IntellectAI Guarantees 40% Back-Office Efficiency Gain via Multi-Agent AI

IntellectAI承诺通过多智能体AI将后台效率提升40% IntellectAI声称，其多智能体AI系统通过以自动化证据收集和调查取代人工检索，可在三个月内为超过100名员工的企业团队减少40%以上的后台工作量。 Aditya Rajashekaran介绍了IntellectAI的Purple Fabric平台，将混乱的后台运营比作蕴藏着巨大潜在价值的青少年凌乱卧室。他用足球运动作比喻来解释多智能体协同机制，并详细分享了英国最大财富管理机构的案例：多智能体系统以超过90%的准确率将投诉处理时间从50天缩短至20分钟。该演讲重点阐述了自动化数据检索如何在全链路可追溯性和模型基准测试的支持下，帮助人类专家做出高达10倍的高杠杆决策。Rajashekaran最后承诺，可在三个月内为规模超过100名全职员工（FTE）的团队实现40%以上的人力效率优化。 IntellectAI 推出 Purple Fabric 平台，主张采用类似足球队协同分工的多智能体（multi-agent）工作流，解决金融机构中后台复杂且高度依赖人工的运营瓶颈。 (3:46) 在英国最大财富管理机构的案例中，多智能体系统成功清理数万起历史投诉，将个案调查时间由 50 天大幅缩减至 20 分钟，实现超 90% 的准确率并降低 50% 运营成本。 (5:53) Purple Fabric 支持将决策追溯至具体文档切片或数据库行，通过自动化数据整合将领域专家的高价值决策吞吐量提升至原来的 10 倍。 (13:14) 平台配备 LLM 优化中心以评估模型准确度与成本，并提供基础设施无关的部署灵活性，针对 100 人以上团队承诺在 3 个月内实现 40% 以上的工作流优化。 (19:33)

IntellectAI's Aditya Rajashekaran argues that automating complex enterprise back offices requires shifting from standalone chatbots to coordinated, multi-agent teams that integrate with legacy infrastructure while keeping human experts in supervisory roles. Coordinated Agent Teams for Complex Operations (3:46) Rajashekaran contends that back- and mid-office environments—frequently characterized by manual, document-heavy, and unstructured workflows—cannot be effectively handled by a single artificial intelligence model. Comparing operations to a soccer match, he argues that relying on an isolated chatbot against complex institutional demands is akin to putting a single goalkeeper on the pitch without support. To manage continuous operational events such as failed transactions, customer onboarding, and sanctions screening, organizations require specialized digital agents operating in defined roles. In this framework, classification agents act as defenders to triage and block incoming friction, analytical agents operate as midfielders connecting disparate systems, and execution agents function as strikers that synthesize tasks into final recommendations, all overseen by human domain managers. Resolving Regulatory Backlogs in Wealth Management (5:53) To demonstrate the multi-agent approach, Rajashekaran cites a case study from the UK's largest wealth manager. Following regulatory changes that allowed clients to retroactively submit complaints spanning a 10-year period, the institution's caseload expanded tenfold from approximately 100 cases per month to 1,000 per week. Evaluating each complaint manually required reviewing a decade of medical histories, investment records, and evolving regulatory standards, averaging 50 days per file. IntellectAI deployed a specialized agent structure comprising a complaint manager, triage classification agents, evidence gatherers querying internal systems, and investigation agents drafting decision recommendations. According to Rajashekaran, the implementation cleared tens of thousands of backlogged claims, compressed the investigation window from 50 days to 20 minutes, reduced operating costs by 50%, and achieved an accuracy rate exceeding 90%. Data Lineage and High-Leverage Human Supervision (9:13) Deploying autonomous workflows in heavily regulated sectors requires verifiable explainability rather than black-box outputs or rigid decision trees. IntellectAI's Purple Fabric platform constructs an investigation knowledge repository aggregating structured databases and unstructured inputs, including handwritten advisor notes, Snowflake, Salesforce, policy administration systems, and Financial Conduct Authority data. To ensure auditability, the system maintains strict source lineage that links every recommendation back to specific database rows, query outputs, or document segments. It also logs execution traces that reflect inter-agent communication. This restructuring eliminates legacy bottlenecks where human experts spent the majority of their time hunting for records across siloed teams; instead, domain specialists review pre-assembled dossiers and recommendations, which Rajashekaran reports increased decision throughput tenfold. Benchmarking, Governance, and Enterprise Scale (15:24) Beyond cost reduction, aggregated operational data can be re-evaluated to identify life-stage triggers, cross-selling possibilities, and customer sentiment to support retention and Net Promoter Scores. To address enterprise adoption hurdles around return on investment and legacy technology debt, Rajashekaran highlights Purple Fabric's optimization hub, which benchmarks agent accuracy, speed, and inference costs across models from Google, OpenAI, Anthropic, or proprietary providers prior to deployment. Rajashekaran states that the architecture is infrastructure-agnostic and engineered for large-scale operations, referencing a client objective to coordinate up to 300,000 active AI agents over five years under central role-based access and auditing controls. Rajashekaran pledges that IntellectAI can optimize operational workflows by over 40% within three months for organizations with more than 100 full-time employees. However, the evidence presented centers primarily on vendor-reported metrics from a single UK wealth management implementation, leaving open how seamlessly multi-agent teams adapt to organizations with different legacy software environments and regulatory burdens.

IntellectAI产品经理Aditya Rajashekaran指出，面对金融与政企中后台杂乱、重度依赖文档的手工流程，单一AI工具无法解决复杂问题，企业必须构建由专业智能体协同运作的多智能体团队，并在受监管的环境中实现降本增效与战略增值。 多智能体协同化解中后台运营复杂性 (3:46) 中后台运营长期受困于非结构化数据和繁琐的手工流程，宛如杂乱的青少年卧室，充满运营摩擦却蕴藏着待开发的潜能。Rajashekaran指出，复杂的日常运营任务（如开户、交易失败、制裁审查等）犹如球场上不断发起攻势的对手，如果仅部署单一AI聊天机器人，就像只派一名守门员上场比赛，即便其能力拔尖也无法赢得整场赛事。针对这一困境，IntellectAI依托服务超过450家全球金融机构与政府组织的经验，推出Purple Fabric平台，以协同多智能体架构应对复杂工作流。 该架构将AI团队划分为明确的分工角色：由“后卫”智能体负责阻截并分流来袭的运营摩擦，“中场”智能体连接跨部门系统并执行分析策略，“前锋”智能体推进多步骤工作流程并交付执行结果，而人类专家则作为教练与管理者全程参与关键决策。通过自然语言交互实现智能体之间及人机协同，企业能够将散乱且充满合规风险的操作转变为条理分明的自动化体系。 财富管理实践：工时与成本跨越式缩减 (5:53) 为验证多智能体体系的实际商业价值，Rajashekaran分享了英国最大财富管理机构的落地案例。受当地追溯期长达10年的监管新规影响，该机构面临历史投诉激增的严峻挑战，月均投诉量从约100件飙升至每周约1,000件。传统人工调查需跨越10年的医疗记录、投资档案、退休规划与法律变迁，单起案件耗时约50天，导致数万起案件严重积压。 IntellectAI通过部署多智能体编排方案化解了这一危机：由投诉经理智能体拦截并记录请求，分类智能体对案件定性，证据收集智能体跨数十个遗留系统汇总档案，最终由调查智能体综合提炼生成决策建议并交由主管审核。该方案将单起案件调查周期从50天压缩至20分钟，在整个多智能体系统中达成超过90%的准确率，不仅清空了历史积压案件，还使整体运营成本下降了50%。 细粒度数据溯源与专家决策生产力跃升 (9:13) 在强监管领域建立对AI决策的信任，核心在于透明的数据溯源与动态推理。该系统深入连接Salesforce、Snowflake、SSP核心系统以及英国金融行为监管局（FCA）监管数据库，直接处理结构化数据与历史手写会谈记录等非结构化材料，形成集中的“调查知识花园”。系统生成的每项建议或判断均能向下精确穿透至具体的文档片段、数据库行或执行查询语句，同时保留类似智能体间邮件往来的执行追踪链路，摒弃了僵化的硬编码决策树。 这种自动化流程彻底重构了原有的四级流转分工模式，消除了业务骨干耗费大量时间手工检索背景资料的瓶颈。在多智能体预先完成证据检索与合成后，业务专家得以将全部精力聚焦于审核案卷与执行高价值裁量。据Rajashekaran介绍，这使得人类专家在相同时间内产出良好决策的通量提升了10倍，其部署逻辑如同使用自然语言指令和访问边界来入职新型数字化员工。 大模型评测治理与战略数据价值释放 (15:24) 多智能体系统的价值不仅限于成本削减，还在于将中后台沉淀的运营数据转化为前台战略驱动力。通过在“调查知识花园”中引入专门智能体，企业能够系统分析历史账户行为，识别客户人生阶段变化、追加销售与交叉销售契机及客户情绪倾向，从而推动超个性化服务，拉动顶线营收并提升客户净推荐值（NPS）。 针对企业在技术债务、投资回报率不确定性及模型评估方面的顾虑，Purple Fabric内置了LLM优化中心（LLM Optimization Hub）。该模块支持在正式部署前对单个或成组智能体进行精度、响应速度与调用成本的全面基准评测，并兼容Google Gemini、OpenAI、Anthropic及专有模型。Rajashekaran提及，有客户规划在未来五年内运行多达30万个活跃AI智能体，这要求底层架构必须具备完善的角色访问控制和全流程审计能力。 Rajashekaran强调其方案采用完全独立于特定基础设施的部署架构，并给出承诺：对于团队规模超过100名全职员工的企业，IntellectAI可在3个月内将其运营工作流优化40%以上。然而，该优化承诺在不同技术堆栈与复杂业务场景下的实际兑现程度，仍有待更长期的行业实践检验。

## Background 本视频由 Ai4 于 2025 年 8 月 26 日在 YouTube 平台上发布，主讲人为 IntellectAI 的高级项目经理 Aditya Rajashekaran。根据页面描述，本次演讲探讨了智能多智能体工作团队如何自主协调、推理并执行企业内部的复杂工作流，展示了企业如何从对话式 AI 转向能够采取行动的运营式 AI。 ## TL;DR 演讲重点探讨了企业中后台业务运营面临的“无结构化、缓慢且复杂”等痛点，并介绍了 IntellectAI 推出的 Purple Fabric 平台。结合公司 30 多年服务顶级金融机构的深厚底蕴，主讲人阐述了多智能体协作如何帮助企业梳理繁琐的手工流程与非结构化文档。最后，主讲人还提出了强有力的商业承诺：针对 100 人以上规模的团队，可在 3 个月内实现 40% 以上的效率提升。 ## Atmosphere & style 演讲整体氛围专业且务实，主讲人采用开门见山的方式与现场观众进行积极互动，调动起热烈而沉浸的现场氛围。通过将混乱的后台业务生动比喻为“青少年的卧室”，演讲者以幽默风趣且形象的叙事风格将复杂的技术概念娓娓道来。整场 Presentation 充满了科技前沿感与商业洞察力，既有对行业痛点的深刻剖析，又展现出强烈的自信与商业说服力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) IntellectAI 推出 Purple Fabric 平台，旨在帮助企业在可信的组织上下文中构建、基准测试、部署与扩展 AI 智能体。 - [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 传统企业的中后台业务运营普遍存在缺乏结构化、处理速度缓慢且极其复杂的痛点。 - [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) IntellectAI 将中后台运营比作“青少年的卧室”——虽然混乱无序并充斥非结构化数据，但蕴含着极大的潜力和整理价值。 - [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) IntellectAI 作为拥有 30 多年历史的金融科技巨头，全球服务超过 450 家顶级金融机构与政府客户。 - 企业 AI 的演进正在从单纯的对话式 AI（Conversational AI）向能够自主执行任务的操作型 AI（Operational AI）转变。 - 多智能体协作系统（Multi-Agent Workforce）相比单智能体和传统聊天机器人，能够自主进行跨业务流程的推理与复杂工作流协调。 - [19:33](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=1173s) Purple Fabric 平台具备独立于具体基础设施的部署优势（Infrastructure-agnostic），能够适应复杂的企业 IT 环境。 - [19:33](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=1173s) 对于拥有 100 名全职员工（FTE）以上的团队，IntellectAI 承诺在 3 个月内实现 40% 以上的效率提升与流程优化。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 混乱的中后台业务与 AI 智能体团队的协作潜力 来自 IntellectAI 的 Purple Fabric 平台产品经理 Aditya Rajashekaran 开门见山地介绍了其平台的核心定位：帮助企业在可信的组织上下文中构建、基准测试、部署和扩展 AI 智能体（AI Agents）。这些智能体旨在解决隐藏在企业“后门”之下的中后台运营难题。Aditya 现场与听众互动，询问大家对中后台业务的印象，听众纷纷给出了“缺乏结构化（Unstructured）”、“速度缓慢（Lack of speed）”以及“极其复杂（Complicated）”等评价。 Aditya 将企业的中后台运营生动地比作“十几岁青少年的卧室”——混乱无序、充斥着大量非结构化文档和繁琐的手工流程，但只要你知道去哪里挖掘，其中就蕴藏着巨大的潜力。现在的关键问题在于：如何从这场“光荣的混乱”中建立秩序，让企业像青少年成长一样绽放出绚丽的光芒。 IntellectAI 并非初创公司，而是一家拥有超过 30 年历史的金融科技巨头，全球服务了超过 450 家顶级金融机构和政府客户。企业在中后台 AI 探索上的投入并非始于 GPT### 后台业务的现状与 Purple Fabric 平台的推出 Aditya Rajashekaran（IntellectAI 的 Purple Fabric 平台高级产品经理）开启了演讲。他首先指出了绝大多数企业面临的共同痛点：隐藏在“门后”的后台（Back-office）与中台（Mid-office）业务运营。现场观众在互动中用“无结构化”、“速度极慢”、“极其复杂”等词汇形容后台流程。Aditya 将企业后台形象地比喻为“十几岁青少年的房间”——混乱不堪、人来人往，但只要知道去哪里寻找，就隐藏着巨大的潜力。 IntellectAI 并非初创公司，而是一家有着 30 多年历史的金融科技巨头，全球 ### [19:33](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=1173s) 部署灵活度与 IntellectAI 的效率提升承诺 演讲者在最后总结了 IntellectAI 在部署架构上的优势，强调其解决方案在基础设施（Infrastructure）方面具备极高的灵活性与兼容性（Infrastructure-agnostic），能够适应不同企业复杂的 IT 环境与部署需求。 紧接着，演讲者面向在场拥有大型团队的企业客户给出了一个极其明确且强有力的商业价值承诺：如果企业拥有超过 100 名全职员工（FTE）的运营或业务团队，IntellectAI 能够保证在短短 3 个月内将该团队的业务流程与人力效率优化提升 40% 以上。最后，演讲者向观众表达了感谢，并邀请有兴趣深入交流或寻求合作的参会者前往 421 号展位（Booth 421）进行线下互动。 - **基础设施兼容性**：支持灵活、独立于特定基础设施的部署（Infrastructure-agnostic）。 - **优化承诺目标**：针对 100+ FTE（全职员工）规模的团队。 - **量化效益与时效**：承诺在 3 个月内实现 40% 以上的效率提升与流程优化。 - **展位信息**：IntellectAI 团队位于 421 号展位（Booth 421）。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 缺乏结构化（Unstructured） — 听众 > [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 速度缓慢（Lack of speed） — 听众 > [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 极其复杂（Complicated） — 听众 > [00:04](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=4s) 十几岁青少年的卧室 — Aditya Rajashekaran > [19:33](https://www.youtube.com/watch?v=XlKGCaoCIPk&t=1173s) 如果企业拥有超过 100 名全职员工（FTE）的运营或业务团队，IntellectAI 能够保证在短短 3 个月内将该团队的业务流程与人力效率优化提升 40% 以上。 — Aditya Rajashekaran ## Mentioned Aditya Rajashekaran, IntellectAI, Purple Fabric, Ai4 ## Video description Aditya Rajashekaran, Senior Project Manager, IntellectAI Abstracts In a world full of chatbots, real business value comes from AI that doesn’t just talk - it thinks, acts, and delivers. This session will showcase intelligent multi-agent workforces that autonomously coordinate, reason, and execute complex workflows across the enterprise. From streamlining operations to accelerating decisions, AI agents are able to drive real business impact through faster resolutions, greater efficiency, and measurable results. This session explores how multi-agent systems differ from traditional chatbots and single-agent models, with real-world examples of how they’re driving business impact. Learn how organizations are shifting from conversational AI to operational AI that takes action. Whatever your role, see how AI is evolving from conversation to execution and transforming business along the way. [Source](https://www.youtube.com/watch?v=XlKGCaoCIPk)

[0:04] 听众: 缺乏结构化（Unstructured） [0:04] 听众: 速度缓慢（Lack of speed） [0:04] 听众: 极其复杂（Complicated） [0:04] Aditya Rajashekaran: 十几岁青少年的卧室 [19:33] Aditya Rajashekaran: 如果企业拥有超过 100 名全职员工（FTE）的运营或业务团队，IntellectAI 能够保证在短短 3 个月内将该团队的业务流程与人力效率优化提升 40% 以上。

[3:46] Aditya Rajashekaran: Think of now your operational landscape as a dynamic soccer match... Put on one goalkeeper or a single AI agent against your entire operational landscape, regardless of if it's the best goalkeeper in the world, you're not going to be able to win.

[0:04] So I'm Adita. I'm a product manager for [0:07] the platform Purple Fabric. And just a [0:10] brief TLDDR, our platform allows you to [0:14] build, benchmark, deploy, and scale [0:17] agents within a trusted enterprise [0:20] context, right? And what are the [0:24] problems that these agents are trying to [0:26] solve? And those problems are hidden [0:29] behind that door. Behind that door are [0:33] your back office and mid office [0:35] operations. So just shout out what do [0:38] you think of when you actually think [0:40] about your back office operations? [0:42] Anybody? [0:44] >> Unstructured. [0:45] >> Unstructured. Yes. Very very [0:47] unstructured. Any other thoughts? [0:51] >> Speed or the lack thereof in some cases. [0:53] Right. Exactly. Anybody else? [0:57] complicated. Yes, exactly. So, speed, [1:01] unstructured, complicated, [1:04] um, lots of documents, manual processes. [1:08] This is really the back office that most [1:10] of us are dealing with. And that's a [1:12] reality today. And for me, it's really [1:15] like a teenager's bedroom. You know, [1:18] it's very messy. You have friends in all [1:21] the time, lots of people, but there's a [1:24] lot of potential if you know where to [1:26] look. So the question is through that [1:30] glorious mess, how do you actually get [1:33] it in order to make sure that the [1:35] teenager blossoms into something really [1:37] amazing. [1:39] A little bit about who we are. We are [1:42] not a startup. We are a fintech giant [1:45] who's been in operation for more than 30 [1:47] years. So we've delivered real solutions [1:51] to over 450 financial service [1:53] institutions around the world and these [1:56] are enterprises and governments. What [1:59] that means is that most of you probably [2:03] are actually customers of our clients [2:06] and we know what it's like to work with [2:11] enterprises that have very very messy [2:14] teenagers bedrooms and back offices that [2:16] are critical to actually making sure [2:19] that you as a customer are getting the [2:21] best possible experience. [2:24] Not only that, these are highly [2:26] regulated industries where every single [2:28] decision is up for scrutiny. [2:32] And when we got onto the AI journey, it [2:36] was not two years ago with the [2:38] popularization of GPD 3.5, we've been [2:42] investing in AI for over 10 years. And [2:45] we've had different streams for data, [2:48] different streams for AI. And what we've [2:50] been doing is solving real problems for [2:53] our real customers as we as we've spent [2:56] tens of millions of dollars on this [2:59] journey. [3:00] That journey has brought us to purple [3:02] fabric which is our open business impact [3:05] AI platform. This is where we've unified [3:08] our AI and data capabilities into a [3:11] single enterprise ready foundation. This [3:14] is what we call wave five and our goal [3:17] is to help enterprises evolve into being [3:21] AI native enterprises. [3:24] So we focus on business impact. We don't [3:28] do AI for the sake of AI because every [3:31] single problem that we've solved has [3:33] been for a customer. And I know that all [3:38] of us are here today because we [3:40] understand the disruptive potential that [3:42] AI agents can actually bring our [3:44] enterprises. [3:46] I want to focus on the last capability, [3:49] the ability for the AI agents to [3:51] collaborate. Because they can talk in [3:53] natural language, they can go beyond [3:56] doing single tasks and actually work [3:59] together not just with each other but [4:02] with humans in the loop. specialized [4:04] domain uh experts to solve big [4:08] complicated problems and they work [4:11] within your evolving and messy [4:14] operations reality. Think of the [4:16] teenager's bedroom. [4:19] This can be illustrated with all of our [4:22] favorite soccer coach Ted Lasso. [4:25] So, think of now your operational [4:28] landscape as a dynamic soccer match. You [4:31] have customers to onboard, failed [4:33] transactions, packages lost in trans uh [4:36] in transit, uh sanction screening that [4:39] you may have to do, right? That is the [4:42] opposition team. That is your [4:44] operational landscape. That's all of the [4:46] challenges that are being thrown towards [4:48] you. And your goal is to make sure that [4:53] not only do you deal with that dynamic [4:56] landscape to stay in the match, but that [4:58] you have a strategy to actually win the [5:00] match by scoring the goals. Now, if you [5:04] put on one goalkeeper or a single AI [5:07] agent against your entire operational [5:10] landscape, regardless of if it's the [5:12] best goalkeeper keeper in the world, [5:14] you're not going to be able to win. [5:17] What you need to do is ensure that you [5:20] actually have a full team of AI agents [5:23] that are specialized defenders that are [5:26] actually blocking things that are coming [5:28] your way. Midfielders that are thinking [5:30] about strategy, how to approach the [5:31] opposing half, and strikers who can [5:33] actually finish it off. These don't have [5:36] to be AI agents entirely. They can also [5:38] be human in the loop. You can have human [5:42] experts that are collaborating with AI [5:44] agents to make sure that you're getting [5:46] business impact. And let me introduce [5:48] you to one of our clients real [5:51] operational landscape. [5:53] We're working with the largest wealth [5:56] manager in the UK. And the operational [6:00] landscape that they were operating in is [6:03] one where the regulations changed which [6:06] allowed customers like you or me to [6:09] retroactively complain for the past 10 [6:12] years about poor advice or poor service [6:15] that I had received on behalf of the [6:18] wealth manager. What this meant was that [6:21] instead of the hundred complaints that [6:23] used to come in every month, I was now [6:26] facing about a thousand a week, I was [6:29] being crushed underneath the complaints [6:31] that were coming in. And these are not [6:34] simple complaints. These complaints take [6:36] 50 days to investigate. The reason is [6:39] the advice that has been given to me [6:42] based on my health history, investment [6:45] history, retirement goals, age, appetite [6:48] for risk is going to be very different [6:51] than the advice given to anybody else. [6:54] And not just that, I need to look at [6:56] dozens of systems, CRM systems, policy [6:59] admin systems, handwritten notes that [7:01] financial adviserss took 10 years ago in [7:04] meeting with clients to actually gather [7:06] the evidence I need to investigate this [7:08] complaint. And over these 10 years, [7:11] regulations have also been changing. [7:14] This is a complicated operational [7:16] landscape I need to take into account [7:18] when investigating a complaint. And what [7:21] we've put into production is a team of [7:24] specialized a specialized AI agents that [7:26] actually collaborate independently and [7:29] with human in the loop to solve the [7:31] problem. So going back to our soccer [7:34] story, we have a team of expert agents, [7:38] a complaint manager that is like the [7:40] goalkeeper. The complaint manager [7:42] actually intercepts a complaint coming [7:44] in from going into the goal. He then [7:47] passes it forward to the defenders who [7:49] understand what type of complaint is [7:51] this? How do we classify it? What [7:54] actions do we need to further take? Then [7:57] they pass it on to the midfielders, the [7:59] evidence gathering agents who have [8:01] access to the dozens of systems that run [8:04] this enterprise where evidence is [8:06] present. The evidence gathering agents [8:09] actually create a case dossier and give [8:11] it to the forwards, the complaints [8:13] investigation agents to actually create [8:16] a decision recommendation that is then [8:18] provided to a human in the loop like uh [8:22] Ted Lassel the manager to make a final [8:24] decision on whether to actually really [8:27] keep the complaint or to not. This is [8:31] something that took the time required to [8:34] actually gather evidence for and [8:37] investigate a complaint from 50 days to [8:39] 20 minutes. So, it was a step change and [8:43] the tens of thousands of complaints that [8:45] our client actually had on backlog were [8:48] addressed because of the multi- aent [8:51] solution that was put into production [8:52] for them. [8:54] All of this was also with a 50% [8:57] reduction in cost and a greater than 90% [9:00] accuracy over the entire multi- aent [9:03] system. Once again, this is not [9:06] theoretical. This is something that [9:07] we've put in production and I'll show [9:09] you bits and pieces of it so that you [9:11] can see it for yourself. [9:13] Apart from the structured information [9:15] that was present in the CRM systems, [9:17] policy admin systems, so on and so [9:19] forth, unstructured evidence is very [9:21] critical to investigate a complaint. [9:24] That is a mock letter that a financial [9:28] adviser has written way back in 2016 [9:31] just talking about what happened during [9:33] this meeting with a client. So instead [9:35] of looking just at different [9:37] transactions [9:38] um you need to really be able to [9:40] understand what happened during these [9:43] meetings as well. And this is for [9:46] complaint letters themselves, factf [9:48] finding documents, handwritten notes, so [9:50] on and so forth. [9:53] Once you actually understand what a [9:55] complaint is about, you need to create [9:57] an investigation knowledge garden. This [10:01] is the AI agents connecting to all of [10:05] your different enterprise platforms [10:07] where you can actually find all of the [10:10] information and data required to [10:13] investigate a complaint. So, our AI [10:15] agents are connected to Salesforce, [10:18] um, to Snowflake, to the FCA, which is a [10:20] regulatory authority in the UK, to SSP, [10:24] and on the right side, you can actually [10:26] see the AI agents going through all of [10:29] these sources to go and gather evidence [10:32] for a complaint that has been raised. [10:37] Now that you've actually gathered the [10:38] evidence, investigation agents [10:41] scrutinize that case dossier and produce [10:44] a final investigation report. So a [10:47] complaint decision recommendation is [10:49] provided with support and evidence that [10:52] me as a complaint handler can then look [10:54] at instead of going and actually finding [10:58] all of the data myself and doing that [11:01] analysis. I get a case dossier with a [11:03] decision recommendation. I can act upon. [11:08] How do I trust that the decision [11:10] recommendation these AI agents have [11:12] given is really sound? I can trust it [11:16] because of data source lineage. On our [11:19] platform, every single decision or [11:22] recommendation or answer that AI agents [11:24] give is always tracked back to the [11:27] source. Not just the document itself, [11:29] but which portion and which chunk of the [11:32] document that information was found. Or [11:34] in case it's an enterprise system or a [11:36] database, you actually know which [11:39] database, what query was run, what [11:41] insights were received for you to [11:43] actually get to that decision. [11:47] When you have a team of over 20 agents [11:50] that are collaborating with each other, [11:52] how can you actually ensure that they're [11:54] doing a good job, that they're actually [11:57] uh working well together? That is where [12:00] we bring traces for explanability. Think [12:02] of it as an email chain between [12:05] different agents. An agent receives a [12:08] problem, they think about how they [12:10] should solve it. Then they understand [12:12] the operational landscape or their [12:14] colleagues landscape available to them. [12:17] They know that they have a colleague [12:19] agent that actually connects with the [12:21] complaints management system and can [12:23] raise a complaint. So instead of doing [12:25] it themselves, they go to the specialist [12:28] agent that actually is connected to the [12:30] complaints management system to raise a [12:31] complaint. This particular agent that [12:34] we've stopped on is a classification [12:36] agent that's reads through a complaint [12:38] and understands what a complaint is [12:41] actually about. And you can see that [12:43] there's two reasons that have been [12:44] identified here. Inappropriate product [12:47] transfer uh and partner service. You can [12:51] see that there's two complaints, not [12:54] just one. This is not rules-based. Just [12:57] like a human would read a complaint [12:59] letter and figure out, okay, you know, [13:01] there's two complaints present here. [13:03] That's what the AI agents do in this [13:05] case. There's no decision tree. There's [13:07] no rules that we tell the agents that [13:09] there will be one or two or three [13:11] complaints per letter. [13:14] This is a picture of what the old [13:16] process looked like. There were four [13:18] teams in total doing the processes that [13:21] you see on the right side. the case [13:22] logging team, the triage team, the data [13:24] gathering team, and the case handling [13:26] team. All of them had a target capacity [13:29] per week. And you can see that the [13:31] subject matter experts, the case [13:33] handlers making the decision was one of [13:35] the largest teams and the slowest team [13:38] because they had so much evidence to [13:40] parse. And in today's world, most of [13:43] these steps were taken over by AI [13:45] agents. And AI agents provide a decision [13:48] recommendation, which means that the [13:51] manual case handling team only needs to [13:54] look at the the decision recommendation [13:57] and the evidence. [14:00] In reality, now what we're able to do is [14:03] optimize the case logging, triage, and [14:05] data gathering teams because they're run [14:08] by AI agents and the subject matter [14:11] experts who are actually in charge and [14:13] own the decisions. They're able to make [14:16] 10 times as many good decisions in the [14:18] same amount of time that they were able [14:20] to previously without the solution and [14:23] operation. [14:26] All of this looks like it's going to be [14:28] very complicating complicated having a [14:31] team of over 20 agents that are [14:33] connected to dozens of different data [14:36] sources with different roles. But I'm [14:38] here to tell you that instructing a [14:41] digital expert is really like inducting [14:43] a team member. You tell them what they [14:46] have to do in English because they [14:48] understand natural language. You give [14:51] them access to relevant data sources so [14:54] that they know that they have the [14:56] relevant guard rails and guidelines in [14:58] place to investigate and make decisions [15:01] within your enterprise context. And you [15:04] also provide them access to the team [15:06] members and tools that are required for [15:08] them to actually do their job much like [15:11] you would do in a new employee [15:12] induction. [15:15] So now you actually have an [15:17] investigation knowledge garden connected [15:20] to your CRM systems, policy admin [15:22] systems. [15:24] What's amazing is that every single use [15:27] case you put into production is not a [15:29] transaction. It is a step towards [15:32] something bigger. What you can do is [15:34] create different agents with different [15:36] personas to ask different questions of [15:39] this investigation knowledge garden. So [15:42] you can actually find life stage [15:45] triggers. You can figure out what are [15:47] the best um cross-ell or upsell [15:50] opportunities for your customer. You can [15:52] see their feedback, their account [15:54] analysis and account activity. And [15:57] what's very important to take away from [15:59] here is this gives you the ability to [16:02] hyperpersonalize your interactions with [16:05] your clients. So instead of just saving [16:08] cost and improving speed like you were [16:10] doing in the previous example, you're [16:12] able to actually improve your topline, [16:15] give your customers better service, [16:18] increase your retention, and also ensure [16:20] that your NPS is climbing northwards. [16:24] I've shown you just a couple of [16:26] enterprise digital experts that are [16:28] already working for our clients. Here [16:31] are just a few more. So, it's really up [16:35] to your imagination to figure out how [16:38] the enterprise digital experts, how [16:40] these AI agents can work within your [16:43] organization [16:45] and what is holding you back, what is [16:47] holding most of our clients back from [16:49] achieving business impact. We've done [16:52] some research and this is what they [16:54] said. Largely, how do you measure if the [16:58] AI agents are actually performing well? [17:00] How can you see if you're getting any [17:02] ROI out of this? Also, there's a lot of [17:05] legacy tech and infrastructure. [17:08] And that is one of the biggest problems [17:10] that we've solved. Our LLM optimization [17:13] hub allows you to benchmark each agent [17:16] in a multi- aent system and a team of [17:20] agents for accuracy, speed, and cost. [17:24] One of the things that we don't hold you [17:27] to is certain LLM models or providers. [17:29] So for one agent, what you can do is [17:32] choose between Gemini, the OpenAI [17:35] models, anthropic models to figure out [17:37] which of those models actually powers [17:39] the agents the best from an accuracy, [17:41] speed, and cost perspective and make [17:44] sure that you know the metrics of what [17:46] you're getting out of it before you put [17:48] them into production. [17:50] This is just one of the features of [17:53] Purple Fabric, the open business impact [17:55] AI platform. And this is a very very [17:59] short note at how we actually allow you [18:02] to build to benchmark to deploy and [18:07] scale AI agents within your enterprise. [18:11] The enterprise knowledge garden really [18:14] allows the platform to connect to all of [18:16] your trusted data sources, be they [18:18] structured or unstructured, to find the [18:21] information that's required and put them [18:23] in a format that is available for AI and [18:26] AI agents to consume. This is all of the [18:30] these are all of the connectors that you [18:32] saw with Salesforce uh so on and so [18:35] forth earlier, right? The enterprise [18:37] digital experts allow you to actually [18:40] build these AI agents and orchestrate [18:42] them so that they're working together [18:44] and with human in the loop. The LLM [18:47] optimization hub is the benchmarking [18:49] screen that you saw before, but what we [18:51] also allow you to do is bring your own [18:54] LLM models and your own LLM keys. So, we [18:56] don't hold you to that. And finally, the [18:59] enterprise governance is really [19:02] important. We're working with a client [19:04] that stated a desire to have 300,000 [19:07] agents in operation over the next 5 [19:10] years. How do you control them? How do [19:12] you monitor them? How do you actually [19:14] find these agents um limit them to the [19:17] data entitlements that they have? How do [19:19] you ensure that the people using these [19:21] agents are not accessing data that they [19:23] don't need? These are the problems that [19:25] we're trying to solve along with [19:27] explanability, traceability that you saw [19:30] on purple fabric. And all of this is [19:33] infraostic in terms of deployment. So [19:36] thank you very much. And all I want to [19:38] say is if you have a 100 plus FTE team, [19:42] the intellect promise as is that we're [19:45] going to be able to optimize it by more [19:47] than 40% in 3 months. Thank you. We're [19:49] at booth 421.