ZL Tech Pushes In-Place Data Virtualization to Accelerate Bank AI

Generative AI… So, Where’s the Data? With ZL Technologies — Ai4

Ai4

Virtualizing unstructured communications in place cuts storage costs by 90% and accelerates AI feeds 1,000-fold while eliminating the legal liabilities of unmanaged data copies.

AI-assisted summary of Generative AI… So, Where’s the Data? With ZL Technologies — original source. Check the original for full context.

Source published: 2025-08-26

Detailed notes

## Background Published by the Ai4 channel on YouTube on August 26, 2025, this presentation features Kon Leong, CEO of ZL Technologies. According to the description, the session examines how large organizations can manage unstructured enterprise data for strict governance requirements while curating a targeted, compliant data feed for corporate AI and analytics.

## TL;DR Kon Leong explains that the primary obstacle to enterprise Generative AI adoption is not algorithms, but managing the massive, legally 'radioactive' corpus of unstructured human communications. Rather than duplicating sensitive files into isolated, ungoverned AI sandboxes, organizations must adopt unified in-place extraction and data virtualization across their entire communication footprint. This architectural shift reduces storage costs by 90%, speeds up analytics pipelines by 1,000x, and enforces reliable data-level privacy guardrails. Ultimately, achieving comprehensive 100% capture unlocks exponential value through modular models and deterministic search, creating a defensible and enduring organizational memory.

## Atmosphere & style The presentation carries the pragmatic, authoritative weight of a veteran technologist dissecting the gritty legal and architectural realities of enterprise data systems. Leong balances high-stakes warnings regarding regulatory subpoenas and massive infrastructure bottlenecks with sharp, memorable metaphors like 'radioactive data,' sandboxes versus beaches, and Sherlock Holmes's silent dog. The delivery is direct, analytical, and grounded, cutting through speculative GenAI hype to address the foundational compliance and engineering challenges faced by modern corporate leaders.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) Unstructured human-generated communications (emails, chats, SMS) are the fundamental fuel for Generative AI and true organizational memory, far outweighing structured tables. - [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) Regulatory mandates in finance requiring comprehensive communication tracking provide an unprecedented opportunity to construct institutional knowledge from existing archives. - [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) Unstructured data is intrinsically 'radioactive' because unmanaged duplicates of unvarnished human conversations can resurface years later as catastrophic legal and regulatory liabilities. - [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) JPMorgan Chase's $2 billion AI deployment highlights that the primary bottleneck in scaling enterprise AI is formatting and managing unstructured data across use cases, not the algorithms. - [07:20](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=440s) Unlike public web search engines that operate under a 'best-efforts' retrieval standard, enterprise compliance and legal systems face devastating penalties if a single relevant document is missed. - [07:20](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=440s) In-place metadata and semantic extraction reduces storage costs by 90%, minimizes subpoena risks, and accelerates analytical pipeline speeds by 1,000x by avoiding redundant file copying. - [10:13](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=613s) Consolidating fragmented governance silos (compliance, eDiscovery, records, and privacy) onto a single unified data platform eliminates isolated point solutions and creates a governed AI data feed. - [10:13](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=613s) Applying data-level guardrails at the repository stage is vastly more dependable than relying on prompt-layer algorithmic filters to prevent private PII leaks in downstream AI applications. - [12:34](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=754s) Enterprises must transition from sampled sandbox experiments to full data virtualization to eliminate sampling bias and enable comprehensive risk and evidence discovery across the organization. - [15:43](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=943s) Analytical insight experiences an exponential 'hockey stick' surge only at 100% data capture, which uniquely enables detection based on the total absence of records ('the dog that didn't bark'). - [15:43](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=943s) Building AI architectures with modular 'puzzle pieces' allows organizations to untrain and expunge specific private records without discarding or retraining monolithic models. - [15:43](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=943s) Extracting semantic essence and metadata prior to defensible deletion preserves institutional memory for AI feeds while safely eliminating the legal discovery risks of raw source files. - [17:44](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1064s) Enterprises must retain deterministic search alongside generative AI because adversarial legal and regulatory proceedings demand transparent, explainable, and reproducible evidence.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) The Reality of Unstructured Data: Fuel for GenAI and Organizational Memory

The primary driving force behind Generative AI and the eventual path toward Artificial General Intelligence (AGI) is not clean, structured database tables, but messy unstructured data. For the past 26 years, ZL Technologies has operated inside the massive data environments of the world's largest financial institutions, managing email archives, file shares, Bloomberg chats, SMS text messages, and collaboration suites. In heavily regulated sectors, the legal, eDiscovery, and regulatory compliance pressures surrounding this data are relentless.

What is often overlooked about unstructured data—whether it lives in Microsoft 365, Teams, or instant messaging platforms—is that it is "100% generated by humans for humans." Because major financial institutions already operate under strict legal mandates to collect, track, and retain every scrap of these communications, this regulatory burden offers an unprecedented opportunity. By harnessing this entire human record, enterprises can assemble genuine organizational memory and institutional knowledge for the very first time.

Whenever a scandal, crisis, or major lawsuit erupts, investigators instinctively gravitate toward human communications. A prominent example occurred immediately following the events of January 6, where the primary investigative focus on January 7 centered on recovering Secret Service SMS text records. Because enterprise systems already capture these granular human interactions, the resulting data sets frequently "know more about you than you think you know of yourself."

---

[03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) The Radioactive Nature of Human Data and JPMC's $2 Billion Lesson

In standard data architectures, structured to unstructured data exists at a ratio of roughly one to five. Yet corporate data strategies often overlook a fundamental truth: unstructured data is intrinsically "radioactive." Because it reflects unvarnished human thoughts and informal conversations, unmanaged copies can easily resurface years later with catastrophic consequences. A poignant illustration was seen on Capitol Hill, where a prominent executive was grilled on the witness stand over decade-old corporate communications that put his company at risk of being dismantled.

```mermaid flowchart TD accTitle: Unified Governance and In-Place Data Pipeline subgraph SOURCEDATA["Enterprise Communications Corpus"] A["M365 / Exchange"] B["Teams / Slack / Chats"] C["Bloomberg / SMS"] D["File Shares / SharePoint"] end

subgraph UNIFIED["ZL Unified In-Place Platform"] E["Deterministic Index & Metadata Extraction"] F["Virtual Repository / Zero Duplication"] G["Unified Schema: Records, Privacy, eDiscovery & Compliance"] end

subgraph OUTPUTS["Downstream Consumer Workflows"] H["GenAI / RAG / SLM Feeds (1000x Speedup)"] I["Defensible Deletion / Retention"] J["Regulatory Production & Auditing"] end

SOURCEDATA --> E E --> F F --> G G --> H G --> I G --> J ```

Because of these severe legal liabilities, enterprise AI initiatives must coordinate directly with the Chief Legal Officer, Chief Compliance Officer, Chief Records Officer, and Chief Privacy Officer. Rather than creating countless unsecured duplicate copies of sensitive data across sandbox data lakes, organizations must implement virtualized repositories where the original raw file remains protected or eliminated, allowing the enterprise to honestly respond to subpoenas with: *"Sorry, I don't have it. The original isn't there."*

This data dilemma was recently underscored by the CEO of JPMorgan Chase (JPMC). Reflecting on their massive AI deployment, he noted that the true bottleneck is neither machine learning algorithms nor generative models:

> *"The hardest part is the data, which, you know, it isn't the AI... it's getting your people to use it properly, but it's the data. I mean, we have—I forgot the number—but it dwarfs everything. So getting the data in the form that it could be usable is the hard part, because it's in different formats and completely unstructured, like just conversations that people had on call centers. We now spend about $2 billion on AI. We have something like 600 actual in-use cases... That number will probably double next year and maybe triple."*

- **Platform vs. Silo**: JPMC achieves 600+ AI use cases by deploying unified platforms that replicate across operations, rather than building isolated, bespoke silos. - **Enterprise Adoption Reality**: An audience poll reveals that while mega-banks are managing hundreds of AI use cases, the vast majority of enterprise practitioners are still in the early stages with fewer than 100 to 200 use cases.

---

[07:20](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=440s) The Scale Paradox: In-Place Extraction and 1000x Analytics Acceleration

Handling enterprise unstructured data requires overcoming three massive technical hurdles: gaining access to the entire data corpus rather than partial slices, ingesting all data types through a single pipeline, and scaling to dozens of petabytes. Tier 1 banks routinely manage more unstructured documents than Google indexes. However, while Google operates on a commercial "best-efforts" retrieval standard, an enterprise bank that loses a single responsive legal document can face devastating penalties.

To solve the dual challenges of storage cost and regulatory liability, modern governance relies on "in-place" processing. Instead of copying low-value files into redundant storage, the system inspects the document in place, extracts its core semantic essence and metadata for AI and analytics, and avoids creating superfluous copies. High-value documents can be archived selectively with full copies, while low-value items remain virtualized.

Implementing in-place extraction across Tier 1 banking environments yields three immediate structural benefits: - **90% Reduction in Storage Costs**: Drastically lowers infrastructure overhead, enabling organizations to retain rich metadata histories over significantly longer periods. - **Subpoena and Risk Minimization**: Eliminates toxic duplicate document stores, insulating the company from discovery overreach. - **1,000x Speedup to Data Analytics**: Eliminates the bottleneck of reaching back into underlying Microsoft 365 or SharePoint repositories to fetch source files, accelerating analytical pipelines by three orders of magnitude.

---

[10:13](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=613s) Unifying Information Governance Silos into an Enterprise Feed

Historically, enterprise governance has been fragmented into separate operational silos: compliance monitoring, legal eDiscovery, records lifecycle management, and privacy compliance (such as GDPR and CCPA). Each silo independently purchased point solutions, maintained isolated indexes, created duplicate data copies, and operated at cross-purposes with other departments.

Unifying these disparate workflows onto a single data platform with a single underlying data schema consolidates licensing and infrastructure costs while unlocking a clean, fully governed data feed for enterprise AI.

Enforcing data-level guardrails is far more effective and dependable than relying on prompt-layer algorithmic filters. By stripping private or non-compliant information directly at the data repository level, an enterprise guarantees that downstream marketing algorithms or automated agents cannot violate privacy laws.

---

[12:34](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=754s) Leaving the Sandbox for the Beach: Virtualization Strategies

Most enterprise analytics, eDiscovery, and Generative AI experiments currently take place inside restricted "sandboxes." Sandbox experimentation is fundamentally flawed: it is slow, relies on sampled data subsets that introduce severe sampling errors, consumes excessive manual labor, and operates entirely outside established governance protocols.

Instead of playing in an ungoverned sandbox, enterprises should "jump out of the sandbox and do the beach." By virtualizing the entire data corpus (the entire beach), governance, data scrubbing, and sampling discrepancies are resolved comprehensively. Instead of finding a few isolated "red shells" or smoking guns in a bucket of sand, analysts can query across the complete environment to uncover massive, multi-hundred-foot piles of relevant evidence.

```mermaid flowchart LR accTitle: Targeted Virtualization Implementation subgraph PHASE1["Phase 1: Targeted Virtualization"] RND["Core R&D (2,000 Staff)"] PROD["Product Management"] GLOBAL["Overseas Engineering"] end

subgraph REPO["Virtual Repository"] EXTRACT["In-Place Metadata Extraction"] end

subgraph SCALE["Phase 2: Enterprise Rollout"] ORG["Global Workforce (100,000+ Staff)"] DEFENSIBLE["Defensible Disposition & AI Feeds"] end

PHASE1 --> EXTRACT EXTRACT --> SCALE ```

The recommended path forward is to start small by virtualizing high-risk or high-value pockets before expanding globally: - **Targeted Deployment Example**: An enterprise facing intellectual property and product liability litigation virtualized only the domestic R&D, product management, and overseas engineering teams—comprising 2,000 employees out of a 100,000-person global workforce. Once validated inside the analytical pipeline, the architecture was seamlessly expanded across the entire company. - **Solving Global Dispersion**: In one Tier 1 global bank, data was scattered across independent country-level silos with zero unified visibility into risk or lifecycle status. Consolidating the index under unified in-place governance enabled the bank to enforce defensible disposition, feed AI sandboxes 1,000 times faster, and close capability gaps inherent in base Microsoft platforms.

---

[15:43](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=943s) The 100% Capture Hockey Stick: Detecting Missing Footprints and Untraining AI

When capturing enterprise data, insight generation does not follow a linear progression. Value grows at a snail's pace through 50%, 70%, and 90% capture rates. However, once coverage reaches 99% to 100%, analytical value experiences a dramatic "hockey stick" spike.

This exponential jump occurs because a 100% comprehensive data capture allows organizations to derive insights from *data that does not exist*. In a partial data capture, observers only see scattered, incomplete footprints. In a guaranteed 100% capture environment, an investigator can definitively prove the total absence of a required record—the classic Sherlock Holmes principle of "the dog that didn't bark." This definitive absence is the bedrock of anti-money laundering (AML), fraud detection, and insider threat identification.

- **Untraining AI via Modular Puzzle Pieces**: Rather than training massive monolithic models that cannot be modified after ingestion, AI architectures should be built from modular "puzzle pieces." When data must be expunged due to privacy or compliance mandates, specific puzzle pieces can be extracted and retrained independently without tearing down the entire model. - **Essence Extraction Prior to Deletion**: Before executing defensible disposition on legacy emails and file shares, organizations should extract their semantic essence and structural metadata. Once the raw file is safely destroyed to eliminate legal discovery risk, the extracted insights remain accessible for future AI applications.

---

[17:44](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1064s) Governing the Future: Deterministic Search and True Organizational Memory

Managing unstructured data requires fundamentally different rules than structured databases. In structured environments, database administrators freely replicate tables and meticulously track lineage; in unstructured systems, organizations must aggressively minimize duplicate files and avoid unnecessary replication.

Furthermore, enterprises must avoid completely replacing legacy deterministic tools with opaque AI models. In legal proceedings and regulatory compliance reviews, there is an inherent lack of trust between opposing litigants or between regulators and regulated institutions. Black-box AI outputs are legally unacceptable in discovery disputes; organizations must maintain deterministic search capabilities to deliver exact, explainable, and reproducible evidentiary results.

When architecting retrieval mechanisms, practitioners must carefully evaluate the cost and latency profiles of GenAI Retrieval-Augmented Generation (RAG) and Agentic RAG workflows. By pairing deterministic search, small language models, and comprehensive in-place virtualization, organizations are doing far more than adopting generative tooling—they are building an enduring, queryable institutional memory that places all enterprise knowledge at their fingertips in real time.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) 100% generated by humans for humans. — Kon Leong

> [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) know more about you than you think you know of yourself. — Kon Leong

> [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) Sorry, I don't have it. The original isn't there. — Kon Leong

> [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) The hardest part is the data, which, you know, it isn't the AI... it's getting your people to use it properly, but it's the data. I mean, we have—I forgot the number—but it dwarfs everything. So getting the data in the form that it could be usable is the hard part, because it's in different formats and completely unstructured, like just conversations that people had on call centers. We now spend about $2 billion on AI. We have something like 600 actual in-use cases... That number will probably double next year and maybe triple. — CEO of JPMorgan Chase

> [12:34](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=754s) jump out of the sandbox and do the beach. — Kon Leong

> [15:43](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=943s) the dog that didn't bark. — Kon Leong

## Mentioned Kon Leong, ZL Technologies, Ai4, ChatGPT, Microsoft 365, Microsoft Teams, Exchange, SharePoint, Bloomberg, Slack, JPMorgan Chase (JPMC), Google, Sherlock Holmes, GDPR, CCPA

## Video description Kon Leong, CEO, ZL Technologies

Abstracts

The momentum from ChatGPT has forced every organization to answer a simple question: “What’s our AI story?” Whereas most generic AI tools today are trained on public data, the real value for corporate use is training AI on enterprise data. However, the sheer scale of corporate data and the need for data governance test the limitations of today’s technology. This session will explore how large organizations are implementing a system to manage enterprise data for governance requirements, while curating and delivering a targeted and relevant data feed for AI and analytics. By tapping into the entire corpus of employee-created data, organizations are able to unlock deep insights into the “human side” of the company.

[Source](https://www.youtube.com/watch?v=miMJ4EVbcWI)

中文简报

ZL Tech推进原位数据虚拟化以加速银行AI 对非结构化通信数据进行原位虚拟化可降低90%的存储成本，并将AI数据馈送加速1000倍，同时消除未受管数据副本带来的法律责任。 Kon Leong解释称，企业采用生成式AI的主要障碍不在于算法，而在于如何管理庞大且在法律层面上具有“放射性”风险的非结构化人类通信语料库。企业不应将敏感文件复制到孤立且缺乏管控的AI沙盒中，而必须在其整个通信足迹中采用统一的原位提取与数据虚拟化。这种架构转变能够降低90%的存储成本，将分析流水线提速1000倍，并落实可靠的数据级隐私防护栏。最终，实现100%的全面捕获将通过模块化模型和确定性搜索释放出指数级价值，构建起具有合规抗辩力且持久的组织记忆。人类生成的非结构化通信数据（如电子邮件、即时聊天、短信）是生成式 AI 和企业组织记忆的核心燃料，其价值与规模远超传统结构化表格。 (0:04) 非结构化数据具有天然的“放射性”，随意散落的副本可能转化为毁灭性的法律风险；摩根大通（JPMC）耗资 20 亿美元的 AI 实践证明，规模化落地的最大瓶颈在于治理和格式化非结构化数据，而非算法模型本身。 (3:30) 通过就地（in-place）元数据与语义提取技术，企业可在不重复复制低价值文件的情况下降低 90% 的存储成本，大幅减少传票诉讼风险，并将分析管道速度提升 1000 倍。 (7:20) 将合规、电子取证（eDiscovery）、记录归档与隐私管理整合到统一的数据治理平台，并在数据层而非 Prompt 提示词层设置防护栏，能够更稳健地杜绝 PII 隐私泄漏。 (10:13) 当数据捕获率达到 100% 时，分析价值将呈现指数级爆发，企业不仅能实现“缺席记录”级别的欺诈审计，还能通过模块化 AI 架构安全地注销敏感数据与执行合规清理。 (15:43)

Reading summary

According to ZL Technologies, unstructured human communications form the essential fuel for enterprise generative AI and institutional memory, yet unmanaged duplication makes this data legally hazardous. Successfully scaling enterprise AI requires shifting from siloed sandbox copies to unified, in-place virtualization that extracts semantic metadata directly from source repositories. The Liability and Value of Human Communications (3:30) Unstructured communications—including emails, chat messages, Bloomberg transcripts, and SMS—outnumber structured database records by roughly five to one and reflect unvarnished human thought. While regulatory retention mandates in sectors like finance allow firms to assemble a comprehensive institutional memory, ZL Technologies warns that this data is inherently "radioactive." Unmanaged copies and informal exchanges frequently resurface years later during lawsuits and regulatory investigations with severe legal consequences. This operational bottleneck was highlighted by the CEO of JPMorgan Chase, who reported that while the firm spends approximately $2 billion on AI across roughly 600 use cases, the primary hurdle is not AI algorithms, but formatting vast volumes of unstructured data like call-center conversations. Managing these liabilities requires integrating AI initiatives directly with corporate legal, compliance, records, and privacy officers rather than creating unsecured sandbox copies. In-Place Extraction Eliminates Redundant Storage (7:20) Tier 1 banks routinely manage document repositories larger than Google's public search index, but unlike search engines operating on a commercial "best-efforts" standard, regulated institutions face severe penalties if a single subpoenaed file is lost. Moving petabytes of sensitive communications into duplicate data lakes compounds both storage costs and legal discovery liabilities. To resolve this dilemma, modern information architecture relies on in-place processing. Rather than duplicating low-value files, the system inspects records where they reside, extracting structural metadata and semantic essence into a virtualized index while leaving raw files protected or slated for defensible disposition. Across major banking environments, this approach delivers a 90% reduction in storage costs, insulates the enterprise from overreaching discovery subpoenas, and accelerates analytical and AI pipeline retrieval speeds by roughly 1,000 times compared to querying raw repositories like Microsoft 365 or SharePoint. Unifying Governance Silos with Virtual Repositories (10:13) Information governance has historically been fractured across discrete operational silos: legal eDiscovery, regulatory compliance, records management, and privacy compliance. Each department traditionally purchased separate point solutions and created redundant data copies. Unifying these functions into a single platform with a shared data schema lowers licensing overhead and establishes data-level guardrails—such as stripping personally identifiable information (PII) before ingestion—which ZL Technologies considers significantly more reliable than prompt-layer filters. Organizations can bypass the sampling errors and manual overhead of isolated sandboxes by progressively virtualizing their entire communications corpus. For instance, an enterprise facing product liability risks initially virtualized only 2,000 R&D and engineering employees out of a global workforce of 100,000. Once validated, this architecture scaled globally across regional silos to reliably feed small language models (SLMs) and retrieval-augmented generation (RAG) workflows. Comprehensive Capture, Absence Detection, and Modular AI (15:43) The analytical utility of enterprise data follows a non-linear curve, experiencing a sharp "hockey stick" surge only when coverage reaches 99% to 100%. Comprehensive capture allows investigators to identify "the dog that didn't bark"—definitively proving the absence of a record or communication, which forms the evidential foundation for anti-money laundering, fraud detection, and insider threat monitoring. To accommodate compliance mandates that require data expungement, enterprise AI architectures must be built from modular "puzzle pieces" rather than opaque, monolithic models. Modular components allow specific training data to be removed and retrained independently. Additionally, extracting the semantic essence of legacy files before defensible disposition preserves institutional knowledge for future AI queries while permanently eliminating the discovery risks of retaining raw documents. ZL Technologies stresses that enterprises must preserve deterministic, explainable search capabilities alongside generative models, as regulators and opposing legal counsel demand reproducible evidence that black-box AI outputs cannot satisfy. A central challenge remains closing the implementation divide, as most enterprises still operate with fewer than 200 AI use cases while navigating complex governance hurdles.

精炼阅读摘要

ZL Technologies 指出，企业生成式人工智能与真正组织记忆的核心驱动力并非结构化数据库，而是高风险、体量庞大且全由人类生成的非结构化通信数据；通过原位虚拟化与统一治理架构打通合规孤岛，企业才能在控制“放射性”法律与隐私风险的前提下释放全量数据价值。非结构化数据的价值与放射性法律风险 (0:04) 在典型企业数据环境中，非结构化数据与结构化数据的体量比例约为五比一。ZL Technologies 强调，电子邮件、Teams、彭博社即时通讯及短信等非结构化记录 100% 由人类为人类而创造，忠实记录了员工的直白思想与非正式交流，构成了机构记忆的基石。然而，这类数据天然具备“放射性”法律风险，若未经严格管理，数年甚至数十年前的无序沟通记录在诉讼调查或国会听证中极易被翻出，对企业构成灾难性打击。例如，在调查危机事件时监管部门往往重点调阅此类人类交互记录。针对当前 AI 落地瓶颈，摩根大通（JPMorgan Chase）首席执行官曾公开表示，企业 AI 面临的最大阻碍并非算法或模型，而是体量远超一切、格式分散且完全非结构化的底层人类对话数据。尽管摩根大通每年投入约 20 亿美元支持约 600 个实际 AI 用例并计划翻倍，但现场行业调研显示，绝大多数企业从业者目前落地的用例仍停留在 100 到 200 个以下，深陷数据整合困境。原位提取与全域虚拟化打破沙盒局限 (7:20) 目前大多数企业的 AI 实验与分析仍局限在隔离的“沙盒”环境中，这种做法不仅依赖抽样数据从而引入严重误差，还耗费大量人工并脱离了现存治理框架。ZL Technologies 主张企业应“跳出沙盒，拥抱整片海滩”，采用全域数据原位（in-place）虚拟化策略。原位处理直接在底层存储系统上提取语义精髓与元数据，无需向沙盒数据湖中复制多余副本，高价值文档按需归档，低价值文档则维持虚拟化。在大型跨国银行的实际落地中，原位提取带来了三重收益：首先，由于消除了无意义的重复文件存储，基础设施成本降低了 90%；其次，大幅收缩攻击面并避免被传票过度调取数据；最后，下游分析流水线无需频繁返回底层 Microsoft 365 或 SharePoint 调取源文件，使数据分析速度提升了 1000 倍。实施路径上，企业可先针对知识产权或产品责任风险最高的部门（例如 10 万人组织中的 2000 名研发与产品人员）开展局部验证，再顺畅推向全公司。打破合规孤岛与构建数据级防护屏障 (10:13) 过去，企业内部的合规监控、电子数据发现（eDiscovery）、记录生命周期管理以及 GDPR/CCPA 等隐私合规往往各自为政。各业务孤岛独立采购点状工具、维护割裂的索引，不仅推高了软件许可与存储开销，更导致敏感数据副本四处流窜。通过在单一架构下统一数据模型，企业能够打破跨部门摩擦，为下游 AI 平台提供标准且受控的数据流。更为关键的是，统一平台允许在底层数据存储库直接施加数据级防护栏（Guardrails），而非单纯依赖提示词层的算法过滤。ZL Technologies 认为，提示词层过滤无法提供绝对的法律保障；只有在数据接入阶段直接剥离个人身份信息（PII）或违规内容，才能彻底确保下游的小型语言模型（SLM）、微模型及自主智能体不会产生合规侵权风险。全量捕获拐点与模块化可解耦AI (15:43) 企业数据捕获的洞察价值释放遵循非线性的“曲棍球棒”效应：当捕获覆盖率从 50% 爬升至 90% 时，价值增长较为平缓；而一旦跨越 99% 达到 100% 全量覆盖，分析价值将出现指数级跃升。其根本原因在于，只有在 100% 覆盖的前提下，分析系统才能可靠地根据“未发生的数据”推导结论——即福尔摩斯名案中“不叫的狗”原则。明确证明某项记录的完全缺失，是反洗钱（AML）、反欺诈与内部威胁调查的底层基石。针对监管部门的数据删除要求与 AI 合规审查，传统一旦训练完成便无法修改的庞大单体模型存在严重缺陷。企业应当将 AI 系统解构为模块化的“拼图”结构，当特定数据因合规指令被依法处置时，只需单独提取并重新训练对应模块，而无需推倒整个模型架构。同时，在对旧邮件或文件执行防卸性销毁（Defensible Disposition）以规避调证风险前，系统应提前提取语义精髓与结构化元数据，在销毁原件的同时留存可用智能资产。确定性检索保障法律可信与组织记忆 (17:44) 管理非结构化数据与管理传统结构化数据库有着本质不同：关系数据库鼓励复制表结构并精细追踪血缘，而非结构化数据管理则必须严格限制不必要的复制传播。此外，在司法诉讼或强监管审查中，对抗各方之间天然缺乏信任，不透明的黑盒 AI 结论无法在法庭质证中被采信。因此，企业不能完全依赖概率性的大语言模型，而必须保留具备可解释、可复现与精准特性的确定性（Deterministic）搜索工具。将确定性检索与小型语言模型、检索增强生成（RAG）以及全域原位虚拟化相结合，企业不仅能优化检索成本与系统延迟，更能将 20 余年来散落在邮件、聊天群和协同套件中的全量人类通信沉淀为真正可实时调用的企业制度性记忆。虽然依托原位虚拟化与 100% 全量捕获构建统一组织记忆在理念上消除了 AI 数据瓶颈与合规冲突，但绝大多数非头部金融机构在跨越“沙盒”、重组跨部门治理架构时面临着高昂的协同成本；此外，如何在保障严密法律确定性的同时兼顾 Agentic RAG 的灵活性，仍是企业全面落地此类架构需要持续验证的问题。

Detailed notes — zh

## Background 该内容由 YouTube 频道 Ai4 于 2025 年 8 月 26 日发布，由 ZL Technologies 首席执行官 Kon Leong 主讲。视频简介指出，企业 AI 的真正价值在于利用企业数据进行训练，本期演讲深入探讨了大型企业如何在满足严苛数据治理与合规要求的同时，管理全量非结构化员工数据并为 GenAI 和分析系统构建高效精准的数据供给通道。 ## TL;DR 演讲指出由人类产生的非结构化数据是 GenAI 和未来 AGI 的基石，但其天然具备法律诉讼层面的“放射性”风险。通过“就地虚拟化管理”（In-Place Management）和统一数据治理模式，企业无需保留多余的高危原始副本即可提炼知识精髓，不仅能将存储成本降低 90%、数据分析流转速度提升千倍，还能打破部门数据孤岛。Kon Leong 强调企业应跳出沙盒抽样陷阱，实现 100% 全量数据捕获以发掘“缺失数据”中的深层线索，并通过模块化模型解耦和保留布尔确定性搜索，构建经得起合规检验的下一代企业组织记忆。 ## Atmosphere & style 整场演讲呈现出一种严谨务实、老练且极具行业警示感的氛围。主讲人以二十余年深耕金融合规一线的实战积淀，用国会山听证会汗流浃背的高管、摩根大通的 600 个 AI 用例以及福尔摩斯“夜间未吠之犬”等生动隐喻，将枯燥的数据架构与惊心动魄的法律诉讼风险紧密交织。听众既能深切感受到海量数据治理的如履薄冰，又能领略到攻克数据底座后如《少数派报告》般人机协同的高远视野。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) 非结构化数据（邮件、聊天、短信等）是 100% 由人类产生并为人服务的数据，不仅是驱动 GenAI 与未来 AGI 的核心基石，也是企业重构“组织记忆”的关键语料库。 - [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) 非结构化数据具有高度的“放射性”，未经妥善治理的陈年敏感沟通极易在监管听证和法律诉讼中反噬企业，因此 AI 数据战略必须由法律、合规、档案与隐私主管严格把关。 - [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) 摩根大通拥有约 600 个落地 AI 用例，其核心突破不在于算法本身，而在于通过统一平台化方法克服了非结构化数据规模庞大、格式各异且高度分散的治理难题。 - [06:17](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=377s) 传统全量复制模式成本高昂且法律风险大，采用“就地虚拟化管理”（In-Place Management）仅提取文本精髓而不留冗余副本，能降低 90% 存储成本并将分析调取速度提升 1000 倍。 - [10:14](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=614s) 将合规、电子数据调查（E-Discovery）、生命周期管理与隐私保护整合到统一数据模式下，可以直接在数据层构建安全护栏，比在算法提示词端设限更有效率。 - [10:14](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=614s) 企业级 AI 分析应跳出存在严重抽样偏差与混乱的局部“沙盒”，转向对全量数据海滩进行虚拟化，以全局视野直接筛选核心线索与证据堆。 - [13:40](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=820s) 数据捕获具有“曲棍球棒效应”，只有实现 100% 全量捕获才能精准侦测反欺诈与反洗钱中“本该存在却缺失的数据”（如福尔摩斯未吠之犬），实现洞察价值的爆发式跃升。 - [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 构建 GenAI 架构应遵循“模块化拼图”解耦设计，以便在需要合规剔除或修正知识时能够对模型进行精准“反向训练（Untrain）”。 - [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 在执行文件生命周期销毁前应预先提炼并归档其知识精髓，既消除原始文件的法律放射性，又为企业未来 AGI 留存知识火种。 - [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 在面对监管合规与司法诉讼时，不可完全依赖黑盒语义模型，必须保留传统的确定性布尔搜索能力以满足法庭要求的可解释性与信任基础。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) 非结构化数据：GenAI 与组织记忆的核心 ZL Technologies 首席执行官 Kon Leong 开篇指出，非结构化数据不仅是当前生成式人工智能（GenAI）的核心驱动力，也将是未来通用人工智能（AGI）的基石。在过去 26 年中，ZL Technologies 始终专注于为全球顶尖的大型金融机构处理超大规模的数据集，业务主要集中在信息治理、电子邮件、文件共享、彭博社（Bloomberg）终端记录、短信（SMS）以及即时通信聊天记录的合规与电子数据挖掘（E-Discovery）。由于金融行业面临极其严苛的监管要求和诉讼压力，如何管理并利用这些非结构化数据成为了银行业生存与发展的关键。 ```mermaid flowchart TD accTitle: 非结构化数据治理与组织记忆流转 subgraph S1["多源非结构化数据"] A["企业通讯 (邮件 / 聊天 / 短信)"] B["协作套件与文档 (SharePoint / M365 / 文件共享)"] C["专业终端数据 (Bloomberg / 呼叫中心录音)"] end subgraph S2["统一就地治理平台 (In-Place)"] D["元数据与精髓提取 (虚拟化数据仓库)"] E["合规 / 诉讼 / 隐私治理 (确定性检索 + 数据护栏)"] end subgraph S3["企业级智能化输出"] F["GenAI / RAG / SLM 数据源 (提速 1000 倍)"] G["组织记忆与 AGI 知识库 (欺诈侦测 / 行为洞察)"] end A --> D B --> D C --> D D --> E D --> F E --> G F --> G ``` 这些非结构化数据有一个最本质的共性：它们是“百分之百由人类产生、并为人类服务的数据”。大型银行原本就负有收集、整理和分析此类通讯的法定义务；而如今，这批庞大的历史数据可以被重新赋能，首次转化为企业级“组织记忆与知识库”（Organizational Memory and Knowledge）。这种深层价值已在无数危机、丑闻和诉讼中显现。例如在 2021 年 1 月 6 日国会山事件发生后的第二天（1 月 7 日），调查人员首先调取审查的就是美国特勤局（Secret Service）的短信记录。面对如此详尽的人际互动动态，数据对员工行为和企业运行的了解程度，甚至常常“超过人们对自身的了解”。 - 覆盖主要非结构化数据源：企业协作套件（M365、SharePoint）、即时通讯工具、短信、彭博通信及呼叫中心记录。 - 监管合规数据具备双重价值：既能满足防范法律风险的法定底线，又能转化为训练企业专属 GenAI 和未来 AGI 的高质量语料库。 --- ### [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) 非结构化数据的“放射性”与摩根大通的 600 个 AI 用例在传统数据湖架构中，结构化数据与非结构化数据的比例大约维持在 1 比 5 左右。但许多企业在拥抱 AI 时往往忽略了一个致命属性：由于非结构化数据高度记录了人类的真实言行，它天然具有“放射性”（Radioactive）。如果控制不当，这些陈年数据很可能成为反噬企业的核弹。Kon Leong 提到几个月前在国会山听证会上的一幕，某高管在证人席上汗流浃背，正是因为十多年前内部写下的一段敏感沟通内容被翻出，导致整个公司如今面临被拆分的巨大法律风险。因此，任何负责任的 AI 数据战略，都必须事先获得首席法律官（CLO）、首席合规官（CCO）、首席档案官（CRO）以及首席隐私官（CPO）的严格把关与签署批准。为了避免法律传票带来的巨大敞口，企业在处理非核心非结构化数据时不应随意复制全文，而应建立“虚拟数据仓库”（Virtual Data Repository），在不留存原始多余副本的前提下提取数据价值，从而在遭遇传票调取时能够合规申明“未持有该冗余原始副本”。Kon Leong 引用了客户之一——摩根大通（JPMorgan Chase）首席执行官 Jamie Dimon 的最新观点来印证这一痛点： > “最困难的部分是数据……它不是 AI 本身，不是机器学习，也不是生成式 AI，甚至不是如何让员工正确使用它，真正的难点在于数据。我们拥有的数据量让一切相形见绌，但要让数据变成可用形式极为困难，因为它们格式各异、完全非结构化，比如日常对话或呼叫中心记录……我们目前在 AI 上的投入约为 20 亿美元，拥有大约 600 个实际投入使用的用例，有些至关重要，有些相对次要，这个数量明年可能会翻倍甚至翻三倍。” 当 Kon Leong 现场询问听众有多少企业拥有 100 个、200 个或 500 个以上 AI 用例时，全场几乎无人举手。摩根大通之所以能将落地用例推升至 600 多个，核心就在于其坚决摒弃了部门数据孤岛，转而采用了能够全行复用的“平台化方法”。 --- ### [06:17](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=377s) 规模化治理突破：就地虚拟化（In-Place）的巨大红利非结构化数据之所以极难驾驭，主要源于三大挑战：第一，必须访问完整的数据而非抽样切片；第二，必须支持所有数据类型，孤立处理不仅效率低下且漏洞百出；第三则是极度苛刻的扩展性（Scale）。大型跨国银行所管理的非结构化数据量高达数十 PB，其文档总量甚至“超过了 Google 索引的网页文档数”。更严峻的是，Google 的搜索与管理只遵循“尽力而为”（Best Efforts）原则，而在强监管的银行体系中，“遗失任何一份关键文档都可能让你彻底完蛋”。面对昂贵且极具法律风险的海量数据，行业最新探索出了一种“就地虚拟化管理”（In-Place Management）机制：对高价值文档保留完整归档副本，而对海量低价值文档则不留冗余副本，仅在原始位置就地提炼抽取其文本精髓（Essence）用于 GenAI 与 AGI 训练。这种部署架构为大型银行带来了三大立竿见影的颠覆性效益： | 核心维度 | 传统全量复制模式 | 就地虚拟化（In-Place）模式 | |:--- |:--- |:--- | | **存储基础设施成本** | 随着历史数据无限膨胀，多副本备份推高开销 | 存储开销**直接下降 90%**，支持更低成本长期留存 | | **法律诉讼风险** | 冗余数据副本极易成为法律传票的攻击靶子 | 消除多余原始副本，显著降低诉讼暴露面 | | **数据分析流转速度** | 每次调用均需反复访问 SharePoint 或 M365 底层 | 绕过原始应用调取，分析速度**惊人提速 1000 倍** | - 数据获取全面性：彻底告别只抽取部分切片的传统做法，实现对全类型非结构化数据的全局穿透。 - 风险与效率解耦：在不保留冗余风险文件实体的同时，将全量知识内容安全输入 AI 分析管线。 --- ### [10:14](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=614s) 统一治理架构与摆脱“沙盒采样”陷阱过去二十年间，企业在应对合规（Compliance）、电子数据调查（E-Discovery）、档案与生命周期管理（Records & Lifecycle）、以及 GDPR/CCPA 隐私保护时，往往各自采购独立系统，形成了严重割裂、成本重复且目标冲突的“烟囱式数据孤岛”。最佳实践是将这些治理模块统一整合到一个统一的数据平台和数据模式（Data Schema）之上，用一套底层处理动作同时满足多方需求，不仅大幅削减了软件授权与运维成本，更构筑起了受控的 GenAI 供给源。这一统一平台使得企业能够基于 100% 完整治理数据投喂小语言模型（SLM）、未来微型模型以及检索增强生成（RAG）系统。相比于在算法端费力编写提示词或规则，直接在“数据层施加安全护栏”（Data Guardrails）更为高效直接——例如从数据源头对隐私数据进行严格管控，即可彻底杜绝营销算法违反隐私法的可能性。此外，Kon Leong 严厉指出了当前企业级 AI 分析的通病：绝大多数团队仍然停留在隔离的“沙盒”（Sandbox）中进行抽样测试。沙盒模式不仅耗时费力、存在严重的抽样偏差，更充斥着缺乏治理的“西部荒野式”混乱。企业应当“跳出沙盒，拥抱整片海滩”（Jump out of the sandbox and do the beach）。通过对整片数据海滩进行虚拟化，企业不再是在沙盒里碰运气寻找少量样本，而是可以直接从全局数据中筛选出高达 300 英尺的“关键贝壳堆”（目标线索与证据）。 - 沙盒模式局限：局部取样误差大、反复清洗浪费人力、缺乏法律与合规护栏约束。 - 虚拟化全量数据海滩：自动解决取样偏差与清洗治理问题，以全局视野实现极速目标数据萃取。 --- ### [13:40](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=820s) 从局部到全局：100% 捕获与“缺席数据”的洞察力量在具体落地策略上，企业应遵循“小步快跑”的原则。以某跨国企业的知识产权与产品责任风险排查项目为例，该团队并没有一开始就全面铺开至全公司 10 万名员工，而是首先虚拟化整合了研发部门、产品管理部门以及海外研发中心的 2,000 名核心人员数据，在虚拟知识库中验证并调优 GenAI 分析管线，确认成效后再逐步推广至全集团。同样，一家全球大型银行此前在世界各地的分支机构各自为政、数据孤岛林立，总部对各地风险资产一无所知；通过引入统一就地治理平台后，不仅梳理清楚了全球资产、实现合规处置（Defensible Disposition），更将 GenAI 数据供给速度提升了千倍以上。在数据捕获与业务洞察的关系上，Kon Leong 揭示了一个“曲棍球棒效应”（Hockey-stick effect）：当数据捕获率在 0% 到 99% 之间缓慢爬升时，企业所能获得的洞察价值增长非常平缓；但一旦跨过 99% 迈向“100% 全量捕获”，洞察的价值将发生爆发式跃升。 ```mermaid flowchart LR accTitle: 100% 数据捕获与曲棍球棒效应 subgraph Part["部分捕获 (0% - 99%)"] P1["存在采样盲区"] --> P2["仅能分析已知足迹"] P2 --> P3["洞察价值增长缓慢 (蜗牛曲线)"] end subgraph Full["100% 全量捕获"] F1["全量行为基线"] --> F2["精准识别『缺失的足迹』 (福尔摩斯未吠之犬)"] F2 --> F3["洞察价值爆发跃升 (曲棍球棒暴增)"] end P3 -.->|突破全量阈值| F3 ``` 其深层原因在于：不完整的数据捕获只能让你分析“已经存在的数据”，但在反洗钱（AML）、反欺诈调查等关键领域，最致命的线索往往隐藏在“不该缺失却缺失的数据”之中。如同福尔摩斯探案集中“夜间未吠的狗”一样，只有在确保 100% 全量捕获的前提下，系统才能确凿判定某些沟通记录或行为足迹遭到了刻意抹除或异常缺失，从而精准锁定制胜的违规证据。 --- ### [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 关键架构原则：解耦训练、预先提炼与确定性搜索在演讲尾声，Kon Leong 为企业构建 GenAI 架构总结了数条关键的避坑准则与经验法则： 1. **模块化解耦与“反向训练”（Untrain GenAI）**：切忌构建铁板一块的单体巨型黑盒模型，而应将知识模型拆分为灵活的“拼图碎片”（Puzzle Pieces），以便在需要剔除错误数据或合规受限知识时，能够像更换拼图一样精准替换或重置特定模块。 2. **在生命周期删除前“提炼精髓”**：当企业准备对过期的电子邮件和非结构化文件执行合规清理时，切莫直接物理销毁。在丢弃原始庞大文件的同时，应预先提取文档中的知识脉络与核心洞察并予以归档，既规避了原始文档的法律放射性，又为未来的 GenAI/AGI 留存了宝贵火种。 3. **分清结构化与非结构化的底层逻辑差异**：结构化数据通常倾向于多建副本并追踪数据血缘（Lineage），而非结构化数据则应极力避免创建任何多余副本。 4. **切勿废弃确定性技术（Deterministic Search）**：虽然 AI/ML 语义检索风头正劲，但在面对监管机构和法庭诉讼时，双方缺乏信任基础，黑盒模型的不透明性无法被司法采信。企业必须同时保留传统的布尔逻辑确定性搜索与 AI 搜索，以满足严格的“可解释性”（Explainability）要求。重构企业组织记忆的意义早已超越了当前的 GenAI 工具范畴。随着近实时数据处理能力的普及，企业距离经典科幻电影《少数派报告》（Minority Report）中汤姆·克鲁斯所展示的“海量数据在指尖近实时流转调取”的人机协同图景，已经近在咫尺。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) 非结构化数据是百分之百由人类产生、并为人类服务的数据。 — Kon Leong > [00:04](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=4s) 面对如此详尽的人际互动动态，数据对员工行为和企业运行的了解程度，甚至常常超过人们对自身的了解。 — Kon Leong > [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) 由于非结构化数据高度记录了人类的真实言行，它天然具有『放射性』。如果控制不当，这些陈年数据很可能成为反噬企业的核弹。 — Kon Leong > [03:30](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=210s) 最困难的部分是数据……它不是 AI 本身，不是机器学习，也不是生成式 AI，甚至不是如何让员工正确使用它，真正的难点在于数据。我们拥有的数据量让一切相形见绌，但要让数据变成可用形式极为困难，因为它们格式各异、完全非结构化，比如日常对话或呼叫中心记录……我们目前在 AI 上的投入约为 20 亿美元，拥有大约 600 个实际投入使用的用例，有些至关重要，有些相对次要，这个数量明年可能会翻倍甚至翻三倍。 — Jamie Dimon > [06:17](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=377s) 大型跨国银行所管理的文档总量甚至超过了 Google 索引的网页文档数。Google 的搜索只遵循『尽力而为』原则，但在强监管的银行体系中，遗失任何一份关键文档都可能让你彻底完蛋。 — Kon Leong > [10:14](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=614s) 跳出沙盒，拥抱整片海滩。企业不再是在沙盒里碰运气寻找少量样本，而是可以直接从全局数据中筛选出高达 300 英尺的关键贝壳堆。 — Kon Leong > [13:40](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=820s) 当数据捕获率在 0% 到 99% 之间缓慢爬升时，企业所能获得的洞察价值增长非常平缓；但一旦跨过 99% 迈向 100% 全量捕获，洞察的价值将发生曲棍球棒式的爆发跃升。 — Kon Leong > [13:40](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=820s) 在反洗钱与反欺诈调查中，最致命的线索往往隐藏在不该缺失却缺失的数据之中，如同福尔摩斯探案集中『夜间未吠的狗』。 — Kon Leong > [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 切忌构建铁板一块的单体巨型黑盒模型，而应将知识模型拆分为灵活的拼图碎片，以便在需要剔除错误数据或合规受限知识时，能够精准进行反向训练与模块替换。 — Kon Leong > [16:51](https://www.youtube.com/watch?v=miMJ4EVbcWI&t=1011s) 虽然 AI 语义检索风头正劲，但在法庭诉讼与监管审查中，黑盒模型的不透明性无法被司法采信，企业必须同时保留传统的确定性搜索以满足可解释性要求。 — Kon Leong ## Mentioned Kon Leong, ZL Technologies, 摩根大通 (JPMorgan Chase), Jamie Dimon, Google, 彭博社 (Bloomberg), 微软 (Microsoft), Microsoft 365 (M365), SharePoint, ChatGPT, 美国特勤局 (Secret Service), 汤姆·克鲁斯 (Tom Cruise), 《少数派报告》 (Minority Report) ## Video description Kon Leong, CEO, ZL Technologies Abstracts The momentum from ChatGPT has forced every organization to answer a simple question: “What’s our AI story?” Whereas most generic AI tools today are trained on public data, the real value for corporate use is training AI on enterprise data. However, the sheer scale of corporate data and the need for data governance test the limitations of today’s technology. This session will explore how large organizations are implementing a system to manage enterprise data for governance requirements, while curating and delivering a targeted and relevant data feed for AI and analytics. By tapping into the entire corpus of employee-created data, organizations are able to unlock deep insights into the “human side” of the company. [Source](https://www.youtube.com/watch?v=miMJ4EVbcWI)

Source quotes

[0:04] Kon Leong: 非结构化数据是百分之百由人类产生、并为人类服务的数据。 [0:04] Kon Leong: 面对如此详尽的人际互动动态，数据对员工行为和企业运行的了解程度，甚至常常超过人们对自身的了解。 [3:30] Kon Leong: 由于非结构化数据高度记录了人类的真实言行，它天然具有『放射性』。如果控制不当，这些陈年数据很可能成为反噬企业的核弹。 [3:30] Jamie Dimon: 最困难的部分是数据……它不是 AI 本身，不是机器学习，也不是生成式 AI，甚至不是如何让员工正确使用它，真正的难点在于数据。我们拥有的数据量让一切相形见绌，但要让数据变成可用形式极为困难，因为它们格式各异、完全非结构化，比如日常对话或呼叫中心记录……我们目前在 AI 上的投入约为 20 亿美元，拥有大约 600 个实际投入使用的用例，有些至关重要，有些相对次要，这个数量明年可能会翻倍甚至翻三倍。 [6:17] Kon Leong: 大型跨国银行所管理的文档总量甚至超过了 Google 索引的网页文档数。Google 的搜索只遵循『尽力而为』原则，但在强监管的银行体系中，遗失任何一份关键文档都可能让你彻底完蛋。 [10:14] Kon Leong: 跳出沙盒，拥抱整片海滩。企业不再是在沙盒里碰运气寻找少量样本，而是可以直接从全局数据中筛选出高达 300 英尺的关键贝壳堆。 [13:40] Kon Leong: 当数据捕获率在 0% 到 99% 之间缓慢爬升时，企业所能获得的洞察价值增长非常平缓；但一旦跨过 99% 迈向 100% 全量捕获，洞察的价值将发生曲棍球棒式的爆发跃升。 [13:40] Kon Leong: 在反洗钱与反欺诈调查中，最致命的线索往往隐藏在不该缺失却缺失的数据之中，如同福尔摩斯探案集中『夜间未吠的狗』。 [16:51] Kon Leong: 切忌构建铁板一块的单体巨型黑盒模型，而应将知识模型拆分为灵活的拼图碎片，以便在需要剔除错误数据或合规受限知识时，能够精准进行反向训练与模块替换。 [16:51] Kon Leong: 虽然 AI 语义检索风头正劲，但在法庭诉讼与监管审查中，黑盒模型的不透明性无法被司法采信，企业必须同时保留传统的确定性搜索以满足可解释性要求。

Source quotes

[0:04] Kon Leong: 100% generated by humans for humans. [0:04] Kon Leong: know more about you than you think you know of yourself. [3:30] Kon Leong: Sorry, I don't have it. The original isn't there. [3:30] CEO of JPMorgan Chase: The hardest part is the data, which, you know, it isn't the AI... it's getting your people to use it properly, but it's the data. I mean, we have—I forgot the number—but it dwarfs everything. So getting the data in the form that it could be usable is the hard part, because it's in different formats and completely unstructured, like just conversations that people had on call centers. We now spend about $2 billion on AI. We have something like 600 actual in-use cases... That number will probably double next year and maybe triple. [12:34] Kon Leong: jump out of the sandbox and do the beach. [15:43] Kon Leong: the dog that didn't bark.

Original transcript

[0:04] Uh my talk today is about unstructured [0:06] data in the main which is h which [0:09] happens to be the driving force behind [0:12] Genai and will be for AGI. [0:17] uh we are coming from the perspective of [0:20] servicing very very large data sets at [0:23] very large banks for primarily [0:26] information governance for the last 26 [0:29] years. So much of our experience is with [0:33] emails and files, fileshares, Bloomberg [0:37] uh SMS text uh chats and so forth. [0:41] And uh as you well know from the [0:44] financial sector, the regulatory and [0:48] eiscocovery [0:49] forces are quite uh considerable. So [0:53] I'll be talking about those two and [0:56] relating them to Genai on unstructured [0:59] data. This is your primary sources of [1:02] unstructured data at any financial [1:04] institution. [1:06] And uh you'll excuse me if I my comments [1:10] focus primarily on the problems faced by [1:14] very large banks [1:16] uh because that's been our experience in [1:18] the last 26 years. [1:21] Uh a lot of it is in the collab uh [1:23] suites and then of course in the IM and [1:26] chats and so forth. [1:28] One thing that people haven't really [1:30] noted is that [1:33] all of this is [1:36] has one thing in common that's all human [1:39] 100% generated by humans for humans. [1:43] And as such, [1:46] since all the banks already have the [1:48] obligation to collect and harness and [1:52] analyze all of this, you can also [1:55] repurpose this [1:57] to form for the first time [2:00] organizational memory and knowledge. [2:04] This actually extends far beyond to [2:07] today's genai and into the future into [2:10] AGI. [2:12] You might want to keep that in mind. [2:15] Obviously, this has already come to pass [2:18] because the last 20 years, we've seen [2:20] everyone focus on this. Every time [2:22] something happens that's significant, a [2:25] crisis, a scandal, uh, litigation, all [2:29] the regulators, litigators, and [2:32] management focus on it. If you recall on [2:36] Jan 7, the day after, the first thing [2:39] they went to was the SMS text by the [2:42] Secret Service. Since we already have [2:46] the obligation [2:48] as banks to track all of these [2:52] media [2:55] based on that, [2:58] there's not a whole lot you don't know [2:59] about every human participant [3:03] and their dynamics. [3:09] It is sometimes a little bit scary, [3:13] but the data may know more about you [3:15] than you think you know of yourself. So, [3:18] everyone's kind of gotten into the uh [3:21] unstructured data space, but it takes a [3:25] fair bit of understanding [3:27] to know what to do there. [3:30] For example, [3:32] this is the data lake concept that's [3:34] often propagated [3:37] uh and championed by many players in the [3:39] field. [3:41] Structured and unstructured is roughly [3:44] between one ratio of 1 to five. [3:48] However, [3:50] what is probably skipped over is that [3:53] unstructured data is human and therefore [3:56] by that nature it's radioactive. [4:00] What do we what do I mean by [4:02] radioactive? [4:04] It can come back to haunt you if you [4:05] don't control it. [4:07] If you remember a few months ago on [4:10] Capitol Hill, [4:13] this chap was sweating on the stand [4:16] and based on some content that was over [4:20] a decade old, his company may well be [4:24] broken up just based on that. When you [4:26] construct your strategy, [4:30] understand that you probably need to get [4:32] a sign off from the chief legal officer [4:35] or the chief compliance officer or the [4:39] chief records officer or the chief [4:41] privacy officer. [4:45] The [4:46] rest of it probably you don't want to [4:49] make a copy of it accessible. [4:52] You probably want to make it a virtual [4:55] data repository where there's no [4:57] original. [4:59] That way, if you get hit by a subpoena, [5:01] you can simply say, "Sorry, I don't have [5:03] it. The original isn't there. If it is [5:07] there, watch out." As I said, we service [5:10] the very large banks, the largest ones, [5:13] including the likes of City Bank and uh [5:19] Wells Fargo and JPMC and so forth. I'd [5:22] like to share with you a recent uh [5:25] comment from the CEO of one of them, [5:29] JPMC. I the hardest part is the data [5:33] which you know yeah it isn't the AI you [5:35] know it's not the machine learning it's [5:36] not the generative AI it's getting your [5:39] people to use it properly but it's the [5:40] data I mean we have I forgot the number [5:43] but it is dwarfs everything so getting [5:45] the data in the form that it could be [5:47] usable is the hard part because it's in [5:50] different formats and completely [5:51] unstructured you know like just [5:53] conversations that people had or uh you [5:56] know on call center something like that [5:57] but now we're be able to access all of [5:59] them to do a better job. We now spend [6:01] about $2 billion on on AI. We have [6:04] something like 600 actual inuse cases. [6:07] Some are really important, some are more [6:08] minor. That probably number will [6:10] probably double next year and maybe [6:12] triple, you know, just it'll just keep [6:13] on growing. [6:17] uh a couple of things to note and that [6:19] is they've come quite a long ways on [6:21] their journey and he said that [6:24] unstructured data is really hard [6:27] and the last one is that they're pumping [6:30] away at over 600 use cases. I'd like to [6:33] take a quick survey to see the [6:34] state-of-the-art of AI implementation. [6:37] How many of you have got over aundred [6:39] use cases? If you raise your hands, [6:43] how many of you have over [6:45] 500 use cases. [6:49] How about 200 use cases? [6:52] No. So most of you are beginning on the [6:55] journey. One thing about JPMC is that [6:58] they very much prefer a platform [6:59] approach because otherwise you you're [7:02] you're stuck with silos and they prefer [7:05] a platform that can replicate across. [7:07] That's how they get to 600. [7:09] But uh I'd like to dwell a few moments [7:12] on why he said unstructured data is so [7:15] difficult. And uh this kind of gives you [7:18] a hint. [7:21] The first order of business is that you [7:23] got to access the unstructured data and [7:27] you need to access all of it not just [7:29] slices of it. [7:32] Two is that you need to access all types [7:35] of data. [7:38] Just doing one and repeating it for [7:40] another is just inefficient and also [7:44] problematic ineffic ineffectiveness. So [7:46] you need to get all the types of data. [7:49] Third one is that you need to scale and [7:51] for here I think most people [7:53] misunderstand the scaling requirements. [7:59] We're talking about tens and tens of [8:01] pabytes and in particularly the big [8:04] banks. Would it surprise you to know [8:07] that they have more unstructured data [8:11] documents than even Google does [8:16] except that the management requirement [8:18] in the bank is a whole lot more [8:20] difficult than at Google because Google [8:22] is just best efforts. [8:25] here you lose a single document, you [8:27] could be toast. [8:30] So that's the scaling issue. The third [8:33] one is that this is a new innovation [8:36] because it's so darn expensive and so [8:39] radioactive. Would be nice if you could [8:42] actually have the option to either keep [8:45] a copy for highv value documents or no [8:48] copy for lowv value documents but [8:51] extract all the essence of the documents [8:53] so that you can use it for geni and agi. [8:59] So that's the in place that came in and [9:01] because in place came in this is what [9:04] many of the big banks realized when we [9:07] deployed [9:09] the benefits are massive. Number one [9:12] your cost the storage cost directly goes [9:16] back goes down by 90%. So that allows [9:19] you to keep a whole lot more data for a [9:21] lot longer. [9:23] Number two, your risk also goes down [9:26] significantly because if they hit you [9:29] with a subpoena for it, you say, "I'm [9:31] sorry, don't have it. Originals are not [9:34] here." [9:36] The last one is serendipitous and that [9:39] is your speed to data to data analytics [9:44] because you do not if you extract all [9:46] the data and you don't have to go back [9:49] to the original SharePoint or 365 or [9:52] what have you your time to data [9:55] analytics is sped up by a thousand [9:58] times. [10:01] That's a massive massive [10:04] differentiation. [10:08] It changes the paradigm of uh analytics. [10:13] Now along with the obligations of [10:16] information governance, you have [10:17] compliance which has mostly been covered [10:19] by uh silos of the last 20 years. And [10:23] then another silo is eiscocovery. [10:25] Another silo would be records and life [10:28] cycle management. Another silo would be [10:30] uh privacy, GDPR, CCPA. [10:33] And these have all been siloed very [10:36] inefficient [10:37] and basically at cross purposes. So the [10:41] way to do a go forward is to actually [10:43] unify them into one [10:48] so that one motion does it all. And not [10:51] only that, go further and unify the data [10:54] platform, the data schema so that you [10:57] just keep one copy. [11:00] That also reduces your cost which then [11:03] means that you can actually combine the [11:05] cost of all of these silos and actually [11:08] save money while increasing your uh [11:12] capabilities. [11:15] Now the gen data feed you have uh one [11:21] huge benefit. It's now governed and you [11:24] have access to 100% of the data. No more [11:26] sampling. Number two, you have also time [11:30] to analytics speeded up by a [11:32] thousandfold. And thirdly, you now have [11:36] additional capabilities such as SLM. I [11:40] think they'll pretty soon we'll have [11:42] micro language models and you have rag [11:45] to augment and reduce our cost and make [11:48] it more relevant. You also can apply the [11:50] data guardrails instead of guardrails by [11:53] algorithms. data is a whole lot more [11:55] direct and faster to get to it. If you [11:59] guard your privacy data, it's unlikely [12:02] that your marketing algorithms will [12:03] violate the privacy laws. [12:06] And then you can also start to untrain [12:09] Genai and I'll talk a bit a little bit [12:12] about that later. [12:16] So you can afford to have both copies [12:21] and no copies archive or in place and [12:24] then you should put it all on one data [12:26] platform. I think you'll find that the [12:29] large institutions are going this route. [12:34] Quick run through in terms of what [12:36] actually happens out in the field. Most [12:39] of analytics today and eiscocovery and [12:41] genai it happens in the sandbox. The [12:44] sandboxes are fundamentally limited. [12:47] It's slow. It's sampling. You always [12:49] have sampling errors. Number two, a lot [12:52] of your labor is wasted. A lot of your [12:53] time is wasted. And if you don't like [12:55] it, rinse and repeat. And lastly, [12:57] there's no governance in it. That cannot [12:59] today. It's the wild west. That cannot [13:02] continue. So, what to do? Well, you [13:06] don't have to stay with the sandbox. [13:08] Jump out of the sandbox and do the [13:10] beach. [13:12] This way. All of the sampling errors, [13:14] scrubbing, and governance are already [13:16] solved automatically. [13:19] Now, instead of having just a small [13:22] little sample of, let's say what you're [13:24] looking for, red shells or smoking guns, [13:28] instead if you scour the beach, you can [13:31] have a pile of shells, your target [13:32] shells, [13:34] 300 ft high if you want. That's [13:37] completely different paradigm. [13:40] What should you do? Well, number one, [13:43] don't invest in duplicating the beach. [13:46] Virtualize the beach. [13:48] Number two, start small. [13:51] For example, someone looking at product [13:54] uh this is an actual case product [13:56] product liability and intellectual [13:58] property. Started just with the R&D [14:01] department, the product management [14:03] department and the overseas R&D [14:05] department. 2,000 people versus 100,000. [14:09] put it all into a virtual repository and [14:12] then feed the sandbox and get all the [14:14] the feeds that you want you like it and [14:17] then expand it to the rest of the data. [14:21] Start small. One takeaway from the big [14:25] bank I was talking about earlier is that [14:28] they had a huge problem. Here's a full [14:30] use case. They had uh data scattered all [14:33] over the place around the world and it [14:36] was siloed. So every country had all [14:38] their own silos. They had no idea what [14:40] was in there. They had no idea what was [14:43] uh uh problematic. They had no idea how [14:46] to fix it. They had no idea how to life [14:48] cycle it. And so the first thing was to [14:52] actually consolidate the management and [14:55] then to actually [14:58] cut cost by reducing it to in place [15:02] and then do all of the governance [15:04] together. [15:06] and then use the access to 100% of the [15:09] data to feed your Genai sandbox as I [15:12] said a thousand times faster and then of [15:16] course uh you can now uh implement your [15:20] life cycle defensible disposition for [15:24] those of you sensitive to that and also [15:27] incorporate governance and genai [15:30] together. [15:33] All of these are [15:36] capabilities that are not available in [15:38] the Microsoft platform. So you need to [15:40] augment them there. [15:43] I'll spend the next few minutes in maybe [15:46] uh giving you a few good hints. [15:49] One is that uh when you get to 100% [15:52] insights [15:54] or capture, the value of your insights [15:58] goes like grows like a snail's uh pace. [16:01] But when you get to the 99% 100% [16:04] something miraculous happens. The value [16:06] of your insights go like a hockey stick. [16:10] And the reason is because your insights [16:12] from data that that exists is that low. [16:16] But when you have a 100% capture, you [16:19] can get insights from data that does not [16:22] exist. [16:25] Because then instead of dealing with [16:27] this, this is a partial capture. There's [16:30] only so much insight you can get from [16:32] the footprints. But if you have a 100% [16:34] capture, you can definitively say, "Yep, [16:38] these data footprints are missing." [16:41] And there in lies all of the fraud [16:43] detection and so forth. Uh the [16:47] anti-money laundering and so forth. [16:51] Uh I won't I'll tell you the details of [16:54] the Sherlock Holmes uh the dog that b [16:56] didn't bark. But the next uh thing is [16:59] that you can also untrain genai by h by [17:03] instead of building a monolithic model [17:06] do it in puzzle pieces and then replace [17:08] the puzzle pieces with genai. [17:11] Third is, and this is important, when [17:14] you're deleting your unstructured data, [17:17] emails, and files, stop. Wait. See if [17:21] you can actually do something with it [17:22] because once you delete it, it's gone. [17:25] Instead, before deletion, if you can use [17:28] a process to capture the essence of the [17:31] documents and then throw the originals, [17:33] you can then go go to your deletion but [17:38] keep the data for future use. [17:44] And lastly, [17:48] the other lessons learned, uh, don't [17:50] apply structure to unstructured. They're [17:52] very, very different. uh copies in [17:55] structure you make all the copies galore [17:58] and unstructured you do not make a [18:00] single copy if you can help it. Lineage [18:03] structured you're worried about that in [18:04] unstructured you're not [18:08] uh be careful about swapping legacy for [18:10] new technologies because in governance [18:13] there's a problem. There's a lack of [18:15] trust between two parties the litigants [18:18] or the regulator and the regulated. So [18:22] do not throw away your deterministic [18:24] technologies like search. You need both [18:28] a IML searches and uh deterministic [18:31] search because when you're asked to [18:33] produce, they want deterministic. They [18:35] don't want a black box. They want [18:37] explanability. [18:39] Lastly, [18:41] look twice at a gente rag especially the [18:44] impact on cost and performance. It can [18:46] make a huge difference. [18:49] Uh basically this is the exercise of [18:51] rebuilding or building for the first [18:53] time your organizational memory and [18:56] knowledge. It goes way beyond geni. [19:00] And I'll just close with a scene from [19:02] Hollywood. We're not too far away from [19:05] the shot from uh what was that Tom Cruz [19:08] movie. Um we'll have all of the data at [19:11] our fingertips and near real time. [19:14] So that's a summary of uh my [19:17] presentation [19:18] and if you want uh please stop by our [19:21] booth. We can answer any questions you [19:22] might want.