Progress Deploys Hybrid Knowledge Graphs to Lift AI Accuracy to 88%

Trusted AI: Unlock Data Value With Progress Software — Ai4

Ai4

Standalone generative models fail in the enterprise because probabilistic algorithms lack proprietary context, making human-curated knowledge graphs essential to stop autonomous agents from multiplying errors at scale.

AI-assisted summary of Trusted AI: Unlock Data Value With Progress Software — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by Ai4 on August 29, 2025, this presentation features Philip Miller, Senior Product Marketing Manager at Progress Software. According to the description, the session explores how Progress Software's Retrieval-Augmented Generation (RAG) solution connects generative AI directly to proprietary enterprise data, enabling business leaders and data strategists to turn siloed knowledge into trusted, automated insights without requiring an engineering background.

## TL;DR The session addresses the core operational pitfalls of unmanaged enterprise generative AI deployments, where high error rates and lack of domain context prevent scalable adoption. Progress Software presents its multi-model data platform and hybrid retrieval framework, combining vector search, knowledge graphs, and human-in-the-loop 'Progressive Graphs' to boost accuracy from 70% to 88%. Demonstrating real-world ROI across financial analysis, multi-source crime investigations, and the agentic capabilities of their Nuclia acquisition, Miller outlines how businesses can transform internal data into an auditable, deterministic competitive moat.

## Atmosphere & style Fast-paced, pragmatic, and outcome-oriented, the presentation delivers the brisk energy of an experienced strategist compressing a deep technical topic into a crisp conference briefing. Miller blends candid realism about the operational dangers of baseline AI hallucinations with an enthusiastic, solutions-driven pitch for contextual data governance. The talk feels highly engaging and accessible, balancing sharp technical architecture diagrams and live investigative demos with approachable humor and an insistent focus on tangible business value.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=AxYew0fzuiA&t=4s) The Progress Data Platform natively unifies structured, unstructured, geospatial, bitemporal, graph, and vector data into enterprise knowledge graphs to translate domain expertise into machine context. - [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) Unmanaged baseline LLMs can fail up to 60% of the time due to lack of domain context, static training data, and hallucination risks, which creates severe liability when scaling autonomous AI agents. - [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) Enterprise AI requires deterministic, accurate, and auditable outputs without demanding continuous manual fact-checking by knowledge workers. - [08:01](https://www.youtube.com/watch?v=AxYew0fzuiA&t=481s) Harmonizing structured data and knowledge graphs saved 4,000 financial analysts $30M annually, while adding GenAI cut regulatory report drafting time from 4 hours to 15 minutes. - [10:06](https://www.youtube.com/watch?v=AxYew0fzuiA&t=606s) The contextual retrieval pipeline executes hybrid search across rank-relevance, vector, and graph layers before applying fine-grained security policies to prompts. - [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) Targeting retrieval at the sentence and paragraph level reduces compute costs and latency while providing verifiable sentence-level citations against the source data. - [12:18](https://www.youtube.com/watch?v=AxYew0fzuiA&t=738s) The 'Progressive Graphs' feature alerts subject matter experts to missing domain concepts, allowing rapid curation that boosted retrieval accuracy from 70% to 88% in a customer proof of concept. - [14:15](https://www.youtube.com/watch?v=AxYew0fzuiA&t=855s) A law enforcement demo illustrated how correlating bitemporal timelines, geospatial records, and 911 audio transcripts can reveal hidden criminal behavioral patterns. - [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) Progress Software acquired Nuclia to integrate agentic RAG capabilities, enabling autonomous agents to orchestrate data classification, video parsing, and unstructured ingestion. - [17:13](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1033s) A sustainable enterprise competitive moat is built by scaling unique proprietary knowledge, logic, and human common sense through generative AI rather than relying on generic models. - [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) Organizations should prepare clear business outcomes, relevant datasets, and API requirements before adopting AI rather than running unguided technology experiments.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=AxYew0fzuiA&t=4s) Introduction to Progress Software and the Enterprise Data Platform

The discussion focuses on moving generative AI beyond experimental research and into production to deliver measurable business outcomes rather than focusing purely on the underlying technology. Progress Software provides a comprehensive infrastructure stack that spans from the core infrastructure management layer, through the enterprise data and information layer, all the way to the user interface (UI) layer. To help organizations scale AI deployments from small departmental groups to hundreds of thousands of concurrent users, Progress offers both an enterprise Retrieval-Augmented Generation (RAG) platform and an agentic RAG solution integrated directly into its portfolio.

At the center of this architecture sits the Progress Data Platform, a multi-model environment built on roughly two decades of engineering development. The platform natively ingests and harmonizes diverse data formats—including structured and unstructured records, geospatial, bitemporal, graph, and vector data. By combining these multimodal search and curation capabilities with metadata and knowledge management tools, the platform constructs knowledge graphs. These knowledge graphs formalize domain-specific business concepts, effectively translating implicit human expertise into machine-readable context at enterprise scale.

- **Full-Stack Portfolio**: Spans infrastructure management, enterprise data and metadata layers, and front-end UI components. - **Multimodal Ingestion**: Native handling of structured, unstructured, geospatial, bitemporal, graph, and vector data. - **Knowledge Graphs**: Bridges human intuition and machine intelligence by expressing organizational terminology and semantic relationships at scale.

---

[02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) The Reality of Generative AI and Enterprise Deployment Pitfalls

Generative AI is no longer a future promise; it has already fundamentally disrupted the landscape. As Philip Miller notes, "This bomb went off in 2022 and no one heard it and it went around the world," leaving organizations to grapple with the unintended consequences of rapidly adopting these models. Real-world customer implementations expose severe operational hurdles when off-the-shelf generative models are deployed without enterprise-grade guardrails. In unmanaged deployments, models have been observed answering queries correctly as little as 40% of the time—a failure rate that makes enterprise scaling impossible, particularly for autonomous AI agents where "the human being can make one mistake" while "your AI agents can make 10,000 mistakes in that time."

These systemic failures stem from several core deficiencies in baseline LLMs: - **Hallucinations and Reliability**: High error rates in out-of-the-box setups break user trust and introduce massive operational liability. - **Data Biases and Lack of Domain Context**: Inability to interpret proprietary information through the lens of specific industry regulations and operational standards. - **Lack of Human Reasoning**: Highly capable generic models falter on specialized internal data, akin to asking a PhD-level physicist to perform open-heart surgery. - **Static Training Cutoffs**: LLM weights remain frozen at the point of training, preventing the model from learning real-time internal updates, new business models, or shifting corporate policies. - **Lack of Explainability**: Both generative models and fragmented legacy data ecosystems act as black boxes, making it difficult to audit why specific data points informed a decision in regulated environments. - **Lack of Robustness**: Demanding deterministic, repeatable enterprise outcomes from a fundamentally non-deterministic, probabilistic tool is comparable to "wanting to roll a dice and it always land on six."

---

[06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) Enterprise Requirements for Scalable, Trusted AI

The fundamental challenge for enterprise AI is bridging the gap between proprietary organizational context and automated language generation. Organizations require systems that can scale human domain knowledge without demanding constant manual intervention. If knowledge workers are forced to repeatedly fact-check, correct, and re-prompt the AI due to inaccurate or unverified responses, the economic and operational value of the deployment collapses.

To deploy AI safely in complex or heavily regulated domains, systems must guarantee strict accuracy, domain expertise, adaptability, compliance, timeliness, and intuitive user experiences. Establishing trusted data foundations ensures that models provide consistent, auditable, and deterministic results, eliminating the need to have a human constantly in the loop fixing routine output errors.

```mermaid flowchart TD accTitle: Enterprise AI Scalability Architecture subgraph Context ["1. Enterprise Context Layer"] A["Proprietary Domain Data"] --> C["Knowledge Graph & Semantics"] B["Regulatory & Business Rules"] --> C end

subgraph Platform ["2. Governance & Data Engine"] C --> D["Progress Data Platform"] D --> E["Context Augmentation & Fine-Grained Security"] end

subgraph Out ["3. Enterprise Outcomes"] E --> F["Deterministic Outputs"] E --> G["Autonomous Agent Guardrails"] E --> H["Auditable Compliance"] end ```

---

[08:01](https://www.youtube.com/watch?v=AxYew0fzuiA&t=481s) Real-World Implementations: Tangible Time and Cost Savings

Enterprise organizations implementing this contextualized data platform approach have demonstrated significant operational returns both before and after integrating generative AI. In one financial services deployment, the organization used knowledge graphs and structured content authoring to harmonize internal research reports across 4,000 analysts. Saving each researcher 1.5 hours per week resulted in annual cost savings of $30 million. Layering generative AI directly on top of this structured foundation enabled the team to save an additional 4 hours per week per analyst, condensing the drafting of complex tax and regulatory impact analyses from 4 hours down to just 15 minutes.

In another instance, a client leveraged Progress Software's UI-driven reference architecture to develop an internal knowledge discovery tool. Within three weeks of receiving the architectural concept, the client successfully shipped an alpha build, which subsequently scaled to serve over 20,000 employees globally. By taking an iterative approach—beginning with focused, accurate pilot programs—enterprises can progressively expand their AI-enabled data platforms into enterprise-wide production rollouts.

- **Financial Services Savings**: 4,000 researchers saved 1.5 hours/week, totaling $30M/year; subsequent GenAI deployment cut 4-hour regulatory analysis tasks to 15 minutes. - **Rapid Global Rollout**: Enterprise customer went from reference architecture to functional alpha in 3 weeks, scaling to 20,000+ global employees. - **Iterative Pilot Strategy**: Validating accuracy and data connectivity in departmental pilots prior to broad enterprise rollouts.

---

[10:06](https://www.youtube.com/watch?v=AxYew0fzuiA&t=606s) The Contextual Retrieval and Governance Pipeline

Progress Software structures the retrieval and generation pipeline to guarantee that generative models receive enriched, governed, and highly accurate internal context. The process begins when an end user or an autonomous agent issues a query to the system. Rather than routing the prompt directly to an external language model, the system interprets the request and executes a multi-layered hybrid retrieval process across rank-relevance search, vector search, and graph search engines.

```mermaid flowchart LR accTitle: Contextual Retrieval and Governance Pipeline subgraph Ingest ["Query & Search"] Q["User / Agent Query"] --> I["Query Interpretation"] I --> S["Multi-Modal Search<br>(Rank, Vector, Graph)"] end

subgraph Gov ["Governance & Security"] S --> A["Context Augmentation"] A --> G["Fine-Grained Security<br>& Data Governance"] end

subgraph Gen ["AI Execution"] G --> LLM["Enriched LLM Context<br>& Output Generation"] end ```

This hybrid search queries both the underlying data and the contextual relationships surrounding it. Before the retrieved context is appended to the prompt, the platform applies fine-grained security and data governance policies. This ensures sensitive proprietary information is protected from unauthorized exposure before the enriched, context-complete prompt is submitted to the AI model for generation.

[10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) Precision Context, Token Efficiency, and Knowledge Graph Governance

Sending an entire thousand-page document into a large language model wastes vast amounts of inference compute, incurs massive token costs, and drags response latency to a crawl. Progress Software circumvents this by pinpointing and extracting data strictly at the sentence or paragraph level, supplying the model with "just what matters." This hyper-targeted extraction minimizes costs and latency while aligning the context directly with the user’s exact prompt.

Once the model generates a response, a rigorous governance layer immediately inspects the output to ensure the LLM has not hallucinated personally identifiable information (PII) or proprietary internal data. The system cross-references the generated context back against the enterprise knowledge graph to confirm that the underlying facts genuinely exist. When validated, the final answer is provided alongside granular citations, allowing both human users and AI agents to immediately verify the specific sentence or paragraph from which the conclusion was drawn. Rather than forcing organizations to duplicate or ingest all enterprise files into a single monolithic store, the platform can index data at source, supporting flexible deployments across both cloud and on-premises environments.

```mermaid flowchart TD accTitle: Progressive Graph RAG Architecture subgraph INGEST ["Data & Query Layer"] A["Enterprise Data Sources<br/>(Index at Source)"] --> B["Context & Metadata Enrichment"] C["User Prompt / Natural Language Query"] --> D["Query Augmentation"] end subgraph HYBRID ["Hybrid Retrieval Engine"] B --> E["Vector Search"] B --> F["Knowledge Graph"] D --> E D --> F F -. "Missing Concept Alert".-> G["SME Review & UI Curation<br/>(Progressive Graphs)"] G -. "Approve / Update".-> F end subgraph GOVERN ["Governance & Output"] E --> H["PII Scrubbing & Graph Validation"] F --> H H --> I["Governed, Sentence-Annotated Answer"] end ```

- **Index at source flexibility:** Ingests only necessary metadata and pointers rather than duplicating entire files, accommodating diverse enterprise architectures. - **Deep metadata enrichment:** Enriches documents far beyond basic authorship or timestamps by indexing the semantic information and concepts buried within unstructured files. - **Sentence-level citations:** Answers are presented with explicit provenance, allowing immediate manual verification against source paragraphs.

---

[12:18](https://www.youtube.com/watch?v=AxYew0fzuiA&t=738s) Progressive Graphs: Human-in-the-Loop Concept Curation and Accuracy Gains

The core of the platform is a hybrid, rules-based retrieval-augmented generation framework combining vector embeddings with knowledge graphs and human oversight. A key differentiator is a proprietary capability termed "Progressive Graphs." If an end user submits a natural language query involving an entity or concept absent from the knowledge graph, the platform does not fail silently; instead, it dispatches an automated notification to a designated subject matter expert (SME). The expert reviews the concept through a dedicated, UI-driven curation console, confirms whether it belongs in the domain model, and updates the graph with a click—enabling the graph to learn and expand adaptively as queries occur.

To illustrate the dramatic impact of this human-in-the-loop mechanism, Miller shares metrics from a customer proof-of-concept (POC). Working alongside two domain SMEs for a single day, the platform programmatically identified 400 potential domain concepts. The SMEs reviewed the candidates and deliberately disregarded 200 that were either irrelevant or likely to degrade traversal performance, approving the remaining 200 into the live knowledge graph. That single pass with Progressive Graphs elevated retrieval accuracy from a baseline of 70%—achieved using a traditional vector-only retrieval approach—up to 88% accuracy. The objective is not generic search, but verifiable "knowledge discovery" intended to trigger automated enterprise action.

- **Baseline vs. Hybrid Accuracy:** Moving from a pure vector search approach (70% accuracy) to a hybrid knowledge graph with one round of Progressive Graphs curation produced 88% accuracy. - **Rapid SME Curation:** In a single day, 2 domain experts triaged 400 suggested concepts, filtering out 200 noise candidates and operationalizing 200 vetted relationships. - **Action-Oriented Retrieval:** Focuses on extracting structured intelligence and decision-ready facts rather than returning a traditional list of document links.

---

[14:15](https://www.youtube.com/watch?v=AxYew0fzuiA&t=855s) Demo in Action: Crime Analysis, Bitemporal Data, and the Purple Jumper

In a rapid 90-second recorded demonstration, Miller walks through an investigative law enforcement scenario where investigators seek to track a criminal suspect. The system has already executed a ranked relevance search to establish baseline hits, after which the investigator queries the generative AI directly using conversational prompts. The platform integrates heterogeneous data types across bitemporal timelines, geospatial mapping, field officer reports, and 911 dispatch audio transcripts.

By synthesizing these diverse streams against natural language questions, the system detects underlying behavioral patterns that would remain invisible in siloed searches. The user interface plots these connections graphically while preserving direct drill-down links to the underlying evidence, allowing investigators to inspect raw 911 transcripts directly. In the demo, the system pinpoints an identifying detail—a "purple jumper" worn by the suspect—and links it across multiple incidents to reveal an observable pattern of criminal behavior across specific geographical coordinates, enabling immediate tactical law enforcement action.

- **Bitemporal & Geospatial Intelligence:** Synchronizes valid-time and transaction-time records alongside geographic coordinates to reveal chronological criminal movement. - **Multi-Source Unstructured Ingestion:** Correlates disparate data formats, including dispatch audio transcripts and typed field incident logs, into unified graphical timelines. - **Traceable Pattern Recognition:** Surfaces behavioral correlations—such as clothing descriptions linked across incidents—with underlying source evidence immediately accessible underneath the visualizations.

---

[16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) Agentic RAG and the Nuclia Acquisition

Miller highlights Progress Software's strategic acquisition of Nuclia, an agentic RAG platform. The acquisition followed months of deep technical co-development between the two engineering teams, during which Progress found that Nuclia’s methodology mirrored their own philosophy, but extended it through autonomous, specialized agents. These autonomous agents can orchestrate models to classify enterprise data, parse video, and process raw unstructured inputs in entirely novel ways.

By incorporating Nuclia’s agentic capabilities, Progress enables organizations ranging from "a one-man band" to large global enterprises to deploy sophisticated RAG pipelines without requiring custom-built infrastructure. The system operationalizes domain-specific human intelligence at machine scale across the organization's existing data footprint, delivering the structured automation and semantic scaffolding that business leaders need to scale generative AI safely.

- **Agentic Model Orchestration:** Nuclia introduces autonomous agents capable of classifying information, ingesting video, and handling unstructured data streams. - **Accessible Scalability:** Designed to support early-stage implementations through enterprise-grade deployments without requiring specialized AI research teams. - **Machine-Scale Human Intelligence:** Embeds human domain knowledge and reasoning rules directly into the automated agent execution loop.

---

[17:13](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1033s) Building the Competitive Moat and Enterprise Value

Reflecting on what makes generative AI sustainable for business—an idea Miller jokes he probably came up with "over like a coffee or a beer"—he stresses that speed of innovation alone is not enough. True competitive differentiation comes from taking the existing contextual knowledge, proprietary logic, and specialized intelligence already locked inside an enterprise and systematically scaling it through machines.

Applying generative models directly to that proprietary knowledge creates a defensive "competitive moat" that generic foundation models cannot replicate. By maintaining a strict human-in-the-loop architecture, organizations supply the essential ingredients that purely algorithmic systems lack: domain context, common sense, and institutional accountability.

> "Speed and of innovation, the contextual knowledge and the intelligence that is in your business and applying that to the machine... that becomes your competitive moat... Let's scale that with the generative AI and let's extend the moat around your business."

---

[18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) Call to Action: 14-Day Free Trial, Data Readiness, and Booth 503

Wrapping up what he notes was an attempt to compress "a three-hour conversation into roughly 20 minutes," Miller invites attendees to begin exploring the platform through a 14-day free trial. However, he cautions organizations against jumping into technology trials without preparation: teams must first define the specific business outcome they want to achieve, gather the internal data relevant to that use case, and identify the required API integrations.

Miller emphasizes that Progress is not interested in running AI experiments merely for the sake of novelty—"I don't want to do this for the sake of an experiment or kind of this is fun, look what I can do... what we want for business is to deliver value from the AI that you've got today." The Progress team works directly with customers to inspect their proprietary data, configure the platform, and teach teams how to prompt domain agents effectively. Attendees are directed to scan the presentation's QR code and visit Miller and his team at Booth 503 to discuss specific enterprise AI implementations.

## Notable quotes > [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) This bomb went off in 2022 and no one heard it and it went around the world — Philip Miller

> [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) the human being can make one mistake — Philip Miller

> [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) your AI agents can make 10,000 mistakes in that time. — Philip Miller

> [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) wanting to roll a dice and it always land on six. — Philip Miller

> [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) just what matters — Philip Miller

> [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) a one-man band — Philip Miller

> [17:13](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1033s) over like a coffee or a beer — Philip Miller

> [17:13](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1033s) Speed and of innovation, the contextual knowledge and the intelligence that is in your business and applying that to the machine... that becomes your competitive moat... Let's scale that with the generative AI and let's extend the moat around your business. — Philip Miller

> [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) a three-hour conversation into roughly 20 minutes — Philip Miller

> [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) I don't want to do this for the sake of an experiment or kind of this is fun, look what I can do... what we want for business is to deliver value from the AI that you've got today. — Philip Miller

## Mentioned Philip Miller, Progress Software, Ai4, Progress Data Platform, Nuclia, Progressive Graphs

## Video description Philip Miller, Product Marketing Manager, Senior, Progress Software

Abstracts

What if you could ask your business’s data any question and get an instant, trusted answer? In this session, I’ll show how Progress Software’s Retrieval-Augmented Generation (RAG) solution can make that scenario a reality by connecting generative AI directly to your proprietary data. By seamlessly linking AI with your enterprise knowledge, this solution turns siloed information into tailored, high-value AI outputs for your organization. I’ll highlight how this approach boosts intelligent data access, enables decision automation, improves operational efficiency and elevates customer experience. From finance to healthcare, organizations across industries are already turning their internal data into a competitive edge with RAG – unlocking new insights and automating decisions that drive innovation. Best of all, no engineering background is required – this session is built for business leaders and data strategists who are ready to unlock AI’s strategic value.

[Source](https://www.youtube.com/watch?v=AxYew0fzuiA)

中文简报

Progress 部署混合知识图谱，将 AI 准确率提升至 88% 独立的生成式模型在企业中难以奏效，原因在于概率算法缺乏专有上下文，因此人工构建的知识图谱对于阻止自主智能体大规模成倍放大错误至关重要。 本场会议探讨了未经管控的企业生成式 AI 部署中的核心运营隐患，即高错误率和领域上下文的缺失阻碍了规模化应用。Progress Software 介绍了其多模型数据平台与混合检索框架，通过将向量搜索、知识图谱与引入人在回路机制的“渐进式图谱”（Progressive Graphs）相结合，将准确率从 70% 提升至 88%。结合在金融分析、多源犯罪调查以及收购 Nuclia 所获智能体能力的实际投资回报率展示，Miller 阐述了企业如何将内部数据转化为具备可审计性与确定性的竞争壁垒。 Progress Data Platform 原生统一融合结构化、非结构化、地理空间、双时态、图与向量数据，构建企业级知识图谱，将领域专业经验高效转化为机器可理解的上下文。 (0:04) 未加管治的基础大模型在实际应用中错误率可高达 60%，因缺乏专有领域知识与实时数据，在规模化扩展自主 AI Agent 时会带来严重的运营与合规风险。 (2:31) 某金融机构通过知识图谱与结构化数据整合，每年为 4,000 名分析师节省 3,000 万美元，叠加生成式 AI 后进一步将监管分析报告的起草时间从 4 小时缩短至 15 分钟。 (8:01) 平台独特的“Progressive Graphs”机制可在遇到未知概念时自动提醒专家审核，在某客户概念验证（PoC）中，仅经一轮专家筛选便将检索准确率从纯向量检索的 70% 提升至 88%。 (12:18)

Reading summary

Progress Software's Philip Miller argues that generic large language models fail in enterprise production due to hallucinations and a lack of proprietary context, requiring organizations to anchor generative tools in governed, multi-model data foundations and knowledge graphs to deliver deterministic business outcomes. Operational Pitfalls of Unmanaged Foundation Models (2:31) Off-the-shelf generative models introduce significant operational risk when deployed without enterprise guardrails. Miller notes that in unmanaged corporate deployments, models have answered queries correctly as little as 40% of the time. While an individual worker might make a single error, unconstrained autonomous AI agents can amplify that failure rate to 10,000 mistakes within the same timeframe, rendering enterprise scaling impossible without strict controls. These deployment failures stem from structural deficiencies in baseline large language models, including static training cutoffs that ignore real-time corporate updates, black-box architectures that impede regulatory auditing, and an inability to interpret proprietary domain standards. Miller compares expecting deterministic, repeatable business results from probabilistic foundation models to demanding that a rolled die always land on six. Hybrid Retrieval and Granular Output Governance (10:06) To bridge proprietary context and automated generation, Progress Software structures queries through a hybrid retrieval pipeline combining rank-relevance, vector search, and graph search. Rather than transmitting entire multi-thousand-page documents into model prompts—which drives up token costs and slows inference latency—the system isolates specific data at the sentence and paragraph level. The platform indexes data at the source across cloud or on-premises environments, avoiding the need to duplicate files into a central repository. Governance occurs both before and after text generation. Enriched prompts are subjected to fine-grained security policies prior to model execution. Once an output is produced, the platform scrubs personally identifiable information, cross-references generated facts against the enterprise knowledge graph to prevent hallucinations, and supplies sentence-level citations so users and agents can trace conclusions directly back to underlying source material. Progressive Graphs and Human Curation (12:18) To maintain semantic domain models without requiring continuous manual maintenance, the architecture uses a human-in-the-loop mechanism called 'Progressive Graphs.' When a user prompt includes an entity or concept missing from the existing knowledge graph, the platform alerts a designated subject matter expert. The expert reviews the concept through a curation console and either approves or rejects its integration into the live domain model. In a customer proof-of-concept cited by Miller, two domain experts evaluated 400 system-suggested concepts in a single day. The experts operationalized 200 relevant concepts and dismissed 200 noisy candidates. This single pass increased retrieval accuracy from a 70% vector-only baseline to 88% under the hybrid graph framework, demonstrating how targeted human oversight stabilizes automated retrieval. Quantifiable Cost Savings and Multi-Modal Deployments (8:01) Implementations of this structured data approach show substantial operational returns across different industries. In financial services, using knowledge graphs and structured content authoring across 4,000 research analysts saved 1.5 hours per analyst weekly, producing $30 million in annual cost savings. Layering generative AI over this curated foundation reduced complex tax and regulatory impact analyses from four hours to 15 minutes. In another deployment, an enterprise shipped an alpha discovery tool within three weeks of receiving the reference architecture, scaling it globally to more than 20,000 staff. The platform also integrates diverse data streams to support investigative workflows. In a law enforcement demonstration, conversational queries synthesized bitemporal timelines, geospatial maps, field incident reports, and audio dispatch transcripts. The platform isolated an identifying 'purple jumper' across disparate incident records, correlating behavioral patterns with underlying source evidence to guide tactical decisions. Through the acquisition of agentic RAG platform Nuclia, Progress aims to scale this contextual automation to unstructured data and video parsing without requiring custom infrastructure builds. Miller concludes that long-term enterprise differentiation stems from embedding proprietary context and human institutional rules into generative architectures, rather than relying on generic models. However, he cautions that organizations must clearly define desired business outcomes, assemble pertinent internal data, and verify API connectivity before entering platform trials.

精炼阅读摘要

Progress Software指出，未经严格治理的基础大语言模型在企业级应用中面临高达60%的失误率与运营责任风险；企业唯有结合多模态知识图谱、精准混合检索及人机协同治理，才能将专有数据转化为可靠的AI输出并构筑不可替代的商业竞争壁垒。 基础模型的企业部署缺陷与运营风险 (2:31) 讨论强调，生成式AI在2022年爆发后已全面颠覆行业格局，但企业直接采用开箱即用的大模型却遭遇了严峻的运维困境。Philip Miller指出，在缺乏企业级护栏的管理松散环境中，通用大模型回答业务查询的准确率有时仅有40%。这种故障率不仅无法支撑企业规模化扩展，在自主AI代理（Agent）的场景中更具破坏性——当单个人类员工犯下一次错误的时间内，未受控的AI代理可能已经连续产生上万次错误。 这类系统性风险根源于基础模型的固有短板：依赖概率统计容易产生虚假幻觉、欠缺解读私有数据与特定行业监管标准的能力、受限于固定的训练时间截点而无法获知企业最新动态，且黑箱机制导致输出难以审计。要求本质上属于概率性工具的模型持续输出具备确定性、可重复的业务结果，直接违背了其底层运行逻辑，给受监管行业带来了巨大的合规与法律隐患。 混合检索管道与句子级精准治理 (10:06) 为确保生成模型获取经过治理的企业内部语境，Progress Data Platform构建了涵盖相关性排序检索、向量检索和知识图谱检索的混合流水线。当用户或AI代理发出查询时，系统会先进行查询意图解析并执行多层检索，全面调取数据实体及其关联关系。在将检索结果附加至模型提示词之前，平台会严格执行细粒度安全与数据治理策略，防止机密数据越权泄露。 为降低推理计算消耗并缓解传输延迟，该架构改变了将整份长文档输入模型的做法，而是将上下文提取精确限制在句子或段落级别。在模型生成响应后，治理层会立即执行检查以清除个人身份信息（PII）或机密内容，并将生成结果与知识图谱进行反向事实核验。最终输出附带细粒度引文标注，支持源头数据索引与多云或本地部署，确保人机均可追溯事实来源。 渐进式图谱与专家协同的精度跃升 (12:18) 平台的核心机制在于将向量嵌入、知识图谱与人机协同监督深度融合的“渐进式图谱”（Progressive Graphs）技术。当终端查询包含现有知识图谱未涵盖的新实体或业务概念时，系统不会静默失败，而是向对应业务领域的专家（SME）发送自动化提醒。专家可通过专用控制台一键审核并确认该概念是否纳入业务模型，使知识图谱能够伴随业务查询自适应动态扩展。 Miller援引的一项客户概念验证（POC）数据显示，两位业务专家在单日内对系统识别出的400个候选概念进行了快速筛选，主动排除了200个可能降低图谱遍历效率的噪音概念，并将剩余200个有效关系正式载入图谱。仅经过这一轮人工协同梳理，系统的检索准确率便从单纯依赖向量检索时的70%基线大幅提升至88%，实现了从通用搜索向驱动业务动作的高精度知识发现转变。 异构数据融合与可量化的降本实效 (8:01) 在实际业务落地中，上下文数据平台展现出清晰的财务与效率回报。在一家金融服务机构的实践中，4000名分析师借助知识图谱和结构化内容编写整合研报，每人每周节省1.5小时，每年直接节约3000万美元成本；进一步引入生成式AI后，分析师每周额外节省4小时，将复杂的税务与法规影响分析编写工作从原先的4小时缩短至15分钟。另有企业客户在获取该参考架构后仅用3周便交付Alpha测试版本，随后顺利拓展至全球超过2万名员工。 在一项针对执法调查场景的技术演示中，该架构融合了双时态数据（有效时间与事务时间）、地理空间坐标、现场报告与911调度音频转录文本。通过对多源非结构化数据进行关联分析，系统成功识别出分散在不同警情中的微小线索（如嫌疑人身着的“紫色毛衣”），在地图与时间轴上清晰还原了犯罪行为的时空轨迹，同时为执法人员保留了点击下钻至原始音频记录的完整证据链。 代理化RAG与专有知识竞争护城河 (16:05) 通过对代理化RAG平台Nuclia的战略收购，Progress Software进一步扩展了自主代理编排能力。Nuclia的自主代理能够对企业内部数据进行智能分类、深度解析视频并处理复杂非结构化输入流，使企业无需自建底层AI基础设施即可运行先进的RAG管道，将人类在特定领域的推理规则直接注入自主代理的自动化执行循环中。 Miller指出，单纯依靠算法创新的速度无法为企业建立长久优势，真正的竞争壁垒在于将企业内部沉淀的专有语境、内部逻辑与专有智能系统化地赋予机器。通用的基础大模型无法自动复制这些独家资产；企业通过保持人机协同架构，能够为自动化系统补足常识判断与机构问责能力，从而在通用技术普及浪潮中构筑坚固的防御性护城河。 Miller告诫企业，在推进技术试点或14天免费试用前，必须首先明确预期的具体商业成果、梳理相关内部数据资产并规划API集成路径，避免为了技术新奇性而开展缺乏价值的AI实验；然而，依赖领域专家介入维护图谱的机制是否能在数据异构且专家资源匮乏的中小企业中同样低成本推广，仍有待实践检验。

Detailed notes — zh

## Background 该内容由行业峰会频道 Ai4 于 2025 年 8 月 29 日发布在 YouTube 平台，主讲人为 Progress Software 高级产品营销经理 Philip Miller。视频简介指出，本场演讲面向企业管理者与数据战略家，重点阐述如何通过 Progress Software 的检索增强生成（RAG）解决方案将生成式 AI 直接连接至企业专有数据，打破数据孤岛并实现即时、可信的高价值决策自动化。 ## TL;DR 演讲直面生成式 AI 在企业落地过程中遭遇的高幻觉率、数据偏差、缺乏业务上下文及黑盒不可解释等核心痛点，提出以业务成效为驱动的治理路径。Progress Software 依托深厚的多模态数据平台与企业知识图谱，通过创新的“渐进式图谱”（Progressive Graphs）与人机协同（Human-in-the-loop）机制，成功将检索准确率由纯向量方案的 70% 提升至 88%。通过展示金融报告自动化与多模态犯罪情报分析等落地案例，结合近期收购 Nuclia 进军 Agentic RAG 的战略举措，主讲人强调唯有将企业专有知识与 AI 深度融合，才能构建起不可替代的长期商业护城河。 ## Atmosphere & style 整场分享风格务实严谨、直切商业痛点，主讲人完全摒弃了浮夸的学术噱头与代码堆砌，展现出极强的企业级解决方案专家气场。从直指原生大模型“如理论物理学家做心脏手术般不可靠”的生动警示，到层层拆解落地架构、量化节省数千万美元的客户案例与直观的情报侦破演示，叙事节奏紧凑而极富说服力。整体观感专业而沉着，为身处 AI 落地焦虑中的企业决策者传递出一种安全、可控且高度笃定的技术落地信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=AxYew0fzuiA&t=4s) 企业级 AI 落地的关键在于追求可衡量的业务成效与价值交付，而非盲目停留在技术实验或概念演示阶段。 - [00:04](https://www.youtube.com/watch?v=AxYew0fzuiA&t=4s) Progress 数据平台融合了多模态数据摄取（涵盖图、向量、地理空间与双时态数据）与知识图谱管理，将员工隐性经验转化为机器可理解的业务上下文。 - [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 开箱即用的原生生成式 AI 存在严重幻觉与低准确率（实测部分仅 40%），且模型静态权重与黑盒属性难以适应企业专有业务与强监管合规需求。 - [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 随着自主 AI Agent 的普及，自动化带来的错误放大风险剧增，必须在系统层面对其施加严格的上下文约束与安全治理。 - [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) 某大型金融机构通过知识图谱与 RAG 方案，将复杂税务政策分析从 4 小时大幅缩减至 15 分钟，实现每年数千万美元的效率收益。 - [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) 借助低门槛、易上手的 UI 工具，企业客户仅耗时 3 周便完成知识发现系统 Alpha 版上线，并顺利推广至全球逾 20,000 名员工。 - [10:06](https://www.youtube.com/watch?v=AxYew0fzuiA&t=606s) 端到端企业 RAG 架构集成了意图解析、多模态检索（相关性、向量与知识图谱）及细粒度权限控制，在请求送达 LLM 前完成安全过滤与上下文增强。 - [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) 句子或段落级别的精准检索能有效降低 Token 消耗与推理延迟，并通过文档来源注释和知识图谱双重验证防止 PII 泄露与虚假生成。 - [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) “渐进式图谱”（Progressive Graphs）实现了人机协同演进，通过领域专家审核新概念，将检索准确率从纯向量方案的 70% 跃升至 88%。 - [14:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=871s) 犯罪情报分析演示证明，结合双时态、地理空间及语音文本等多模态数据，AI 能够跨时空挖掘隐蔽行为模式并提供可执行的决策支持。 - [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) Progress 收购 Nuclia 进军 Agentic RAG，增强了利用专门 AI Agent 解析与分类视频等复杂非结构化数据的能力。 - [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) 单纯的模型技术无法构成竞争壁垒，将企业私有专有知识与人机协同（Human-in-the-loop）结合并借助 AI 规模化放大，才是企业的核心护城河。 - [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) 企业推进 AI 转型应以具体用例与专有数据准备为起点，采取小步快跑策略，借助 API 与试用机制快速验证商业回报。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=AxYew0fzuiA&t=4s) 驱动业务成效：Progress Software 的企业级与 Agentic RAG 布局 演讲开篇即明确了本次分享的重点并非晦涩的技术底层细节，而是紧紧围绕企业管理者和数据战略家最为关心的“业务成效”与“价值交付”。核心命题在于如何将企业内部的 AI 项目从早期的实验研究阶段平稳拓展并落地到实际生产环境中。Progress Software 具备贯穿整个企业架构的软件栈，从底层的基础设施管理、中层的数据与信息管理，一直延伸到前端的用户交互界面（UI）。 本次演讲的核心聚焦于检索增强生成（RAG）技术。Progress 提供了成熟的企业级 RAG 方案，并通过最新收购的 Agentic RAG（智能体 RAG）方案进一步完善了产品矩阵。这一组合让客户既能满足单用户或部门级轻量 AI 应用的需求，又能无缝扩展至覆盖数千乃至数十万用户的超大规模企业级场景。 底层支撑这些方案的是 Progress 数据平台（Progress Data Platform），它历经约 20 年的技术积淀，主要包含两大核心能力： 1. **多模态数据平台（Multi-model Data Platform）**：可摄取并统一搜索任何形式的数据，无论是结构化、非结构化数据，还是地理空间（Geospatial）、双时态（Bitemporal）、图数据（Graph）与向量数据（Vector）。平台能够对这些数据进行清洗、协调与管理，精准交付给下游的 AI 应用、交易系统或商业智能（BI）分析工具。 2. **元数据与知识管理系统**：用于构建企业知识图谱（Knowledge Graphs）。知识图谱的作用在于将企业专属的业务语言和员工头脑中的隐性知识（Implicit Knowledge）形式化，使机器能够在系统层面理解并以规模化方式应用这些业务背景。 --- ### [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 生成式 AI 带来的严峻挑战与现实落地障碍 生成式 AI 并非“将要”改变世界，而是“已经”彻底改变了世界。2022 年那场悄然引爆的技术变革席卷全球，如今企业必须直面随之而来的各种意料之外的挑战。演讲者强调，这些问题并非出自纸上谈兵的学术研究，而是 Progress 在帮助客户部署生成式 AI 应用时遇到的真实痛点： - **严重幻觉与低准确率**：在开箱即用的原生状态下，部分客户部署的生成式 AI 方案准确回答率仅有 40%，这种水平根本无法在企业内规模化推广。随着自主运行的 AI Agent（智能体）兴起，风险进一步放大——人类员工一次只能犯一个错误，而自动化 Agent 在同样的时间里可能会连续犯下 10,000 次错误，因此对其施加严格管控更为迫切。 - **数据偏差（Data Biases）**：缺乏上下文关联的数据容易引入未知偏差。通过将数据关联并置于具体行业、监管规则与业务标准的语境中，才能精准识别并修正这些潜在偏差。 - **缺乏人类推理能力与专有领域认知**：即使是具备推理能力的大模型，在面对企业内部专有数据时也极易产生混乱。这就像要求一位顶尖的“理论物理学博士去主持开胸心脏手术”——尽管他智商极高，但没有专门的临床知识，谁也不敢把生命托付给他。 - **知识截止与静态权重（Data Cut-offs）**：大模型训练完成后参数权重即固定，无法自主获取最新知识；而企业每天都在推出新方案、新模型和新业务，这些增量信息无法被静态模型感知。 - **缺乏可解释性（Lack of Explainability）**：AI 本质上是一个黑盒，而企业原有的复杂数据生态往往也是黑盒（许多企业说不清某项数据为何用于特定场景）。在受到严格监管的行业中，平台必须能够清晰解释数据流转及决策生成的完整因果链条。 - **缺乏确定性与鲁棒性（Lack of Robustness）**：企业需要一个非确定性（Non-deterministic）的 AI 工具给出绝对确定、一致且可复现的结果。这好比“掷骰子却必须每次都精确掷出 6 点”一样充满挑战。 这些技术瓶颈直接阻碍了终端用户建立对 AI 输出结果的信任，进而导致推广阻力重重。 --- ### [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) 企业 AI 成功的核心要素与客户落地成效 企业级 AI 若要成功跨越试验期并在复杂或强监管行业规模化落地，核心问题在于：“如何将企业的业务语境、领域知识以及人类专家的隐性经验精准转化为大模型能理解的上下文？”为此，方案必须同时满足以下关键属性：准确性（Accuracy）、专业深度（Expertise）、业务适应性（Adaptability）、合规性（Compliance）、时效性（Timeliness）以及优秀的用户体验（UX/UI）。如果做不到每次都返回正确答案，企业就不得不反复安排人工复核纠错，规模化便无从谈起。 Progress 数据平台与 RAG 方案在多个行业客户中展现了显著的量化价值： - **某大型金融服务机构（内容创作提效）**：在生成式 AI 普及之前，该机构就采用 Progress 数据平台构建知识图谱来沉淀并关联数据。他们开发了结构化内容编写方案，为 4,000 名撰写行业报告的研究员每周每人节省 1.5 小时，每年直接节省 3,000 万美元。 - **复杂税务与法规研报生成**：基于前期的数据基础，该金融机构进一步叠加了生成式 AI 方案。原本分析新出台的税法和复杂监管政策对业务的影响需要耗费研究人员整整 4 个小时，如今在 AI 辅助下仅需 15 分钟即可完成，每周为每位员工额外节省 4 个小时。 - **20,000+ 员工规模的企业级知识发现系统**：另一家客户在获取 Progress 提供的参考架构后，依托直观易用的 UI 驱动工具，仅用 3 周时间就自主研发并上线了 Alpha 版本。该知识管理与智能发现工具目前已稳定服务于其全球 20,000 多名员工。 这些成功案例表明，企业应采取小步快跑、迭代推进的策略，先通过轻量级试点验证数据平台的支撑力与 AI 准确性，再逐步推广至全局大规模落地。 --- ### [10:06](https://www.youtube.com/watch?v=AxYew0fzuiA&t=606s) 架构解析：从用户查询到受控生成的端到端数据流 为了直观呈现技术运转机制，演讲者用最精炼的逻辑梳理了查询处理流程：从用户或自主 Agent 发起提问，到多模态检索与安全治理，再到最终将高度相关的上下文注入大模型。 ```mermaid flowchart TD accTitle: Enterprise RAG Architecture Flow subgraph InputStage ["1. 请求输入与意图解析"] A["用户 / AI Agent 提出查询"] --> B["系统意图解析与检索分发"] end subgraph RetrievalStage ["2. 多模态检索与上下文增强"] B --> C1["相关性排序检索 (Rank Relevance)"] B --> C2["向量检索 (Vector Search)"] B --> C3["知识图谱检索 (Graph Search)"] C1 & C2 & C3 --> D["聚合业务数据与上下文信息"] end subgraph GovernanceStage ["3. 安全治理与大模型生成"] D --> E["细粒度安全与权限控制<br/>(防止专有敏感数据泄露)"] E --> F["输入带有精准上下文的提示词"] F --> G["企业级 LLM 生成高可信结果"] end ``` 该流程确保在提问抵达大语言模型之前，就已经经过了精准上下文的补充与企业级安全合规策略的严格过滤，从而让大模型能够在完全理解业务边界的前提下输出高度精准的答案。 ### [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) 精准检索、数据治理与 Progressive Graphs 渐进式图谱 将包含上千页文档的庞大数据一股脑丢给 AI 进行推理不仅会消耗海量的 Token，还会导致模型响应速度大幅变慢。Progress Software 的解决方案聚焦于在句子或段落级别精准定位关键内容，仅将最核心的信息提供给大模型，在降低推理成本与延迟的同时，实现对用户 Prompt 的高度定制化。更关键的是在输出端实施严格的数据治理，防止专有数据或 PII（个人身份信息）发生幻觉泄露。系统通过知识图谱对生成结果进行二次验证，确保事实在图谱中存在，并自动为答案附带精确的文档来源注释，使用户或 AI Agent 可以一眼锁定答案所在的具体段落或句子。 在底层架构上，系统能够直接在数据源头进行索引，支持云端和本地（On-Premise）等多种部署形态，无需将全部企业数据整体迁移到单一库中。通过提取元数据与文档内深层信息来丰富数据上下文，结合知识图谱与向量检索（Vector Search）构建规则化、可扩展的 RAG 体系。该架构的核心亮点在于名为“渐进式图谱”（Progressive Graphs）的专有特性：当用户的查询涉及到图谱中尚不存在的新概念时，系统不会直接报错，而是会自动向领域专家（SME）发送通知；专家通过 UI 界面审核确认后，新概念即可自动并入知识图谱，实现“边搜索边学习”的人机协同进化。 ```mermaid flowchart TD accTitle: Progressive Graphs RAG Workflow subgraph QueryStage["查询与检索阶段"] U["用户/Agent 查询"] --> RAG["增强检索 (Vectors + KG)"] RAG --> CCheck{"概念是否存在于图谱？"} end subgraph SMEStage["渐进式图谱学习"] CCheck -- "否" --> SME["通知领域专家 (SME)"] SME --> UI["UI 界面审核/筛选概念"] UI --> Update["自动更新知识图谱"] Update -.-> RAG end subgraph OutputStage["推理与治理阶段"] CCheck -- "是" --> LLM["LLM 精准推理 (仅传递核心上下文)"] LLM --> Gov["输出治理与合规验证 (防 PII / 幻觉)"] Gov --> Result["带段落溯源注释的事实答案"] end ``` - **渐进式图谱 POC 实测数据**：在仅投入 2 名领域专家工作 1 天的 POC 测试中，系统通过程序化方式自动识别出 400 个新概念；专家剔除了可能影响图谱性能或不相关的 200 个，将剩余 200 个并入知识图谱。 - **准确率飞跃**：仅通过一次 Progressive Graphs 的迭代处理，模型检索准确率就从纯向量检索（Vector-only）方案的约 70% 显著提升至 88%。 - **核心定位**：这绝非普通的关键词搜索，而是面向业务决策的“可落地知识发现”（Applicable Knowledge Discovery），将企业数据真正转化为具有行动力的洞察。 --- ### [14:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=871s) 犯罪情报分析实战演示：多模态数据与行为模式挖掘 在一段现场展示的实操 Demo 中，系统展现了如何超越传统的类似 Google 的相关性排名检索，协助调查人员侦破复杂的犯罪案件。面对海量的异构情报数据，调查人员直接使用自然语言向生成式 AI 提问。系统整合了双时态视图（Bi-temporal Views）、地理空间信息（Geospatial Information）、外勤人员实地报告以及 911 报警电话录音文本等多元数据源，在多层数据中快速发现隐藏的行为模式。 整个分析过程以直观的图形化界面呈现关联网络，同时在下方清晰列出底层原始数据来源，方便调查人员随时对 911 报警记录等原始证据进行人工核验。在演示案例中，系统根据零散线索（包括目击记录中提及的特定“紫色毛衣/套头衫”特征）串联起不同时空的案件细节，精准识别出同一嫌疑人的一致性行为模式，为执法人员提供了可以直接采取行动的破案依据。 --- ### [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) 收购 Nuclia 进军 Agentic RAG 与构建企业竞争壁垒 为了进一步拓展多智能体能力，Progress Software 正式宣布收购了专注于 Agent 驱动型 RAG 解决方案的公司 Nuclia。双方在前期已共同进行了数月的产品深度集成与合作。Nuclia 的技术范式能够指派专门的 AI Agent 负责数据分类与多维度解析，尤其擅长处理视频等复杂非结构化数据。这一收购让从初创团队到大型跨国企业的各类型组织，都能以机器级的规模（Machine Scale）无缝融入人类专业智慧，全面落地 Agentic RAG。 将生成式 AI 工具与企业内部私有的上下文知识结合，是构建商业护城河的关键所在。单纯追求大模型技术本身并不能形成长期壁垒，唯有将企业多年积累的专有知识通过 AI 进行规模化放大，才能形成不可替代的竞争优势。而“人机协同”（Human-in-the-loop）机制正是确保这一系统具备业务上下文感知、常识判断与合规问责制（Accountability）的核心基石。 > “速度与创新、企业内部沉淀的上下文知识与智慧，并将它们应用于机器，这才是你的竞争护城河……通过生成式 AI 规模化放大企业已有知识，以此拓展企业的防护护城河。” --- ### [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) 14 天免费试用与从业务价值出发的落地建议 在演讲尾声，主讲人呼吁企业管理者不要仅仅为了“做实验”或“好玩炫技”而盲目投入 AI，市场上已经充斥了太多华而不实的演示。企业应将重心放在实际业务成果与商业价值的交付上。他建议企业首先梳理与特定业务场景紧密相关的专有数据，准备好所需连接的 API 接口，然后通过 Progress Software 提供的 14 天免费试用或现场预约 Demo（Ai4 大会 503 展位）开启转型之旅，由技术专家团队协助指导如何针对具体用例对 Agent 进行高效 Prompt 编排与系统对接。 - **快速启动路径**：现场扫描二维码即可开通 14 天免费试用，支持全 API 驱动接入。 - **展位互动与咨询**：主讲人及团队在 Ai4 现场的 503 号展位接待参会者，针对企业的具体用例与数据架构提供一对一的 AI 投资价值变现咨询。 ## Notable quotes > [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 生成式 AI 并非“将要”改变世界，而是“已经”彻底改变了世界。 — Philip Miller > [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 人类员工一次只能犯一个错误，而自动化 Agent 在同样的时间里可能会连续犯下 10,000 次错误，因此对其施加严格管控更为迫切。 — Philip Miller > [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 这就像要求一位顶尖的“理论物理学博士去主持开胸心脏手术”——尽管他智商极高，但没有专门的临床知识，谁也不敢把生命托付给他。 — Philip Miller > [02:31](https://www.youtube.com/watch?v=AxYew0fzuiA&t=151s) 企业需要一个非确定性（Non-deterministic）的 AI 工具给出绝对确定、一致且可复现的结果。这好比“掷骰子却必须每次都精确掷出 6 点”一样充满挑战。 — Philip Miller > [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) 如何将企业的业务语境、领域知识以及人类专家的隐性经验精准转化为大模型能理解的上下文？ — Philip Miller > [06:34](https://www.youtube.com/watch?v=AxYew0fzuiA&t=394s) 如果做不到每次都返回正确答案，企业就不得不反复安排人工复核纠错，规模化便无从谈起。 — Philip Miller > [10:59](https://www.youtube.com/watch?v=AxYew0fzuiA&t=659s) 这绝非普通的关键词搜索，而是面向业务决策的“可落地知识发现”（Applicable Knowledge Discovery），将企业数据真正转化为具有行动力的洞察。 — Philip Miller > [16:05](https://www.youtube.com/watch?v=AxYew0fzuiA&t=965s) 速度与创新、企业内部沉淀的上下文知识与智慧，并将它们应用于机器，这才是你的竞争护城河……通过生成式 AI 规模化放大企业已有知识，以此拓展企业的防护护城河。 — Philip Miller > [18:03](https://www.youtube.com/watch?v=AxYew0fzuiA&t=1083s) 不要仅仅为了“做实验”或“好玩炫技”而盲目投入 AI，市场上已经充斥了太多华而不实的演示。企业应将重心放在实际业务成果与商业价值的交付上。 — Philip Miller ## Mentioned Philip Miller, Progress Software, Nuclia, Ai4, Progress Data Platform, Retrieval-Augmented Generation (RAG), Agentic RAG, Progressive Graphs, Knowledge Graphs (知识图谱), LLM (大语言模型) ## Video description Philip Miller, Product Marketing Manager, Senior, Progress Software Abstracts What if you could ask your business’s data any question and get an instant, trusted answer? In this session, I’ll show how Progress Software’s Retrieval-Augmented Generation (RAG) solution can make that scenario a reality by connecting generative AI directly to your proprietary data. By seamlessly linking AI with your enterprise knowledge, this solution turns siloed information into tailored, high-value AI outputs for your organization. I’ll highlight how this approach boosts intelligent data access, enables decision automation, improves operational efficiency and elevates customer experience. From finance to healthcare, organizations across industries are already turning their internal data into a competitive edge with RAG – unlocking new insights and automating decisions that drive innovation. Best of all, no engineering background is required – this session is built for business leaders and data strategists who are ready to unlock AI’s strategic value. [Source](https://www.youtube.com/watch?v=AxYew0fzuiA)

Source quotes

[2:31] Philip Miller: 生成式 AI 并非“将要”改变世界，而是“已经”彻底改变了世界。 [2:31] Philip Miller: 人类员工一次只能犯一个错误，而自动化 Agent 在同样的时间里可能会连续犯下 10,000 次错误，因此对其施加严格管控更为迫切。 [2:31] Philip Miller: 这就像要求一位顶尖的“理论物理学博士去主持开胸心脏手术”——尽管他智商极高，但没有专门的临床知识，谁也不敢把生命托付给他。 [2:31] Philip Miller: 企业需要一个非确定性（Non-deterministic）的 AI 工具给出绝对确定、一致且可复现的结果。这好比“掷骰子却必须每次都精确掷出 6 点”一样充满挑战。 [6:34] Philip Miller: 如何将企业的业务语境、领域知识以及人类专家的隐性经验精准转化为大模型能理解的上下文？ [6:34] Philip Miller: 如果做不到每次都返回正确答案，企业就不得不反复安排人工复核纠错，规模化便无从谈起。 [10:59] Philip Miller: 这绝非普通的关键词搜索，而是面向业务决策的“可落地知识发现”（Applicable Knowledge Discovery），将企业数据真正转化为具有行动力的洞察。 [16:05] Philip Miller: 速度与创新、企业内部沉淀的上下文知识与智慧，并将它们应用于机器，这才是你的竞争护城河……通过生成式 AI 规模化放大企业已有知识，以此拓展企业的防护护城河。 [18:03] Philip Miller: 不要仅仅为了“做实验”或“好玩炫技”而盲目投入 AI，市场上已经充斥了太多华而不实的演示。企业应将重心放在实际业务成果与商业价值的交付上。

Source quotes

[2:31] Philip Miller: This bomb went off in 2022 and no one heard it and it went around the world [2:31] Philip Miller: the human being can make one mistake [2:31] Philip Miller: your AI agents can make 10,000 mistakes in that time. [2:31] Philip Miller: wanting to roll a dice and it always land on six. [10:59] Philip Miller: just what matters [16:05] Philip Miller: a one-man band [17:13] Philip Miller: over like a coffee or a beer [17:13] Philip Miller: Speed and of innovation, the contextual knowledge and the intelligence that is in your business and applying that to the machine... that becomes your competitive moat... Let's scale that with the generative AI and let's extend the moat around your business. [18:03] Philip Miller: a three-hour conversation into roughly 20 minutes [18:03] Philip Miller: I don't want to do this for the sake of an experiment or kind of this is fun, look what I can do... what we want for business is to deliver value from the AI that you've got today.

Original transcript

[0:04] I'm going to be talking about AI, but [0:06] I'm not going to dive into the tech or [0:08] try not to dive into the tech. I'm going [0:10] to try and talk to you about the [0:12] outcomes that matters the most to [0:14] people, right? The value that AI can [0:17] deliver in the business and how to kind [0:20] of scale your AI from production or [0:22] sorry research into production. [0:25] I'd be remiss if I didn't talk about the [0:27] pro progress for the company where I'm [0:30] from. Um, we have an entire stack of [0:35] infrastructure software that goes right [0:38] from the the infrastructure management [0:40] layer right through the data and [0:43] information layer in your business right [0:45] through to the UI layer. Today I'm going [0:47] to be talking to you about uh rag. I'm [0:50] going to be talking to you about our [0:53] enterprise solution as well as the new [0:55] agentic rag solution that we acquired uh [0:58] recently that fits really nicely within [1:00] our portfolio and allows people to go [1:02] from kind of those departmental level [1:05] AIs kind of single users or multiple [1:08] users right through to then our solution [1:10] in that space which scales out to uh to [1:13] kind of thousands of people hundreds of [1:15] thousands of people that are using this [1:17] technology today. [1:19] This is the progress data platform. It's [1:21] made up of two fundamental technologies. [1:23] One is a multimodel data platform. That [1:26] essentially means we can have any form [1:28] of data that you want to ingest in [1:30] there. Whether that's structured or [1:32] unstructured data, geospatial, bmporal. [1:35] Sorry, I did say I wouldn't get [1:37] technical, but um we've got graph data, [1:40] we have vector data. We're able to [1:42] search all of that data in the platform. [1:44] We're able to harmonize and curate that [1:47] data so we can pass it downstream to [1:50] your AI applications or any other [1:52] transaction or BI applications that you [1:54] might have in your business. It's all [1:57] underpinned by various different [1:59] functions within the data platform that [2:01] have been built over roughly 20 years. [2:04] Right? It's a combination of that with [2:07] our metadata management and knowledge [2:09] management tool. It helps create [2:11] knowledge graphs. Knowledge graphs for [2:14] anybody who doesn't know is a way of [2:16] expressing the language that's in your [2:18] business to the machine. It's a way to [2:22] take the implicit knowledge that the [2:23] human beings have within your [2:25] organization and apply it at a machine [2:28] scale. Right? [2:31] The reason is is because Genai isn't [2:35] going to change the world. It has [2:38] changed the world. Right? I get a lot of [2:40] people I hear a lot of talks and it's oh [2:42] it's going to do this and it's going to [2:43] do that. It already has. This bomb went [2:47] off in 2022 and no one heard it and it [2:51] went around the world, right? And we're [2:54] dealing with the unintended consequences [2:57] in some cases, right? Where businesses [2:59] are now trying to leverage AI in their [3:02] organizations. [3:04] The problem with that is it comes with a [3:07] number of different challenges. Now, [3:09] these aren't challenges that I kind of [3:11] think of or they're kind of research. [3:13] These are from our customers. These are [3:14] real world challenges that we've [3:16] experienced when trying to deploy [3:18] generative AI solutions and applications [3:21] in our customers. So, we've got the [3:23] hallucinations, the big one. You know, [3:25] out of the box, we've had customers who [3:28] have deployed a generative AI solution [3:30] and it is only answering the question [3:32] 40% of the time correctly. You cannot [3:35] possibly scale that in your [3:36] organization. And this includes agentic [3:39] AI, right? So that they're one and the [3:42] same, right? You need to have the same [3:44] controls as your generative AI solution, [3:47] even more so with your agents, right? [3:49] Because these are autonomous. The human [3:51] being can make one mistake. Your AI [3:54] agents can make 10,000 mistakes in that [3:57] time, right? So then we've got data [4:00] biases, right? We bring this data [4:02] together and we connect and [4:04] contextualize the data because it helps [4:06] us understand our data. It helps us [4:09] understand the data in the frame of your [4:11] business in the domain that you operate [4:14] in, the regulations, the standards that [4:16] might be applicable to your data and [4:18] helps you identify those data biases, [4:21] the the unintended biases that are [4:24] within your data. [4:26] There's a lack of human reasoning with [4:28] generative AI tools at the moment, [4:30] right? Even though we've got reasoning [4:32] models, they sometimes get confused. [4:34] Now, often that's because it's [4:35] proprietary data that it's dealing with. [4:38] It is like taking a PhD level, let's [4:41] say, physicist and say, "Do open heart [4:44] surgery. I don't want that person doing [4:46] open heart surgery on me, right? I know [4:48] he's very bright, but still I don't want [4:50] that happening." There's data cut offs. [4:53] So at the moment large language models [4:56] are trained to a certain point and then [4:59] that data is static. Those weights and [5:04] models are the same, right? Um and as a [5:07] result it can't learn. Now as a business [5:11] you are learning all the time. You're [5:12] deploying new solutions, new models, new [5:16] applications in your business. But that [5:18] won't be in the generative AI's model. [5:21] There's a lack of explanability, right? [5:24] Why did the AI make this decision? Now, [5:26] AI is a black box, but so is most of the [5:30] enterprise data ecosystem. Why did you [5:33] do this with this data? Why is this data [5:36] used in this use case? Well, lots of [5:39] people can't explain that because it's [5:40] so disconnected and there's no context [5:43] around that data within our platform. [5:45] It's used in regulatory spaces because [5:48] we can explain what happens to the data [5:50] and why. And finally, there's a lack of [5:54] robustness. [5:55] I need the generative AI tool, this [5:59] non-deterministic tool to give me a [6:02] determined result, which is insane, [6:05] right? Like I want to roll a dice and it [6:08] always land on six. Well, that dice [6:11] might round on five or four or three or [6:14] whatever, right? And so we need those [6:17] repeatable results within the system. So [6:20] these are just some of the challenges [6:22] that we have. There's human challenges [6:24] as well, by the way. So getting people [6:26] to use these things and stuff like that, [6:28] but they won't do that until they can [6:30] trust the output from the system. [6:34] the problem statement really and there's [6:37] loads of different questions that you [6:38] could put up there but it's how do I get [6:40] the context of my business my [6:42] organization the domain that I work in [6:45] the implicit knowledge that I have about [6:48] the data that's in my system to scale to [6:52] get that human expertise to scale to the [6:55] generative AI and so [6:59] this is what the enterprise needs it [7:02] needs accuracy [7:04] expertise, [7:05] adaptability, right? It needs to be able [7:09] to change and evolve as the business [7:11] does that. It needs the compliance, the [7:14] timeliness and then to provide these new [7:17] UX and UI experiences to people within [7:20] the business. It needs to build the [7:23] value that people are looking for from [7:25] their investments today in AI. If you [7:28] can't get these things sorted, if you [7:31] can't tick these boxes, it's very [7:33] difficult to scale the solution, [7:35] especially in enterprise or very complex [7:39] or regulatory spaces, right? It's very [7:42] difficult to scale those solutions [7:43] because you're constantly having to get [7:45] that human back in there and say, "Oh, [7:47] no, wait a minute. This is wrong. This [7:49] is wrong. This is wrong." Right? Repeat [7:51] that. Try and do it again. Right? What [7:53] we want is that answer to come back and [7:55] be correct every time. [7:59] So these are some of the results. I'm [8:01] not going to go through all of them. Um [8:02] I'll just give you a couple of examples. [8:04] So I've got a financial services [8:07] organization that took our data platform [8:10] approach before generative AI. They took [8:13] our data platform approach connecting [8:15] and contextualizing that data with those [8:17] knowledge graphs and they built a [8:20] solution where they were able to um save [8:24] their researchers who are writing [8:27] reports for their business in a [8:29] structured authoring content solution. [8:31] They were able to save them an hour and [8:32] a half a week. They deployed that to [8:35] 4,000 people and save $30 million a [8:38] year. [8:39] They're now building a generative AI [8:42] solution on top of that where they're [8:44] saving four hours a week with those same [8:47] people where they're taking very complex [8:50] tax and legislation reports uh about new [8:54] regulations and things that are coming [8:55] in and how it's going to affect the [8:56] business. They normally take four hours. [8:59] They're able to do this in about 15 [9:01] minutes, right? So, they're they're [9:03] realizing these savings today. We had [9:05] one customer again they were using our [9:08] data platform approach and they were [9:10] very happy with it and we came along and [9:12] we gave them the referential [9:13] architecture of our solution. We said, [9:16] "Here's the concept." Right? Now, they [9:19] were really good at using our tech, [9:21] right? But because it's all driven by [9:23] UI, right? So, but we came back to them [9:26] 3 weeks later thinking, well, you know, [9:28] we'll we'll see what they think. And [9:30] they'd already released an alpha version [9:33] of the product and now that knowledge [9:35] management tool, that that discovery [9:37] tool in their business works for the [9:40] 20,000 plus employees worldwide now, [9:42] right? And the list goes on. the the the [9:46] idea is to get you to iteratively build [9:49] this data platform, this AI enabled data [9:53] platform so that you can provide data [9:56] and improve the accuracy as you go [9:59] through these smaller pilots so that it [10:01] gets to the stage where you can roll [10:02] these out to larger implementations. [10:06] This is how we do it. It's really like [10:08] this is the simplest way I can put this [10:10] up without diving into the tech, but [10:12] I'll just walk you through it very [10:13] quickly. [10:14] The user makes a query and or agent by [10:17] the way because agents will make queries [10:19] of your data as well. Um it's then [10:21] interpreted by the system. The system [10:24] looks for a rank relevance search, a [10:26] vector search, a graph search, [10:28] essentially looking at data and the [10:30] information around that data, the [10:32] context of that data and augments that [10:34] to the users query. It then governs it [10:38] using the fine security controls that we [10:41] have within the platform before sending [10:42] it to the AI making sure that it's not [10:45] sharing proprietary information you [10:47] don't want to leave your business. [10:50] We then send it to the AI. Now the AI [10:52] has the context of which your business [10:54] operates in and the and the the [10:56] information that is very relevant. And [10:57] when I say very relevant, I mean down to [10:59] the sentence or paragraph level within a [11:02] document. It's great sending the whole [11:04] document, a thousand pages to the AI. [11:06] That kind of means that the thousand [11:08] pages is going to use a hell of a lot of [11:10] tokens to kind of inference. It's going [11:12] to slow down the response. Whereas this [11:15] way, we give it just what matters. So, [11:18] we're saving you that inference time, [11:19] that token cost, but we're also making [11:22] sure it's hyper realized to to the exact [11:25] prompt that the user has. When it comes [11:27] out, we govern it again. We make sure it [11:30] didn't hallucinate PII information or [11:32] your proprietary data for whatever [11:33] reason, right? We we ensure that we can [11:35] share the answer with the user as well [11:38] as the the the agent. Um we validate it [11:41] again. We look with inside that data for [11:43] the context. We go to the the knowledge [11:45] graph and say does this exist? Yes, it [11:48] does. Okay, here's the answer. And then [11:51] we annotate that. So we we make sure [11:54] that it gets that answer, but you can [11:56] very quickly go, "Oh, there's the [11:58] answer. It's right on this paragraph or [12:00] it's in this sentence [12:03] very quickly. All we're doing is we're [12:06] connecting your data into a single [12:08] repository. We don't need to pull all of [12:10] your data in. We can index at source." [12:12] So there's loads of different ways of [12:13] deploying our solution, whether it's in [12:15] cloud or on prem. [12:18] We then enrich and contextualize your [12:21] data. We put the metadata about your [12:23] data. Not just the kind of person, [12:26] author, timestamps, all of this, but [12:28] actually the data and the information [12:30] that's inside your documentation. [12:34] We then enhance the query. We augment [12:37] that query with the data that matters to [12:39] you, that matters to your prompt, that [12:42] matters to the output that you're [12:44] looking for. [12:46] And then we retrieve those fact-based [12:48] responses. We can do all of this in the [12:51] enterprise in very complex use cases [12:54] because we put the human in the loop. [12:56] We're basically creating a rulesbased [12:58] system. It's a retrieval augmented [13:01] generation with knowledge graphs with [13:03] vectors, right? Um it allows you to be [13:07] very scalable, be very adaptable. this [13:10] system we have a proprietary feature in [13:12] there called progressive graphs where [13:14] the uh user makes a query the the the [13:17] concept doesn't exist in the graph. [13:19] Okay, no problem. We'll send a [13:22] notification to a subject matter expert [13:23] that can access the system via a [13:26] UIdriven environment can look at that [13:29] concept and say okay that should exist [13:31] in the graph and then it automatically [13:33] updates that system. It learns as the [13:35] user searches. We did a P with somebody. [13:38] We spent one one day um with two subject [13:42] matter experts. We identified 400 [13:44] concepts programmatically. [13:47] They disregarded 200 of them because [13:50] actually that would impact the [13:51] performance of the graph or they weren't [13:53] relevant or whatever, but they were able [13:54] to put 200 into their knowledge graph. [13:57] that one aspect of the the thing got it [14:01] from the vector only research which was [14:03] I think or vector only approach which [14:05] was 70%. And they got into I think it [14:09] was 88% accuracy with just one pass of [14:12] progressive graphs [14:15] and what we're doing is we're sending it [14:17] for answer or action. This isn't a [14:19] search. We're not searching for data. [14:22] We're trying to retrieve knowledge, [14:24] information, and insight within your [14:26] business to give you applicable [14:28] knowledge discovery. [14:31] Let's visualize this in a quick demo. [14:33] And I'm going to really quick. It's like [14:35] one and a half minutes, right? I'm going [14:36] to try and talk to you as much as I can. [14:38] Right? In the demo, you're going to see [14:40] that they've already done their rank [14:42] relevant search. They've retrieved the [14:43] normal results that you're going to get [14:44] from a Google like search. Um, and now [14:48] they're asking the generative AI, okay, [14:50] I want to know more about this. It's a [14:52] crime. There's some crime going on, [14:54] right? And they're trying to look for [14:56] the the the bad guy, right? And they've [14:59] got lots and lots of information. You [15:00] can see here. There's bite temporal [15:02] views. There's geospatial information. [15:04] There are reports from the field. There [15:07] is uh what's it 911 transcripts and [15:10] things like that. It's passing all of [15:12] this information through the natural [15:14] language questions that the user is [15:16] asking it for and it's identifying where [15:19] it can see patterns in that information. [15:21] And the user is then getting that [15:23] information graphically displayed as [15:25] well as then displaying the raw [15:27] information and where it came from [15:28] underneath. So they can actually go in [15:30] and check the 911 transcripts manually, [15:33] right? And they find some information. [15:35] And I think there's a purple jumper [15:36] involved because of course there is, you [15:38] know, like in San Francisco or wherever [15:40] there is purple's bad or whatever. Um, [15:42] and they're looking through it and [15:44] they're finding out that, okay, wait a [15:46] minute, this same person has committed [15:48] these crimes in this pattern of [15:50] observable behavior and then they can [15:54] take action with this. That's just one [15:57] example of like kind of a rapid P that [16:00] we did with someone just to demonstrate [16:02] the value of what can be returned. [16:05] So this is it. It's a gentic now because [16:10] we now have a new acquisition um called [16:14] nuclear. Now this is um an agentic [16:17] ragbased solution and the reason we [16:19] acquired these people is because we [16:22] worked with them for a number of months [16:24] on building out some of our solution [16:26] with them and their solution with us. We [16:29] really love their approach. Right. It's [16:31] the same approach that we have but it [16:33] agentically. So they are instructing [16:36] models to classify data to look at data [16:38] in new ways. They can take video the [16:40] same unstructured information that we [16:42] talked about earlier and they can have [16:44] agents do this and it really helps [16:47] people get started on that track of [16:49] building out a ragbased solution. [16:51] Whether you are a oneman band or whether [16:53] you are an enterprise company, right? we [16:55] now can take you on this rag journey [16:58] where you can put human intelligence at [17:00] a machine scale um across your data. Um [17:06] again it's because this is what the [17:08] business needs in order to scale these [17:10] solutions. Right? [17:13] This is I I kind of like this. I think I [17:15] came up with it. I don't know probably [17:17] over like a coffee or a beer or [17:19] something like that. I got very confused [17:20] and wrote it down and I kind of liked [17:22] it. But I think it's [17:26] it's all about this competitive [17:27] advantage, right? Using generative AI [17:29] and tools and speed absolutely is key, [17:33] right? Speed and of innovation, the [17:36] contextual knowledge and the [17:38] intelligence that that that is in your [17:40] business and applying that to the [17:42] machine. Um that becomes your [17:44] competitive moat, right? the knowledge [17:47] that you have in your business already. [17:49] Let's scale that with the generative AI [17:52] and let's extend the moat around your [17:54] business. [17:56] It's human in the loop. It provides [17:59] context, common sense, and [18:01] accountability. [18:03] If you want to get started, you can go [18:05] and book a 14-day free trial. Don't [18:08] yet. You can book that 14-day free trial [18:11] today and you can get started on the [18:12] journey. But again, I go back to you [18:16] want the outcome. You want to deliver [18:18] value and and insight into your [18:20] business. So go and gather the data that [18:22] matters to that use case. Go and get the [18:25] APIs that you need to connect because [18:26] this is all APIdriven and then go and [18:29] book your demo because these guys will [18:32] talk you through your data. They'll talk [18:34] you through the system. They'll talk you [18:36] how to prompt these agents effectively [18:39] based on your use case. because again I [18:42] don't want to do this for the sake of an [18:44] experiment or kind of this is fun look [18:47] what I can do right we've had a lot of [18:49] those already you see a lot of them [18:51] online there I've done some of them [18:53] right um what we want for business is to [18:56] deliver value from the AI that you've [18:58] got today now I've tried to fit a [19:00] three-hour conversation into roughly [19:03] what are we 20 minutes are we the 20 [19:05] minute mark are we five minutes away [19:07] what have we got one more minute right [19:10] Okay, I'm at boo booth 503. I've already [19:13] had lots of conversations uh with some [19:15] of you in the room. I've had lots of [19:17] conversations um with people that aren't [19:19] in the room. I'd love for you to come [19:20] and ask questions of me and my team. Um [19:24] if you want to get started, like I say, [19:26] you can book that free trial, take a [19:28] picture of that QR code or scan that QR [19:30] code. Um but we would love to hear from [19:32] you. We would love to talk to you about [19:34] your use case, what matters to you, and [19:36] how we can drive value from your AI [19:39] investments today. Thank you very much [19:41] for your time, and thanks very much for [19:42] coming.