & Observability）。
- [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) 为解决 Agent 上线信任危机，必须在投产前利用历史真实对话数据进行全量回放仿真，并配置灰度切流、版本控制与安全熔断机制。
- [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) 企业引入 AI 应从重复痛点和业务基线度量出发，杜绝为了跟风而盲目套用 AI。
- [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) 当基础智能泛滥时，竞争优势将转向“组织智能”——即企业专属的数据资产、业务知识沉淀以及人机协同的持续学习能力。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) Google Voice 基因与智能丰裕时代的降临
Dialpad 联合创始人兼首席技术官 Brian Peterson 开场回顾了自己的创业背景。他与团队早期均来自 Google Voice，在拥有约 1,000 万用户的消费级通讯产品积累了丰富经验。当时团队敏锐地察觉到企业通信与客户体验（CX）领域存在巨大痛点，本希望在谷歌内部推出企业级版本，但随着谷歌将战略重心转向社交网络 Google+，团队决定离场创业。如今，Dialpad 已服务全球超过 50 个国家的 70,000 多家企业，涵盖 Netflix、Uber、Motorola 以及印第安纳波利斯小马队等专业体育俱乐部，为企业提供集 AI 与人类沟通于一体的全套解决方案。
针对展会上铺天盖地的代理（Agent）宣传，Brian Peterson 提出了反直觉的行业洞察。他指出，正如过去的每次技术革命都伴随着某种要素的极度丰裕与低廉化——互联网让内容发布变得空前廉价，云计算极大地丰富了算力与基础设施，开源软件则让高性能代码随手可得——AI 的核心本质在于**让“智能”本身变得丰裕且廉价**。当某种要素从稀缺走向丰裕，它就不再是企业的核心壁垒，各类单体 AI Agent 终将同质化为通用的大宗商品（Commodities）。
* **技术变革历史演进规律**：
* **互联网时代**：内容发布丰裕化，竞争优势向**平台与受众**转移。
* **云计算时代**：基础设施丰裕化，竞争优势向**软件系统**转移（创业成本大幅降低，原本需要 100 人的团队在云端仅需 10 人即可搭建完备架构）。
* **开源软件时代**：高性能代码丰裕化，竞争优势向**产品体验（UX）与场景**转移。
* **AI 时代**：智能丰裕化，单体 Agent 迅速商品化，竞争优势向**系统级智能层与持续学习机制**转移。
### [00:06](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=6s) 告别断裂的 Copilot 与脆弱的单点自动化
纵观当下 AI 在客户体验领域的落地过程，行业经历了明显的演进路径。起初是各类 SaaS 软件纷纷推出的侧边栏 AI 助手（Co-pilots），但大多数用户几乎从不使用；紧接着是传统网页 Chatbot，其糟糕的体验屡屡让人抓狂——用户在聊天框详细描述问题后，机器人却因权限不足只能回复基础答案，并强制转接人工，而接通后的人工客服又会一无所知地重新要求验证身份和重述问题，数据与上下文完全断裂。
当下众多企业正尝试构建 Agentic 自动化，但绝大多数仍停留在极其脆弱的“单点自动化”（Point Automations）层面。这类系统缺乏共享上下文、统一内存与协同能力，只要底层流程稍有微调就会彻底崩溃。不仅部署成本极高，甚至逼得行业催生出“前线部署工程师”（Forward Deployed Engineers）这类专门维护 AI 的岗位。缺少自我演进机制的 AI 只能依赖人工不断调试，难以提供连续、高品质的客户体验。
> "I ask them, 'What are you accomplishing with that AI?' They start with the AI because they have pressure from their board, but they don't know what problem they're actually trying to solve."
### [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) 商业胜出的核心：具备自我进化能力的智能系统
面对单体 Agent 的困境，Brian Peterson 明确指出，AI 绝非仅仅是 Chat GPT 式的“心理咨询师”或单一的操作代理，未来的真正优势在于搭建**强化学习飞轮（Reinforcement Learning Flywheel）**。无论是用于企业内部流程自动化，还是外部客户服务与销售，赢家系统都必须构建在统一的智能框架之上，将上下文、长短期记忆、安全治理与多模态通道紧密串联。
```mermaid
accTitle: Dialpad Reinforcement Learning System Architecture
flowchart LR
subgraph MultiChannel["多通道交互层"]
A["语音 / Voice"]
B["文本与邮件 / Chat & Email"]
C["会议 / Meetings"]
end
subgraph IntelligenceCore["系统智能核心层"]
D["统一上下文与记忆<br/>Context & Memory"]
E["动态调度与工作流<br/>Orchestration & Workflows"]
F["安全与企业治理<br/>Security & Governance"]
end
subgraph LearningLoop["强化学习飞轮"]
G["人机协同与数据采集<br/>Human Interaction Capture"]
H["领域专属小模型微调<br/>Domain SLM Tuning"]
I["连续自我优化<br/>Continuous Optimization"]
end
MultiChannel --> IntelligenceCore
IntelligenceCore --> LearningLoop
LearningLoop --> MultiChannel
```
在此体系中，大语言模型（LLM）与针对特定数据微调的微小语言模型（SLMs）相互配合。通用 LLM 虽然能在初期提供宽泛的能力，但由于其无法直接根据企业私有数据实现实时持续学习，且容易产生幻觉，企业必须通过强化学习飞轮将客户交互中的成功经验反哺给 AI，实现系统的自驱式进化。
### [13:48](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=828s) 建立企业私有数据壁垒与人机协作机制
对于 AI 是否会彻底替代人类工作的焦虑，Brian Peterson 持否定态度。他强调人类将始终处于系统闭环之中（Human-in-the-loop），且人类真实的交互行为正是训练 AI 最核心的优质养料。在 Dialpad 的系统架构中，所有由人类客服处理的真实对话都会实时流入平台，系统会自动识别出人类在解决特定复杂难题时最能赢得客户满意的最佳实践，并直接用这些优质样本训练 AI Agent。
这种基于真实人类互动的学习机制，远比直接向通用大模型提问有效得多。在 AI 垂直化深入发展的阶段，公有大模型（如 OpenAI）无法触及企业内部私有的客户历史、业务偏好及独特流程，因此**捕获并沉淀私有交互数据**将构成企业最坚固的技术护城河。
* **优秀 AI 系统的关键支撑维度**：
* **全量数据捕获（Data Capture）**：安全地跨语音、文字、会议等一切渠道沉淀私有数据，严格防护个人身份信息（PII）。
* **系统深度互联（Full Connectivity）**：无缝链接 CRM、知识库与各类业务工具，保障信息的实时同步与一致性。
* **人机协同编排（AI-Human Coordination）**：以人类的高质量服务示范作为 AI 迭代进化和能力提升的基准。
* **个体级深度个性化（Personalized SLMs）**：未来将实现针对具体客户的细粒度模型服务，当客户致电时，系统无需反复验证即可秒级调出其全量历史偏好与行为记录。
### [16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) 为什么大语言模型只占 Agentic AI 系统的 5% 到 10%
优秀的 AI 绝不是静态的，而是需要随着持续学习不断改变自身行为并逐步迭代进化。当前许多企业对 AI 的理解过于单一，往往只是在底层替换一个稍好一点的模型，但模型本身的提升幅度有限，且仅拥有通用的公开数据。市面上充斥着大量令人惊叹的 Demo，这是因为演示往往是针对特定工作流进行硬编码（Hard-coded）定制的，看起来神奇却无法直接用于生产环境。真实的生产系统需要极其复杂的配套组件，这正是 Agentic AI 领域迎来爆发的根本原因。
从实际构建自动化工作流的经验来看，底层的大语言模型（LLM）充其量只占整个 Agentic 系统的 5% 到 10%。虽然 LLM 是一把至关重要的钥匙，是使一切成为可能的革命性突破，但剩下的 90% 才是真正最难攻克的工程与系统难题。为此，Dialpad 针对内部系统的扩展、自学习与稳定性，提炼并构建了关键的基础设施，这些能力也是企业在采购或评估任何 AI 方案时必须具备的核心要素。
```mermaid
flowchart TD
accTitle: Agentic AI System Composition
subgraph System["Agentic AI 生产系统 (100%)"]
direction TB
LLM["LLM 核心底座 (5% - 10%)<br/>通识推理与生成能力"]
Eng["生产级工程架构 (90% - 95%)<br/>上下文管理 / 编排 / 仿真评估 / 安全治理 / 持续学习"]
end
LLM --- Eng
```
---
### [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) 生产级 Agentic 系统的四大支柱：技能挖掘、开发、仿真验证与治理
构建一套能够自我学习且具备可扩展性的 Agentic 系统极其复杂，普通企业切忌在没有深厚积累的情况下完全从头自研。以 Dialpad 的实践为例，一个成熟的 Agentic 平台通常由四大核心模块构成：
1. **技能挖掘（Skills Mining）**：实现人机学习的闭环。系统自动旁听真实业务通话并转录文本，分析人工坐席的高频重复操作，进而向 Agent 系统提出自动化建议与指令。
2. **智能体工坊（Agent Studio）**：由于直接构建 Agent 逻辑极度繁琐，该模块引入 AI 辅助人类更直观地设计、配置与生成 Agent 工作流。
3. **仿真试验场（Proving Ground）**：Agent 上线后最大的痛点是“信任危机”。一旦直接切入生产环境，模型可能会向客户输出违规言论、在 CRM 中误改数据，甚至出现诸如“用户成功忽悠车行 AI 聊天机器人以 1 美元买下一辆卡车”的知名事故。Proving Ground 允许在 Agent 正式上线前，基于历史所有真实的人工与 AI 对话数据进行全量回放仿真，量化评估改动效果，避免盲目切流。
4. **安全与治理（Governance & Observability）**：在生产发布时绝不能一刀切直接上线，必须严格遵循版本控制（Versioning）、灰度比例切流（Percentage Rollout）以及全方位可观测性，并配备能够在 Agent 越界时即时阻断的熔断机制。
```mermaid
flowchart LR
accTitle: Dialpad Agentic System Core Pillars
subgraph Pipeline["Dialpad Agentic 核心链路"]
direction LR
A["技能挖掘<br/>Skills Mining<br/>(分析通话/提炼重复操作)"] --> B["智能体工坊<br/>Agent Studio<br/>(AI 辅助设计 Agent)"]
B --> C["仿真试验场<br/>Proving Ground<br/>(历史全量对话回放验证)"]
C --> D["安全治理与灰度<br/>Governance & Rollout<br/>(版本控制/灰度发布/熔断)"]
end
```
- **实战部署铁律**：严禁直接全量上线 Agent，必须配置可观测监控、灰度切流机制与安全熔断策略。
- **历史数据回放**：通过将 Agent 变更置于历史全部客户交互记录中做模拟测试，能够在不冒犯真实客户的前提下验证其可靠性。
---
### [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) 打造组织智能：当智能泛滥，价值将重归数据与持续学习
理想的客户体验应当由人类与 AI 深度协作构成，双方在业务流中相互赋能、持续学习。企业在引入 AI 时，必须转变思维方式：不要为了用 AI 而强行套用，而应首先自问“业务中哪些流程是高度重复的？哪些环节有提升空间？”，再审视 AI 如何有效切入。同时，企业必须在引入 AI 之前就建立基线度量，否则上线后往往陷入无法评估其实际业务成效的尴尬境地。
当底层的大模型技术逐渐沦为大宗商品（Commodity）、所有竞争对手在明天都能调用相同的顶尖模型时，能够拉开企业差距的唯有**组织智能（Organizational Intelligence）**。技术变革的规律在于：当某种原本昂贵的要素变得廉价且泛滥时，商业价值的锚点就会发生转移。AI 正在让“基础智能”变得极为充沛，但真正的护城河将转移到企业专有的数据资产、沉淀的专业知识体系，以及人机协同的持续学习能力上。理解这一规律并围绕其重构系统的企业，才能在 AI 时代赢得长远胜利。
- **业务衡量原则**：在上线 AI 方案前必须完成基准指标测量，杜绝“无法衡量 ROI”的盲目落地。
- **价值转移规律**：基础智能走向大宗商品化，数据壁垒、专有知识网络与自迭代飞轮构成企业最终的核心竞争力。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=4s) 正如过去的每次技术革命都伴随着某种要素的极度丰裕与低廉化……AI 的核心本质在于让“智能”本身变得丰裕且廉价。当某种要素从稀缺走向丰裕，它就不再是企业的核心壁垒，各类单体 AI Agent 终将同质化为通用的大宗商品。 — Brian Peterson
> [00:06](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=6s) I ask them, 'What are you accomplishing with that AI?' They start with the AI because they have pressure from their board, but they don't know what problem they're actually trying to solve. — Brian Peterson
> [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) AI 绝非仅仅是 Chat GPT 式的“心理咨询师”或单一的操作代理，未来的真正优势在于搭建强化学习飞轮。 — Brian Peterson
> [08:40](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=520s) 赢家系统都必须构建在统一的智能框架之上，将上下文、长短期记忆、安全治理与多模态通道紧密串联。 — Brian Peterson
> [13:48](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=828s) 人类将始终处于系统闭环之中，且人类真实的交互行为正是训练 AI 最核心的优质养料。 — Brian Peterson
> [13:48](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=828s) 在 AI 垂直化深入发展的阶段，公有大模型无法触及企业内部私有的客户历史、业务偏好及独特流程，因此捕获并沉淀私有交互数据将构成企业最坚固的技术护城河。 — Brian Peterson
> [16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) 优秀的 AI 绝不是静态的，而是需要随着持续学习不断改变自身行为并逐步迭代进化。 — Brian Peterson
> [16:35](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=995s) 底层的大语言模型（LLM）充其量只占整个 Agentic 系统的 5% 到 10%。虽然 LLM 是一把至关重要的钥匙，是使一切成为可能的革命性突破，但剩下的 90% 才是真正最难攻克的工程与系统难题。 — Brian Peterson
> [17:39](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1059s) 在生产发布时绝不能一刀切直接上线，必须严格遵循版本控制、灰度比例切流以及全方位可观测性，并配备能够在 Agent 越界时即时阻断的熔断机制。 — Brian Peterson
> [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) 不要为了用 AI 而强行套用，而应首先自问“业务中哪些流程是高度重复的？哪些环节有提升空间？”，再审视 AI 如何有效切入。 — Brian Peterson
> [19:43](https://www.youtube.com/watch?v=TjTVZ-o28a4&t=1183s) 当某种原本昂贵的要素变得廉价且泛滥时，商业价值的锚点就会发生转移。AI 正在让“基础智能”变得极为充沛，但真正的护城河将转移到企业专有的数据资产、沉淀的专业知识体系，以及人机协同的持续学习能力上。 — Brian Peterson
## Mentioned
Brian Peterson, Dialpad, Google Voice, Google, Google+, Netflix, Uber, Motorola, 印第安纳波利斯小马队 (Indianapolis Colts), OpenAI, ChatGPT, CRM, Skills Mining, Agent Studio, Proving Ground
## Video description
Brian Peterson - CTO, Founder, Dialpad
Every major technology shift makes something expensive cheap—and when that happens, competitive advantage moves. AI is making intelligence abundant, but most organizations are still treating AI as a feature instead of redesigning the systems that create value. In this session, Dialpad CTO Brian Peterson explores why the future belongs not to standalone AI agents, but to organizations that build AI systems around context, orchestration, governance, and continuous learning.
[Source](https://www.youtube.com/watch?v=TjTVZ-o28a4)
Source quotes
[0:04] Brian Peterson: 正如过去的每次技术革命都伴随着某种要素的极度丰裕与低廉化……AI 的核心本质在于让“智能”本身变得丰裕且廉价。当某种要素从稀缺走向丰裕，它就不再是企业的核心壁垒，各类单体 AI Agent 终将同质化为通用的大宗商品。
[0:06] Brian Peterson: I ask them, 'What are you accomplishing with that AI?' They start with the AI because they have pressure from their board, but they don't know what problem they're actually trying to solve.
[8:40] Brian Peterson: AI 绝非仅仅是 Chat GPT 式的“心理咨询师”或单一的操作代理，未来的真正优势在于搭建强化学习飞轮。
[8:40] Brian Peterson: 赢家系统都必须构建在统一的智能框架之上，将上下文、长短期记忆、安全治理与多模态通道紧密串联。
[13:48] Brian Peterson: 人类将始终处于系统闭环之中，且人类真实的交互行为正是训练 AI 最核心的优质养料。
[13:48] Brian Peterson: 在 AI 垂直化深入发展的阶段，公有大模型无法触及企业内部私有的客户历史、业务偏好及独特流程，因此捕获并沉淀私有交互数据将构成企业最坚固的技术护城河。
[16:35] Brian Peterson: 优秀的 AI 绝不是静态的，而是需要随着持续学习不断改变自身行为并逐步迭代进化。
[16:35] Brian Peterson: 底层的大语言模型（LLM）充其量只占整个 Agentic 系统的 5% 到 10%。虽然 LLM 是一把至关重要的钥匙，是使一切成为可能的革命性突破，但剩下的 90% 才是真正最难攻克的工程与系统难题。
[17:39] Brian Peterson: 在生产发布时绝不能一刀切直接上线，必须严格遵循版本控制、灰度比例切流以及全方位可观测性，并配备能够在 Agent 越界时即时阻断的熔断机制。
[19:43] Brian Peterson: 不要为了用 AI 而强行套用，而应首先自问“业务中哪些流程是高度重复的？哪些环节有提升空间？”，再审视 AI 如何有效切入。
[19:43] Brian Peterson: 当某种原本昂贵的要素变得廉价且泛滥时，商业价值的锚点就会发生转移。AI 正在让“基础智能”变得极为充沛，但真正的护城河将转移到企业专有的数据资产、沉淀的专业知识体系，以及人机协同的持续学习能力上。
Source quotes
[0:04] Brian Peterson: we finally figured out agents and agents are going to take over the world
[0:04] Brian Peterson: scale agents
[4:45] Brian Peterson: sidebar co-pilot that like you probably never use
[4:45] Brian Peterson: What are you calling us about? And who are you?
[4:45] Brian Peterson: forward deployed engineers
[17:39] Brian Peterson: Please don't try to build agentic yourself. It is so complicated... Don't try it at home yourself.
[17:39] Brian Peterson: sell a customer a truck for one dollar
[19:43] Brian Peterson: I need to apply AI
[19:43] Brian Peterson: What do I do repetitively or could be better?
[19:43] Brian Peterson: If every competitor had access to the exact same AI tomorrow, what would make your customer experience different?
Original transcript
[0:04] Hello everyone. So, I am Brian Peterson.
I'm co-founder and CTO of Dialpad.
[0:10] We are a AI platform for customer experience and
my background was and our team's background was we
[0:19] we were the former Google Voice team. So, we came
out of Google and we thought everything else kind
[0:23] of sucked out there and we were like, we can do
this better. We thought we could do this better.
[0:26] Um there's a lot of just like things that didn't
really work well in business communications and
[0:30] customer experience and we said, "Hey, this
makes sense. We should be able to do this for
[0:35] businesses." Google Voice was a consumer company.
It was we would had about 10 million users
[0:41] uh and we said, "Hey, this this makes sense.
We could do this for um for businesses."
[0:45] We were going to do it at Google, but then Google
decided they wanted to work on Google Plus,
[0:50] if you remember the social network. Um and
so, they said, "We'd rather not do that."
[0:54] Uh and this is kind of how we ended up here.
Um but uh so, Dialpad though, we are in over 50
[1:00] countries. We have over 70,000 companies using us.
We have um uh all different types of users like
[1:10] uh Netflix, Uber, Motorola, uh Indianapolis
Colts, other professional sports teams. Um
[1:18] and we basically provide everything you need
for AI human communications into one thing. But,
[1:23] just for my interest slide, so I walked around
the whole the whole thing and everyone's
[1:28] so, everyone's pitch seems to be about the
same thing. So, I'm going to go a little bit
[1:31] counter to what the pitches you're seeing all
over the world and from AI. And AI went from
[1:37] ChatGPT and LLMs to you know, agents and
agentic. And everyone out there's like,
[1:44] "Well, we finally figured out agents and agents
are going to take over the world." And it's And
[1:48] then most of the taglines are like, "We can scale
agents." And and I think agents are just going to
[1:53] be a means to an end. I think it goes LLM first,
it goes agents, and then what I'm going to talk
[1:58] about is where I think everything is going, which
is the intelligence layer in a system, which some
[2:04] people are calling harnesses these days, but it's
better than that, it's bigger than that. Um So, uh
[2:12] when we go back and where we're in this new
phase is every single technology revolution
[2:17] is something that started with abundance.
So, we had the internet made publishing
[2:23] um abundant, right? You could blog posts, you
could anyone could push stuff to the World Wide
[2:28] Web with the press of a button. Um then
the cloud made infrastructure um abundant.
[2:35] And that then became an explosion of which I'll
show you next, open source made code abundant,
[2:41] um explosion of open source. Now everyone has
free code that they can use to do some really
[2:44] complicated things. And then the one thing that
people don't realize with I think with AI is
[2:49] it's now made intelligence abundant. And by
abundance usually means cheap, too, right? So,
[2:55] um it's That is the thing that now used
to be super expensive was intelligence,
[3:00] and now it's really available. Um and what's that
mean? Cuz the intelligence isn't going to be the
[3:06] thing that's actually the differentiator anymore.
You can tell, no offense to the companies, our
[3:11] companies does the same thing sometimes,
too. But everyone building these AI agents,
[3:16] they're just going to slowly become commodities.
Cuz this We're not there yet. This is the means to
[3:21] the end wasn't just to build automation or basic
automation that's repeatable over and over again,
[3:26] it's actually to um to learn. So, a lot
of what I think where everything's going,
[3:32] and not just with like Dialpad of where
we are we've been investing in, but I
[3:35] I want to say anytime you talk to anyone about
AI and you want to go through anyone's business,
[3:39] whether it's for internal adoption, whether
it's for uh anything you want to automate,
[3:44] the thing that's going to take it to the
next level isn't just agentic systems
[3:48] that are really still hard to configure and really
flaky, it's going to be that intelligence layer of
[3:52] like of reinforcement learning. So, um but
you can see like here's the evolution of
[3:59] of technology in the big boom that we've had. So,
the internet now audiences and platforms become um
[4:07] you know, where the value is, right? It wasn't
just in publishing cuz now publishing is easy.
[4:11] Then the cloud, now with infrastructure,
what happened? Now software becomes the
[4:17] moat. Software becomes the big thing because
infrastructure kind of becomes a commodity.
[4:21] When we started Dialpad, we could do it with 10
people at the very beginning what it would have
[4:24] taken us 100 people just a few years before that.
But the cloud made this an easy thing to do. Um
[4:30] and then, you know, open source it became, okay,
now everyone has access to pretty powerful code.
[4:35] It's going to be about the product experience.
Like, what is your UX user experience like? What
[4:39] is the product actually like? Um and then now
AI, I think people think it's AI was the like
[4:45] chat GPT, I'm your therapist moment, and I like
a really good therapist, and I can write things
[4:51] for you and summarize things. And then they
said, "No, no, no, it's not that. It's actually
[4:56] agentic. Oh, agentic is where it's going to
be. It's going to be the thing that, you know,
[5:00] oh, it's going to automate all these things." But
the problem is is that stuff is, one, everyone
[5:05] in the world's trying to build that. Everyone.
Google, Amazon, Microsoft, right? Everyone's going
[5:10] to have an agentic system and governance and, you
know, and all these things to security and, um and
[5:16] workflows and state diagrams. But that's the next
level and where it's going to end up is where the
[5:22] history has always told us it's going to end up.
Um it's pretty obvious. Like, it's going to end up
[5:27] in what is this like system that then controls all
this stuff and gets better over time. Cuz we talk
[5:33] about artificial intelligence, what what good
is it if it doesn't get better? And if you look
[5:37] at a lot of the artificial intelligence now, you
don't have a lot of control over getting better.
[5:41] Um even the agentic stuff still requires a lot
of humans tweaking it all all time. Um so, we are
[5:48] obviously in customer experience. That's our bread
and butter. We do customer communications. We run,
[5:53] you know, all forms of contact center and
omni channel and AI and humans together. Um
[6:00] You've seen probably everyone in and, you know,
their brother thinking, "Okay, chat GPT came
[6:05] out. Customer experience is like the easiest
thing. It can like sound human. It can answer
[6:09] questions really well." And there's now a million
startups trying to do customer support with AI. Um
[6:15] Because it is kind of the obvious thing that it
could have the biggest advantage over this. So,
[6:19] that is where a lot of this intelligence
is shifting, but just being agentic isn't
[6:22] enough. Like it needs to learn from itself. Um
So, what you ended up with was a lot of companies
[6:27] doing these one-off things. Cuz this is just the
general evolution of where we're going with AI.
[6:31] You started with the co-pilots where
every single product you have now
[6:35] was a SaaS product that now has a sidebar
co-pilot that like you probably never use.
[6:39] Um And it like kind of answers questions. It
kind of has some context. It's better than,
[6:44] you know, your traditional search. It's something.
Chatbots have been a thing around the web, but
[6:49] then you all know the frustration with chatbots.
I can't stand it when I go to a chatbot on a web.
[6:53] I ask my tell them my problem, try to get it to
solve it, and then what do they do? They're like,
[6:57] "We can't solve this problem cuz all I can
do is give you basic answers. Would you like
[7:00] to talk to a representative?" Then I end up
talking to a representative, and they're like,
[7:04] "What do you What are you What are you reaching
us about? What are you calling us about? Or and
[7:08] who are you?" And I need to reauthenticate you.
It's like so freaking frustrating to have to like
[7:13] repeat yourself, and nothing seems to work well
together. Um And then you just see a lot of this
[7:17] stuff. I'm sure you're all experiencing this in
your own company. Like these point automations.
[7:21] You're taking like one thing and like really
trying to get this agentic thing to work to
[7:24] automate the thing repetitively, but if like one
thing changes in it, it's fragile and breaks. Um
[7:31] Those things obviously aren't a good customer
experience yet. Um And this goes into my whole
[7:35] theme of where I think AI and agentic is going.
And again, it's not just for Dialpad. So, if you
[7:41] take away one thing from this, what I'm telling
you is super important for your businesses. It's
[7:45] super important for any any software you use,
not just us as a customer experience platform,
[7:51] or not just for your customer support, not just
for your sales. Like, even internal automations,
[7:56] you're going to have to build this like
reinforcement learning in this like flywheel,
[7:59] as we talk about it. But, because the these point
solutions, these little things, like they're
[8:03] not building a better uh combined system, right?
They're all just jointed. They don't talk to each
[8:08] other well. Maybe they're using different models.
They don't share data. They don't share memory.
[8:13] Um they don't share context, obviously, is the the
typical term. Um Customer experience is obviously
[8:20] all over the place. And it goes for anything you
do with anything you want to automate. So, um
[8:25] Where we're at now is a lot of these AI agentic
systems, and they answer questions. They can do
[8:29] some basic stuff. Um but they're not learning.
They're not learning because they don't have the
[8:33] ability to learn yet. And that's where
things are going to keep going. Uh So,
[8:40] this is what a winning company's
learning system looks like. And again,
[8:45] I don't want to pitch too much for Dialpad. I
would say this applies to anything you are doing
[8:51] with AI. Internal, external, pick any
SaaS product. If they say they have AI,
[8:57] this is what they need to be set up for. If you
have a business building an AI, this is what you
[9:01] got to get towards. And this works, obviously,
specifically for for for customer support and
[9:06] sales, like we're talking about channels, like
voice and chat and emails and meetings. But,
[9:11] whatever that is. That could be your channels
could be every tool you use at your company
[9:16] working together, right? On some AI system. Um
and for us, like we have a bunch of AI agents
[9:22] coming together, and then you have all this stuff
which you see highlighted in the middle, which is
[9:27] These are the things that a lot of the companies
out here are talking about. And we're build we're
[9:30] we've been building this, too. So, it is
necessary. But, this is just the next phase. So,
[9:35] you are going to need better context.
You are going to need to have memory.
[9:39] We are going to have to have security and
governance so it doesn't say something, you know,
[9:44] that your company wouldn't want it to say, right?
Or do, as you all know if you've probably used,
[9:49] right, ChatGPT, you ask the same question twice,
it gives you a different answer. Um sometimes it
[9:54] gives you something that sounds very confident,
but it's based off of the entire World Wide Web
[9:59] of fake news and whatever else, so you can't
really trust it. That's great again for your
[10:03] personal maybe therapist if that's the thing you
like, but but uh but it's not good for businesses
[10:09] trying to make things happen consistently over and
over again and and know their customers and know
[10:14] their how to operate as a business. Um and this
stuff has to constantly be optimized, too. So, um
[10:20] it has to connect all this stuff to actually work.
So, once these conversations become understanding,
[10:28] understanding becomes better decisions. So, and
you can see in then understanding like better
[10:34] decisions create better customer experience,
better experiences create more conversations,
[10:39] which creates more understanding. It just
goes in this little automation loop. So,
[10:43] this is the compounding learning loop.
I'll just keep repeating, but this
[10:49] is something that everyone's going to need whether
you are talking about our area or any AI area.
[10:56] Um especially as new models get introduced,
you can't trust them as much. Um
[11:01] you're going to start to see small language models
incorporated more that are specific to your data.
[11:06] Um you're going to start seeing um all of it
continuously [clears throat] improve itself.
[11:11] Because if you look at most of the where we're
at with Agentic, it it's great at the beginning,
[11:17] but then it's it just starts to slowly break over
time. And you got to keep, you know, twisting and
[11:22] turning the knob and adjusting it. Um and it it's
just really hard. If you look at even deploying
[11:28] um what we've first version Agentic, I'll
call it like the first level of AI agentic.
[11:33] It's a lot of work. Like there's a reason why you
we created roles called forward deployed engineers
[11:38] just to get something shipped. You didn't
need that with SAS products. It's cuz it's
[11:43] not necessarily there yet. It doesn't get
better, so it needs constant maintenance.
[11:47] Um so this is something that everyone's
going to eventually need to have. Um
[11:53] So the other thing that keeps happening is and I
I see it on the taglines again of every booth out
[12:00] there. Cuz I think AI or agentic or production
agents is one of the first things you read.
[12:06] And I only see a few booths that are about What
do you What do you get out of that? And I get a
[12:11] lot of questions. I advise a lot of people on how
to like take advantage of AI in their workplace.
[12:15] And they're like, "How do I going to use AI?"
I'm like, "I don't know. Like what do you need
[12:19] to solve? What's your actual end goal?" They
like start with the AI cuz they have pressure
[12:23] to use AI. They need to show their board. They
need to go like, "Oh, I used AI. Look at me."
[12:28] But it's like, "Well, what are you
accomplishing with that AI?" Like is it
[12:31] What is your What is What are you What are they
What is the means to the end you need to get to?
[12:36] Well, for us obviously it's the customer
experience. Like if you have to reauthenticate,
[12:39] if you have to repeat yourself, if you
call in a month later and the person
[12:44] doesn't know who you are or can't get better
at knowing you, then something's broken. Um so
[12:53] this is where the advantage for
anyone doing any AI will become.
[12:59] Obviously in my biased opinion, personal opinion.
Um if anyone actually tells you they know where AI
[13:04] is going in 10 years, they're probably They're
just all guessing. Just Just letting you know.
[13:09] They're all guessing. We're all guessing. Um
A lot of people say AI is going to take over
[13:13] 100% of our jobs. I don't think that's going
to happen. Um So but this is this is where that
[13:18] I think the obvious part is going towards. And
there's still going to be humans in the loop.
[13:22] The humans are actually going to help train the
AI. That's the other thing people don't don't
[13:26] remember or don't think about is that everyone
assumes it's the AI building itself upon itself.
[13:31] But the AI is also going to get better by learning
from the human interactions. So, if you are
[13:36] sending an email internally or if you are doing a
certain workflow a certain way when it gets really
[13:41] complicated, the AI needs to know that and learn
from that. Um and that's across the board. So,
[13:48] So, here's what is going make a great AI system
overall. You need obviously capture all of this
[13:56] interaction, this data. So, if it's not an
interaction, maybe it's just data. Um a lot of the
[14:00] moat people are going to have you as your business
when we get verticalized in AI it's going to be
[14:04] your data. Like because because Open AI and Sam
Altman only have so much of the data out there,
[14:11] they can't get to your private data. So, this is
going to be your moat at some point is going to be
[14:17] capturing all that data and saving it and using
it. Customers are already getting used to you
[14:22] using their data. Be careful with it. Make sure
not you know, personal identifiable information,
[14:28] but it is going to be something that people are
already used to. You need to then connect to
[14:32] everything. So, everyone's going to have a bunch
of systems. I'm not saying you're not going to. I
[14:36] don't think there's some magical brain system
that's going to take over your whole company.
[14:41] You are going to have to connect to your CRMs.
You are going to have to connect um to all your
[14:45] different knowledge bases and stuff like that. But
it needs to constantly be able to sync and work.
[14:49] Um but this is like the basic stuff. Like this
is what you are already doing probably with with
[14:53] AI agents or at least what the companies that
are out there are trying to get you to do. Um
[14:58] So, and here's the thing that's a little different
what I think is happening which um not enough
[15:02] people are talking about is the coordination. I
do think AI and humans will always coordinate. Um
[15:09] good news is I think we'll always have jobs. Uh
and the good news is they the AI needs us still.
[15:16] They it's going to continually feed on itself.
So, one of the things we've been doing in the
[15:21] the advantage we've had at least on our
system is we also run every single human
[15:25] conversation through our platform. So that human
conversation is actually training the AI agents
[15:30] on what are the repetitive things you do and
how do you do them better because I know when
[15:34] a human says this answer that the customer loves
it. Well, that's the best way to train the AI.
[15:41] That's way to train the AI isn't use
like a large language model and ask it
[15:45] what do you think about this customer because it's
going to not be tailored to that customer. It's
[15:48] not going to be tailored to your business. It's
actually from the human interactions are going
[15:52] to train the AI if you do it right. Um and then of
course this is like the whole theme of everything
[15:56] is is the learning and adapting. This is the
only thing missing still and I think it's the
[16:01] it's the end of the line of like this phase of AI
is systems that continuously learn on themselves.
[16:06] Because these LLMs can't can't learn. That's
the problem. You can pass it all the context
[16:10] you want in the world, but it's going to be
slow and you can only have so much context.
[16:14] Like the future is going to be LLMs or small
language models down to even the individual
[16:19] customer. So when I call into the airline,
the airline like knows exactly who I am.
[16:24] Like has all my history, has what I
like, what's my favorite sports team,
[16:28] like what how many times have I traveled last
year without me having to tell it who I am. Um
[16:35] and then it's going to obviously continue to
change its own behavior from this learning
[16:39] and get better over time. Cuz great AI isn't
going to be static. And I think right now it is.
[16:46] I think you can replace the model underneath
and the model gets a little bit better,
[16:49] but the models aren't getting that much
better. Like and they only have their own data.
[16:54] Um so what you're seeing is these like the
differences is we have a lot of impressive
[17:00] demos because they hard code, you know, here's
this exact workflow and it sounds magical. Um
[17:05] but like these production systems, the real things
need all these other pieces. That's why there's
[17:10] such a boom in agentic AI because if you actually
I've been we've been doing this for a long time.
[17:15] If you look at what takes to do automation like
workflows, the LLM is probably about 5 or 10%
[17:21] of actually the agentic system. It is the key that
made it possible. I think it's a huge revelation,
[17:29] but it's only the 5 or 10% that's going to get
you there. The 90% is the last hardest part of it.
[17:39] These are just four examples of something we built
[17:43] um for our system. To make it scale, to
make it work, to make it learn from itself.
[17:50] I'm telling you these are from Dialpad side of
view point of view, but you're going to see some
[17:54] version of this or at least want some version
of this from anything you buy that has AI.
[17:59] So, if you go to buy something for an internal
automation, which I'll also say please don't try
[18:03] to build agentic yourself. It is so complicated.
Please buy someone. Not even if it's not us.
[18:10] Don't don't don't try it at home yourself. But
like here's an example. We have something called
[18:14] skills mining. That's the human learning.
It automatically listens in on calls,
[18:19] transcribes them, finds out what your humans are
doing repetitively, and then instructs the agentic
[18:25] system, here's what you can do to automate
it, and suggest it. Like we have agent studio,
[18:29] which is like building agents are complicated.
You have AI helping you build the agents.
[18:34] You have proving ground. The hardest thing with
these agents is once you put them in production,
[18:39] you can't trust them. It's really hard, right?
Cuz you do these things, and you don't want to
[18:42] just flip the switch, and then oh shoot, like it
just said this inappropriate thing to a customer,
[18:48] or it made a change in my CRM it shouldn't have
done, or if you ever heard the famous story of the
[18:54] the person who convinced the LLM chatbot to
like sell them a truck for like a dollar,
[18:58] right? And they had they had to have like
uphold it cuz the it was from their own website.
[19:02] Like these are the things you can't just like
flip on. So, how do you know that these agents
[19:06] are going to work at scale? Proving ground
was a way to say, we'll simulate your agent
[19:10] changes before they go live on every single
historical conversation you've ever had
[19:15] with the AI agent or the human agent. And
then we'll tell you whether or not it would
[19:19] actually do well when you go live. And
even then when you go live, which is a
[19:23] moral of the story for anything you do in Agentic
is never flip the whole switch. Like you got to
[19:28] do versioning. You've got to do percentage
roll out. You have to have observability. Um
[19:33] and then the last thing which is you'll hear
governance and security and stuff and all that. Um
[19:39] You need something to overlook, make sure it's not
stepping outside the bounds and can shut it down.
[19:43] Um So, this is what a customer should
actually should experience if it's done right.
[19:50] Um it's a combination of humans and AI together
um continuously learning off of each other.
[19:59] And of course, back to whatever problem you're
trying to solve, stop trying to say I need to
[20:03] apply AI. Start saying, "What do I What do
I do repetitively or could be better?" And
[20:08] then how does AI apply to it? And make sure
you measure it before you put a AI in versus
[20:13] after cuz a lot of people then do the AI and then
they're like, "Did it work?" And they're like,
[20:16] "I don't know. I think it did." Well, you needed
to know when you actually put in the AI. So, um
[20:22] if every competitor had access to the exact
same AI tomorrow, what would make your customer
[20:25] experience different? And again, apply this
to anything. Um because a lot of this stuff is
[20:31] becoming commodities. Um but this is this is what
we believe the future is. This is what um becomes
[20:38] organizational intelligence. And it compounds
and it replicates um and it feeds itself. So,
[20:47] AI is making intelligence abundance
abundance and it is wonderful, but um
[20:53] that abundance always changes where value
lives. And the value is going to be in
[20:57] the data. It's going to be in the knowledge.
It's going to be in the continuous learning.
[21:01] And so, companies that understand this and build
around this are going to are going to succeed. So,
[21:06] uh thank you. Uh that was really quick
for me. Okay. Uh, I went a lot of slides.
[21:12] Um, appreciate it. If you have any more questions
about anything, and I can talk about anything in
[21:16] AI. Love talking about anything in AI. I'll
be outside. Um, but thank you. All right.
