Dyna.Ai Sees Big Banks Shift to Internal AI Over Governance Hurdles

From Banking to Building the Future of AI: Real Use, Real Impact With Dyna.Ai — Ai4

Ai4

Large financial institutions are pivoting away from high-ROI customer-facing voice agents to internal employee copilots because enterprise governance and regulatory hurdles prove far more difficult to resolve than underlying technical challenges.

AI-assisted summary of From Banking to Building the Future of AI: Real Use, Real Impact With Dyna.Ai — original source. Check the original for full context.

Source published: 2025-08-29

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 29, 2025, this presentation features Tomas Skoumal, Chairman and Co-Founder of Dyna.Ai. According to the description, the session explores how enterprise-grade, human-like AI solutions can drive intelligent transformation across daily banking operations, sharing real-world case studies and practical applications to enhance both customer experience and frontline employee productivity.

## TL;DR Drawing from over twenty years in commercial banking, Tomas Skoumal explains how financial institutions navigate generative AI adoption across customer experience (CX) and employee experience (EX) tracks. While smaller institutions rapidly deploy autonomous customer-facing voice agents for high sales and collections ROI, larger enterprise banks typically start with internal copilots—such as automated document-parsing credit assistants—to bypass governance friction and preserve human accountability. The presentation highlights real-world deployments in Japan, the UAE, and Hong Kong, illustrating how phased rollouts and human-in-the-loop workflows bridge generative AI outputs with traditional discriminative risk models.

## Atmosphere & style The presentation carries a pragmatic, conversational, and refreshingly candid tone, with Skoumal playfully noting that only his talk stands between the audience and the opening of the bar downstairs. Rather than relying on abstract hype or dense slides, his delivery feels like an insider's debrief from a veteran banker who understands the practical anxieties of risk and compliance committees. It balances grounded operational realism with dynamic case studies, resulting in a brisk, relatable, and solutions-focused industry talk.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) Artificial intelligence is reshaping banking along two parallel paths: external customer experience (CX) and internal employee experience (EX). - [01:33](https://www.youtube.com/watch?v=wog7P677xMI&t=93s) Smaller financial institutions adopt customer-facing agents quickly for immediate ROI, while large tier-one banks pivot to internal copilots due to heavy governance and regulatory friction. - [04:12](https://www.youtube.com/watch?v=wog7P677xMI&t=252s) Enterprise-grade voice agents require on-premise speech processing, persistent cross-channel memory, and strict workflow management to enforce regulatory identity verification. - [08:04](https://www.youtube.com/watch?v=wog7P677xMI&t=484s) Internal credit copilots rely on advanced document parsing engines to digest unstructured balance sheets, P&L statements, and tax filings directly into standardized credit memos. - [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) Maintaining a human-in-the-loop workflow preserves regulatory accountability, leaving legal responsibility for loan approval submissions with the human relationship manager. - [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) AI credit agents reduce document compilation from several days to 15 minutes of generation followed by 30 to 60 minutes of human review. - [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) Human modifications to AI drafts create behavioral training data for legacy discriminative AI models to detect optimistic sales bias and refine risk scoring algorithms. - [13:25](https://www.youtube.com/watch?v=wog7P677xMI&t=805s) AI voice agents achieve 100% script adherence, enabling true champion-challenger (A/B) testing and real-time conversational pivot triggers that are impossible with human operators. - [13:25](https://www.youtube.com/watch?v=wog7P677xMI&t=805s) Customer-facing conversational solutions typically yield a faster and higher return on investment than internal employee tools due to direct top-line revenue generation. - [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) Institutions can navigate regulatory risk via a four-stage rollout: internal knowledge base, real-time agent cues, full operator scripts, and finally autonomous voice interaction.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) An Ex-Banker's Practical View on Banking AI

Tomas Skoumal opens by acknowledging his role as the final speaker standing between the audience and the opening of the bar downstairs, promising a grounded and practical session rather than high-level, sophisticated slide decks. Having spent over 20 years working inside commercial banks before deciding that "enough is enough" and co-founding Dyna.Ai, he approaches artificial intelligence through the pragmatic lens of an experienced banker: evaluating what financial institutions actually like, what they fear, and where genuine business value lies.

Banks are currently being reshaped along two distinct axes: how they interact with external customers and how they execute work internally. While financial institutions have relied on traditional, "old-fashioned" machine learning for years, generative AI alters the playing field by enabling conversational customer interactions and generating far more nuanced, sophisticated operational outputs.

```mermaid flowchart LR accTitle: Two Paths for AI Adoption in Financial Institutions subgraph CX["Customer Experience (CX) Path"] A["External Engagement"] --> B["Sales, Servicing & Collections"] B --> C["High ROI / Stricter Governance Barriers"] end subgraph EX["Employee Experience (EX) Path"] D["Internal Operations"] --> E["Document Parsing & Credit Analysis"] E --> F["Lower Governance Friction / Human-in-the-Loop"] end ```

[01:33](https://www.youtube.com/watch?v=wog7P677xMI&t=93s) Customer Experience vs. Employee Experience: The Size-Governance Divide

Over the past year and a half pitching AI solutions across the financial sector, Skoumal observed a universal pattern: nearly every financial company initially claims they want to deploy AI to talk directly to their customers. Their rationale centers on two immediate drivers: cutting operational costs and increasing conversion rates by combining real-time customer data with continuously optimized sales scripts. However, while smaller institutions follow through and deploy customer-facing agents for sales, servicing, and collections, large tier-one institutions encounter massive regulatory and governance friction. Changing an enterprise bank's governance framework proves to be a far more grueling hurdle than solving the underlying engineering problems.

As a result, larger banks pivot away from direct customer-facing deployments toward internal employee experience (EX) agents. These internal copilots generate specialized reports, draft loan recommendation memos, and support credit analysts. Although internal tools may show a less immediate return on investment compared to direct customer-facing sales bots, they provide large enterprises with a compliant, low-risk entry point to begin their AI journey, build institutional literacy, and gradually modernize internal workflows.

- **Small Institutions:** Quicker to adopt customer experience (CX) agents due to agile governance and clear ROI in frontline sales, debt collections, and servicing. - **Large Institutions:** Typically retreat from external bots to internal employee experience (EX) agents to avoid direct regulatory exposure and maintain existing accountability chains.

[04:12](https://www.youtube.com/watch?v=wog7P677xMI&t=252s) Anatomy of a Customer-Facing AI Agent

Building an effective customer-facing voice agent requires more than just calling a standard Large Language Model. Enterprise-grade deployments require an on-premise Automatic Speech Recognition (ASR) engine to comply with banking privacy standards, avoiding exposure to public clouds, paired with high-quality Text-to-Speech (TTS) on the output end. At the core, the LLM must be paired with an enterprise memory architecture that tracks every historical interaction a customer has had with the bank across all touchpoints. In human call centers, customers frequently suffer the frustration of repeating their story every time they are transferred to a new representative; a unified memory layer allows the AI agent to retain cross-channel context and outperform fragmented human operations.

```mermaid flowchart TD accTitle: Customer-Facing AI Agent Architecture subgraph InputOutput["Audio Processing Layer"] In["Customer Voice Call"] --> ASR["On-Premise ASR"] TTS["Text-to-Speech Engine"] --> Out["Spoken Response"] end subgraph CoreEngine["Core Intelligence & Control"] ASR --> WF["Workflow Management Engine"] Mem[("Historical Interaction Memory")] <--> LLM["Enterprise LLM Core"] WF <--> LLM LLM --> TTS end ```

Achieving this level of continuous context requires banks to digitize and connect entire end-to-end customer journeys rather than deploying narrow, siloed bots for isolated tasks like credit card limit upgrades. Furthermore, the AI cannot be left unconstrained to answer customer prompts freely. It must be strictly guided by an underlying workflow management system that enforces mandatory regulatory compliance. For instance, if an agitated customer calls demanding to know why their credit card was declined, the workflow engine forces the LLM to politely acknowledge the frustration while strictly executing mandatory identity verification protocols—such as confirming dates of birth or existing account holdings—before disclosing sensitive financial details.

[08:04](https://www.youtube.com/watch?v=wog7P677xMI&t=484s) Internal Agents and Document Parsing for Credit Risk

Transitioning from customer-facing agents to internal employee agents requires nearly identical core architecture and infrastructure investments—including on-premise LLMs—with one major functional addition: advanced document parsing engines. While customer service interactions deal with short conversational transcripts, internal agents must digest dense, complex financial paperwork such as corporate balance sheets, P&L statements, and multi-page bank statements to feed structured data into underwriting workflows.

A prime use case is assisting Relationship Managers (RMs) and credit analysts handling commercial lending. In traditional banking setups, high-value corporate RMs frequently lose two full working days sitting at desks manually assembling credit memos and preparing loan documentation for the risk department instead of sourcing clients in the field. An internal credit agent streamlines this burden by parsing financial statements, conducting automated background research, and drafting the standardized loan recommendation report.

Crucially, this design keeps institutional governance and accountability intact. Because the agent compiles the dossier for the RM rather than submitting it directly to credit committee systems, the human relationship manager retains full legal responsibility to review, sign, and submit the proposal for credit committee approval on facilities such as a $5 million corporate loan.

- **Document Ingestion:** Ingests and parses unstructured financial statements, tax filings, and account statements directly into standardized credit evaluation templates. - **Frontline Productivity:** Frees corporate relationship managers from spending multiple days on administrative memo preparation, redirecting their bandwidth toward business development. - **Human-in-the-Loop Governance:** Preserves existing regulatory and risk structures by requiring human RMs to review, validate, and sign off on all AI-drafted loan submissions.

[10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) Governance, Human Accountability, and the Synergy Between GenAI and Discriminative AI

Maintaining standard banking governance is critical when deploying artificial intelligence in financial institutions. Because the legally and regulatory responsible party remains a human being rather than the algorithm, bank legal and compliance departments can comfortably approve AI implementations. In practice, a relationship manager (RM) who previously spent roughly three days compiling a complex financial assessment can have an initial draft generated by an internal AI agent in about fifteen minutes. The RM then spends thirty minutes to an hour reviewing the document, verifying its accuracy, and either signing off or making manual adjustments.

This human-in-the-loop workflow produces a valuable secondary data stream for risk departments. Because an RM is naturally incentivized to sell loans, any retrospective edits they make—such as tweaking two or three lines to portray a borrower's profitability more favorably—provide behavioral signals. While an AI might make unintentional, technical errors, humans often introduce intentional biases to push deals through. These human modifications serve as fresh training data fed directly into traditional discriminative AI and legacy machine learning models to refine credit risk scoring and approval algorithms. Similarly, customer-facing AI agents retain persistent conversational memory across calls, generating rich interaction datasets that discriminative scoring models use to evaluate call quality and customer satisfaction.

- **Preparation Time Reduction:** Document compilation drops from ~3 days of manual preparation down to ~15 minutes with an AI agent, followed by 30–60 minutes of human review. - **Intentional vs. Unintentional Error Tracking:** Discriminative AI cross-references RM edits against raw data to detect intentional optimistic biases in credit memorandums. - **GenAI and Machine Learning Symbiosis:** Call transcripts and RM behavioral edits feed structured telemetry back into legacy predictive models to continuously upgrade automated risk assessment.

[13:25](https://www.youtube.com/watch?v=wog7P677xMI&t=805s) Global Implementations: Standardized Voice in Japan and Credit Agents in the UAE

Real-world deployments across different geographies illustrate how conversational and internal AI agents solve distinct operational challenges. In Japan, an insurance client implemented voice agents to dramatically improve reach rates, customer communication, and sales conversion. A major advantage of conversational AI over human call operators is strict adherence to scripts: "every script which is given, it's followed 100%." Human operators naturally introduce wide conversational variance, making it nearly impossible to run reliable champion-challenger (A/B) testing between two script variants. With an AI voice agent, institutions can systematically test workflow tweaks, adjust voice tone, analyze customer reactions, and detect specific phrasing to immediately pivot prospects toward alternative insurance products when they show disinterest.

```mermaid flowchart TD accTitle: Internal Credit Agent Workflow in the UAE subgraph S1["Data Ingestion & Synthesis"] A["Client P&L & Financial Docs"] B["Internet Search & Market Data"] end subgraph S2["AI Generation"] C["Dyna.Ai Credit Agent"] end subgraph S3["Human Review & Risk Feedback"] D["Relationship Manager Review"] E["Risk Scoring & Discriminative AI"] end A --> C B --> C C --> D D --> E ```

In the United Arab Emirates, Dyna.Ai deployed an internal credit agent for a leading bank to assist RMs handling commercial lending. Instead of spending up to three weeks gathering internet research, parsing P&L statements, and drafting reports, RMs use the credit agent to generate comprehensive financial dossiers in minutes with high accuracy. While internal productivity tools deliver significant operational efficiency and high resolution for staff, customer-facing conversational solutions typically yield a faster and higher overall return on investment due to their direct impact on top-line revenue generation.

- **Script Adherence in Japan:** 100% compliance with conversational flows enables clean champion-challenger testing and instant keyword detection for product cross-selling. - **Credit Dossier Turnaround in UAE:** Slashed preparation time from up to 3 weeks down to a few minutes per corporate client. - **ROI Disparity:** Customer-facing voice agents directly driving sales offer a higher and more immediate ROI than back-office employee augmentation tools.

[16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) Phased Rollout Strategies: Knowledge Management in Hong Kong to Full AI Autonomy

For financial institutions navigating strict regulatory constraints, AI adoption can be paced through a phased, multi-step roadmap rather than an immediate jump to fully autonomous customer interaction. At a financial institution in Hong Kong, this transformation began with an internal knowledge management platform. Step one consolidates all internal documentation, product descriptions, and training manuals into an enterprise chatbot that human sales agents query during live calls to answer customer questions accurately.

```mermaid flowchart LR accTitle: Phased AI Adoption Roadmap for Contact Centers subgraph P1["Assisted Stages"] S1["Step 1: Internal Knowledge Chatbot"] --> S2["Step 2: Real-Time Bullet Prompts"] S2 --> S3["Step 3: Pre-Written Script for Operator"] end subgraph P2["Autonomous Stage"] S3 --> S4["Step 4: Fully Autonomous Voice AI"] end ```

As the system matures, step two enables the AI to listen to the call in real time and automatically surface bullet-point cues on the agent's screen. In step three, the AI drafts the complete verbal response for the operator to read directly, substantially accelerating newcomer onboarding and cutting training costs. Once the bank records several weeks or months of production data proving human operators read the generated scripts with zero manual corrections, internal risk and compliance committees have the empirical evidence required to approve step four: allowing the voice AI to speak directly with customers without a human intermediary. Beyond customer operations, this unified knowledge base can scale enterprise-wide to eliminate internal friction, replacing scattered wikis and intranet files with a single conversational interface.

Concluding the presentation before the audience heads to the networking reception, Tomas directs attendees to an on-screen QR code linking to an interactive digital avatar business card, inviting attendees to "call me and talk to my avatar" to explore Dyna.Ai's solutions firsthand.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) enough is enough — Tomas Skoumal

> [01:33](https://www.youtube.com/watch?v=wog7P677xMI&t=93s) Changing an enterprise bank's governance framework proves to be a far more grueling hurdle than solving the underlying engineering problems. — Tomas Skoumal

> [04:12](https://www.youtube.com/watch?v=wog7P677xMI&t=252s) Building an effective customer-facing voice agent requires more than just calling a standard Large Language Model. — Tomas Skoumal

> [08:04](https://www.youtube.com/watch?v=wog7P677xMI&t=484s) Because the agent compiles the dossier for the RM rather than submitting it directly to credit committee systems, the human relationship manager retains full legal responsibility to review, sign, and submit the proposal — Tomas Skoumal

> [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) Because the legally and regulatory responsible party remains a human being rather than the algorithm, bank legal and compliance departments can comfortably approve AI implementations. — Tomas Skoumal

> [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) While an AI might make unintentional, technical errors, humans often introduce intentional biases to push deals through. — Tomas Skoumal

> [13:25](https://www.youtube.com/watch?v=wog7P677xMI&t=805s) every script which is given, it's followed 100% — Tomas Skoumal

> [13:25](https://www.youtube.com/watch?v=wog7P677xMI&t=805s) customer-facing conversational solutions typically yield a faster and higher overall return on investment due to their direct impact on top-line revenue generation. — Tomas Skoumal

> [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) call me and talk to my avatar — Tomas Skoumal

## Mentioned Tomas Skoumal, Dyna.Ai, Ai4, Automatic Speech Recognition (ASR), Text-to-Speech (TTS), Large Language Models (LLMs), Generative AI, Discriminative AI, Relationship Managers (RMs), Japan, United Arab Emirates, Hong Kong

## Video description Tomas Skoumal, Chairman & Co-Founder, Dyna.Ai

Abstract

AI is fundamentally reshaping the future of banking — are you ready to lead this transformation? Banks today face the dual challenge of enhancing both customer experience and employee productivity. At Dyna.Ai, we deliver enterprise-grade, human-like AI solutions that drive intelligent transformation across daily banking operations.

In this session, we will share real-world data and success stories showing how we co-develop practical AI applications that deliver measurable impact for both customers and frontline teams. Join us to discover how banks can make AI a core competitive edge — and lead the next wave of smart, scalable banking innovation.

[Source](https://www.youtube.com/watch?v=wog7P677xMI)

中文简报

Dyna.Ai：治理阻力促使大型银行转向内部AI 大型金融机构正从高投资回报率的面向客户语音智能体转向内部员工副驾驶系统，因为事实证明，企业治理和监管障碍远比底层技术挑战更难解决。 Tomas Skoumal结合其在商业银行领域二十多年的经验，阐释了金融机构如何在客户体验（CX）与员工体验（EX）两条路径上推进生成式AI的落地应用。小型机构能够快速部署面向客户的自主语音智能体，以获取销售和催收方面的高投资回报率；而大型银行通常从内部副驾驶（例如自动化文档解析信贷助手）起步，以规避治理阻力并保留人工问责机制。演讲重点介绍了在日本、阿联酋和香港的实际部署案例，展现了分阶段推广与“人在回路”（human-in-the-loop）工作流如何将生成式AI的输出与传统判别式风控模型有效衔接。小型金融机构因治理敏捷而迅速部署面向客户的 AI 代理以追求即时 ROI，而大型一线银行受严格合规与监管摩擦影响，普遍转向部署内部员工协作成型（EX）工具。 (1:33) 面向客户的企业级语音 AI 必须配备本地化 ASR、端到端跨渠道交互记忆，以及严格的工作流引擎，以确保在披露敏感财务信息前强制执行法定身份验证。 (4:12) 内部信贷 AI 代理通过非结构化财报解析，将原本耗时数天的信贷备忘录起草压缩至 15 分钟生成加 30 至 60 分钟人工复核，同时保留客户经理（RM）签字提交以维持合规责任链。 (8:04) 客户经理对 AI 草稿的人工修改构成了行为训练数据，可反哺传统判别式 AI 模型，用于检测销售人员的主观乐观偏差并优化风险评分算法。 (10:45) 受监管限制的机构可通过四阶段平稳落地 AI：从内部知识库检索、实时要点提示、坐席逐字脚本辅助，最终在积累充分验证数据后平滑过渡至全自动语音交互。 (16:10)

Reading summary

Tomas Skoumal of Dyna.Ai argues that while financial institutions seek artificial intelligence to lower costs and increase revenue, governance and regulatory hurdles dictate whether they implement external customer agents or internal employee copilots. The Governance Divide Between CX and EX (1:33) According to Skoumal, financial institutions initially seek AI to speak directly with customers to lower operational costs and lift conversion rates. However, organizational scale dictates the eventual deployment path. Smaller institutions act quickly on customer experience (CX) tools across sales, servicing, and debt collections due to agile governance and direct commercial returns. In contrast, large tier-one banks face significant regulatory hurdles when attempting customer-facing deployments. To mitigate regulatory exposure, larger enterprises pivot toward employee experience (EX) copilots. While internal operational assistants yield a less immediate return on investment than direct sales bots, they provide large enterprises with a compliant mechanism to build institutional literacy, modernize workflows, and maintain strict risk controls. Architecting Compliant Customer-Facing Voice Agents (4:12) Deploying customer-facing conversational agents in banking requires infrastructure beyond standard public cloud foundation models. Skoumal stresses that implementations need on-premise Automatic Speech Recognition (ASR) to meet data privacy mandates, paired with high-quality Text-to-Speech engines and a unified memory architecture that tracks customer interactions across all channels to avoid conversational fragmentation. Crucially, language models cannot run unconstrained. They must be governed by an underlying workflow engine that enforces mandatory compliance protocols. For example, if a customer inquires about a declined transaction, the workflow engine requires the model to perform identity verification before disclosing financial details. Real-world implementations underscore operational precision: an insurance deployment in Japan achieved 100% script adherence, eliminating human conversational variance to enable systematic champion-challenger script testing. Streamlining Commercial Lending with Credit Agents (8:04) Internal employee agents utilize similar core architectures but rely heavily on document parsing engines to digest unstructured corporate balance sheets, profit-and-loss statements, and multi-page bank statements. In traditional corporate banking, high-value relationship managers often spend two full working days compiling credit memos for the risk department rather than developing client relationships. Skoumal notes that an internal credit agent can parse financial statements and draft standardized loan recommendations in approximately fifteen minutes, requiring only thirty to sixty minutes of human review. In a deployment for a leading bank in the United Arab Emirates, preparation times fell from up to three weeks to a few minutes. Crucially, governance remains intact because the human manager reviews, validates, and signs off on submissions to the credit committee, ensuring institutional and legal accountability remains with human personnel. Predictive Symbiosis and Phased Autonomous Adoption (10:45) Human oversight in internal credit workflows generates valuable telemetry for risk departments. Because commercial managers face sales incentives to close loans, manual edits to AI drafts often reveal intentional optimistic bias. Financial institutions feed these human modifications and conversational transcripts back into traditional discriminative machine learning models to refine credit scoring algorithms and evaluate call quality. For risk-averse institutions, Skoumal outlines a four-step phased adoption model demonstrated at a Hong Kong institution. Organizations begin with an internal knowledge management chatbot, advance to real-time onscreen bullet prompts, and progress to pre-written scripts read by human operators. Only after accumulating empirical operational data demonstrating zero operator corrections do risk committees permit the transition to fully autonomous voice interaction. Customer-facing conversational agents offer higher and more immediate financial returns, but operationalizing artificial intelligence in banking requires adapting deployment strategies to institutional scale, maintaining legal accountability through human-in-the-loop workflows, and building compliance evidence through structured, phased roadmaps.

精炼阅读摘要

Dyna.Ai 联合创始人 Tomas Skoumal 结合其二十余年商业银行从业经验指出，人工智能在银行业的落地瓶颈主要在于企业治理框架与合规审查，而非底层工程技术。面对客户体验（CX）与员工体验（EX）两条路径，大型银行倾向于先部署保留人类法律责任的内部助手以规避监管摩擦，而中小型机构则更快速地采用直接触达客户的销售与服务智能体。机构规模与治理约束决定落地路径 (1:33) Skoumal 观察到，金融机构在规划 AI 方案时普遍倾向于直接面向客户，希望借助实时数据与动态话术削减成本并拉升转化率。然而在实际落地中，不同体量的机构出现了显著分化。中小型机构依托灵活的治理机制，能够迅速在销售、客服与催收前端部署客户体验（CX）智能体；大型一级银行则面临严苛的监管与治理约束，改变企业既有治理框架的阻力远超工程研发难度。因此，大型银行多将重心转向内部员工体验（EX），部署辅助信贷分析、撰写贷款建议书与专业报表的内部副驾驶工具。尽管内部赋能工具带来的直接投资回报率（ROI）不如前端销售工具立竿见影，但它为大型机构提供了合规、低风险的切入点，有助于循序渐进地积累 AI 能力并推动内部流程现代化。客户语音智能体的架构与合规控制 (4:12) 构建企业级客户交互语音智能体不能仅依赖调用标准大语言模型。由于银行存在严格的隐私规范，系统输出端须搭配高品质文本转语音（TTS），输入端则须在本地私有环境（On-premise）部署自动语音识别（ASR）引擎，避免音频流暴露给公共云。同时，大语言模型必须接入记录客户全触点交互历史的企业级统一记忆架构，彻底解决人工客服转接时客户反复说明背景的痛点。为防范合规风险，智能体绝不能脱离控制自由作答，必须由底层工作流管理引擎严格引导。例如，当持卡人因信用卡被拒致电质询时，工作流引擎会强制要求模型在安抚情绪的同时，严格执行核对出生日期或账户资产等法定身份核验步骤，未通过验证前坚决禁止透露敏感财务信息。信贷文档解析与人机闭环责任制 (8:04) 与处理简短对话的客服智能体相比，用于支持信贷风险分析的内部智能体必须搭载复杂的文档解析引擎，用于消化企业资产负债表、损益表及多页流水等非结构化文件。在传统商业贷款场景中，高价值客户经理（RM）常需耗费数天甚至两至三周（如阿联酋某银行案例）搜集背景信息并手动编制备忘录；而信贷智能体可在数分钟内生成标准化报告初稿，将文档起草时间从大约 3 天压缩至 15 分钟左右。该架构的关键在于完整保留了既有的银行风控与法律问责体系。智能体仅为客户经理准备材料，无权直接向信贷委员会系统提交申请；面对如 500 万美元的授信项目，客户经理仍须花费 30 至 60 分钟核查初稿，并作为法定责任人签字提交。这一“人机在环”（Human-in-the-Loop）设计使合规与法律部门能够安心放行 AI 技术的业务嵌入。生成式协同与人类销售偏差监测 (10:45) 人机协同流程不仅大幅压缩了信贷材料准备周期，还为风险部门提供了宝贵的次级行为数据。客户经理具备推动贷款成交的业绩倾向，其对 AI 初稿所作的人工修改——如调整措辞以优化借款人盈利表现——反映了人类销售的主观倾向。与大模型可能出现的技术性无意错误不同，人类修改往往带有促成交易的主观意图，这些修改记录被回传给传统判别式 AI 与既有机器学习模型，用于持续迭代信用风险评分与审批算法。此外，在日本保险机构的落地中，语音智能体实现了预设话术脚本的 100% 严格执行，彻底消除了人工坐席表达不一的偏差，使机构得以开展严谨的对照测试（Champion-Challenger），并在检测到客户特定拒绝意向时瞬间切换推荐方案，实现了生成式交互与判别式分析的紧密协同。四阶段演进路径化解自主决策风险 (16:10) 针对监管严苛的金融机构，Skoumal 提出以循序渐进的四阶段路线图替代激进的完全自主化改造。在香港某金融机构的实践中，第一阶段先搭建整合分散制度与产品手册的企业级知识库问答助手，供人工坐席在通话中实时查询；第二阶段实现模型实时监听通话并在坐席屏幕上动态推送要点提示；第三阶段由系统直接生成完整的逐字应答脚本，坐席无需额外组织语言即可直接宣读，大幅降低新员工培训成本。当生产环境积累了数周或数月坐席零差错朗读脚本的运行数据后，风控与合规委员会便获得了充分的实证依据，从而放行第四阶段：允许语音 AI 脱离人工中介，直接与客户展开端到端交互。这种分步落地策略通过可验证的运行表现逐步建立组织信任，为传统机构提供了一条风险可控的智能化实施路径。 Tomas Skoumal 总结指出，人工智能在银行业的实质价值兑现依赖于分阶段推进与审慎的权责结构设计。尽管直接面向客户的销售型语音智能体能更快带动顶层营收增长并展现更高 ROI，但在治理要求极高的大型金融机构中，以深度文档解析、知识协同及保留人类法定责任为核心的内部辅助工具，构成了平衡业务创新与金融合规的基础路径。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 29 日发布，主讲人为 Dyna.Ai 联合创始人兼董事长 Tomas Skoumal。视频简介指出，本次演讲旨在探讨人工智能如何从重塑客户体验与提升员工生产力两方面深度变革银行业，并结合 Dyna.Ai 的实战数据与成功案例，分享银行如何将 AI 转化为核心竞争壁垒。 ## TL;DR Tomas Skoumal 结合自身 20 余年的银行业经验，剖析了金融机构在落地生成式 AI 时的两种主要路径：面对强监管和治理阻力的大型银行倾向于从对公信贷分析、文档解析等内部员工赋能场景切入，以保持既有权责链条不变；而小型机构则更倾向于直接部署面向终端客户的销售、客服与催收 Agent。此外，AI 在信贷报告编制中生成的中立基准与客户经理的人工修改记录，还能反哺传统判别式风控模型以捕捉人为偏差。最后，演讲通过日本保险外呼标准化、阿联酋信贷自动化以及香港机构的四阶段渐进式落地案例，系统呈现了银行在合规前提下稳妥引入 AI 的可行策略。 ## Atmosphere & style 整场演讲风格务实接地气，开场以“自己是挡在听众与酒吧之间的最后一人”幽默破题，迅速拉近了与现场观众的距离。Skoumal 摒弃了晦涩的技术堆砌，全程以“老银行人”的内行视角直击金融合规与内部治理的核心痛点，语调亲切自然又兼具严谨度。从宏观架构分析到全球实战案例层层推进，尾声通过展示会讲笑话的个人数字人名片收尾，让整场充满干货的专业分享在轻松愉悦的氛围中圆满落幕。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 生成式 AI 正在从“外部客户体验”与“内部员工体验”两个维度深刻重塑银行业日常业务流程。 - [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 金融机构引入 AI 的初衷普遍希望直达销售、客服与催收等客户交互场景，以降低人力成本并动态标准化话术。 - [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 小型金融机构因决策敏捷多直接落地客户交互 Agent，而大型银行受制于复杂的监管治理与合规流程，通常优先选择内部员工 Agent 切入。 - [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 客户交互 Agent 需支持本地化（On-premise）部署，其核心竞争力在于具备跨周期、调取历史全量通话的记忆能力。 - [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 在金融级交互中，AI Agent 必须受严密的工作流引擎约束，确保在回答业务问题前优先完成身份核验（KYC）等合规动作。 - [08:08](https://www.youtube.com/watch?v=wog7P677xMI&t=488s) 内部员工 Agent（如对公信贷分析）深度依赖复杂长文档解析工具，用于处理损益表（P&L）与银行对账单等非结构化数据。 - [08:08](https://www.youtube.com/watch?v=wog7P677xMI&t=488s) 内部信贷 Agent 采用“人机协同（Human-in-the-loop）”设计，由客户经理最终签字确认，完全保留了原有的治理架构与责任主体。 - [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) AI 辅助可将客户经理编写贷款评估报告的时间从 3 天大幅压缩至 15 分钟，人工复核仅需 30 至 60 分钟。 - [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) 客户经理对客观 AI 初稿的主观修改痕迹会被记录并喂入传统判别式风控模型，从而反哺并提升信贷审批风控精度。 - [13:26](https://www.youtube.com/watch?v=wog7P677xMI&t=806s) 日本保险公司案例表明，AI 严格执行脚本的特性使金融机构能够开展 Champion-Challenger A/B 测试，精准优化成单率与险种推荐。 - [13:26](https://www.youtube.com/watch?v=wog7P677xMI&t=806s) 阿联酋重点银行案例显示，信贷 Agent 将跨网络搜集财报与撰写信用报告的周期从 3 周压缩至数分钟。 - [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) 银行无需一步到位上线面向客户的 AI，可通过“内部知识库检索 -> 屏幕实时提示 -> 完整拟稿朗读 -> 全自主交互”四阶段渐进式落地。 - [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) 基于 AI 的知识管理体系大幅缩短了新员工培训周期，并能沉淀扩展为全行统一的智能检索底座。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 银行业落地 AI 的双重路径：客户体验与员工体验演讲者 Tomas Skoumal 开场幽默地表示，自己是“挡在大家与楼下酒吧开放之间的最后一个人”，因此会尽量长话短说，摒弃那些空洞复杂的技术套话，从一名拥有 20 多年银行业经验的“老银行人”视角，分享金融企业究竟如何务实地应用 AI，以及金融机构在落地 AI 时喜欢与顾虑的核心痛点。Skoumal 在传统银行工作了二十余年，最终决定投身 AI 赛道创立 Dyna.Ai。他指出，生成式 AI（Generative AI）正在从两个维度重塑银行业：一是如何与客户互动（外部客户体验），二是如何在内部完成日常业务（内部员工体验）。尽管银行在过去几年里已经广泛使用传统机器学习（Machine Learning），但生成式 AI 带来了与客户实现拟人化直接交互，以及产出高复杂度、定制化业务决策的新可能。在过去一年半与金融机构打交道的过程中，Skoumal 发现几乎所有金融客户在初次接触时，诉求都高度一致——他们第一反应都是“希望用 AI 直接与客户沟通”，涵盖销售、日常客服以及催收（Collections）等场景。这背后的商业驱动力非常直接：首先是显著降低人力成本；其次是将 AI 与银行已有的客户数据相结合，根据对话反馈动态优化与标准化沟通话术，从而直接拉动销售业绩。然而，在实际落地中，规模不同的金融机构很快展现出截然不同的发展路径。小型金融机构往往决策灵活、包袱较轻，能够直接将 AI 落地于面向终端客户的交互场景；而大型金融机构在深入评估后会意识到，改变一家巨型机构的组织治理（Governance）和合规流程，其难度远超技术本身的实现。因此，大型银行通常会退而求其次，选择从员工体验（Employee Experience）或内部 Agent（Internal Agents）切入。例如，利用 AI 生成非单纯堆砌数字、而是包含具体建议与信贷推荐的标准化报告（如信贷审批 Agent）。虽然内部 Agent 在短期的投资回报率（ROI）上可能不如直接替代人工坐席的客户交互 Agent 那样立竿见影，但它能让大型机构在合规安全的前提下启动 AI 旅程，逐步积累内部对 AI 的认知与技术储备。 ```mermaid flowchart TD accTitle: 金融机构 AI 落地路径决策 subgraph Institution["金融机构规模与诉求差异"] A["金融机构引入 AI 诉求"] --> B{"机构规模与治理复杂度"} end subgraph SmallBank["小型金融机构"] B -- "小型机构：决策敏捷 / 治理负担轻" --> C["客户体验 Agent (外向型)"] C --> C1["销售转化 / 额度提升"] C --> C2["客户服务与业务咨询"] C --> C3["智能催收 (Collections)"] end subgraph BigBank["大型金融机构"] B -- "大型银行：强监管 / 治理变革阻力大" --> D["员工体验 Agent (内向型)"] D --> D1["HR 招聘与入职辅助"] D --> D2["信贷分析报告生成"] D --> D3["对账单与 P&L 文档解析"] end ``` - **客户交互 Agent 的核心动力**：通过自动化降低呼叫中心与坐席成本，并基于客户历史数据实时标准化话术，以提升销售转化。 - **大型机构的现实妥协**：技术实施不是最大阻碍，监管合规与治理架构的调整才是瓶颈，促使大行优先部署风险可控的内部赋能工具。 --- ### [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 客户交互 Agent 的核心架构：本地化部署、多轮记忆与工作流控制一个典型的客户交互 Agent 架构远不止大语言模型（LLM）本身。对于大型金融机构而言，出于数据安全与监管合规的严格要求，语音识别（ASR）组件必须支持本地化部署（On-premise），而非直接接入公共云服务；同时后端需配备精准的语音合成（Text to Speech, TTS）技术。位于核心层的大模型，其最关键的能力在于管理并调用“全量历史记忆”（Memories of all previous calls）。这正是 AI 相比传统人工坐席最具颠覆性的优势所在。在现实生活中，用户拨打银行客服或接到外呼电话时，经常会遇到换一个坐席就需要把背景情况“重新解释一遍”的糟糕体验；而 AI 能够完整记录并调取客户与机构之间发生的每一次历史沟通细节，从而提供连贯的个性化服务。要真正发挥这种记忆优势，银行必须推动端到端的全流程 AI 改造。如果银行仅将 AI 局限在单一环节（例如仅用于某一张信用卡的额度提升销售，或是单个新卡种的获客），那么 AI 所掌握的客户记忆就会被切割在特定产品或特定流程中，一旦客户涉及其他业务场景，记忆便会出现断层。客户体验的本质提升，是随着 AI 在更多业务线与流程中的持续铺设而不断累积放大的。除了记忆机制，AI Agent 必须受到严密的“工作流管理”（Workflow Management）约束。在金融业务场景中，绝对不能让大模型像开放聊天那样随心所欲地自由回答，而必须由既定工作流来强行引导沟通节奏。Skoumal 举例说明：当客户一接通电话就情绪激动地抛出问题“我的信用卡为什么刷不了？”时，即便客户在追问具体故障，AI 的首要任务依然是遵循合规流程完成身份验证。AI 必须在话术上既照顾客户情绪，又牢牢掌控流程：“先生，我非常理解您遇到的问题，但为了能够帮您处理，我首先需要核实您的出生日期（或核对您在我们名下的产品信息）。”这种工作流控制看似“严格甚至有些不近人情”，但它是金融合规不可逾越的底线，AI 必须在安抚客户与遵循监管验证步骤之间取得精确平衡。 - **基础设施合规性**：金融级 ASR 与 LLM 需具备 On-premise 本地化部署能力，以符合金融监管要求。 - **跨周期全局记忆**：克服人工坐席信息割裂的痛点，记录每一次历史交互，避免客户重复陈述背景。 - **受控的工作流引擎**：大模型生成能力必须被合规规则锁死，确保身份验证（KYC）等必要步骤优先于问题解答。 --- ### [08:08](https://www.youtube.com/watch?v=wog7P677xMI&t=488s) 内部员工 Agent：降低治理门槛与文档解析赋能将视角转向内部员工体验 Agent（如 HR 招聘助手、对公信贷分析 Agent 等）时，其底层技术架构与面向客户的 Agent 惊人地相似：同样需要本地部署的大模型底座与结构化的工作流控制。然而，内部 Agent 在输入端增加了一个极其关键的能力模块——复杂长文档解析工具（Document Parsing Tool）。面向终端消费者的客户交互通常不需要处理过于冗长复杂的材料，但对公信贷分析等内部场景涉及大量的企业损益表（P&L）、银行对账单（Account Statements）以及财务报告。AI 必须逐一解析并理解这些文档中的复杂数据，进而提炼并更新至信贷审批专用的标准化报告中。从投资角度来看，内部 Agent 与客户交互 Agent 的软硬件基础设施投入相差无几，但大型金融机构在推行内部 Agent 时阻力却小得多。其核心原因在于内部 Agent 完全不需要改变银行现有的治理架构与责任认定体系。Skoumal 以对公客户经理（RM）的日常工作为例进行了深度剖析：在传统流程中，一名对公客户经理如果想要为一家企业客户申请一笔 500 万美元的贷款，他往往需要离开市场一线，连续整整两天坐在电脑前进行繁琐的研究、翻阅账单，并手动撰写提交给风控与审批部门的信贷申报材料。引入内部信贷分析 Agent 之后，原本由客户经理在脑海中自我审问、反复核对的撰写逻辑，被完整固化进了 AI 的工作流引擎中；AI 自动完成文档解析并生成结构化的信贷建议报告。最关键的逻辑在于：**这份报告是生成给客户经理本人的，而不是直接越过人工提交给银行系统的。** 客户经理依然需要对这份报告进行人工复核、签字确认，并最终提交给风险部门。这种“人机协同（Human-in-the-loop）”的设计，使得银行原有的权责链条与风控责任主体没有任何改变，既帮前线业务人员从繁重文书工作中解脱出来以专注客户拓展，又为大型机构平稳引入 AI 铺平了道路。 ```mermaid flowchart LR accTitle: 内部信贷分析 Agent 人机协同工作流 subgraph Ingestion["文档输入与解析"] A["企业财务文档\n(损益表 P&L / 银行对账单)"] --> B["文档解析工具\n(Document Parser)"] B --> C["企业级 LLM & 工作流分析"] end subgraph Draft["报告生成"] C --> D["标准化信贷分析草案\n(含贷款建议与风控提示)"] end subgraph HumanApproval["责任闭环 (治理保持不变)"] D --> E["客户经理 (RM)\n人工复核与签署"] E --> F["风险审批部门\n(Risk Department)"] end ``` - **文档解析核心组件**：针对 P&L、财务报表、流水账单等非结构化与半结构化长文档提供精准语义提取。 - **权责机制保持原样**：AI 输出物仅作为员工的工作辅助材料，最终由业务人员审核并签署，不破坏银行既有的合规风控责任链。 - **解放一线生产力**：将客户经理从数天的材料撰写中解放出来，重新聚焦于客户拓展与高价值服务。 ### [10:45](https://www.youtube.com/watch?v=wog7P677xMI&t=645s) 治理合规与双重校验：生成式 AI 如何反哺判别式风控模型在金融机构部署 AI 解决方案时，最大的阻力往往来自监管与治理合规。Tomas Skoumal 强调，Dyna.Ai 的方案并没有改变银行现有的治理架构——法律与监管层面的最终责任人依然是自然人（如客户经理 RM），因此银行的法务与合规部门非常乐意接受这种模式。过去客户经理需要耗费长达 3 天时间撰写的贷款评估报告，现在由 AI Agent 在短短 15 分钟内即可自动生成；客户经理只需花 30 分钟到 1 小时通读全文，确认无误后签字确认，或者针对具体数据进行追溯修改。这一人机协同流程还催生了一种意想不到的风控增益机制。由于客户经理的核心诉求是把贷款推销出去，他们往往倾向于美化客户的财务状况。AI 作为一个中立系统，其生成的初稿是客观基准（哪怕有技术性失误，也是无意之失）。当客户经理在 AI 报告的基础上修改了两三行数据让报表显得更漂亮时，这一修改行为本身就构成了极具价值的特征数据。这些修改记录会被直接喂入后端的判别式 AI（传统机器学习风控模型）中，帮助风险管理部门准确捕捉客户经理的人为主观偏差，从而极大优化机构的信贷审批决策。 ```mermaid flowchart TD accTitle: 人机协同与双重风控反馈流程 subgraph DataGen["1. 生成式 AI 初稿"] A["多源客户数据与历史通话"] --> B["AI Agent 生成初步信贷报告"] end subgraph Review["2. 人工审核与修改"] B --> C["客户经理 (RM) 审核"] C --> D{"是否有主观修改?"} end subgraph RiskFeedback["3. 判别式模型风控"] D -- "记录修改痕迹与偏差" --> E["判别式 AI / 传统 ML 模型"] E --> F["输出高精度风控审批评分"] end ``` - **生成效率跃升**：信贷报告编制时间从原本人工耗费的 3 天缩短至 AI 生成的约 15 分钟，人工复核仅需 30 至 60 分钟。 - **GenAI 反哺 Discriminative AI**：AI Agent 留存的完整通话记录与人工修改痕迹，转化为高价值结构化数据，持续提升传统机器学习模型的判别能力。 - **无意错误 vs 有意美化**：AI 的错误属于技术性无意失误，而人类客户经理常有倾向性地改动利润数字；两者的差异成为风控部门的关键审查信号。 --- ### [13:26](https://www.youtube.com/watch?v=wog7P677xMI&t=806s) 全球落地实战：日本保险外呼标准化与中东信贷自动化为了避免占用大家前往茶歇区的时间，Tomas 结合地图快速展示了几个典型落地案例。第一个案例位于日本的一家保险公司，其核心痛点在于电话外呼的触达率、沟通质效与销售转化。相比人类坐席在通话中极易夹杂个人语气和随意发挥，AI 能够做到 100% 严格执行话术脚本。这使得金融机构首次具备了进行严格“赛马机制”（Champion-Challenger A/B Test）的能力——通过微调话术流程或语气语调，精准评估不同脚本对客户满意度和成单率的影响，甚至能智能捕捉关键词并即时推荐替代险种。第二个案例来自阿联酋（UAE）的一家重点银行。该行客户经理原本需要耗费 3 周时间在全网搜集企业损益表（P&L）与公开数据以撰写信用报告，Dyna.Ai 部署的信贷 Agent 将这一周期大幅压缩至数分钟以内，同时显著提高了数据精度。不过 Tomas 也客观指出，面向内部员工效率提升的解决方案（如信贷助手），其投资回报率（ROI）通常不如直接面向外部客户、拉动业务增长的交互型 AI Agent 那样立竿见影。 - **日本保险公司案例**：利用 AI 100% 执行标准化话术的特性，实现 Champion-Challenger 脚本对比测试，大幅优化成单率与险种推荐准确度。 - **阿联酋银行案例**：实现跨网络数据抓取与 P&L 财务分析自动化，将长达 3 周的报告准备期压缩至数分钟。 - **ROI 差异考量**：面向内部员工的效率类 Agent 虽然节省工时，但面向前端客户的体验与销售类 Agent 带来的直接商业回报往往更高。 --- ### [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) 香港案例：知识管理体系与四阶段渐进式落地路径在展示香港金融机构的知识管理系统（Knowledge Base）时，Tomas 详细拆解了银行在面对严格治理约束时，如何分步将生成式 AI 引入客户体验（CX）全流程。他指出，银行无需一步到位直接让 AI 与终端客户对话，而是可以通过四个渐进阶段逐步建立组织信任并完成内部合规审批。 ```mermaid flowchart LR accTitle: 银行客服 AI 渐进式部署四阶段 subgraph Step1["阶段 1: 辅助查询"] S1["内部知识库 Chatbot<br/>辅助坐席检索答复"] end subgraph Step2["阶段 2: 实时提炼"] S2["识别客户问题<br/>屏幕即时提示要点"] end subgraph Step3["阶段 3: 完整拟稿"] S3["生成完整回答内容<br/>人工坐席直接朗读"] end subgraph Step4["阶段 4: 自主交互"] S4["全自动化合规接管<br/>AI 直接与客户对话"] end Step1 --> Step2 --> Step3 --> Step4 ``` 香港机构最初从内部知识库起步，将分散的培训教材与产品手册沉淀至 AI 问答平台中，不仅客服能用，还能扩展为全员共享的内网知识引擎，免去在维基（Wiki）和混乱文档中翻找的繁琐。当系统运行一段时间且坐席几乎不再对 AI 的建议答案做任何修正时，法务与风控部门便拥有了充分的置信数据，得以批准 AI 直接面向终端客户提供服务。在演讲尾声，Tomas 幽默地展示了屏幕上的个人专属二维码：“这是一个很有趣的电子名片工具，大家扫码后可以给我打电话，直接与我的数字人 Avatar 对话交流。我平时聊天非常风趣，大家觉得无聊时随时欢迎找我的 Avatar 聊天。”随后他向全场观众致谢并结束了本次分享。 - **员工培训成本骤降**：当 AI 能够直接在屏幕上为坐席生成提纲或完整讲稿时，新员工的入职培训周期大幅缩短，人力成本显著降低。 - **全行级知识底座**：知识库不仅服务于前端客服与销售，还能向中后台泛化为全行统一的智能检索中枢。 - **互动名片**：现场提供包含个人 Avatar 语音交互功能的数字二维码，供参会者现场体验 Dyna.Ai 的人机交互技术。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 挡在大家与楼下酒吧开放之间的最后一个人 — Tomas Skoumal > [00:04](https://www.youtube.com/watch?v=wog7P677xMI&t=4s) 希望用 AI 直接与客户沟通 — Tomas Skoumal > [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 全量历史记忆 — Tomas Skoumal > [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 重新解释一遍 — Tomas Skoumal > [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 我的信用卡为什么刷不了？ — 客户 > [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 先生，我非常理解您遇到的问题，但为了能够帮您处理，我首先需要核实您的出生日期（或核对您在我们名下的产品信息）。 — Tomas Skoumal > [04:13](https://www.youtube.com/watch?v=wog7P677xMI&t=253s) 严格甚至有些不近人情 — Tomas Skoumal > [08:08](https://www.youtube.com/watch?v=wog7P677xMI&t=488s) 这份报告是生成给客户经理本人的，而不是直接越过人工提交给银行系统的。 — Tomas Skoumal > [16:10](https://www.youtube.com/watch?v=wog7P677xMI&t=970s) 这是一个很有趣的电子名片工具，大家扫码后可以给我打电话，直接与我的数字人 Avatar 对话交流。我平时聊天非常风趣，大家觉得无聊时随时欢迎找我的 Avatar 聊天。 — Tomas Skoumal ## Mentioned Tomas Skoumal, Dyna.Ai, Ai4, ASR（语音识别）, TTS（语音合成）, LLM（大语言模型）, Generative AI（生成式 AI）, Discriminative AI（判别式 AI）, Champion-Challenger 赛马机制, P&L（损益表）, KYC（了解你的客户）, 日本保险公司, 阿联酋重点银行, 香港金融机构 ## Video description Tomas Skoumal, Chairman & Co-Founder, Dyna.Ai Abstract AI is fundamentally reshaping the future of banking — are you ready to lead this transformation? Banks today face the dual challenge of enhancing both customer experience and employee productivity. At Dyna.Ai, we deliver enterprise-grade, human-like AI solutions that drive intelligent transformation across daily banking operations. In this session, we will share real-world data and success stories showing how we co-develop practical AI applications that deliver measurable impact for both customers and frontline teams. Join us to discover how banks can make AI a core competitive edge — and lead the next wave of smart, scalable banking innovation. [Source](https://www.youtube.com/watch?v=wog7P677xMI)

Source quotes

[0:04] Tomas Skoumal: 挡在大家与楼下酒吧开放之间的最后一个人 [0:04] Tomas Skoumal: 希望用 AI 直接与客户沟通 [4:13] Tomas Skoumal: 全量历史记忆 [4:13] Tomas Skoumal: 重新解释一遍 [4:13] 客户: 我的信用卡为什么刷不了？ [4:13] Tomas Skoumal: 先生，我非常理解您遇到的问题，但为了能够帮您处理，我首先需要核实您的出生日期（或核对您在我们名下的产品信息）。 [4:13] Tomas Skoumal: 严格甚至有些不近人情 [8:08] Tomas Skoumal: 这份报告是生成给客户经理本人的，而不是直接越过人工提交给银行系统的。 [16:10] Tomas Skoumal: 这是一个很有趣的电子名片工具，大家扫码后可以给我打电话，直接与我的数字人 Avatar 对话交流。我平时聊天非常风趣，大家觉得无聊时随时欢迎找我的 Avatar 聊天。

Source quotes

[0:04] Tomas Skoumal: enough is enough [1:33] Tomas Skoumal: Changing an enterprise bank's governance framework proves to be a far more grueling hurdle than solving the underlying engineering problems. [4:12] Tomas Skoumal: Building an effective customer-facing voice agent requires more than just calling a standard Large Language Model. [8:04] Tomas Skoumal: Because the agent compiles the dossier for the RM rather than submitting it directly to credit committee systems, the human relationship manager retains full legal responsibility to review, sign, and submit the proposal [10:45] Tomas Skoumal: Because the legally and regulatory responsible party remains a human being rather than the algorithm, bank legal and compliance departments can comfortably approve AI implementations. [10:45] Tomas Skoumal: While an AI might make unintentional, technical errors, humans often introduce intentional biases to push deals through. [13:25] Tomas Skoumal: every script which is given, it's followed 100% [13:25] Tomas Skoumal: customer-facing conversational solutions typically yield a faster and higher overall return on investment due to their direct impact on top-line revenue generation. [16:10] Tomas Skoumal: call me and talk to my avatar

Original transcript

[0:04] Hello everybody. I will I'm the last one [0:07] standing between you and the bar opening [0:09] downstairs. So I will try to be brief. I [0:11] will try to make it bit of more [0:13] practical than than than the you know [0:16] those sophisticated uh presentation [0:18] which you definitely went through today [0:20] like uh I will talk more like about the [0:22] use cases or like what I what I believe [0:25] uh from perspective of X banker uh what [0:29] might be the good use for for uh for the [0:32] companies how to use AI and and what is [0:35] like for the companies which are running [0:37] the AI what the finance like and or [0:40] don't like on on the AI. So yeah, I was [0:43] working for bank for 20 years uh plus [0:46] and then decided uh that uh you know [0:50] enough is enough and and and then uh you [0:52] know started the AI company. what what [0:55] really AI is now uh changing in in [0:58] banking is I would say like a like a [1:01] both let's say two perspective uh and [1:04] and the one perspective is like how they [1:07] deal with the customers or the second [1:09] perspective is like how they do the [1:11] stuff internally [1:13] even though internally they used to use [1:16] this uh old-fashioned machine learning [1:18] for quite few years the this generative [1:21] AI really comes to perspective on on on [1:23] being able to communicate with the [1:25] customers or let's say produce much more [1:28] selective or much more uh sophisticated [1:31] solutions. You know what what I I find [1:35] out in in let's say last year and a half [1:37] of of approaching the customers in a [1:40] financial industry is that when you come [1:42] to them every single company starts with [1:45] we would like to talk to our customers [1:47] through AI for two or three reasons. [1:50] First of all it will be cheaper. Second [1:52] of all, if we combine that with some [1:55] additional information we know about [1:57] customer and we are able to kind of keep [1:59] the script standard or keep the script [2:02] updated based on the result, we will [2:04] improve the sales. And this is a kind of [2:08] rule which applies to any finance. If [2:11] you start talking to them first time [2:13] about like your solution or first time [2:15] about like what the AI can do for them [2:17] and only the smaller ones will stay on [2:20] that top line which is the customer [2:22] experience agents such as sales such as [2:25] servicing such as collections. uh but [2:29] when it comes to the big institution [2:31] then they realize and you've heard many [2:34] of that in this previous forum here [2:36] talking about regulation that you know [2:38] the changing the governance of an [2:40] institution it is uh it is much bigger [2:44] challenge than the technical [2:45] implementation and then it kind of moves [2:48] down to the second option of the AI [2:51] which is let's say uh can we call like [2:53] an employee experience or internal [2:56] agents uh typically producing some some [3:00] reports uh but not the standard report [3:02] like of a numbers but reports with the [3:04] suggestions standardized report of of of [3:07] recommendation of a loan uh like a [3:10] credit agent. So, so those are let's say [3:13] the solution which are maybe not doesn't [3:15] have so much of the return and [3:18] investment but still keep the uh uh [3:21] implementation of an AI into institution [3:24] which then helps to understand and and [3:27] evolve. So I would say my year and a [3:30] half experience of trying to sell AI [3:32] solution is that smaller the institution [3:34] is the bigger the likelihood is that [3:36] they will be ready like mentally and and [3:39] governance wise ready to go for the for [3:42] the customer experience agent while the [3:46] bigger the constitution is that will [3:48] most likely end up on the employee [3:51] experience agent to to kind of start the [3:53] way start the journey do a lot of [3:56] investment which maybe from retro [3:58] internal investment doesn't make so much [3:59] sense like a customer experience agent [4:02] uh but does make sense from perspective [4:04] of moving that step forward and start [4:07] building the internal knowledge of an [4:09] AI. So how the typical agent uh pretty [4:13] much looks like uh I I guess like you [4:15] probably seen like much more [4:17] sophisticated solution this like a [4:19] charts to describing the solution but [4:21] here let's stick with the what is in the [4:25] in the middle is the LLM but it's it's [4:27] not really what plays the the the only [4:31] role of of solution. It's like obviously [4:33] you have to have the ASR to understand [4:36] the customer well and and for the bigger [4:38] institution has to be ASR which is [4:40] possible to put on premise uh not not [4:43] some cloud solution like a public cloud [4:45] solution and obviously the text to [4:47] speech at the end to be understanding [4:48] but in the middle the LLM has to work [4:51] with the memories uh of of all the [4:54] previous calls that that's what the AI [4:57] brings additional over the standard [5:00] solution with the human operators [5:02] as you can feel yourself that when you [5:05] are calling you are called by [5:07] institution like two times three times [5:09] with a different operator or you are [5:11] calling the institution then then you [5:13] have to reexplain yourself so so the AI [5:16] to to be able to to kind of demonstrate [5:19] the the benefits and and to be able to [5:23] to kind of fight the human operators it [5:26] has to kind of carry that memories to be [5:28] able to remember every single historical [5:30] discussion [5:31] between the customer and the [5:33] institution. Obviously that requires a [5:35] second step which is against the typical [5:38] way how the institution are installing [5:40] the solution and that second step is [5:43] that you have to whole entire process [5:45] because the moment you do only let's say [5:47] the sales part or sales part of one [5:49] product then the memories what the AI is [5:52] is having about the customer is only [5:54] about the calls of that particular [5:56] product or have that particular part of [5:59] the process and then obviously there [6:01] might be the calls which are happening [6:03] aside side about different products or [6:05] about different uh in a different part [6:07] of the process of a customer. So that [6:10] that that only builds on as as the [6:12] implementation is happening because [6:14] normally the institution tend to [6:15] implement the AI solution for let's say [6:18] only sales of upgrade of a credit card [6:20] limit and then only like for the new [6:22] credit cards. So then then obviously the [6:25] information which the memories are [6:26] having and then the customer experience [6:29] is improving as the rollout is [6:31] happening. Then obviously the AI agent [6:34] has to be uh trained through something [6:36] which is called workflow management and [6:38] it's kind of typical when when the in [6:40] financial institution you are training [6:42] the the the operator how to sell the [6:45] product or how to service that [6:48] particular issue then you have to not [6:51] let LLM when you have an agent to answer [6:54] the question freely but you have to [6:57] drive the answer regardless what [6:59] customer is asking for or saying [7:02] typically regardless that the customer [7:04] is shooting the question my credit card [7:06] is not working the first you have to [7:08] verify the customer so the workflow is [7:10] managing the flow the perspective yes [7:12] sir I understand your issue however to [7:15] be able to help you I first have to ask [7:17] you what is your date of birth or how [7:19] many products you have with us whatever [7:21] the compliance rules for verifying [7:24] customers verifying customers are so [7:26] those rule have to follow so that's [7:28] typically that you know that you and [7:30] then you might let your LLM to give the [7:32] some basic answer uh before getting the [7:35] second step. But technically the [7:37] workflow has to manage uh manage the the [7:40] flow of answering regardless of of [7:42] question of a customer. It it sounds [7:45] brutal but I I don't mean regardless but [7:47] I mean I mean that you know the answer [7:50] should somewhat please the customer but [7:52] yet goes through the flow and and manage [7:55] the the understanding. So this is uh I I [7:57] I don't have much time to go very deep [8:00] through all the solutions but uh through [8:02] all the parts of it but I I would like [8:04] to mention that if I will go now to the [8:06] next slide it will look almost the same [8:08] and that's the internal agents the [8:10] customer experience agent what only [8:13] changed here is like pretty much like [8:14] the name of the agents on the top like [8:16] that they are kind of doing the internal [8:18] job like an HR you know like the like [8:21] the HR agent which might be helping the [8:23] HR to hire the people or or the agent of [8:26] credit analyst which is analyzing the [8:29] corporate loans and and and then helping [8:31] the the the bank of uh solving the [8:34] problem of that the RM instead of going [8:37] outside and and finding the customer or [8:40] helping the customer does nothing but [8:42] for two days in a row setting and [8:44] preparing the document for credit [8:45] department uh to kind of submit the [8:48] application for the loan. So that's [8:50] that's what that can happen. But what [8:52] happened there also on the top is the [8:54] tool which is parsing the document. Uh [8:57] because obviously when you have the [8:58] customer experience agent you don't [9:00] really work with very sophisticated and [9:02] very long documents. However, when you [9:05] come to internal agent, then you already [9:07] are working with the documents which are [9:10] you know P&L or or or or account [9:14] statement which you have to parse [9:16] through LLM to understanding every [9:19] single part of it and then update the [9:21] reports uh which is prepared for the [9:23] credit for that example. So otherwise it [9:26] looks exactly the same. That's that [9:28] comes to the point what I said at the [9:30] beginning is that the investment for the [9:32] fine solution will be similar. They have [9:35] to have the LLM most likely as well on [9:37] premise for the regulatory perspective. [9:39] They have to have all the solution. [9:41] However, then the agent is not talking [9:44] to the customer but it's preparing some [9:46] some task for uh supporting the [9:49] employees which is easier for the bigger [9:52] institution to adopt because uh [9:54] obviously it doesn't change the [9:55] governance of the bank because uh just [9:57] just an example to explain what I mean [10:00] by that is that if as of now I am an RM [10:04] and I'm supposed to prepare the document [10:06] for the risk department to evaluate the [10:09] company pay a for the loan of $5 [10:12] million. [10:13] Now I'm doing it myself through standard [10:16] research. [10:18] When the agent is done then the workflow [10:20] doesn't mean anything there in this in [10:21] this chart that I mean like the [10:23] questions which I would ask me in [10:26] preparing document the workflow would [10:28] ask exactly the same question and [10:30] prepare the document for me and that's [10:32] important. It's for me it's not for the [10:35] institution. So then I still I'm [10:38] responsible for reviewing the document [10:40] and sign it and send it to the risk [10:41] department and that's kind of mean that [10:45] I'm not changing the governance of the [10:47] bank the regulatory legally responsible [10:50] is still a human being and therefore [10:53] it's easier to implement the AI solution [10:56] for the finance solution where the [10:59] responsible person remains unchanged. So [11:01] the legal and compliance of the bank are [11:03] happy. However, instead of me sitting [11:06] and preparing it for like 3 days, it's [11:09] prepared by the AI agent probably by is [11:11] in 15 minutes. Uh, and then I will spend [11:14] like half an hour, 1 hour to read the [11:16] document clearly and and then sign I [11:19] agree with everything what is stated [11:21] here or I can go and change backwards [11:24] data and then obviously risk is getting [11:27] additional information because me as an [11:29] RM I'm interested of of selling the [11:32] loan. uh so then all of a sudden there [11:34] is a comparison what I changed and [11:37] whether I'm changing it in favor of a [11:38] customer. So then obviously then the [11:41] discriminative AI which will come on the [11:43] next slide is like that you know [11:46] discriminative AI like the old fashioned [11:48] machine learning has additional [11:50] information. So what I'm trying to say [11:53] here is that uh you know that AI is [11:57] having like several kind of ways of [12:00] usage but the old fashioned machine [12:02] learning it can be benefiting from the [12:04] generative AI because as of now as I [12:07] said like the AI will have these [12:09] memories and remember all the calls and [12:10] discussion with the customer but that [12:13] obviously is a set of data which can [12:15] then be used in old-fashioned machine [12:17] learning and and help the help the [12:19] customer to be satisfied by by [12:22] evaluating every single core and score [12:24] every call and feed in the database of [12:27] the company in the case of external [12:29] agent. In the case of internal agent [12:32] again if I produce the report as an AI [12:35] and then RM changes something that [12:37] changes what RM did is an additional [12:40] information for the risk department to [12:42] make a decision about the loan because [12:44] AI make can can make mistakes but they [12:47] are unintentional mistakes. There will [12:49] be the mistakes like a technical mistake [12:51] but me as a human being I can make many [12:54] intentional mistake that I will try to [12:56] interpret the profit slightly better to [12:59] show that the loan should be granted. So [13:02] while if I after AI change two three [13:05] lines to make it sound more positive [13:07] that's an additional information for the [13:08] discriminative AI as another set of data [13:11] coming into that to uh to kind of uh [13:17] improve the approval process of an [13:21] institution. Uh so that's uh pretty much [13:24] the stories and now let's let me talk [13:26] about some typical examples. five [13:29] minutes. No. Uh so you know then I will [13:33] not hold you from the bar. So then the [13:35] the typical example of of a solution is [13:38] like the the blue and light blue shape [13:41] on the chart it's demonstrating which [13:43] country it was implemented. So that's a [13:46] map of uh of the Japan. So for like [13:49] insurance company in Japan uh there is [13:52] like improvement of uh uh the the reach [13:55] rate and improvement of communication [13:58] and obviously improvement of of the [14:00] sales simply through uh perspective of [14:04] uh of being able to have standardized [14:07] calls and being able to have a champion [14:10] challenger on perspective of the calls [14:14] because AI has like another benefit. I [14:17] mean like any [14:19] uh machine has a benefit compared to [14:21] human that every script which is given [14:24] it's followed 100%. Uh human brings all [14:28] that kind of variety in human operator [14:31] brings all that variety into the phone [14:33] call. So it's very difficult to evaluate [14:35] script A against script B. While if it's [14:37] done through the AI and you do the [14:40] workflow, you change the bit of [14:42] workflow, you change the tone that can [14:44] see like how customers are reacting and [14:47] it can improve the sales or can improve [14:49] the customer satisfaction because you [14:51] can easily recognize some words which [14:54] based on which you understand that the [14:55] customer is not interested for that type [14:57] of insurance but a different type of [14:59] insurance. So that that that's one [15:01] example. The second example is the [15:04] credit agent which I was talking about [15:07] like a bit more I guess the map you know [15:10] let's take it a quiz over here but it is [15:12] map of UAE. Uh so that's that's like as [15:15] I said like the agent like the RM who is [15:18] supposed to mostly go and sell. So for [15:21] the theme bank in in Middle East, we we [15:24] prepare something which is from uh three [15:27] weeks of preparing this document and [15:28] collecting and searching the internet [15:30] and collecting all those P&Ls created [15:32] report in just few minutes and obviously [15:34] with much better accuracy and and and [15:37] and and with the with the perfect [15:39] resolution of uh of an employee [15:42] solution. However, as I said, revenue [15:45] for the company like revenue on [15:46] investment, the return on investment for [15:48] this type of solution is not not as [15:50] great as the previous one when you talk [15:53] to the customer. Uh there is a video but [15:56] since I have few minutes I will skip [15:58] over it and let's see if we if I can go [16:00] back afterwards. But there's a video of [16:02] like how the the solution of reading of [16:04] a documents is working compared to the [16:07] uh compared to the other solutions. uh [16:10] and then the another solution here this [16:12] is uh Hong Kong and this is uh like [16:16] let's say knowledge base or knowledge uh [16:19] management uh and and that's uh mostly [16:23] like again another solution for customer [16:25] experience you can see that you know I [16:27] have more examples for customer [16:28] experience agent specifically because [16:30] the big fine solution really like that [16:33] and this is more like that if you kind [16:35] of have a intention to to to talk to the [16:39] customers through AI I but as as a [16:42] governance perspective you need the [16:43] implementation time then you can pace [16:46] that implementation into several steps. [16:50] Uh like the very first step is that you [16:52] will take all documents which you might [16:55] have as uh as as as as training [16:58] materials as product descriptions and [17:01] and then support your sales team simply [17:04] that they during the call then they can [17:06] when the customer ask the question that [17:08] they can use the chatboard of an [17:10] internal knowledge to help them to get [17:12] the better answer. Then the second [17:15] perspective and this is the solution [17:17] like a knowledge base. The next step of [17:18] that is that when the customer ask the [17:21] question, it might already the agent [17:23] might already understand it and put on a [17:25] screen uh for the operators like a [17:28] bullet points what should be the answer [17:30] looking like that. The next step might [17:33] be that you already kind of have an [17:36] answer fully written for the customer [17:39] and then then you for the operator and [17:42] operator is reading it through. So then [17:45] technically it helps you from [17:47] perspective of training because then [17:48] obviously like you are coming to the [17:50] perspective and the operator is already [17:53] seeing the answer then the training [17:55] requirements for the for the newcomers [17:57] is is uh by far by far faster and and [18:01] less less costly. And obviously the next [18:04] the final step is if you already have [18:06] like a one months of uh of of kind of uh [18:09] testing that everything what was put on [18:11] a screen by the by the agent AI agent is [18:15] only simply [18:19] read by the operator to the customers [18:22] with no corrections no no changes that [18:25] can help you in in the in the company to [18:27] kind of get an internal approval for the [18:30] for the governance change and that the [18:32] AI I can directly read it to the [18:34] customer straight. Yeah. So that's uh [18:37] that's let's say the pacing of an [18:39] implementation and this is the solution [18:40] of the let's say step one for the for [18:42] the one fination in Hong Kong where we [18:45] started with exactly that knowledge base [18:47] which doesn't necessarily need to stand [18:49] only for the operators. It can then [18:52] extend for for entire in uh entire [18:54] company. So then you have nothing but [18:56] the let's say chatbot for the for your [18:59] employees to to gain like all the [19:02] information they need otherwise they [19:04] would have to search through some wiki [19:05] or or any any internal the documents. So [19:09] that's uh let's say what what we do and [19:12] and uh and then what we can have as a [19:16] solution and here you can whoever wants [19:19] to can scan the QR code at the end and [19:21] uh talk to me virtually about our [19:24] solution or about us. It's a it's a fun [19:26] tool like of a business card where you [19:29] can call me and talk to me uh talk to my [19:32] avatar. Yeah. So [19:35] uh I'm I'm I'm very fun to talk to. So [19:38] whenever you feel bored you can just do [19:39] so. Yeah. [19:41] >> Thank you. Thank you so much and uh [19:43] enjoy the rest of the conference. Yeah.