RSM Tells CIOs to Scrap Task-Level AI for Multi-Agent Orchestration

Orchestrate, Don't Automate: How CIO's Build AI Systems That Scale, Decide, and Execute with RSM US — Ai4

Ai4

Chasing 100% automation inflates operational unit costs on edge cases, requiring enterprises to deploy multi-agent orchestration with strict human-in-the-loop risk thresholds to secure scalable returns.

AI-assisted summary of Orchestrate, Don't Automate: How CIO's Build AI Systems That Scale, Decide, and Execute with RSM US — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 21, 2026, this presentation features Tyler Howard, Director at RSM US LLP. According to the description, the session provides a practitioner-level blueprint for CIOs and CTOs to move beyond single-model point solutions toward production-grade, multi-AI orchestration platforms that unify enterprise data, incorporate human oversight, and scale across operational functions.

## TL;DR Tyler Howard argues that enterprise AI initiatives frequently stall when organizations pursue isolated task automation and single-use point tools instead of an intentional orchestration platform. Drawing on case studies from high-volume healthcare operations and multi-agent sales lead prioritization, Howard illustrates how activating dormant historical data and coordinating specialized models dramatically enhances human decision-making. He cautions against the costly trap of pursuing 100% touchless automation, advocating instead for strict financial thresholds, role-based observability, and granular transaction-level cost tracking. To achieve sustainable production deployment, organizations must anchor their AI architectures in four core tenets: outcome-driven design, grounded data foundations, proactive governance, and reusable, scalable orchestration.

## Atmosphere & style Delivered from the stage at a Las Vegas industry conference, the talk carries the pragmatic, no-nonsense tone of an experienced implementation consultant speaking directly to enterprise technology leaders. Howard cuts through foundational model hype to focus on the unglamorous realities of data plumbing, unit economics, and operational risk. The delivery is structured, analytical, and authoritative, offering a clear-eyed roadmap that balances ambitious enterprise transformation with practical governance constraints.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) Organizations must move beyond isolated point solutions and bots, which create functional silos, toward an enterprise-grade orchestration platform that unifies workflows across business boundaries. - [01:39](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=99s) Treating AI as ad-hoc task automation traps teams in an inefficient cycle of constantly rebuilding data plumbing from scratch, resulting in zero scalability and wasted capital. - [02:05](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=125s) Enterprise AI initiatives typically stall across four critical pillars: poor data quality and architecture, cloud security and governance policies, technical talent gaps, and a lack of long-term sustainable funding. - [04:49](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=289s) Prompts and standalone models cannot scale on their own; enterprise value is unlocked only by orchestrating data models, prompt layers, business logic, and automated checks into end-to-end workflows. - [05:30](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=330s) Orchestration eliminates manual cross-system swivel-chairing by synthesizing disparate data and dynamically populating downstream operational work queues across multiple teams. - [08:27](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=507s) Unlocking dormant enterprise assets, such as years of untranscribed call recordings combined with CRM data, provides the context necessary to power real-time predictive lead scoring and next-best-action guidance. - [10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) AI delivers the highest ROI when designed to augment human domain expertise and accelerate onboarding, allowing novice workers to perform like top-tier revenue generators using structured playbooks. - [10:45](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=645s) Multi-agent systems can coordinate engagement strategies based on customer demographics, routing younger audiences to digital channels like SMS and reserving voice outreach for high-value prospects. - [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) Automating 100% of transactions is an unrealistic goal; systems should enforce strict financial and risk thresholds that route edge cases to humans with pre-synthesized case context. - [12:55](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=775s) Long-term production success requires persona-based visibility, delivering daily task flags for operational analysts, team velocity metrics for directors, and aggregate ROI and compliance tracking for CIOs. - [15:08](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=908s) Attempting to automate the final tail of complex exceptions creates diminishing returns; organizations should automate high-volume processes while deliberately routing edge cases to human operators. - [15:35](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=935s) Managing operational unit economics requires tracking granular runtime costs per transaction and per task, rather than fixating on foundational model hype or broad annual platform licensing fees. - [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) A scalable enterprise AI architecture rests on four core pillars: being outcome-driven, remaining data-grounded, embedding governance and auditability by design, and orchestrating modular components at scale.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) Moving Beyond Point Solutions to Enterprise Transformation

Tyler Howard, Director of the Custom Applications AI team within RSM US LLP's consulting practice, introduces his focus at the tail end of a long conference week in Las Vegas. While RSM is broadly known as an audit, tax, and advisory firm, Howard's specialized practice concentrates on designing and building complex, custom AI workflows rather than delivering pre-packaged point solutions or standalone bot products. Drawing on insights from RSM's annual client survey, where artificial intelligence took center stage, Howard points out a glaring disconnect across industry leadership: while enterprise appetite for transformation is exceptionally high, organizations routinely struggle to bridge the gap between initial interest and scalable enterprise execution.

The core challenge stems from treating AI as an isolated, single-use-case tool. When organizations deploy one-off point solutions, they trap themselves in functional silos that fail to scale or cross team boundaries. Howard emphasizes that the objective of modern AI initiatives must shift toward creating an enterprise-grade orchestration platform that unifies workflows, drives ongoing engagement, and fundamentally transforms how the business operates.

- **Survey Context:** Annual RSM client survey highlighted strong executive appetite for AI, yet revealed persistent execution hurdles. - **Consulting Focus:** Building production-grade custom architectures and multi-system workflows rather than off-the-shelf point tools.

---

[01:39](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=99s) The Trap of Task Automation and Why AI Deployments Stall

Reflecting on the history of enterprise technology, Howard observes that organizations have been discussing task automation since the early 2000s—progressing through automated scripting and basic workflow tools without achieving true enterprise scalability. Deploying isolated point solutions forces teams into an inefficient cycle: whenever a new use case arises or an existing tool fails to meet the specific requirements of another group, engineering teams are forced to start from scratch. This "zero scalability, zero investment" dynamic wastes valuable capital because teams constantly rebuild the underlying data plumbing rather than leveraging an intentional, reusable data foundation.

```mermaid flowchart TD accTitle: Enterprise AI Deployment Foundation vs Stalling Factors subgraph StallingFactors["Core Roadblocks to AI Deployment"] A["Poor Data Quality & Pipeline Architecture"] B["Cloud Security & Policy Restrictions"] C["Talent & Operational Skill Gaps"] D["Lack of Long-Term Sustainability"] end

subgraph SolutionPlatform["Orchestrated Foundation"] E["Unified Enterprise Data Layer"] F["Shared Audit & Governance Framework"] G["Cross-Functional Scalable Architecture"] end

A --> E B --> F C --> G D --> G ```

Howard outlines the four critical pillars where deployments typically stall, warning that if an organization cannot confidently address all four, projects lose momentum, executive support, and organizational interest:

- **Data Layer and Data Quality:** Failing to construct a clean, reliable data architecture before attempting to deploy models. - **Security, Governance, and Policies:** Managing cloud environments, preventing proprietary data leakage, and monitoring the specific types of data moving through external models. - **Talent and Skill Gaps:** Navigating the difference between staff who have the appetite to engage with AI and specialists who possess the technical skills to maintain and manage solutions in production. - **Sustainability and Growth Confidence:** Maintaining organizational focus and funding beyond proof-of-concept experimentation.

---

[04:49](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=289s) The Orchestration Imperative: Connecting End-to-End Enterprise Workflows

While generative, predictive, and natural language models are widely adopted for daily tasks—such as internal HR policy chatbots or support query assistants—executives frequently express an "all-in mentality" demanding measurable business transformation and direct revenue generation. Howard argues that simple prompts and agents cannot achieve scalability on their own; value is unlocked exclusively through orchestration. Orchestration coordinates data models, prompt layers, business logic, and automated checks to deliver end-to-end workflow validation, bringing enterprises closer to reliable, touchless automation while maintaining critical checks and balances.

```mermaid flowchart LR accTitle: Multi-Team Orchestrated Healthcare Workflow subgraph Ingestion["Intake & Processing"] T["20M Annual Transactions"] --> M["Variable Data Models"] end

subgraph OrchestrationEngine["AI Orchestration Engine"] M --> D["Decision Logic & Context Synthesizer"] D --> H{"Human Review Threshold Met?"} end

subgraph Execution["Cross-Functional Routing"] H -- "Yes: High Risk / High Value" --> E["Human Escalation & Validation"] H -- "No: Touchless Execution" --> Q["15 Work Queues Across 4 Downstream Teams"] E --> Q end ```

To illustrate this cross-functional value, Howard shares a case study from a healthcare client. The project initially targeted a single workflow processing 20 million transactions in its first year. Instead of confining the AI to that initial silo, RSM engineered an orchestration layer that synthesized transaction outputs, restructured the underlying data, and dynamically populated 15 separate work queues across four distinct downstream operational teams. By extracting analysts from the friction of manually juggling four disparate internal and external core systems, the AI constructed an integrated narrative for each case, substantially improving decision accuracy and downstream speed.

- **System Interoperability:** Eliminates the need for operational analysts to cross-reference multiple disconnected systems to compile case histories. - **Multi-Model Adaptability:** Embraces variable model usage suited to specific tasks rather than locking the enterprise into a single rigid model or proprietary drag-and-drop platform burdened by inflexible licensing and hosting constraints. - **Human-in-the-Loop Safeguards:** Places human oversight at critical business thresholds and high-value revenue checkpoints to satisfy strict auditability and risk management standards.

---

[08:27](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=507s) Case Study: Transforming Sales Lead Orchestration and Unlocking Dormant Data

Howard turns to a detailed case study of a sales organization comprising several hundred licensed agents. The client was ingesting leads from multiple sources—including multi-channel marketing campaigns, email outreach, digital inbound funnels, and purchased third-party prospect lists. Historically, all incoming prospects were dumped into a generic pool, forcing agents to cold-call down a flat list with minimal context. When RSM evaluated the workflow, they recognized that the sales force possessed deep domain expertise, with many agents holding over 25 years of industry experience. Rather than attempting to replace human sales skills with AI, the goal was to arm these high-performing agents with comprehensive context and intelligence before they initiated a call.

RSM tackled the challenge by uniting 15 years of structured CRM data with external demographic indicators and uncovering a massive, untapped asset: 10 years of archived customer phone calls and outreach recordings that had never been transcribed or leveraged for machine learning. By transcribing and training models on this dormant audio data, the orchestration platform began analyzing incoming leads in real time, diagnosing customer intent, identifying product affinities, and surfacing the exact next best action for each licensed professional.

- **Empowering Expert Human Talent:** Focusing AI on data synthesis and contextual lead prioritization so seasoned agents can focus on high-touch client conversations. - **Activating Dormant Data Assets:** Converting 10 years of un-transcribed historical voice recordings into actionable training data and behavioral insights. - **Multi-Source Data Fusion:** Combining third-party purchased leads, inbound marketing signals, demographic details, and deep CRM histories into a prioritized lead orchestration pipeline.

[10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) Tailored Engagement and Multi-Agent Sales Orchestration

When engaging leads, determining the right contact channel is critical. Demographics dictate interaction preferences: making a cold phone call to a 22-year-old rarely succeeds because "they're going to hang up," whereas reaching them via marketing emails, text messages, or digital product information yields far better results. Conversely, for high-value or experienced leads, the orchestration engine equips sales representatives with rich, contextual guidance. Instead of operating with no targets or structure beyond a vague directive to sell products, reps receive an AI-generated structured playbook detailing the exact narrative to pitch, optimal call times, availability windows, and specific product recommendations.

This capability stems from a strong data foundation paired with continuous human-in-the-loop controls. Because customer interactions involve sensitive personal information, strict end-to-end data management ensures no proprietary data leaks externally. Crucially, the solution avoids relying on a single model or isolated prediction. It coordinates multiple autonomous agents working alongside human analysts. This system delivers immense ROI: new sales hires coming off the street with zero prior experience transition from earning a flat hourly rate to driving exponential revenue growth for the organization.

```mermaid flowchart TD accTitle: Multi-Agent Lead Orchestration Architecture subgraph INTAKE["Data Foundation & Intake"] D1["Demographic & Lead Data"] D2["Interaction History"] end

subgraph AGENTS["Multi-Agent Orchestration Layer"] A1["Channel Selector\n(Email / Text vs. Voice)"] A2["Timing & Script Generator\n(Story, Call Time, Focus)"] end

subgraph EXECUTION["Human-in-the-Loop Execution"] H1["Sales Analyst Playbook"] H2["Financial Threshold Checks"] H3["Customer Engagement & Revenue"] end

D1 --> A1 D2 --> A2 A1 --> H1 A2 --> H1 H1 --> H2 H2 --> H3 ```

- **Demographic Channel Matching:** Directs younger demographics toward automated SMS and email sequences while reserving voice calls for high-touch prospects. - **Dynamic Sales Scripting:** Provides novice reps with bespoke narratives, timing guidance, and targeted product focus. - **Multi-Step Agent Architecture:** Replaces brittle single-prompt models with specialized, cooperating agents. - **Fast Ramp-Up:** Transforms entry-level workers into top-performing revenue generators rapidly.

---

[12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) Risk Thresholds, Exceptions, and Persona-Based Visibility

Designing enterprise AI systems requires addressing risk, financial exposure, and inevitable workflow exceptions upfront. Automated systems should never claim to "handle 100% of your transactions 100% of the time perfectly"—edge cases and high-stakes scenarios will occur. Establishing strict financial thresholds ensures that whenever an action exceeds a specific monetary value, the system automatically routes the case to a human approval workflow, arming the reviewer with fully contextualized data to make an informed decision.

```mermaid flowchart LR accTitle: Enterprise AI Persona-Based Tracking subgraph AGENT["AI Orchestration Engine"] E1["Execution & Event Logging"] end

subgraph PERSONAS["Role-Specific Dashboards"] P1["Analyst View\nDaily tasks & operational flags"] P2["Director View\nWorkflow health & team velocity"] P3["CIO View\nSystem ROI & compliance posture"] end

E1 --> P1 E1 --> P2 E1 --> P3 ```

To prevent AI initiatives from stalling in isolated pilot stages, organizations must maintain holistic operational visibility structured around distinct user personas: - **Analyst Persona:** Daily operational insights, immediate flags, and contextual tasks. - **Manager / Director Persona:** Team throughput, approval queues, and workflow bottlenecks. - **CIO / Executive Persona:** Macro governance, aggregate ROI, cross-system performance, and total cost visibility.

Many organizations complete a proof-of-concept (POC) believing that simply checking security boxes is sufficient. However, long-term success demands active governance frameworks and clear people-centric processes to prevent teams from feeling hesitant or taking steps backward when transitioning into production.

---

[15:08](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=908s) Operational Unit Economics and the 100% Automation Trap

A common failure mode in enterprise AI adoption is obsessing over the latest foundational model without understanding operational unit economics. As usage scales, operational costs increase dramatically. Organizations must look past high-level annual software licenses or cloud hosting fees to track the granular cost per task and runtime cost per transaction.

Trying to automate the final tail of complex exceptions leads to diminishing returns and bloated operational spending. Rather than chasing a 100% automated output, organizations should identify high-volume, low-hanging fruit for automated execution and deliberately route edge cases to human operators. This intentional design keeps the payback period short and ensures cost structures remain sustainable.

| Metric Focus | Low-Maturity Approach | Scalable Orchestration Approach | |:--- |:--- |:--- | | **Automation Target** | 100% full automation (unrealistic & costly) | High-volume automation + intentional human-in-the-loop exception routing | | **Cost Tracking** | Broad annual hosting / SaaS platform fees | Granular runtime cost per transaction and unit cost per task | | **Model Selection** | Chasing the newest, largest model indiscriminately | Right-sized models and reusable prompt modules matched to task complexity | | **Exception Handling** | Undefined, causing workflow failures | Automated escalation triggered by financial and risk thresholds |

---

[16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) Blueprint for Scaling: Four Core Tenets of Enterprise Orchestration

Building enterprise AI that scales reliably requires shifting focus from isolated task automation to an overarching architectural framework. Four core pillars define this blueprint:

```mermaid flowchart LR accTitle: Four Pillars of Enterprise AI Orchestration P1["1. Outcome-Driven\nAligns personas & secures executive buy-in"] P2["2. Data-Grounded\nCloud integrations as real-time assets"] P3["3. Governed by Design\nAuditability & risk management from Day 1"] P4["4. Orchestrated at Scale\nReusable prompts, agents, and modular workflows"]

P1 --> P2 --> P3 --> P4 ```

1. **Outcome-Driven:** Ground every project in measurable business results and clear personas. Delivering concrete value secures stakeholder buy-in, user adoption, and ongoing executive sponsorship. 2. **Data-Grounded:** Leverage modern cloud platforms and integration layers so enterprise data acts as a continuous operational asset rather than a friction point, delivering information at the speed and time required. 3. **Governed by Design:** Embed governance, tool compliance, operational observability, and auditability directly into the platform architecture from day one rather than treating compliance as a late add-on. 4. **Orchestrated at Scale:** Build systems using modular, reusable building blocks—including specialized agents, master prompts, and standardized workflow templates—that can be repurposed across departments from sales to finance and operations.

## Notable quotes > Most organizations are automating tasks. The leaders are orchestrating intelligence. — Tyler Howard

> This session is not about a single tool or industry. It’s a blueprint CIOs and CTOs can apply as AI moves from curiosity to core enterprise capability. — Tyler Howard

> [01:39](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=99s) zero scalability, zero investment — Tyler Howard

> [04:49](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=289s) all-in mentality — Tyler Howard

> [10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) they're going to hang up — Tyler Howard

> [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) handle 100% of your transactions 100% of the time perfectly — Tyler Howard

> [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) Outcome-Driven — Tyler Howard

> [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) Data-Grounded — Tyler Howard

> [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) Governed by Design — Tyler Howard

> [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) Orchestrated at Scale — Tyler Howard

## Mentioned Tyler Howard, RSM US LLP, Ai4, Las Vegas

## Video description Tyler Howard - Director, RSM US LLP

Most organizations are automating tasks. The leaders are orchestrating intelligence.

In this session, I will share how we designed and delivered a production grade AI orchestration platform that moves beyond single model point solutions to a fully integrated, multi-AI architecture. One that unifies data, reasons across context, inserts human judgment where required, and executes decisions reliably across real enterprise workflows.

While customer engagement is a powerful example, the architecture and lessons are broader. This is a practitioner level look at what separates proof of concept from production including how to design data foundations for orchestration, not experimentation, where automation breaks down and human oversight must remain and the architectural patterns that allow AI systems to scale across functions — from customer engagement to operations, finance, and IT.

This session is not about a single tool or industry. It’s a blueprint CIOs and CTOs can apply as AI moves from curiosity to core enterprise capability.

[Source](https://www.youtube.com/watch?v=uDbeZWPFqr8)

中文简报

RSM建议CIO弃用任务级AI，转向多智能体编排 追求100%自动化会在边缘用例上推高运营单位成本，企业需部署具备严格“人在回路”风险阈值的多智能体编排，以确保可扩展的回报。 Tyler Howard认为，当企业追求孤立的任务自动化和一次性单点工具，而非有规划的编排平台时，企业AI项目往往会陷入停滞。结合大规模医疗运营和多智能体销售线索优先级排序的案例，Howard展示了激活沉睡的历史数据并协调专业化模型如何能大幅提升人类决策效率。他告诫企业警惕追求100%免干预自动化的昂贵陷阱，主张建立严格的财务阈值、基于角色的可观测性以及细粒度的事务级成本追踪。为实现可持续的生产级部署，企业必须将AI架构建立在四大核心原则之上：结果驱动的设计、扎实的数据基础、前瞻性的治理，以及可复用、可扩展的编排能力。 企业必须超越导致职能孤岛的孤立单点方案与机器人，转向能够跨业务边界统一工作流的企业级编排平台。 (0:04) 将AI视为临时的任务自动化会使团队陷入不断从零重建数据管道的低效循环，导致毫无可扩展性并浪费资金。 (1:39) 企业AI项目通常在四大关键支柱上面临停滞：数据质量与架构薄弱、云安全与治理策略限制、技术人才短缺，以及缺乏长期可持续的资金支持。 (2:05) 提示词和独立模型自身无法实现规模化；唯有将数据模型、提示词层、业务逻辑和自动化检查编排为端到端的工作流，才能释放企业价值。 (4:49) 编排通过综合分散的数据并动态填充跨多个团队的下游运营工作队列，消除了跨系统的人工“转椅式”操作。 (5:30)

Reading summary

Tyler Howard of RSM US argues that enterprise AI initiatives stall when treated as isolated point automations, contending that scalable transformation requires an orchestrated platform combining reusable data foundations, multi-agent workflows, and strict operational governance. The Limits of Isolated Task Automation (1:39) Howard observes that despite high executive interest highlighted in RSM's annual client survey, organizations routinely struggle to bridge the gap between initial experimentation and production-grade execution. Treating AI as a one-off point solution creates functional silos that force engineering teams to rebuild foundational data plumbing from scratch for every new use case, resulting in wasted capital. According to Howard, enterprise deployments stall when leaders fail to address four critical vulnerabilities: unreliable data layer architecture and data quality; security, cloud governance, and proprietary data leakage risks; the talent gap between staff interest and production engineering capability; and fading organizational funding once proof-of-concept tests conclude. Cross-Functional Workflows and Enterprise Orchestration (4:49) Howard asserts that standalone prompts and agents cannot achieve scalability on their own; enterprise value is unlocked only by orchestrating data models, business logic, and automated checks across end-to-end workflows. In a healthcare case study processing 20 million annual transactions, RSM engineered an orchestration layer that synthesized transaction outputs and dynamically populated 15 distinct work queues across four downstream operational teams. Instead of requiring analysts to manually cross-reference four disconnected internal and external systems, the platform generated consolidated case narratives. Howard notes that this approach improves decision speed while avoiding lock-in to single, rigid vendor models or proprietary platforms. Unlocking Dormant Data Through Multi-Agent Systems (8:27) In a sales case study of several hundred licensed agents, RSM combined 15 years of structured CRM data and demographic indicators with a previously untapped asset: 10 years of archived, un-transcribed customer call recordings. Rather than replacing human sales specialists, the orchestration platform uses cooperating autonomous agents to diagnose intent and surface next best actions. The system dynamically matches outreach channels to demographic preferences—routing younger leads to text and email sequences while reserving voice calls for high-touch prospects—and generates structured pitch playbooks and timing windows, allowing entry-level sales hires to rapidly transition into top revenue generators. Operational Unit Economics and Human Exception Handling (15:08) Howard cautions against falling into the trap of attempting 100% touchless automation, warning that resolving complex edge cases with AI produces diminishing returns and inflated operational spending. Organizations must track granular runtime costs per transaction and per task rather than focusing exclusively on annual SaaS fees. To maintain sustainable unit economics, enterprises should automate high-volume processes while intentionally routing edge cases and high-stakes actions exceeding defined financial thresholds to human reviewers. Howard emphasizes that operational visibility must be tailored to specific personas, providing analysts with operational flags, directors with queue velocity metrics, and CIOs with macro ROI and compliance oversight. Howard concludes that sustainable enterprise AI requires four core pillars: outcome-driven alignment, real-time data grounding, day-one governance, and modular cross-department orchestration. However, long-term success remains contingent on avoiding the 100% automation trap, managing granular runtime unit economics, and maintaining rigorous human-in-the-loop thresholds for high-risk actions.

精炼阅读摘要

Tyler Howard 指出，企业在推进人工智能落地时，常因陷入孤立的单点任务自动化而面临扩展瓶颈；唯有构建统一的端到端编排平台，整合可复用的数据底座与多智能体协作，才能将初始兴趣转化为具备经济可持续性与风险把控力的业务转型。 告别单点工具：突破AI停滞四大阻碍 (1:39) RSM US 的年度客户调查显示，尽管企业高管对人工智能转型抱有强烈兴趣，但多数组织在从概念验证迈向规模化落地的过程中遭遇显著脱节。Tyler Howard 分析认为，自2000年代初以来，企业普遍陷入孤立任务自动化与预封装单点解决方案的怪圈。一旦业务场景发生变化或现有工具无法满足其他团队的需求，工程团队就不得不推倒重来，陷入零扩展性与重复投资的资源浪费，无法积累可复用的底层数据架构。 Howard 明确指出了导致 AI 部署停滞的四大支柱性痛点：一是底层数据架构混乱与数据质量欠佳，在未建立可靠数据管道前盲目部署模型；二是云端安全性、数据泄露防护与合规监管机制不足；三是技术运营人才断层，缺乏在生产环境中长期维护和管理解决方案的专业人员；四是长期可持续性与投入信心不足，导致项目在完成概念验证后丧失资金支持与组织关注。 端到端编排机制：打破壁垒与跨系统集成 (4:49) 针对日常任务中常见的生成式或对话类 AI，Howard 强调简单的提示词和单体智能体无法直接驱动深层业务转型与营收增长，真正的价值必须通过编排来释放。编排机制通过协调数据模型、提示词层、业务逻辑和自动化校验，实现端到端的业务流程闭环，在保障严格审计与合规检查的前提下推进自动化。 在 RSM 协助实施的一个医疗行业案例中，客户最初仅针对首年处理量达2000万笔交易的单一流程开展 AI 项目。RSM 并未将其局限在初始业务孤岛，而是构建了一个编排层，用于综合分析交易输出、重构底层数据，并动态分发至下游4个运营团队的15个独立工作队列中。该方案使分析师摆脱了在4个分散系统之间手动比对案卷的繁琐流程，系统自动生成集成化的案情描述，大幅提高了决策准确性与流转效率，同时在关键风险节点嵌入人工审核机制。 盘活沉睡资产：多智能体驱动销售赋能 (8:27) 在涉及数百名持牌销售代表的案例中，客户长期将多渠道营销活动、外呼邮件、数字漏斗及第三方采购名单汇入通用线索池，经验丰富的销售代表只能对照扁平名单盲目拨打冷电话。RSM 的方案并非用 AI 替代高绩效员工，而是将15年的 CRM 结构化数据、外部人口统计指标以及长达10年从未被转录或利用的客户通话录音全面整合。通过对这些沉睡音频数据的转录与模型训练，系统实现了对新线索意图与产品偏好的实时诊断。 该系统采用多智能体协作架构，针对不同客群实施差异化触达策略。例如，针对22岁等年轻客群自动分配短信和电子邮件序列，而将高价值线索导向电话沟通，并为销售代表生成包含叙事脚本、最佳呼叫时间窗口及推荐产品的标准化作战手册。Howard 指出，这种人机协作机制与数据隔离防护相结合，不仅缩短了新员工的培训周期，还使无销售经验的基层员工迅速成长为高产值的业务骨干。 警惕全自动化陷阱：单元经济学与治理协同 (15:08) Howard 提醒企业必须审慎对待 AI 的运营单元经济学，切莫盲目追求最新基础模型或追求100%全自动化。随着调用规模扩大，云端算力消耗与交易运行成本会急剧增加，试图用算法覆盖所有长尾边缘案例往往会导致边际效益递减和成本失控。企业应将自动化聚焦于高频、标准化的高价值场景，并在触及特定财务金额或高风险阈值时，自动将任务路由至人工审批流程。 为了保障系统长期稳定运行，企业需要构建围绕分析师、部门主管和 CIO 的多角色运营监控面板，实时洞察异常指标、流程吞吐量与整体投资回报率。Howard 总结了规模化编排的四大核心支柱：坚持以业务结果为导向以争取持续支持、依托云平台实现实时数据驱动、从底层原生嵌入合规治理与审计能力、以及构建模块化可复用的多智能体工作流。 Howard 总结认为，企业 CIO 和技术领导者必须从底层重构 AI 实施路径，停止在割裂的单点脚本上重复投资。未来的核心挑战不再是单一模型的推理能力，而在于企业能否在统一的数据底座之上，平衡好细粒度运行成本、业务风险阈值与人机协同机制，最终实现可持续的规模化运营。

Detailed notes — zh

## Background 本视频由 Ai4 频道于 2026 年 8 月 21 日在 YouTube 平台发布，主讲人为来自全美知名咨询机构 RSM US 的自定义应用与 AI 团队负责人 Tyler Howard。内容聚焦于企业如何跨越单点任务自动化的局限，构建集数据统一、多模型推理、人机协同与可靠执行于一体的生产级企业 AI 智能编排平台架构。 ## TL;DR 分享直击企业从 PoC 走向实际投产时普遍面临的数据底座缺失、安全合规及人才断层等瓶颈，指出盲目追求 100% 免干预自动化的弊端。Tyler 结合医疗健康跨队列分发及激活沉睡十年通话录音的销售赋能实战案例，深入阐述了多 Agent 协作、人在回路（Human-in-the-Loop）与精细化单任务运行成本核算的必要性。最后，他系统提出了以结果为导向、以数据为基石、原生内嵌治理与规模化编排的四大核心支柱，为 CIO 和技术决策者提供了可落地的企业级 AI 架构蓝图。 ## Atmosphere & style 演讲整体呈现出务实、严谨且极具实战指导性的技术咨询专家风格，毫无浮夸的概念炒作，处处透露着一线架构师的敏锐洞察。Tyler 的语调从容沉稳、条理清晰，在剖析企业常见误区时直击痛点，在拆解落地案例时又层层递进，既有高屋建瓴的架构推演，又充满了对一线业务场景与成本账本的细致关照，带给观众极强的专业信服感与启发性。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) 企业在 AI 落地中的瓶颈在于沿用传统的“单点任务自动化”思维，导致方案缺乏可扩展性与资产积累，亟需转向平台级的智能编排架构。 - [02:55](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=175s) AI 项目从概念验证走向生产受阻的四大核心症结在于：底层数据质量欠缺、安全合规红线、专业运维人才断层以及缺乏长期扩展规划。 - [02:55](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=175s) 企业必须超越日常效率工具的表层应用，以业务结果与投资回报率（ROI）为锚点，将 AI 深度融入核心业务流程以直接驱动营收增长。 - [04:42](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=282s) 智能编排（Orchestration）的本质在于动态调度多模型并实现跨团队任务流转，打破各部门之间在不同业务系统间反复切换的孤岛状态。 - [04:42](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=282s) 在医疗案例中，编排引擎基于年处理 2000 万笔交易的底座，自动向 4 个团队的 15 个业务队列分发研判结果，极大精简了人工多系统核验流程。 - [08:20](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=500s) 激活企业沉睡的多模态资产（如 10 年未转写的电话录音与 15 年 CRM 数据）能为专属模型提供深度上下文，从而精准重塑销售转化漏斗。 - [10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) 通过多 Agent 协同分析客群特征，系统能在触达渠道、沟通话术与最佳时机上为一线人员提供针对性建议，推动新人业绩实现指数级增长。 - [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) 企业应警惕追求 100% 全自动化的陷阱，在涉及财务风险与合规的关键节点设定明确的人工介入阈值与审批责任链。 - [13:16](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=796s) 构建覆盖一线分析师、业务主管与 CIO 的差异化度量视图，是确保编排系统具备全链路可视化并支撑未来横向扩展的基础。 - [15:08](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=908s) 评估 AI 财务回报需从宏观的云托管费细化至“单任务运行成本”与“单笔交易运行时成本”，以保证系统在经济上的长期可行性。 - [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) 规模化落地 AI 必须立足四大核心支柱：以结果为导向、以数据为基石、原生内嵌治理以及通过模块化 Prompt 与 Agent 实现规模化系统编排。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) 从单点工具到平台化：告别孤立的任务自动化 在拉斯维加斯 Ai4 大会的现场，来自全美知名审计、税务和咨询公司 RSM US 的自定义应用与 AI 团队负责人 Tyler Howard 带来了这场关于企业级 AI 架构演进的分享。Tyler 直截了当地指出，团队的工作重点并非向客户兜售现成的标准化 Agent、现成软件或局部的“单点解决方案”（point solutions），而是深入企业一线，为其构建定制化的端到端工作流与高复杂度业务架构。RSM 每年都会面向全体客户开展企业调研，今年的核心主题正是 AI 的实际落地情况。调研数据显示，虽然各家企业对 AI 抱有极高热情与转型渴望，但在实践中却普遍陷入困境：多数组织仍然停留在零散的用例上，未能建立起能够跨团队协同、持续驱动业务增长的平台级架构。 企业在 AI 应用上面临的最大瓶颈，在于依然沿用自 2000 年代以来的“任务自动化”（task automation）与脚本化思维。单点工具往往带来极差的可扩展性与高昂的试错成本——当业务部门出现新需求时，现有的单点工具无法复用，往往只能推倒重来，陷入“零扩展性、零资产积累”的恶性循环。真正的数字化转型需要在底层架构上具备高度的“设计前瞻性与自觉性”（intentionality in design and build），在不重复造轮子、不反复重构数据管道的前提下，灵活接入并微调定制模型，同时确保系统具备清晰的风控与全流程可审计性，将 AI 真正打造为驱动企业发展的战略支柱。 ```mermaid flowchart TD accTitle: 从单点自动化到企业级智能编排架构演进 subgraph Legacy["传统单点自动化模式 (Point Solutions)"] A["局部任务/独立脚本"] --> B["单一部门孤岛"] B --> C["零扩展性 / 新需求推倒重来"] end subgraph Modern["RSM 智能编排平台 (AI Orchestration)"] D["统一企业数据底座"] --> E["动态多模型编排引擎"] E --> F["端到端跨业务工作流"] F --> G["人机协同与可审计决策"] end ``` - **团队定位与使命**：RSM 自定义应用 AI 团队专注于端到端定制工作流开发，摒弃拼凑孤立工具的做法。 - **单点方案弊端**：缺乏通用数据底座与管道支撑，导致跨部门迁移能力几乎为零，推高重复建设成本。 - **架构核心诉求**：兼顾定制化模型的灵活性、数据底座的稳定性以及风控与合规审计能力。 --- ### [02:55](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=175s) AI 落地停滞的四大症结与业务结果导向 企业 AI 项目在概念验证（PoC）之后往往难以顺利投产甚至停滞不前，Tyler 总结了导致部署受阻的四大核心要素。首先是**数据底座（Data Layer）**与**数据质量**的缺失，缺乏高质量、结构化的上下文输入，上层模型便无从发挥；其次是**安全与合规策略**，在全面上云的背景下，如何确保敏感业务数据不出内网、严格管控模型处理的数据流向是企业绕不开的红线；最后是**人才与技能断层**，不仅体现为员工对新技术的接纳意愿，更关键的是企业内部是否具备能够长期管理、运维这套复杂 AI 系统的工程与业务复合型人才。如果企业无法在规划初期对这四项挑战给出清晰且具可持续性的答案，项目势必会遭遇动力流失、推进受阻并最终被搁置。 面对技术采纳，Tyler 强调不能仅仅满足于日常的通用生成式工具或基础问答（例如查阅 HR 制度的客服 Agent），而必须将目光紧锁在具体的**商业产出（Business Outcomes）**与营收增长上。尽管许多高管在口头上表达了“全力以赴”（all-in）拥抱 AI 的决心，但很多落地项目并未实质性反映在生产力跃升或收入增加等硬性指标上。企业必须以业务回报为锚点，将 AI 深度融入核心业务流程，使之成为切实的营收驱动力，而非仅仅停留在一个新奇的效率玩具阶段。 - **AI 部署停滞的四大诱因**： - 数据架构与底层质量不达标 - 严格的企业级数据安全与合规治理要求 - 缺乏维护与管理复杂 AI 系统的专业技能人才 - 缺乏可持续演进与扩展的长期规划 - **价值衡量转变**：超越单纯的“日常效率辅助”，转向能直接衡量投资回报率（ROI）、生产力跃迁和收入增长的业务指标。 --- ### [04:42](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=282s) 编排的本质：跨团队协同、多模型调度与人在回路 深入到架构设计的核心，Tyler 阐释了为什么必须强调“智能编排”（Orchestration）而非单纯的“Agent + 提示词”。单点 Agent 与 Prompt 工程虽然容易上手，但没有全局编排，系统就无法实现真正的横向扩展与端到端闭环验证。很多客户期望实现“完全免干预（Touchless）的 AI 方案”，但在真实企业运营中，盲目追求全自动往往并不现实且充满风险；真正高价值的系统在于通过编排引擎实现严密的制衡与跨团队联动。Tyler 分享了一家医疗健康行业客户的实际案例：系统最初仅服务于单个团队，年处理交易量达 2000 万笔；团队随后对其输出进行重构与智能编排，自动向另外 4 个团队的 15 个工作队列分发任务，彻底打破了过去分析师需要在 4 套不同业务系统间来回切换、多方核验、反复层层上报审批的繁琐流程。 ```mermaid flowchart LR accTitle: 医疗健康客户跨系统编排与协同工作流 subgraph Ingest["输入与分析"] A["2000万笔年交易量"] --> B["AI 综合研判与上下文生成"] end subgraph Dispatch["智能分发与编排"] B --> C{"多模型编排引擎"} C --> D["跨4个团队 / 15个业务队列"] end subgraph Execution["人机决策与执行"] D --> E["业务人员人机协同复核"] E --> F["系统自动化闭环执行"] end ``` 这种平台级架构的威力还体现在“动态模型选择”与“人在回路（Human-in-the-Loop）”的有机结合上。企业内部不应绑定单一数据模型，而应根据工作流中的具体环节灵活调度不同模型的能力。更关键的是，架构设计必须在底色上将人类的深度介入视为核心一环：在关键决策点设置阈值监控、合规复核与审计关卡，将 AI 生成的洞察转化为比纯人工分析师更全面、更具故事性的决策依据，从而让员工将精力集中在价值更高、更能创造营收的环节。此外，编排架构赋予了企业极强的自适应演进能力，使企业摆脱低代码/拖拽式工具在 License 授权费、托管环境和功能僵化上的束缚，随时紧跟企业战略变化调整战术。 - **医疗案例成效**：依托 2000 万笔交易的处理底座，AI 自动完成多系统交叉验证与线索研判，无缝填充 15 个下游工作队列。 - **动态模型选择**：不拘泥于单一 LLM，依据不同任务复杂度、成本与时延动态切换最优模型。 - **人在回路设计**：在关键风险节点嵌入合规阈值与人工介入机制，兼顾自动化效率与业务审计安全。 --- ### [08:20](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=500s) 销售赋能实战：唤醒沉睡十年数据与重塑转化链路 为了生动说明自适应编排架构的落地过程，Tyler 分享了一家拥有数百名专业销售人员的企业实操案例。该客户长期以来投入巨资通过营销推广活动、邮件触达、在线获客以及向第三方采购等多种渠道获取大量销售线索（Leads）。然而，原先的工作流极其粗放：每天将海量未经深度加工的线索直接打包丢入公海池，销售人员每天机械地逐个拨打电话。RSM 团队深入调研后发现，该团队拥有大量持证且具备长达 25 年丰富行业经验的资深业务代表，他们与客户建立信任和沟通的专业能力远超 AI；系统真正需要做的，不是替代他们打电话，而是利用智能化编排为他们提供最精准的线索优先级排序与全方位的客户画像故事。 在梳理客户资产时，RSM 发现了巨大的沉睡数据宝库：企业坐拥长达 15 年的 CRM 历史业务数据，以及近 10 年积累的录音外呼电话，但这些录音此前“从未经过转写，也从未被任何算法或业务流程用于学习和优化”，白白沉淀在存储介质中。RSM 团队以此为突破口构建专属数据模型，从这些沉睡的多模态数据中提炼历史成交模式与有效沟通话术，并与外部人口统计学及产品偏好数据相结合。在线索进入系统的一瞬间，编排系统便能迅速还原潜在客户的立体故事，预测最佳下一步行动（Next Best Action），并在销售拨出电话前为其提供详尽的上下文支撑，彻底重塑了销售转化漏斗。 - **客户团队现状**：数百人规模的专业销售团队，拥有深厚行业经验与持证资质，原先依赖粗放的线索池进行盲打。 - **沉睡资产盘活**：成功激活并转写了沉睡长达 10 年的历史通话录音，结合 15 年的 CRM 结构化数据进行联合训练与建模。 - **智能化赋能路径**：从简单的线索分发转变为自动提取上下文、绘制客户画像、建议最佳行动，大幅提升销售触达转化率。 ### [10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) 赋能销售团队：从无序脚本到多 Agent 智能编排落地 销售触达的渠道决策需要高度差异化。例如针对 22 岁左右的年轻客群，直接拨打电话的接通率和转化率极低，他们通常会直接挂断；相反，向其发送产品信息、营销邮件或短信往往更有效。但对于高价值且经验要求高的潜在客户线索，销售团队需要系统性赋能：AI 编排层能够明确告诉销售代表“该讲什么样的故事”、“在一天中的哪个最佳时间段致电”、“客户何时方便接听”以及“重点推介哪些产品”。 这彻底改变了传统销售人员没有结构化支持、缺乏精准目标、仅被告知“向某类人推销这些产品”的困境。依托坚实的数据基础，系统将人机协同（Human-in-the-Loop）机制嵌入业务流程，同时设置严格的财务风险阈值与端到端数据管理，确保个人隐私数据绝对不外泄。整个系统并非依赖单一模型的单次预测，而是由多个自主运行的 Agent 协同处理不同阶段的任务，最终赋能业务分析师与销售人员。 ```mermaid flowchart TD accTitle: Sales Orchestration Architecture subgraph DataSec["数据基础与安全治理"] D1["底层数据整合"] --> D2["端到端数据隐私管理"] D2 --> D3["动态模型持续调优"] end subgraph MultiAgent["多 Agent 协同决策"] A1["客群属性分析 Agent"] A2["渠道策略 Agent<br/>(短信/邮件/电话)"] A3["话术与时机建议 Agent"] end subgraph Execution["人机协同执行与业务成效"] E1["财务阈值风险拦截"] E2["一线销售执行<br/>(个性化故事/推荐)"] E3["营收指数级增长"] end DataSec --> MultiAgent MultiAgent --> E1 E1 --> E2 E2 --> E3 ``` - **全员认同与业绩增长**：通过直接帮助一线销售提高个人业绩，新入职、毫无行业经验的员工能迅速从“按时薪计酬但无法为公司创造价值”转变为实现业绩的“指数级增长”。 - **多模型协作架构**：摒弃单一模型输出的脆弱方案，采用多步骤处理与独立 Agent 运行机制，根据持续输入的新业务数据动态调整模型。 --- ### [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) 风险防范与人机协同：不要追求 100% 自动化的陷阱 在系统落地过程中，必须清晰界定风险与合规边界，特别是财务风险敞口、客户体验影响以及异常情况的处理机制。企业不应抱有“100% 的交易在 100% 的时间内完全由 AI 完美处理”的不切实际预期，这种目标既不可行，也会导致投资回报失衡。 对于涉及资金流转的业务场景，必须设定明确的人工介入阈值。一旦交易金额或风险敞口超过设定上限，系统应立即触发人工介入流程，并向审批人提供完整的上下文信息与决策依据。 - **审批工作流与权责明确**：在编排系统中全程追踪异常处理链路，明确由哪个具体角色执行审批。 - **抓取“低垂的果实”**：专注于高确定性、高回报的典型场景，而不是试图覆盖所有极端边缘案例（Edge Cases），避免因过度追求全量自动化而使项目停滞不前。 --- ### [13:16](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=796s) 面向角色的全链路度量体系与安全治理 单点工具（Point Solutions）之所以容易导致投资失败且无法横向扩展，核心原因在于缺乏全流程的可视化指标。为了确保编排系统长期稳定运行并支撑未来用例拓展，必须针对不同业务角色（Personas）构建定制化视图： | 角色层级 | 关注重点与视图需求 | 价值体现 | |:--- |:--- |:--- | | **一线分析师 (Analyst)** | 日常任务状态、当下需处理的线索、即时上下文信息 | 提升单兵作业效率与执行准确率 | | **业务主管/总监 (Manager/Director)** | 团队整体吞吐量、业务瓶颈、审批队列健康度 | 优化团队资源调配与流程运转 | | **首席信息官 (CIO)** | 端到端全景视图、ROI、安全合规与系统可扩展性 | 支撑战略决策与跨部门能力复用 | 许多企业在开展 AI 试点（Pilot）时缺乏系统性规划与统筹聚焦。安全与风控固然是必须满足的“底线预期”，但更关键的是配套的治理规范与人员协作机制，避免陷入“勾选了安全合规项就万事大吉，实则无法推进变革”的防御性停滞状态。 --- ### [15:08](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=908s) 精细化财务核算：从算力持有成本到单任务运行成本 在评估 AI 系统的财务回报时，不能仅停留在云端托管费（Hosting Fee）或年度云资源消耗额等宏观数字上，必须深入核算“单次任务成本”（Cost per Task）与“单笔交易运行时成本”（Runtime by Transaction）。 当前许多企业盲目追求部署最新、最大的基础模型，但随着调用频率和任务复杂度剧增，运营成本急剧攀升。如果无法明确每一笔交易的实际处理开销及其投资回报周期（Payback Period），系统就难以具备长期经济可行性。通过有意识的人机协同设计，将算力精准投放至能够产生高业务价值的触节点，是保障系统可扩展性的关键所在。 --- ### [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) 规模化编排的四大核心支柱 为实现 AI 系统的规模化落地，企业需围绕四大核心支柱构建端到端的编排架构，实现从“启动 AI 项目”向“规模化扩展 AI”的跨越： ```mermaid flowchart LR accTitle: Four Pillars of Enterprise AI Orchestration subgraph CorePillars["企业级 AI 规模化编排四大支柱"] direction TB P1["1. 以结果为导向<br/>(Outcome-Based)"] P2["2. 以数据为基石<br/>(Data-Grounded)"] P3["3. 原生内嵌治理<br/>(Governed by Design)"] P4["4. 规模化系统编排<br/>(Orchestrated at Scale)"] end P1 --> P2 P2 --> P3 P3 --> P4 ``` - **以结果为导向（Outcome-Based）**：以最终业务成效定义需求，据此构建各角色视图，从而赢得一线员工的使用意愿、团队的时间投入以及管理层的持续资金支持。 - **以数据为基石（Data-Grounded）**：利用成熟的云原生集成平台与应用，将数据从阻碍 AI 落地的痛点转化为核心赋能资产，确保系统随时以极高响应速度获取上下文。 - **原生内嵌治理（Governed by Design）**：从项目启动第一天起建立涵盖工具栈、可操作性、可审计性（Auditability）的完整治理体系，杜绝后期的合规风险。 - **规模化系统编排（Orchestrated at Scale）**：将底层 Prompt 沉淀为可跨工作流复用的“核心主 Prompt（Master Prompts）”与通用 Agent 模块，像积木一样拼装复杂业务流，实现能力的跨部门快速复用。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) 团队的工作重点并非向客户兜售现成的标准化 Agent、现成软件或局部的‘单点解决方案’，而是深入企业一线，为其构建定制化的端到端工作流与高复杂度业务架构。 — Tyler Howard > [00:04](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=4s) 企业在 AI 应用上面临的最大瓶颈，在于依然沿用自 2000 年代以来的‘任务自动化’与脚本化思维。 — Tyler Howard > [02:55](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=175s) 企业必须以业务回报为锚点，将 AI 深度融入核心业务流程，使之成为切实的营收驱动力，而非仅仅停留在一个新奇的效率玩具阶段。 — Tyler Howard > [04:42](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=282s) 很多客户期望实现‘完全免干预的 AI 方案’，但在真实企业运营中，盲目追求全自动往往并不现实且充满风险；真正高价值的系统在于通过编排引擎实现严密的制衡与跨团队联动。 — Tyler Howard > [04:42](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=282s) 架构设计必须在底色上将人类的深度介入视为核心一环：在关键决策点设置阈值监控、合规复核与审计关卡。 — Tyler Howard > [08:20](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=500s) 系统真正需要做的，不是替代他们打电话，而是利用智能化编排为他们提供最精准的线索优先级排序与全方位的客户画像故事。 — Tyler Howard > [10:17](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=617s) 新入职、毫无行业经验的员工能迅速从‘按时薪计酬但无法为公司创造价值’转变为实现业绩的‘指数级增长’。 — Tyler Howard > [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) 企业不应抱有‘100% 的交易在 100% 的时间内完全由 AI 完美处理’的不切实际预期，这种目标既不可行，也会导致投资回报失衡。 — Tyler Howard > [12:19](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=739s) 专注于高确定性、高回报的典型场景，而不是试图覆盖所有极端边缘案例，避免因过度追求全量自动化而使项目停滞不前。 — Tyler Howard > [15:08](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=908s) 在评估 AI 系统的财务回报时，不能仅停留在云端托管费或年度云资源消耗额等宏观数字上，必须深入核算‘单次任务成本’与‘单笔交易运行时成本’。 — Tyler Howard > [16:37](https://www.youtube.com/watch?v=uDbeZWPFqr8&t=997s) 将底层 Prompt 沉淀为可跨工作流复用的‘核心主 Prompt’与通用 Agent 模块，像积木一样拼装复杂业务流，实现能力的跨部门快速复用。 — Tyler Howard ## Mentioned Tyler Howard, RSM US LLP, Ai4, CRM, LLM, Agent, CIO, CTO ## Video description Tyler Howard - Director, RSM US LLP Most organizations are automating tasks. The leaders are orchestrating intelligence. In this session, I will share how we designed and delivered a production grade AI orchestration platform that moves beyond single model point solutions to a fully integrated, multi-AI architecture. One that unifies data, reasons across context, inserts human judgment where required, and executes decisions reliably across real enterprise workflows. While customer engagement is a powerful example, the architecture and lessons are broader. This is a practitioner level look at what separates proof of concept from production including how to design data foundations for orchestration, not experimentation, where automation breaks down and human oversight must remain and the architectural patterns that allow AI systems to scale across functions — from customer engagement to operations, finance, and IT. This session is not about a single tool or industry. It’s a blueprint CIOs and CTOs can apply as AI moves from curiosity to core enterprise capability. [Source](https://www.youtube.com/watch?v=uDbeZWPFqr8)

Source quotes

[0:04] Tyler Howard: 团队的工作重点并非向客户兜售现成的标准化 Agent、现成软件或局部的‘单点解决方案’，而是深入企业一线，为其构建定制化的端到端工作流与高复杂度业务架构。 [0:04] Tyler Howard: 企业在 AI 应用上面临的最大瓶颈，在于依然沿用自 2000 年代以来的‘任务自动化’与脚本化思维。 [2:55] Tyler Howard: 企业必须以业务回报为锚点，将 AI 深度融入核心业务流程，使之成为切实的营收驱动力，而非仅仅停留在一个新奇的效率玩具阶段。 [4:42] Tyler Howard: 很多客户期望实现‘完全免干预的 AI 方案’，但在真实企业运营中，盲目追求全自动往往并不现实且充满风险；真正高价值的系统在于通过编排引擎实现严密的制衡与跨团队联动。 [4:42] Tyler Howard: 架构设计必须在底色上将人类的深度介入视为核心一环：在关键决策点设置阈值监控、合规复核与审计关卡。 [8:20] Tyler Howard: 系统真正需要做的，不是替代他们打电话，而是利用智能化编排为他们提供最精准的线索优先级排序与全方位的客户画像故事。 [10:17] Tyler Howard: 新入职、毫无行业经验的员工能迅速从‘按时薪计酬但无法为公司创造价值’转变为实现业绩的‘指数级增长’。 [12:19] Tyler Howard: 企业不应抱有‘100% 的交易在 100% 的时间内完全由 AI 完美处理’的不切实际预期，这种目标既不可行，也会导致投资回报失衡。 [12:19] Tyler Howard: 专注于高确定性、高回报的典型场景，而不是试图覆盖所有极端边缘案例，避免因过度追求全量自动化而使项目停滞不前。 [15:08] Tyler Howard: 在评估 AI 系统的财务回报时，不能仅停留在云端托管费或年度云资源消耗额等宏观数字上，必须深入核算‘单次任务成本’与‘单笔交易运行时成本’。 [16:37] Tyler Howard: 将底层 Prompt 沉淀为可跨工作流复用的‘核心主 Prompt’与通用 Agent 模块，像积木一样拼装复杂业务流，实现能力的跨部门快速复用。

Source quotes

Tyler Howard: Most organizations are automating tasks. The leaders are orchestrating intelligence. Tyler Howard: This session is not about a single tool or industry. It’s a blueprint CIOs and CTOs can apply as AI moves from curiosity to core enterprise capability. [1:39] Tyler Howard: zero scalability, zero investment [4:49] Tyler Howard: all-in mentality [10:17] Tyler Howard: they're going to hang up [12:19] Tyler Howard: handle 100% of your transactions 100% of the time perfectly [16:37] Tyler Howard: Outcome-Driven [16:37] Tyler Howard: Data-Grounded [16:37] Tyler Howard: Governed by Design [16:37] Tyler Howard: Orchestrated at Scale

Original transcript

[0:04] Thanks everyone. Good to be here. Hope everybody's  having a good week. I know it's the end today as [0:10] we get into Thursday, a long week in Vegas. Um, so  my name is Tyler Howard with RSM. Uh, so we're an [0:16] audit tax advisory firm. So within our con kind  of consulting services team, I lead our custom [0:21] applications AI team. So within that my focus  is less on how do we give you products agents [0:28] and kind of point solutions for you is actually  how do we build and develop custom workflows and [0:35] more complex solutions for you. So I'll start off  kind of with a few things that kind of excited to [0:40] highlight as we go in and this is really what help  build our message of how we approach AI and how we [0:45] deliver it. So we have a survey that goes out to  all of our clients annually. A big portion of that [0:51] this year was AI of course. So how are they using  it? And I don't want to go through all of these [0:56] today, but the biggest thing that I want to focus  on, what's been really critical for my teams and [1:01] my clients success is how do we transform with AI?  How do we make sure that this is not a single use [1:08] case, a single point solution that you're using  and building for your teams or your organizations [1:14] and actually build something that can be a  platform across team workflow have engagement and [1:20] kind of build and grow your business beyond that.  And that's been a big struggle because if you look [1:24] at kind of some of the other numbers you can see,  okay, there's transformation, there's appetite, [1:28] we want to build and grow off of this, but are  we doing that? So it's how do we go and kind of [1:33] level up? we become scalable and build more. So  focusing a little bit more on that point solution [1:39] piece. I'll highlight just a few of these, but  it's how do we get away from we've been talking [1:44] task automation since 2000. How do we go from  automated scripting? How do we have automation [1:49] tools? How do we build something that's going  to be actually scalable for us? So it's what is [1:55] going to provide that platform that gives us that.  There's nothing worse than coming in with a single [2:00] solution for your team and taking that and saying  okay we have another use case let's start over or [2:05] our point solution doesn't do what we would like  to do in this team or this group. So you're kind [2:10] of starting with this zero scalability zero  investment to actually build and grow. So from [2:16] that transformation aspect, you can kind of have  that customization and still have those custom [2:21] models that you're looking and building on within  those workflows, but you're not feeling like [2:26] you're rebuilding that plumbing or that starting  point from your kind of data foundation for you. [2:31] So what is that kind of like having that  intentionality in your design and build and [2:36] kind of pulling all of that together so that you  have that clear risk and auditability that comes [2:41] with all of that because I want to have my clients  have many use cases and build and feel that AI is [2:48] a strategic pillar for their business much less  than anything else. So just another one is why [2:55] do your deployments stall? What is causing you  issues? I think none of this should be surprising. [3:00] I talked about data layer. I talked about the  data quality and we know there's always going to [3:04] be security. We don't want any data outside your  systems. Everyone's working in the cloud now. What [3:09] sort of policies do we have? And especially,  you know, what sort of data is running through [3:12] these models that's going through and of course  talent skill gaps. That one's a little different, [3:17] but kind of you can have a talent of who's has the  appetite to engage, but also who has the skill to [3:23] own and manage some of these solutions. So really  important to think about as you go through some [3:27] of these just data quality components. So, what is  that foundation that you're going to have? Because [3:31] if you can't answer all four of these and feel  that you have a confidence of sustainability and [3:36] growth that's going to come with it, you're going  to feel that you're going to pause, lose momentum, [3:41] lose interest, and just lose anyone's effort  that's going to help build these things. Um, [3:49] and then just from just adoption, I've been  talking a little bit about the people. What is [3:53] going to get people to use AI? Everyone's using we  all know gen language predictive AI we have those [4:00] components that's in our wheelhouse but that's  day-to-day that can be me having a simple support [4:05] agent for questions or queries that I have call  it HR policies but the biggest thing is is what [4:10] are your business outcomes what are we building  towards and growing to because leaders really want [4:15] that that everybody at an executive level keeps  saying we want this all-in mentality we want to [4:20] build and grow as we continue to move forward but  are we achieving those with that. That's where [4:26] that kind of transformation is so critical for  you to really focus on how are we making sure that [4:31] this is an actual revenue driver for you because  they're going to I'm a little surprised we don't [4:37] see that like revenue and productivity so much  higher because that should be an assumption as we [4:42] move through some of these kind of focus points.  So why do we why am I here focused and talking [4:49] about orchestration? And it's I I know like some  of it on here I think we're all aware of like [4:54] the high level orchestration your agentic your  prompt and your enterprise data foundation but [4:59] the biggest two for me is if you have this point  solution you're you got your agents you have your [5:04] prompts great that can be what you're initially  focused on but you can't have that scalability [5:09] with that or without that orchestration. So,  how do you think from moving beyond just that [5:15] kind of simple task or what we're doing? Because  I want all of my clients to feel like they have a [5:20] touch touchless AI solution for themselves. Um,  that's not always achievable, not always right, [5:26] but they I want that feel of those checks  and balances and what's coming through. So, [5:30] from a complete workflow validation, what are we  going through end to end that is actually working [5:35] for them? Um, so that you can feel like it's  transforming. And that's why I was saying before [5:40] those cross team and cross function solutions are  where you find your value. A lot of what I do is [5:46] maybe start with one workflow, one team, but how  does that impact? I have a client that we started [5:51] we had 20 million transactions year one. How did  we take that and then transition into I have 15 [5:58] work cues that I'm populating for four other teams  and actually impact them. I changed my coding. I [6:05] changed some of that. So now I have agents that  are taking the outputs, rebuilding them and moving [6:10] into that next workflow. So it's a lot of that  intentional design that's we're transforming it. [6:15] So what is what are teams doing now? Are they  outside of their four different systems that [6:20] they have to check and validate? Especially this  was a healthcare solution for them. So how do we [6:25] get them out of I have to do everything in the  core system? I have to go to external systems. [6:30] I have to build this story and case to go  through and then anything that gets elevated, [6:35] escalated and reviewed again that causes more  challenges and friction as they go through. But [6:40] also AI can tell them a better story than most  agents like agent analysts can. So I want to [6:46] make sure that I'm informing them better than what  we actually have in that current process. So that [6:51] workflow is now rebuilt and stronger for them to  have. Um I talked about the variable model usage. [6:57] I think it's really critical just to think that  you don't have one data model. You can have the [7:02] flexibility to it and that's something that will  adjust. And then intentionally the same color on [7:07] this slide is design with the human involvement  in mind. A lot of the things we want to do are [7:13] so critical for business operations and they  sound great but there's checks, thresholds, [7:20] just auditability that has to be included in  there. So where do we need them to review? Where [7:25] do we need the followup? but also strategically,  is there something that I can generate more [7:29] revenue for? And I'll talk more specifically on  that one. And then all of this kind of leads to [7:35] that adaptive AI as we evolve. We want to make  sure that this we have all of our pain points [7:40] and challenges. That's where our use cases and  starting points really come from. But for me, [7:45] it's I want to make sure that's adaptable. We're  going to find more challenges as we have those [7:49] models learning, collecting even better data than  we had before, and growing what that looks like. [7:55] So what is that next step going to be? So where  do we have that flexibility? So we don't feel [7:59] like it's great to say we have a drag and drop  solution. We all love that. It's great. But then [8:04] you're stuck with your license cost, your hosting,  your team, and you're kind of tied in there of [8:09] what else is happening. What are our organizations  new strategic initiatives is that we can actually [8:15] move and build to meet those because that's where  those that adjustability and flexibility is just [8:20] so critical to anything that you are happening or  working on. So I'll kind of talk through a little [8:27] bit of just a story of like one of my clients and  how we had a little more flexibility building them [8:32] a solution. So specifically they came to us and  said we have a sales organization of hund of of [8:39] a few hundred that are coming in and building  these solutions or building kind of our sales [8:44] leads. They're actually having a pool of different  marketing campaigns. They have email contacts, [8:50] uh, online outreach, but they're also paying for  customer leads as they come in. We started with [8:56] their workflow and we're like working with their  agents. How do you do go through your sales cycle? [9:01] What is happening? Go to find out we they have  licensed agents. They have agents with 25 years of [9:05] industry experience and they already know who they  are. So, how do we have that targeted list and [9:11] say, you guys know this better than AI will. You  know how to have a customer conversation. How do [9:16] I build and prioritize everything that we see with  15 years of CRM data, external demographic detail, [9:22] and build that for you and actually come in and  really start to tell that story. So instead of [9:28] just pooling all of these sales leads that came  in daily and just saying here's a work, make a [9:33] phone call. We were able to have that output come  in structure take these intake leads say what sort [9:42] of data or story can we start to tell about this  customer. So that's that data model we built on [9:47] their data their historical success find out they  have 10 years of recorded calls and outreach. None [9:54] of those are transcribed. None of those are used  in their processor learning. So they've just been [9:59] sitting there and stored. So we build our model  based on that storage capability. started to [10:04] go through and say what is going to be our next  action. How are we going to enable the agent with [10:10] this detail? What external sources can help build  that demographic of product focus? Um even type of [10:17] contacts. So, are we emailing them? Are we texting  them? Are we making a phone call? Um especially [10:22] from an age and population. Making a phone call  to a 22-year-old is not really going to have high [10:28] success for you. They're going to hang on. But  we can send them product information, marketing [10:32] emails or text over that. So that's great to have.  But for our pri our kind of experienced leads, [10:40] how do these structured sales leaders have that  information and say here's the story we think you [10:46] should tell. Here's the time of day to call them.  Here's where they should be available and here's [10:50] the product focus and also all of that. So they  can take their now script for a for a job that [10:57] they had no structure, no target other than sell  these products to these type of people. This is [11:02] where it works. And that's was really cool for  us because starting with that data foundation, [11:07] we're able to come through and actually build  into how do we have this human in the loop. We [11:12] had thresholds uh on the financial side. We had  endtoend data management. So nothing was leaving [11:18] because there's you know personal information  that's involved as well. And then you can [11:21] actually have that monitoring to go through.  And the goal with that is like I said before, [11:26] we have new data, we have new information.  So how do we keep adjusting that model as [11:31] part of this new application for them that was  using AI throughout but as isn't just reliable [11:37] on one model, one prediction, one output. So we  have multiple steps within that process. Agents [11:42] running on their own. And then we have your  analysts that are much more successful. So [11:46] you you have complete buyin. We're working with a  sales team to increase their personal revenue. So, [11:52] they're going to be successful at their job. So,  somebody that's coming off the street with no [11:56] experience is going from zero dollars earned for  our organization at an hourly rate to actually [12:02] exponential growth on what they can succeed with.  So I think that one's really just exciting to talk [12:08] about on what we did a little bit differently  and how we focused on that orchestration layer [12:13] that really brought it all together so that we can  use AI effectively but also be very adaptable to [12:19] their process. So where it kind of breaks down in  my opinion um as we go through this what sort of [12:26] risk or compliance do we have to think about when  we design these solutions the financial exposure [12:32] impact on the customer side I think that one's  pretty clear but also where do we have exceptions [12:37] I don't want to sit down and tell everyone of I'm  going to handle 100% of your transactions 100% of [12:43] the time perfectly it's not going to happen but  also where's my investment going on that one I [12:49] talked a little bit financial exposure thresholds.  I have almost every client that has a financial [12:55] component to it. I want to put a a high number  out there of who needs human touch immediately. [12:59] Where do I need to expose this out? But also  inform that touch point so we have that and then [13:05] throughout this have all of that kind of tracked  so we have we have some form of approval workflow. [13:10] What is that going through and who specifically  is going to be tasked with doing that for me? [13:16] So, how do you actually make sure that your  orchestration's working? This one is kind [13:22] of part of it of like I want to track that we  have all of this. I talked the point solutions [13:28] fail for your investment for your path forward  where you're structured, but we're not going to [13:34] move forward and we're not going to get to that  next use case, that next expansion if we don't [13:38] actually have visibility into what we're doing.  So, if we're starting with what's our target, [13:43] how do we kind of narrow that down? But also,  I like to build personas for my teams. Who is [13:49] what does the analyst care about and need to know  on a daily basis? What does my manager director [13:55] team need to know and have visibility into on  their end that they're going through? And then [13:59] what is that CIO level view? The CIO wants all  of it. But how do I structure and cater to each [14:04] of those? Because those stories are going to be  how I build my data, how I use my data, and what [14:09] I'm tracking throughout. so that we can actually  have that reporting because we want to have that [14:13] transformational field that this is isn't just a  startup complete oneandone opportunity that's a [14:21] lot of teams that I talk with have said we've gone  through an AI implementation and it's a P it's a [14:28] small pilot and they had no structure intentional  focus on these concepts to actually build those [14:34] together and it's it the risk the security side is  always there that's I call that an expectation of [14:41] this, but what is the governance and people  process that goes with that risk around it? [14:46] Because I want to make sure that you're actually  moving forward with it instead of kind of feeling [14:52] like, okay, I checked all my security boxes. I'm  good. I'm fine for now. But I want that to be [14:58] something that you feel like you're not opening  doors. This is a new solution, a change. And [15:02] that's where a lot of people kind of pause or feel  like they're taking a bit of a step back for you. [15:08] Um and then hopefully it's readable on the screen.  The financial side is of course there. Um I want [15:14] to I like to focus on the financial side is the  actual cost of operating and cost per task. Uh [15:21] it's becoming more and more important that I've  seen customers they want the latest greatest [15:27] model. They want to have flexibility in what  they're doing and executing. But that cost is [15:32] going up significantly because you you continue  to do more. you continue to like operate but it's [15:39] what is that what do we have for that what is the  payback period so we can actually make sure it's [15:43] being effective for us uh as we go through because  it's I I build these kind of custom solutions so [15:50] you're really focused on what is my hosting fee  what is my annual consumption that's great but [15:56] what is the actual runtime by transaction and  getting to that deeper level so you can kind [16:00] of validate what else you're doing and that goes  back to what I was saying I don't want you to feel [16:04] like you have a 100% target or output on what  you're processing. It's really not effective [16:10] as you kind of sit there, review it because  you're just going to try to capture too much, [16:15] get through those lowhanging fruits, feel like  it's something that's moving you forward instead [16:19] of like taking kind of a half step back because  you're trying to do too much with it. So, how do [16:25] it's that human in the loop, it's that intentional  design. What touch points are we going to be [16:30] successful with on my are we capable of handling  there? And this all goes down to the scalability. [16:37] So really to kind of bring it all together for us  here is building at scale. For me that's oneonone [16:46] with our orchestration. I want every single  client that I have to feel like they are not just [16:51] starting AI but they are scaling it. It doesn't  mean I have to be there holding their hand the [16:55] whole time but what are we doing to get there? Um  I've talked a lot of this time on outcomes. what [17:02] do you want to get to? And that's where when we  build those personas, if we start outcome based, [17:07] that's where that visibility and what they care  about matters. That's where I get my people buyin, [17:12] that's where I get my kind of time commitment to  support that. And then of course the investment [17:17] from your executives and leadership to get things  started and move forward because that's always [17:22] a very hard and difficult part for everyone. And  then data grounded. How are we going to make sure [17:28] that our data is going to be a strength of what  we're doing? instead of a painoint. We talked so [17:34] much for years of I can't get started in AI. Now  you've got all these integration platforms, tools, [17:39] you're working with now kind of cloud-based  applications. And that's something that is [17:44] a baseline that I see almost everywhere now. So I  don't want to focus and talk about that too much, [17:49] but it should be enabling you. I want to have  access. I want to make sure everything's available [17:53] at the speed and time that I need it as well.  And then govern by design. making sure that day [17:58] one you're talking governance and it's not just  governance from the people side, it's the tools, [18:03] it's the operability, auditability and all of  those components that come in so that you do [18:08] not enter into that risk territory as things move  forward and then as I said orchestrated at scale I [18:15] want everything connected agents reusable prompts  you can have your workflows that can eventually [18:21] be reusable for you as well. So what are the  core kind of master prompts or capabilities [18:28] that we have that we can take and reuse throughout  additional workflows because they should be your [18:32] building blocks. It I'm not going to say they're  drag and drop for you, but it's how do they all [18:37] come together and build that. So hopefully  everybody leaves today feeling like they kind [18:41] of have a better understanding and think more  design first and focus at scale for you. So it [18:46] gives you an opportunity to move forward. So thank  everybody for the time. I appreciate it today.