## Background
This video was published on YouTube by the channel Ai4 on August 15, 2026, featuring Pinaki Saha, Founder & CEO of Anshar Labs. The talk addresses why 79% of enterprises adopt AI agents while only 11% successfully run them in production, offering a practical playbook to transition agent demos into scalable, real-world applications.
## TL;DR
Pinaki Saha explores the chasm between enterprise AI agent pilots and actual production deployment, highlighting how data unreadiness, lack of single-owner governance, and isolation from real workflows cause most pilots to fail. Drawing from four concrete case studies—automotive security surveillance, financial schema modeling, live media moderation, and automated video course generation—he outlines specific technical bottlenecks such as retrieval limits and unintegrated data feeds. Ultimately, he presents the playbook shared by the surviving 11% of enterprise implementations, emphasizing unified data ownership, designing for model failure, robust citation oversight, and single-pipeline governance.
## Atmosphere & style
The presentation carries an authoritative, pragmatic, and analytical tone typical of a seasoned tech founder speaking directly to enterprise leaders. Saha delivers a candid post-mortem on the real-world friction points of AI deployments, replacing industry hype with concrete architectural realities. The atmosphere is structured, educational, and grounded in field experience, creating a focused arc from systemic failure modes to clear, actionable rules for success.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) There is a massive gap in enterprise AI deployment: while over 75% of corporations have deployed AI agent pilots, only about 11% successfully reach production scale.
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) Data unreadiness, such as dirty, unoptimized, or fragmented data across departments, is a primary reason AI agent pilots fail to reach production.
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) A lack of cross-departmental oversight and governance prevents teams from sharing data or taking ownership of the AI governance pipeline.
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) Demos often fail to transition to production because stakeholders fear integrating agents into real workflows due to security and data privacy concerns.
- [05:40](https://www.youtube.com/watch?v=UTKYepbglsI&t=340s) An automotive security surveillance system using Agentic GraphRAG and VLMs reduced monitoring staff from 25 people down to 3 operators handling real-time events across 500 CCTV feeds.
- [05:40](https://www.youtube.com/watch?v=UTKYepbglsI&t=340s) In a financial agent deployment, the agent had to be rebuilt four times due to schema misalignment and stakeholder refusal to provide optimized data sources for the knowledge graph.
- [06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) A live Twitch stream moderation agent failed and produced false flags because isolated data threads forced the system to rely on downstream assumptions rather than a unified knowledge graph.
- [06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) A multi-agent pipeline using four specialized agents (Extraction, SOP Writing, Citation Check, Media Production) struggled when missing training data caused VLM asset generation to diverge from domain standards.
- [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Technical bottlenecks vary by domain, such as real-time retrieval in surveillance, schema references in finance, isolated data flow in moderation, and model optimization in training video creation.
- [11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) Successful enterprise deployments require collective data ownership within a clear sovereign governance framework so agents can query a unified knowledge graph fabric.
- [11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) Surviving projects treat AI agent deployment as an iterative process and design for model failure rather than abandoning pilots after initial errors.
- [11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) Assigning a single universal stakeholder to own the end-to-end multi-agent pipeline prevents the fragmented failure modes of isolated agent implementations.
- [11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) While less than 5% of enterprise applications currently run production agents, adopting structured execution practices could drive enterprise production adoption up to 40%.
## Detailed walkthrough
Here is a detailed, faithful walkthrough of the transcript:
[00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) The Enterprise Agent Gap: Why Pilots Die on the Vine
Pinaki Saha, Founder and CEO of Anshar AI (a boutique knowledge graph and agentic deployment company serving mid-market to Fortune 1000 enterprises), introduces the central challenge facing the industry today: while "agentics" is the current buzzword and agent deployments are everywhere, very few survive past the pilot stage to reach production scale. Market research shows that over three-quarters (over 75%) of corporations have deployed AI agents in some shape or form, yet only 1 in 9 (roughly 11%) actually reach production. Most organizations simply play with the technology in a sandbox, solve a couple of isolated use cases, and then allow the initiative to "die on the vine."
Saha highlights three primary root causes behind why enterprise AI agents fail to cross the chasm from pilot to production:
* **Data Unreadiness:** The classic "garbage in, garbage out" (GIGO) principle holds true for modern AI. Regardless of the graph databases, LLM layers, or synthesis frameworks used, inference and prediction models fail if the underlying data is dirty, unoptimized, or fragmented across corporate departments.
* **Lack of Oversight and Governance:** Cross-departmental stakeholders often lack trust in the process. Without single-owner oversight, teams refuse to share their data or take ownership of the governance pipeline.
* **Failure to Touch Real Workflows:** Most demos succeed only because they answer questions inside an isolated sandbox. Once moved beyond the sandbox, stakeholders balk at integrating the agent into real workflows due to fears that sensitive data will leave their departmental enclave or cross security firewalls.
[00:340] Case Studies: Surveillance Cameras and Financial Data Schemas
Drawing from field reports across four real-world deployments that Anshar AI pushed across the pilot-to-production chasm, Saha breaks down the technical lessons learned from initial failures:
* **Automotive Security Surveillance:** Deployed across acres of assembly lot parking in countries like Vietnam, India, and the Middle East, this platform ingests live feeds from 500 CCTV cameras (including drone-level video) to track theft and pillage before finished cars ship to dealerships. By implementing an agentic platform using Agentic GraphRAG, Visual Language Models (VLMs), and GPU-level detection, the company reduced the required monitoring staff from 25 people down to 3 operators handling roughly 25 monitor panels to trigger real-time security events.
* **Financial Agent Schema Misalignment:** In a financial application, Anshar AI ended up having to rebuild an agent four separate times. The demo failed repeatedly because key stakeholders refused to provide the right datasets. The lead manager failed to understand which data sources needed to be optimized to feed the knowledge graph fabric, proving that improper data cleansing and restructuring prior to LLM ingestion inevitably leads to broken outcomes.
[06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) Case Studies: Live Media Moderation and Automated Course Generation
Saha shares two additional complex deployment examples:
* **Live Stream Moderation (Twitch):** Working with live video streams on Twitch, the goal was to perform real-time audio and video analysis to detect egregious statements or comments within a 3-to-4 second window, instantly pulling brand sponsor ads before re-inserting them once the offensive frame passed. However, the system acted as a complete black box. Because second- and third-party stakeholders failed to disclose their data in a single unified thread, the agent made wrong assumptions, resulting in incorrect flags. This taught Anshar AI that multi-party data must be fully integrated into a single knowledge graph fabric rather than relying on downstream hunches.
* **Unstructured Data to Training Media:** This multi-agent pipeline ingested unstructured PDFs, manuals, podcasts, and long-form videos to automatically build structured educational video curricula. The architecture utilized four specialized agents operating in serial and parallel:
Despite the design, the system broke down when the VLM generated assets that were out of tune with the underlying training data. Because stakeholders were unwilling to share complete learning datasets to properly train the ML models, the VLM failed to accurately interpolate the data.
[10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Technical Summary of Bottlenecks
Saha summarizes the specific technical bottlenecks discovered across all four domains:
- **Surveillance:** *Retrieval* is the primary bottleneck when handling large, real-time camera feeds without a sophisticated streaming pipeline.
- **Finance:** Incorrect *schema references* and missing data from reluctant stakeholders break agent threading.
- **Live Stream Moderation:** Lack of unified multi-party data flow forces agents to operate on downstream assumptions, causing false flags.
- **Training Video Creation:** Insufficient domain training optimization for VLMs causes media outputs to fall short of quality benchmarks.
[11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) The Playbook of the Surviving 11% and the Path to 2026
The top 11% of enterprises that successfully take agent prototypes into production share four key habits:
1. **Data Ownership:** All stakeholders collectively own and share data within a clear sovereign governance premise, allowing agents to query the unified knowledge graph fabric.
2. **Design for Failure & Iteration:** Successful teams expect initial failure and treat agent deployment as an iterative process rather than canceling projects at the first setback.
3. **Robust Governance & Citation Oversight:** They build clear oversight for data governance, ethics (especially for legal data), and citation quality.
4. **Single Pipeline Ownership:** A single universal stakeholder owns the entire multi-agent pipeline, avoiding the fragmented failure mode of separate stakeholders running isolated agents.
Saha notes that looking ahead in 2026, enterprise adoption is shifting gears. While less than 5% of enterprise applications currently run production agents, the market represents an $8.5 billion pipeline opportunity that could scale 5x to 6x—reaching up to 40% enterprise production adoption—if companies adopt these structured execution rules. As Saha concludes, "Anybody can build a demo... The hardest part is to make it work in real-time scale."
## Notable quotes
> [11:47](https://www.youtube.com/watch?v=UTKYepbglsI&t=707s) Anybody can build a demo... The hardest part is to make it work in real-time scale. — Pinaki Saha
## Video description
Pinaki Saha - Founder & CEO, Anshar Labs
Why do 79% of enterprises adopt AI agents while only 11% actually run them in production? I built and shipped four agents that crossed that gap, in security surveillance, finance, live content moderation, and training video generation, and each one taught the same lesson: the model was never the hard part. In this talk, I'll break down what actually kills a demo before it reaches production (data that isn't ready, no one owning oversight, a workflow that never touches the real system) and the four practices the surviving 11% share, from owning the data pipeline to designing for the model being confidently wrong. Come if you want a concrete playbook for turning a demo that worked once into an agent that runs unattended, at scale, in the real world.
According to Anshar AI founder Pinaki Saha, more than 75% of corporations have experimented with AI agents, but only about 11% survive past the pilot stage to reach production scale. This failure stems primarily from fragmented enterprise data, missing governance ownership, and reluctance to integrate agents into live, mission-critical workflows.
Why Enterprise Pilots Die on the Vine (0:04)
Saha observes that while agent deployment is an industry buzzword, less than 5% of enterprise applications currently run production agents. Most corporate initiatives remain confined to sandboxes, solving isolated use cases before being abandoned. This drop-off is driven by three main operational roadblocks: data unreadiness, absent governance, and isolation from actual workflows.
First, underlying data is frequently dirty, unoptimized, and trapped in departmental silos, causing inference models to fail on basic "garbage in, garbage out" dynamics regardless of the sophisticated graph or LLM architectures used. Second, enterprise departments often distrust cross-functional data sharing, leading to pipelines with no single accountable owner. Finally, while demos function smoothly in secure, artificial sandboxes, organizational stakeholders balk at moving them into live operational workflows due to fears of leaking proprietary data outside departmental firewalls.
Field Lessons and Technical Failure Modes
Drawing on deployments Anshar AI moved into production, Saha details the distinct technical bottlenecks that cause agents to fail across four specific enterprise domains. In automotive security, an agentic system monitored 500 CCTV and drone feeds across assembly lots in Vietnam, India, and the Middle East to prevent theft. Utilizing Agentic GraphRAG, Visual Language Models (VLMs), and GPU-level detection, the platform successfully cut monitoring staff from 25 people to 3 operators managing 25 panels, identifying real-time data retrieval as the primary technical hurdle.
Conversely, other pilots faltered due to schema and transparency failures. A financial agent had to be rebuilt four times because business managers failed to identify required data sources, and stakeholders refused to supply correctly structured datasets. In live stream moderation on Twitch, an agent intended to pull brand ads within a 3-to-4 second window of offensive speech generated repeated false flags; without a unified data thread connecting second- and third-party stakeholders, the agent relied on flawed downstream assumptions. Similarly, an automated four-agent educational pipeline—spanning extraction, curriculum writing, citation checking, and media generation—broke down because stakeholders withheld key domain learning sets, causing the final VLM to output distorted media assets.
Execution Rules of the Surviving 11% (11:47)
Enterprises that successfully transition agents from prototype to production follow four operational principles, according to Saha. First, they establish unified data ownership, enabling agents to query an integrated knowledge graph fabric governed by clear sovereign data rules. Second, successful organizations treat agent deployment as an iterative process and design for initial failure rather than abandoning projects after early setbacks.
Third, viable deployments maintain strict oversight pipelines covering data ethics, compliance (particularly with legal materials), and automated citation verification. Fourth, organizations assign a single universal stakeholder to own the entire multi-agent pipeline from end to end, preventing the breakdowns that occur when disparate teams run uncoordinated, isolated agent components.
Saha emphasizes that building functional demos is easy compared to operating agents at real-time scale. He projects that while current production penetration remains under 5%, the current $8.5 billion enterprise pipeline could expand five-to-six-fold to reach roughly 40% enterprise adoption by 2026 if organizations adopt rigorous data governance and unified architecture.
## Background
该视频由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布。视频记录了 Anshar Labs / Anshar AI 创始人兼 CEO Pinaki Saha 的主题演讲，主要探讨为何高达 79% 的企业尝试部署 AI 智能体却仅有 11% 能够成功推向生产环境，并揭示了阻碍 Demo 规模化落地的核心根因与突围方案。
## TL;DR
Pinaki Saha 指出，大模型本身并非智能体落地的最大瓶颈，数据未准备就绪、监管权属缺失以及脱离真实业务工作流才是扼杀试点项目的核心根源。他结合安防监控、金融数据分析、Twitch 实时内容风控及培训视频生成等四大真实落地案例，深入剖析了高并发实时检索、多源黑盒数据断层及 Schema 引用失真等关键痛点。针对成功跨越鸿沟的 11% 企业，他总结了共同掌控数据权属、接纳失败持续迭代、建立完备监管体系与统一流水线负责人四大核心法则。演讲强调唯有夯实底层数据管道与协同架构，才能实现智能体在企业生产环境中的规模化稳定运行。
## Atmosphere & style
整场演讲风格硬核务实、直击痛点，毫无浮夸的技术炒作，洋溢着浓厚的一线工程实战气息。讲者语调沉稳且富有洞见，通过坦诚复盘项目经历四次重构的挫折与教训，展现出极高的专业说服力与工程反思精神。叙事节奏层层递进，从揭示残酷的行业淘汰现实自然过渡到系统性的解题方法论，为听众带来了一场兼具警示意义与实操价值的企业级 AI 落地复盘。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) 企业智能体落地面临巨大落差：近 79% 的企业正在测试或部署 AI 智能体原型，但最终仅有约 11% 能够成功推向生产环境。
- [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) 阻碍智能体走出试点进入生产环境的核心瓶颈不在于模型算法本身，而在于跨越 Demo 与生产之间的工程与治理鸿沟。
- [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) 智能体死于试点阶段的第一大根因是底层数据未准备就绪，垃圾数据输入（GIGO）直接导致上层模型推理与预测崩溃。
- [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) 监管与治理权属缺失是智能体夭折的第二大根因，跨部门信任缺失与责任主体模糊导致数据无法打通、风险无人承担。
- [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) 脱离真实业务工作流是第三大根因，仅在隔离沙箱中运行的 Demo 一旦面临防火墙与跨部门协同阻力便难以推向实战。
- [03:40](https://www.youtube.com/watch?v=UTKYepbglsI&t=220s) 在安防监控落地中，GraphRAG 与视觉语言模型（VLM）结合将监控人力由 25 人锐减至 3 人，但高并发视频流的实时检索管道成为技术关键瓶颈。
- [03:40](https://www.youtube.com/watch?v=UTKYepbglsI&t=220s) 金融场景智能体四次重构的教训表明，若底层数据 Schema 混乱且缺乏统一知识图谱结构，直接接入 LLM 会导致结果产生严重偏差。
- [06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) Twitch 实时内容风控场景展示出多方数据孤岛与黑盒处理会引发大量误报，多源数据必须在统一的知识图谱下进行协同编排。
- [06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) 多智能体流水线在处理复杂任务（如培训视频生成）时，各节点间的数据授权与模型表现若无法达到基准线，将导致整个链路断裂。
- [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) 成功突围生产环境的核心法则之一是全员共同掌控数据权属，各方主动开放数据并沉淀在统一的知识图谱中。
- [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) 跨越试点的企业善于接纳失败并持续迭代，而不是在系统出现单次推断错误时就盲目砍掉项目。
- [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) 必须建立端到端的数据监管体系，统一数据治理标准并严格把控敏感业务中的 AI 合规边界。
- [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) 多智能体系统必须确立单一的全局流水线负责人，彻底避免多部门各自掌管孤立智能体导致的多头割裂局面。
- [13:49](https://www.youtube.com/watch?v=UTKYepbglsI&t=829s) 随着数据治理与工程化落地规范的成熟，企业生产级智能体上线率有望从不足 5% 提升至 40%，释放巨大的商业价值。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) 智能体落地的“死谷”：从概念验证到生产环境的鸿沟
在当今企业技术领域，“智能体（Agentic）”已然成为最热门的行业关键词，各类 AI Agent 的试点项目如雨后春笋般涌现。然而，Anshar AI（Anshar Labs）创始人兼 CEO Pinaki Saha 指出，行业目前正面临一个极其严峻的现实：绝大多数智能体都在试点（Pilot）阶段走向消亡，无法跨入真正的生产环境。市场研究数据显示，超过四分之三（近 79%）的企业在不同形式上部署或测试了 AI 智能体，但最终能够真正推向生产环境、实现规模化运行的比例仅有九分之一（约 11%）。多数企业仅仅是在沙箱中做尝试、解决一两个边缘用例，随后项目便“在藤蔓上枯萎凋零”（die on the vine）。
作为一家专注于为中型企业及财富 1000 强企业提供知识图谱与智能体落地服务的精品咨询公司，Anshar AI 在安防监控、金融、实时流媒体风控以及培训视频生成等四个实际业务场景中，完整经历了从试点到生产环境的跨越。Pinaki Saha 强调，模型的算法本身从来不是最难的部分，真正的挑战在于跨越从 Demo 到生产的“鸿沟（Chasm）”。
- **市场采纳反差**：超过 75%（约 79%）的企业正在部署 AI 智能体原型，但仅有约 11%（1/9）能够成功上线投产。
- **核心定位**：Anshar AI 专注服务中型企业到 Fortune 1000 企业，深耕知识图谱与多智能体落地。
---
### [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) 扼杀试点 Demo 的三大致命根源
智能体之所以难以走出试点阶段，Pinaki Saha 总结了三大核心根因。第一是“数据未准备就绪（Data isn't ready）”。传统的“垃圾进，垃圾出”（Garbage in, garbage out, GIGO）定律在 AI 时代依然成立。无论企业引入多么先进的图数据库（Graph Database）、多么强大的大语言模型（LLM）层或复杂的数据合成技术，只要底层数据不够干净、没有结构化就绪，任何在其上构建的预测与推理模型都无法正常运转，尤其在跨多部门协同的企业环境中，数据断层尤为严重。
第二是“缺乏监管与治理权属（No one owns oversight）”。这直接关乎企业内部的治理机制与信任壁垒。各业务部门之间往往缺乏对 AI 流程的互信，导致大家都不愿意完全共享自身的数据资产，也没有人愿意为智能体的端到端输出质量与风险承担责任。
第三是“从未触及真实业务工作流（Never touched the real workflow）”。许多 Demo 仅能在隔离的沙箱（Sandbox）中回答预设问题，一旦尝试向真实业务迈进，其他部门便会因担忧数据泄露或跨越防火墙而拒绝配合。这导致智能体始终无法融入真正需要多部门深度协同的复杂工作流中。
```mermaid
flowchart TD
accTitle: 阻碍智能体从试点走向生产的三大根因
subgraph S1["数据准备不足 (Data Readiness)"]
A1["底层数据未清洗 / 未结构化"] --> A2["触发 GIGO 效应，推理与预测崩溃"]
end
subgraph S2["监管与权属缺失 (Oversight & Governance)"]
B1["跨部门互信不足 / 数据孤岛"] --> B2["无人为最终结果与风险负责"]
end
subgraph S3["脱离实际工作流 (Workflow Integration)"]
C1["仅在隔离沙箱中运行"] --> C2["面临防火墙与合规阻力，无法打通跨部门链路"]
end
```
---
### [03:40](https://www.youtube.com/watch?v=UTKYepbglsI&t=220s) 实战案例一与二：安防监控与金融知识图谱的落地教训
在安防监控场景中，团队为一家位于越南、印度及中东等地的汽车制造企业部署了实时智能体平台。该工厂在装配线外拥有数英亩的露天停车场，成千上万辆新车在交付经销商前停放于此，面临盗窃、零部件拆解（pillage）及车辆异常移动的高风险。现场部署了 500 多路 CCTV 摄像头及无人机巡检视角，监控室内原本需要 25 名安保人员时刻紧盯近 25 块屏幕，人力严重不足。团队通过结合知识图谱 RAG（GraphRAG）与视觉语言模型（VLM），在 GPU 端实现了实时的视觉活动检测与事件触发，成功将值班监控人力从 25 人大幅缩减至 3 人。然而该项目在技术上遭遇的重大瓶颈在于“实时检索（Retrieval）”，当高通量实时视频流涌入时，若底层缺乏专门设计的管道，智能体便无法保持各节点的同步运作。
在金融数据分析场景中，团队遭遇了更为曲折的挫折，该智能体系统前前后后重构了整整四次。失败的根源在于数据 Schema 的引用严重失真，数据层未能有效整合进知识图谱的网络之中。业务部门的主管无法理清哪些数据源能为前端用户提供最优价值，导致各个利益相关方提供的数据集碎片化且不完整。这给团队带来了极为深刻的教训：在将数据喂入 LLM 或构建多智能体应用之前，若不彻底进行数据重构与清洗，智能体给出的结果将与预期产生南辕北辙的偏差。
- **汽车制造安防项目覆盖范围**：涵盖越南、印度和中东工厂，管理 500+ 路 CCTV 与无人机视频流。
- **人效提升指标**：监控室内原本负责盯防 25 块监视屏的人员编制从 25 人精简至 3 人。
- **技术栈支撑**：GraphRAG + VLM（视觉语言模型）+ 边缘/云端 GPU 实时事件检测。
---
### [06:32](https://www.youtube.com/watch?v=UTKYepbglsI&t=392s) 实战案例三与四：Twitch 实时内容风控与多智能体培训视频生成
在与知名流媒体平台 Twitch 合作的实时内容审核场景中，品牌广告赞助商希望将广告精准投放在直播活动旁，但一旦主播出现违规言论或极端不良行为，广告系统必须在短短 3 到 4 秒的极窄窗口内迅速撤下广告以保护品牌声誉，并在内容恢复正常后重新加载。该系统涉及极度复杂的实时音视频联合分析，且上下游呈现典型的“黑盒”状态：一方提供实时数据流，中间层进行计算处理，Anshar AI 则在前端运行智能体。由于涉及第二方与第三方数据，一旦合作方未在统一通道内披露完整上下文，智能体就只能依赖推测，从而产生大量的误报（Wrong flags）。这表明，多方数据必须在统一的知识图谱架构下深度协同编排，否则智能体必然崩溃。
另一个代表性项目是将非结构化企业资产（PDF 手册、SOP 文档、播客音频、长视频等）自动转化为系统化的培训课程与多媒体视频资产。该系统由四个智能体组成协同流水线：
1. **提取智能体**：负责抽取非结构化内容并建立实体关系；
2. **课程编写智能体**：依据 SOP 与材料梳理教学大纲与培训课程；
3. **校验与引用智能体**：核验所有生成内容的合法性与引用准确性；
4. **媒体制作智能体**：将文本与结构化数据渲染为最终的培训多媒体内容。
在准实时生成媒体内容的过程中，系统面临的卡点是视觉语言模型（VLM）的输出有时会与原始训练数据脱节。由于业务利益相关方未完全授权核心教学数据用于模型微调与训练，导致智能体在进行数据推断与媒体生成时无法达到基准要求，最终出现流程断裂。
```mermaid
flowchart LR
accTitle: 非结构化数据转培训多媒体的多智能体串并联工作流
subgraph Input["原始非结构化数据"]
A["PDF / SOP 手册 / 播客 / 视频"]
end
subgraph Pipeline["四智能体协同流水线"]
B["1. 提取智能体<br/>(抽取实体与关系)"]
C["2. 课程编写智能体<br/>(生成大纲与课程)"]
D["3. 校验引用智能体<br/>(合规审查与引用核对)"]
E["4. 媒体制作智能体<br/>(多媒体资产渲染)"]
end
subgraph Output["最终交付"]
F["结构化培训视频与课程资产"]
end
A --> B
B --> C
C --> D
D --> E
E --> F
```
---
### [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) 四大场景技术瓶颈盘点与成功跨越鸿沟的四大法则
Pinaki Saha 对上述四个真实场景的技术痛点进行了系统复盘：
- **安防监控**：瓶颈在于高并发视频流的“实时检索（Retrieval）”链路；
- **金融场景**：瓶颈在于底层“数据 Schema 引用”混乱与数据不完整；
- **实时风控（Twitch）**：瓶颈在于二方/三方多源数据断层导致的“上下文缺失与黑盒推断”；
- **视频生成**：瓶颈在于视觉语言模型（VLM）学习优化不足，生成表现跌落基准线。
针对成功突围试点的 11% 幸存企业，Pinaki Saha 提炼出了四项核心共性实践法则：
| 成功实践法则 | 核心内涵与执行要点 |
|:--- |:--- |
| **1. 共同掌控数据权属 (Own the data)** | 所有业务相关方必须在数据主权与治理框架下达成一致，主动开放数据资产，确保智能体能够在统一的知识图谱中无缝检索与调用数据。 |
| **2. 接纳失败并坚持迭代 (Embrace failure & iterate)** | 摒弃“失败一次就直接砍掉（Can it）”的短视心态。生产级智能体必须建立在反复试错、持续迭代与对落地流程的高度信任之上。 |
| **3. 建立完备的数据监管体系 (Build oversight & governance)** | 明确数据治理与引用质量标准，厘清监管责任归属，并在涉及法律法规等敏感业务时，严格把控 AI 伦理与合规边界。 |
| **4. 统一端到端流水线负责人 (Single pipeline ownership)** | 必须由单一的全局负责人统筹所有智能体的数据流水线，坚决避免“四个业务方各自掌管四个孤立智能体”的多头割裂局面。 |
---
### [13:49](https://www.youtube.com/watch?v=UTKYepbglsI&t=829s) 生产级智能体的规模化前景与总结
展望行业前景，Pinaki Saha 指出，预计到 2026 年，将有海量智能体项目加速从概念验证冲向生产环境。目前全行业积压的试点项目规模高达约 85 亿美元，随着落地工程方法的成熟，推向生产后所释放的商业价值将迎来 5 到 6 倍的爆发式增长。现阶段企业级应用中成功上线生产级智能体的比例不足 5%，但如果企业能够严格遵循上述数据治理、统一管线与迭代架构等实施规范，这一比例有望跃升至 40%。
Pinaki Saha 在总结时强调：“任何人都能做一个炫酷的 Demo，任何人都能跑通一个孤立用例；但最艰难的部分永远是如何在真实的生产规模下稳定运行。”面对复杂的跨部门诉求、多变的处理速度要求以及对错误高容忍度的系统架构设计，唯有从底层数据管道扎实做起，才能打造出真正经得起生产检验的企业级 AI 智能体。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) Why do 79% of enterprises adopt AI agents while only 11% actually run them in production? — Pinaki Saha
> [00:04](https://www.youtube.com/watch?v=UTKYepbglsI&t=4s) the model was never the hard part. — Pinaki Saha
> [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) Garbage in, garbage out — Pinaki Saha
> [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) Data isn't ready — Pinaki Saha
> [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) No one owns oversight — Pinaki Saha
> [01:53](https://www.youtube.com/watch?v=UTKYepbglsI&t=113s) Never touched the real workflow — Pinaki Saha
> [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Own the data — Pinaki Saha
> [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Embrace failure & iterate — Pinaki Saha
> [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Build oversight & governance — Pinaki Saha
> [10:27](https://www.youtube.com/watch?v=UTKYepbglsI&t=627s) Single pipeline ownership — Pinaki Saha
> [13:49](https://www.youtube.com/watch?v=UTKYepbglsI&t=829s) 任何人都能做一个炫酷的 Demo，任何人都能跑通一个孤立用例；但最艰难的部分永远是如何在真实的生产规模下稳定运行。 — Pinaki Saha
## Mentioned
Pinaki Saha, Anshar Labs, Anshar AI, Twitch, Fortune 1000, GraphRAG, VLM
## Video description
Pinaki Saha - Founder & CEO, Anshar Labs
Why do 79% of enterprises adopt AI agents while only 11% actually run them in production? I built and shipped four agents that crossed that gap, in security surveillance, finance, live content moderation, and training video generation, and each one taught the same lesson: the model was never the hard part. In this talk, I'll break down what actually kills a demo before it reaches production (data that isn't ready, no one owning oversight, a workflow that never touches the real system) and the four practices the surviving 11% share, from owning the data pipeline to designing for the model being confidently wrong. Come if you want a concrete playbook for turning a demo that worked once into an agent that runs unattended, at scale, in the real world.
[Source](https://www.youtube.com/watch?v=UTKYepbglsI)
Source quotes
[0:04] Pinaki Saha: Why do 79% of enterprises adopt AI agents while only 11% actually run them in production?
[0:04] Pinaki Saha: the model was never the hard part.
[1:53] Pinaki Saha: Garbage in, garbage out
[1:53] Pinaki Saha: Data isn't ready
[1:53] Pinaki Saha: No one owns oversight
[1:53] Pinaki Saha: Never touched the real workflow
[10:27] Pinaki Saha: Own the data
[10:27] Pinaki Saha: Embrace failure & iterate
[10:27] Pinaki Saha: Build oversight & governance
[10:27] Pinaki Saha: Single pipeline ownership
[13:49] Pinaki Saha: 任何人都能做一个炫酷的 Demo，任何人都能跑通一个孤立用例；但最艰难的部分永远是如何在真实的生产规模下稳定运行。
Source quotes
[11:47] Pinaki Saha: Anybody can build a demo... The hardest part is to make it work in real-time scale.
Original transcript
[0:04] We'll talk today a little bit about an interesting
concept which you're seeing in the industry today
[0:08] is that a lot of agents being deployed all over
the place. Agentics is the word of the day.
[0:14] Uh but then how many goes beyond pilot and goes
to production level? And the challenge that we're
[0:21] seeing that agents are dying beyond a pilot. Um
and we'll talk about that through four use cases
[0:27] that we have deployed. Uh my name is Pinaki Saha
once again. I represent I'm the founder and CEO
[0:33] of company called Anshor AI. And we are a boutique
uh knowledge graph and agentic deployment company
[0:41] for mid to Fortune 1000 companies. So, this is a
field report from four deployments we did where we
[0:49] first did the pilot and then we have now moved
into production and we saw that where the pilots
[0:57] have highest propensity to fail. And I will also
share and discuss what measures we took took um
[1:05] in conjunction with the stakeholders to make those
pilots real in a production environment. So, from
[1:12] a market study itself, uh we're seeing that almost
um more than qua three quarter of all deployments
[1:23] um in the corporations today have some degree
some shape or form AI agentics AI agents have
[1:29] been deployed. But only just one out of nine of
them are going to production. Beyond just um any
[1:40] uh scale. They're not reaching scale. Uh they
just want to test it out. They play with it. They
[1:46] kind of solve couple of use cases
and then they just die on the vine.
[1:53] So, the three key areas the three reasons we
are saying that how these things are failing
[1:59] to go beyond a pilot is that the data isn't
ready. Um garbage in, garbage out. We remember
[2:05] the GIGO days. And it's true for AI also. Uh no
matter what kind of graph database, what kind of
[2:13] layer of LLMs, uh or synthesis of data,
whatever you bring on, if the data is not clean,
[2:19] if the data is not ready, uh you just cannot
run a prediction model or a inference model
[2:25] on top of it. So, in many cases we're seeing the
data it really is not ready. And definitely not in
[2:30] corporations where they have multiple stakeholders
across different depart- different departments.
[2:35] No one owns oversight. So, there's a governance
issue, there's a trust issue that's involved where
[2:41] different stakeholders are not trusting the
process and not sharing all of their data
[2:46] and not owning up. And the third one is that
it um never touched the real workflow. Uh
[2:54] the demo answers a question in a sandbox. And when
they decide that we have to go beyond sandbox,
[3:00] uh then the other stakeholders are
not likely to participate because
[3:04] they feel that data is not in their
enclave um and will cross the firewall.
[3:09] So, uh the it never reaches a real workflow where
multi-department act- activities are integrated
[3:16] today. Four live deployments we did, um each
one crossed the gap, we call it the chasm,
[3:24] uh for pilot to production, and each one taught
us some lesson uh the hard way where we failed
[3:31] uh multiple times and then we were able to do
some interesting things to take it to production.
[3:40] The first um the agent, the first one we'll talk
about is a security surveillance surveillance
[3:46] system that we have built uh where we are taking
feed from 500 CCTVs and some of them are drone
[3:54] level TVs also cameras that we implemented across
an operator's location where they're manufacturing
[4:01] automotive cars. And when these cars come
out of the assembly line, they come in the
[4:07] parking lot which are acres of land before they
go to the dealerships. And these are deployed in
[4:12] countries like Vietnam, India, and Middle East.
And these cameras feed into uh live feed into uh
[4:22] tons of uh panels in the security office where
they want to track if there is any theft,
[4:27] the propensity of pillage, moving things out of
the cars or cars are leaving the uh the field.
[4:34] And they want to see that what activity's
going on, but they don't have enough manpower
[4:39] to look at every camera and every display.
So, we implemented the whole agentic platform
[4:45] real time from the CCTV. And uh trying to bring
it down from 25 people to three people addressing
[4:54] close to 25 panels uh monitors and trigger
events if there is some activity going on.
[5:02] Uh there are some agent graph rag and a VLM
the visual language model we implemented.
[5:09] And then we took it to the GPU level uh
for the visual detection. And then we
[5:14] were triggering events based on activities that
were trained um earlier through different feeds.
[5:25] The next one we implemented an agent we ended
up building four times. And this was about
[5:35] a kind of a the demo that we built on different
kind of aspects of data layers that were not
[5:43] integrated together to feed into the fabric of
the knowledge graph. And this one all the time
[5:51] it failed because the different stakeholder were
not giving us the right data set because the main
[5:56] lead stakeholder, the manager, was not able to
integrate understand that which data sources
[6:01] will be most optimized to run those that single
agent to satisfy the needs of the front-end users.
[6:10] And that's where we came to know that the data is
key. And if the data is not properly designed and
[6:18] restructured and cleansed before you feed into the
LLM or you build multiple agents to run on top of
[6:25] it uh the outcome will obviously be completely
different from what you uh decided to build.
[6:32] Now, we did another one from a media standpoint
where uh we were working with a company probably
[6:39] you have heard about Twitch. And we're taking
live streams to detect uh where the sponsors would
[6:45] come on on the live streams and they would put
their ads right next to the activities going on.
[6:51] But if there is some egregious statements
or comments that going on that the
[6:56] the ad placement per the entity will just pull
out of that stream because they don't want to
[7:01] they don't want to associate themselves with any
kind of activities from a comments or flag it
[7:06] flagging that's coming out. They don't want to
participate their placements right next to it.
[7:11] And it's real time. You may have only 3 to 4
seconds in a Twitch stream that's going on and you
[7:16] have to build a whole assessment based on that and
pull the ad out when that event happens and then
[7:23] bring it back on when the event has kind of gone
through the frame on top and you have getting more
[7:28] content out. And that needed a real-time audio
video analysis. And it was complete black box
[7:35] because there's a stream coming in from one end,
there's a processor in the middle, and there's a
[7:40] front end where we are running the agents on top
of it. And what we learned the hard way is that
[7:47] often the third party or the second party
in that case cuz we are the third party
[7:52] if they don't disclose the data to you in a single
thread and you are kind of making a hunch that
[7:57] okay, this data can be like that coming through
the stream then your agents will fail because
[8:01] majority of the time you might end up giving
different flags and wrong flags where you are
[8:06] not liable for that problem, but you still in the
downstream of that data. So, we learned that that
[8:13] every party has to stay in an integrated way and
bring that integration of activities where second
[8:18] party, third party data, they all should flow
in and come in one fabric of the knowledge graph
[8:23] that needs to be orchestrated beyond a
pilot. Next one we did some work on turning
[8:30] unstructured data into media. And that was another
project we did in pilot first where we were
[8:35] taking on tons of unstructured information PDFs,
manuals, long-form videos, podcasts, audio files
[8:43] and we will feed that into the LLM and based on
the ontology of the information will convert them
[8:50] into structured content and media assets that will
be more like uh educational curriculum training.
[8:57] And the way this works is that it's a multi-agent
thread agent credit system where one agent
[9:03] extracts the data and creates a relationship.
The second agent writes the curriculum from that
[9:08] material SOPs and all that stuff and creates a
training curriculum. The third agent goes through
[9:13] all that data sets that come out of that and kind
of make sure that they are cited properly and they
[9:19] are legit. And then the fourth agent goes out and
does all the media production from that content.
[9:26] So, there are four agents in tandem, sometimes in
serial, sometimes in parallel and working together
[9:33] but then what happens in creating this media
in almost like I would say quasi real time
[9:42] we found out that the visual language model, the
language model that creates the media assets of it
[9:48] sometimes may not be in
tune with the training data.
[9:53] And that's where the agent breaks down
because we have not trained that agent
[9:57] or the machine learning side of it enough
yet that it can create kind of a extension
[10:04] or interpolation of the data and build the whole
media out of that. So, the reason being is that
[10:10] again, all the stakeholders are not fully signed
up to give their learning data to us that we can
[10:16] train the system where we can build that final
media out of that unstructured data in one flow.
[10:27] The technical learnings we found from
surveillance, finance, live moderation, and
[10:32] training video creation is that for surveillance
we found out the retrieval is the bottleneck.
[10:39] And when you are using models and large
real-time data sets coming through
[10:44] and if you do not have enough sophisticated
well-designed pipeline to feed that real-time
[10:49] retrieval of the real-time data from a
graphics like a TV cameras and all that
[10:53] um your agents will not be able to operate in
a single sync. From the finance standpoint,
[11:00] we figured out that the schema reference of the
data was not right. Again, as I mentioned to you
[11:05] earlier, the stakeholders were not giving the
right sets of data to us so we could not find
[11:10] out that how these agents could be threaded
together on a data set that was not complete.
[11:16] Um the Twitch element I just mentioned to you,
there's a third-party, second-party data involved
[11:20] and these those stakeholders were not
participating together to put it all in one
[11:25] uh place where our agents could not perform
uh as per the mandate. And from the training
[11:31] video standpoint, the was not enough. Uh the
learning was not optimized uh for the visual
[11:36] language model to process the media creation, and
that's why we are falling off uh the benchmarks.
[11:47] So, the successful or the one that succeed
today, what we found out that 11% of what
[11:53] I mentioned earlier that who take this
pilot prototypes to production level,
[11:57] they do four things right. Like they really do
four things that help them to go to production
[12:03] level. First of all, they own the data. All
stakeholders have to own the data collectively
[12:09] so that they all are willing to provide that
data set within the governance and within that
[12:16] um sovereign premise that they will be
able to answer to the data, fetch the data,
[12:23] and communicate with the data through the agents
in the fabric of knowledge graph. Number two,
[12:30] there will be failure. They should accept
failure, and it should be an iterative model
[12:36] uh of implementation. Um most often times, we
get use cases, we are brought in to build it,
[12:43] and if it fails, they just can it. They say, "We
don't care. Next one, we'll figure out later."
[12:48] Uh they're not tuned into failure. They're not
saying that we can fail, we can learn from it,
[12:52] we can keep on building more use cases,
and eventually we'll go to production. So,
[12:55] there's a lack of um I would say um a trust in the
process. And uh and it needs to be an iterative
[13:04] process. The third one is it's uh the building the
oversight of the data, the governance of the data,
[13:11] the citation quality of the data,
where the data is being gov- governed,
[13:16] and who is in charge of the data. And what's
the ethics side of it also comes in the picture
[13:22] something Sometimes when you're doing legal
work, legal data, and all all stuff. And then
[13:27] um who owns the pipeline? And if there's a
one stakeholder, one universal stakeholder
[13:34] who owns the pipeline of all the agents as opposed
to four stakeholders for four agents, then there's
[13:39] a possibility of success way more than having four
stakeholders having their own agents separately.
[13:49] So now as we are saying that I think
2026 we're seeing more and more
[13:53] pilots are moving into production. And as you
can see that there's a really good scale of this
[13:59] pilots moving to production that we're seeing $8.5
billion worth of pilots can move into production.
[14:04] It could be almost like six times, five to six
times of that number that can go to production.
[14:10] And right now they're less than 5% of the
enterprise apps that have gone to production
[14:16] with agents. And that can move into probably by
40% if we shift the gear and move it to follow
[14:23] those rules and structure of implementation.
And it can really reach some real good scale.
[14:33] As I mentioned that earlier in the beginning
of this presentation, anybody can build a demo,
[14:36] anybody can do one use case. The hardest
part is to make it work in real-time scale.
[14:41] Uh where are different stakeholders,
different business needs,
[14:44] and different speeds needed. And
um and it's optimized for failure.
[14:53] These are the sources. Uh I can share that
with you. Can come back to me and we can
[14:57] where we got the data, but um again,
thank you very much. Um if I if you
[15:02] have any questions, I'll be here for a few
more minutes. We can talk about it. Uh but
[15:07] yes, once again, thank you so much for your time.
And I appreciate sharing the details with you.
