Vellum Blames Siloed Workflows for $11 Billion Enterprise AI Failure
Ai4
Enterprise AI initiatives fail at an 85% rate because legacy release cycles block domain experts from directly optimizing non-deterministic models, trapping engineering talent in maintenance backlogs.
## Background
Published on YouTube by Ai4 on August 27, 2025, this presentation features Rasam Tooloee, GTM at Vellum, discussing why 85% of enterprise AI projects fail to reach production. The video description highlights how an $11 billion waste in annual AI spend can be reversed using collaborative AI platforms that reduce per-application costs by 70-80% and generate compounding returns.
## TL;DR
Rasam Tooloee explains that the high failure rate in enterprise AI stems not from technical model limitations, but from applying rigid legacy development lifecycles to non-deterministic systems. To capture compounding ROI instead of mounting linear costs, organizations must shift scarce engineering talent toward building composable, reusable infrastructure while providing domain experts with visual interfaces for continuous, independent prompt refinement. Adopting this collaborative platform model enables businesses to support significantly more AI initiatives, slash per-application maintenance costs, and achieve same-day market responsiveness.
## Atmosphere & style
The presentation maintains a pragmatic, fast-paced, and solution-driven conference atmosphere, blending analytical economic breakdowns with relatable enterprise realities. Despite starting with flickering screen hiccups and facing a strict moderator who cuts off the implementation framework for a rapid 30-second wrap-up, the speaker stays upbeat, composed, and engaging throughout the brief stage session.
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) Enterprise AI ROI is fundamentally driven by operational agility—speed advantage, market responsiveness, and innovation velocity—rather than model technology alone.
- [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) Approximately 85% of enterprise AI projects fail to reach production, squandering over $11 billion annually due to early-stage prototyping silos.
- [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) Applying rigid, legacy software development cycles to non-deterministic AI models creates a severe collaboration gap between technical teams and business operators.
- [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) In engineering-focused AI models, 60% to 70% of scarce technical talent bandwidth is trapped in routine maintenance, coordination friction, and prompt tweaking.
- [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) Collaborative AI platforms reduce per-application development costs by 70% to 80% and allow teams to support 5x to 10x more initiatives per engineer.
- [09:05](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=545s) Engineers should focus on building reusable, composable architecture and data pipelines rather than one-off, single-use AI tools.
- [09:05](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=545s) Giving domain experts direct, governed access to refine prompts and evaluations generates 'compound intelligence' without engineering backlog bottlenecks.
- [12:18](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=738s) Augmenting human institutional knowledge with accessible AI tooling creates a defensible competitive moat, outpacing simple worker-replacement strategies.
- [12:18](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=738s) Adopting an AI-native operational model enables same-day iteration to address shifting market demands instead of slow quarterly development cycles.
- [15:01](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=901s) Production platform infrastructure requires bi-directional sync, allowing engineers to work in code while business stakeholders inspect and optimize via visual interfaces.
- [16:16](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=976s) A phased 'crawl, walk, run' adoption strategy starts with establishing core data pipelines and modular components that feed initial departmental MVPs.
- [17:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=1028s) Achieving compounding AI returns depends on deploying onto an accessible, centralized platform foundation that bridges the technical-business divide across the entire organization.
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) The AI ROI Reality Gap and the Three Drivers of Compound Returns
Opening the session after a brief technical hiccup with flickering projector screens—jokingly compared to the inconsistency of AI ROI itself—Rasam Tooloee from Vellum frames the fundamental dilemma facing modern enterprises. Organizations across the mid-market and enterprise landscape are pouring millions of dollars into artificial intelligence initiatives, yet the vast majority see only nominal returns on their investments. The core imperative is addressing this imbalance before competitors figure it out, because the longer an organization waits to solve its structural AI delivery issues, the more the cost inefficiencies compound against them.
AI ROI is rarely determined purely by having access to the newest or best underlying technology. While technical capabilities matter, three core operational factors truly dictate whether AI investments produce compounding returns rather than linear costs:
- **Speed Advantage**: The capability to deploy AI solutions rapidly into real-world production environments.
- **Market Responsiveness**: The agility to adapt and fine-tune deployed AI systems dynamically as market conditions shift.
- **Innovation Velocity**: The capacity to continuously iterate and improve upon live models once they scale in production.
```mermaid
flowchart LR
accTitle: Three Drivers of AI Compound Value
subgraph Drivers["Core ROI Drivers"]
direction TB
S["Speed Advantage<br/>(Rapid Deployment)"]
M["Market Response<br/>(Dynamic Adaptation)"]
I["Innovation Velocity<br/>(Continuous Iteration)"]
end
subgraph Outcome["Business Impact"]
V["Compounding Value &<br/>Exponential Returns"]
end
S --> V
M --> V
I --> V
```
[02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) The $11B Waste: Why Legacy Collaboration Models Break AI Economics
In 2024, enterprise spending on AI reached nearly $14 billion, representing a massive six-fold surge compared to the previous year. However, approximately 85% of these initiatives fail to ever reach production, resulting in over $11 billion in squandered corporate capital. Only 1% of enterprise executives describe their AI deployments as truly "mature," with the overwhelming majority of projects remaining perpetually stranded in early-stage prototyping and disjointed proofs of concept.
This failure rate is not driven by model limitations; foundational LLMs, frameworks, and core architectures are continuously advancing. Instead, the failure stems from applying legacy software engineering life cycles—defined by rigid 4-to-6-week or 4-to-6-month release pipelines—to non-deterministic AI models. Non-deterministic systems require ongoing optimization and continuous refinement. Applying legacy waterfall and sprint mechanics creates an unsustainable "collaboration gap" between technical builders and business operators, breaking the fundamental economics of the platform.
```mermaid
flowchart TD
accTitle: The AI Reality Gap Breakdown
Spend["$13.8B - $14B Total AI Spend (2024)"] --> Fail["85% Fail to Reach Production<br/>(~$11B+ Wasted Capital)"]
Spend --> Prod["15% Production Deployments<br/>(Only 1% Mature)"]
Fail --> RootCause["Root Cause: Collaboration Gap<br/>Legacy 4-6 Month Lifecycles applied to Non-Deterministic AI"]
```
[04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) Engineering-Focused vs. Collaborative AI Platform Economics
Examining how organizations allocate resources reveals the structural differences between traditional engineering-siloed AI development and a unified, collaborative platform approach:
| Dimension | Engineering-Focused Approach | Collaborative Platform Approach |
|:--- |:--- |:--- |
| **Cost Structure** | High per-use-case build costs plus escalating annual maintenance per application. | Unified platform infrastructure supporting reusable components across teams. |
| **Resource Allocation** | 60% to 70% of scarce AI engineering bandwidth consumed by coordination overhead, prompt tweaking, and maintenance. | Engineering bandwidth redirected away from minor prompt adjustments toward core strategic differentiators. |
| **Scaling Advantage** | Low; scaling demands linear headcount growth for scarce, expensive AI talent. | High; 5x to 10x more AI initiatives supported per engineering resource. |
| **Cost Trajectory** | Accumulating technical debt and engineering bottlenecks with every new deployment. | 70% to 80% reduction in per-application development and iteration costs over time. |
In a realistic mid-market scenario featuring a team of 60 engineers at a loaded annual cost of $220,000 each:
- **Direct Engineering Waste**: With 60% of engineering time consumed by maintenance and coordination overhead across disconnected prototypes (averaging $85k non-labor costs per use case), nearly $8 million of payroll is locked up in operational friction.
- **Opportunity Cost**: Project delays averaging 4.5 months across 8 planned initiatives (at an estimated $800k monthly potential business value) incur nearly $4 million in delayed and forfeited revenue.
- **Collaborative Reclamation**: By implementing shared platform infrastructure and composable components, reclaiming just 30% of engineering efficiency returns approximately $4 million in direct engineering capacity, alongside $3.6 million in realized value by eliminating launch delays.
Transforming AI investments into compounding returns relies on four structural pillars. The first two establish technical leverage and operational autonomy:
1. **Pillar 1: Exponential Engineering Scalability**
Scarce engineering talent should not be dedicated to building isolated, one-off AI tools that support single business lines. Instead, engineers must focus on building reusable, composable architecture—standardized integration patterns, secure data pipelines, and optimized infrastructure—that acts as appreciating corporate assets supporting dozens of internal stakeholders across varied business units.
2. **Pillar 2: Independent Domain Expert Optimization**
To escape maintenance bottlenecks, domain experts must have direct, safe access to continuously optimize AI systems without writing back-end code. Customer service leaders possess thousands of hours of interaction nuance; financial analysts possess sophisticated forecasting context; supply chain operators understand real-world operational bottlenecks. If domain experts must filter requirements through engineering backlogs to test adjustments, context is lost and cycle times stall. Enabling domain experts to refine prompts and evaluations within governed guardrails unlocks "compound intelligence," making the AI system smarter after every interaction.
```mermaid
flowchart LR
accTitle: Pillar 1 and Pillar 2 Operating Model
subgraph Eng["Pillar 1: Engineering Core"]
E["Engineers"] -->|Build & Secure| P["Reusable Infrastructure & Guardrails"]
end
subgraph Domain["Pillar 2: Business Experts"]
P -->|Accessible via UI| UI["Visual Interface & Evaluation"]
D["Domain Specialists<br/>(Support, Finance, Ops)"] -->|Direct Iteration| UI
UI -->|Feedback Loop| Smart["Appreciating, Smarter Systems<br/>(Compound Intelligence)"]
end
```
The remaining two pillars shift the organization from treating AI as a discrete tooling initiative to operating as an AI-native business:
- **Pillar 3: Workforce Transformation & Moat Building**: Rather than focusing strictly on worker replacement through vertical AI agents, enterprise value is maximized by augmenting and scaling human institutional knowledge. Equipping internal talent with accessible AI tooling turns proprietary workflows into a defensible competitive moat and generates sustainable first-mover advantages in market share.
- **Pillar 4: Real-Time Market Responsiveness**: Traditional corporate development pipelines require quarterly cycles to identify opportunities, write specifications, build logic, and push fixes—by which time market windows have closed. An AI-native operating model achieves "same-day optimization versus quarterly cycles," enabling teams to counter market changes instantly, capture fresh operational data, and innovate ahead of competitors who remain bogged down in retrospective planning.
[15:01](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=901s) Pragmatic Implementation: The Three Core Platform Requirements
Transitioning from abstract strategy to production execution requires establishing three non-negotiable platform foundations:
- **Secure, Scalable Infrastructure with Bi-Directional Sync**: A unified foundation offering enterprise-grade security, production monitoring, and true bi-directional synchronization. Engineers must be able to work purely in code while business stakeholders inspect, manipulate, and optimize the exact same logic through a visual UI.
- **Visual Interfaces for Real-Time Domain Feedback**: Intuitive, governed interfaces that empower domain experts to test, run evaluations, and observe live model execution metrics directly without technical gatekeeping.
- **Enterprise-Wide Operational Integration**: Institutional commitment that embeds AI directly across core operational workflows, treating adoption as a holistic business transformation rather than a siloed technical experiment.
[16:16](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=976s) The Implementation Framework: Phase One and Engineering Foundations
To translate collaborative platform economics into operational reality across an enterprise, organizations must adopt a structured implementation framework. Rasam explains that the journey inherently follows a traditional "crawl, walk, run" progression spanning three distinct phases. Phase one centers entirely on establishing systematic, compounding value anchored in engineering excellence. Rather than siloing developers in one-off task queues, the technical team's mandate during this initial phase is to lay the architectural groundwork: building out the central platform and wiring up the necessary enterprise integrations so that organizational data pipelines can seamlessly feed into emerging AI use cases.
During this foundational stage, engineers focus on constructing custom, reusable components and orchestrating initial workflows. These core modules serve as the building blocks for minimum viable products (MVPs) and early field trials rolled out across various business departments. By front-loading this infrastructural legwork, the engineering organization sets up the shared ecosystem that establishes favorable platform economics, shifting the cost curve so that every subsequent deployment requires drastically less overhead.
```mermaid
flowchart TD
accTitle: Phase One Implementation Framework
subgraph Engineering ["Phase 1: Engineering Excellence"]
A["Enterprise Data Pipelines"] --> B["Central AI Platform & Integrations"]
B --> C["Custom Reusable Components"]
end
subgraph Rollout ["Departmental Deployment"]
C --> D["Initial Workflows & MVPs"]
D --> E["Cross-Department Field Trials"]
E --> F["Compounding Platform Economics"]
end
```
- **Crawl, Walk, Run**: A three-stage implementation model designed to systematically scale AI capabilities across the enterprise.
- **Phase One Focus**: Engineering excellence centered on enterprise integrations, data flow connectivity, and core architectural components.
- **Reusable Artifacts**: Developers produce modular components that power early departmental MVPs and field trials, establishing the foundation for long-term compounding returns.
---
[17:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=1028s) The 30-Second Wrap-Up and Final Takeaways
As Rasam begins detailing the framework, he checks in on his remaining session time, only to be interrupted by the session moderator: "Unfortunately, that is time." The host notes that the speaker had already been granted an extra couple of minutes to compensate for earlier technical difficulties. In a lighthearted bit of stage negotiation, Rasam asks, "Can I just have two minutes to wrap?", to which the moderator firmly counters, "I can give you 30 seconds."
Accepting the rapid countdown, Rasam pivots swiftly to invite audience members who want to hear the rest of the story to visit the Vellum team at Booth 348, located right in the center of the expo floor. He points out his counterpart, Nico, waiting in the corner of the room, and promises to stick around after the session for follow-up discussions. In his final seconds on stage, Rasam leaves the room with a core thesis: the ultimate driver of AI ROI is not merely the underlying models, but the organizational approach. Deploying AI onto a shared foundation that is directly accessible to the entire enterprise creates the operational velocity required to fundamentally transform the economic calculations behind enterprise AI investments.
- **Expo Location**: Booth 348, situated in the middle of the Ai4 expo floor.
- **Core Conclusion**: Providing an accessible, centralized platform foundation across all technical and non-technical teams accelerates deployment velocity and unlocks exponential, compounding AI returns.
## Notable quotes
> The root cause isn't technology—it's the collaboration gap that creates linear costs instead of exponential value. — Rasam Tooloee
> [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) collaboration gap — Rasam Tooloee
With 85% of AI projects failing to reach production, enterprises are wasting $11 billion of their $13.8 billion annual AI spend. The root cause isn't technology—it's the collaboration gap that creates linear costs instead of exponential value. Discover how collaborative AI platforms transform this waste into compound returns: engineering teams support 5-10x more AI initiatives without proportional headcount increases, per-application costs drop 70-80%, and business teams optimize AI in hours rather than weeks—multiplying ROI with each deployment.
Vellum将企业AI的110亿美元损失归咎于孤岛式工作流
企业AI项目的失败率达85%，原因在于传统发布周期阻碍了领域专家直接优化非确定性模型，导致工程人才深陷维护积压。
Rasam Tooloee解释称，企业AI的高失败率并非源于模型本身的技术局限，而是因为将僵化的传统开发生命周期套用于非确定性系统。为了获得复利式投资回报而非不断攀升的线性成本，企业必须将稀缺的工程人才转向构建可组合、可复用的基础设施，同时为领域专家提供可视化界面以实现持续、独立的提示词优化。采用这种协作平台模式使企业能够支持大幅增加的AI项目，显著降低单个应用的维护成本，并实现当日级的市场响应速度。
企业在 2024 年的 AI 支出接近 140 亿美元，但约 85% 的项目未能进入生产环境，导致超 110 亿美元资金浪费，其核心根源在于将传统的刚性软件开发周期套用在非确定性 AI 模型上，造成了严重的技术与业务协作脱节。 (2:41)
在传统以纯工程为主导的模式下，60% 至 70% 的技术人员精力被琐碎的维护、提示词调优和沟通摩擦占用；而采用协作式 AI 平台可使单个应用开发成本降低 70% 至 80%，并让每位工程师支持的 AI 项目数量提升 5 到 10 倍。 (4:29)
工程师应专注于构建可复用的组件与数据管道，并将受控的提示词与评估优化权限直接赋予业务专家，从而在避免研发积压的同时产生系统自我迭代的“复合智能”。 (9:05)
通过向内部员工提供可访问的 AI 工具来放大组织既有的业务知识与工作流，能构建比单纯依靠垂直 Agent 替代人工更具防御性的竞争壁垒，实现当日级优化的市场响应速度。 (12:18)
Reading summary
Rasam Tooloee of Vellum argues that enterprise AI investments largely fail to deliver economic returns because organizations misapply rigid, legacy software engineering lifecycles to non-deterministic models rather than adopting collaborative, platform-based operating models.
The AI Reality Gap and Lifecycle Mismatches (2:41)
In 2024, enterprise spending on AI reached approximately $14 billion—a sixfold increase over the prior year. Despite this capital surge, Tooloee reports that roughly 85% of initiatives fail to reach production, wasting an estimated $11 billion, while only 1% of executives describe their enterprise deployments as mature. Most corporate projects remain perpetually stalled in early-stage prototyping and disjointed proofs of concept.
Tooloee contends that this failure is driven by operational mismatch rather than model limitations. Applying traditional 4-to-6-week or 4-to-6-month engineering release cycles to non-deterministic systems creates an unsustainable collaboration gap between technical developers and operational teams. Because non-deterministic AI models require ongoing optimization, rigid waterfall mechanics introduce structural inefficiencies that compound against the enterprise over time.
Comparative Economics of Siloed vs. Platform AI (4:29)
Under traditional engineering-siloed setups, technical teams spend 60% to 70% of their bandwidth on maintenance, prompt adjustments, and coordination overhead. Scaling requires linear headcount growth for scarce technical talent. In an illustrative mid-market scenario of 60 engineers costing $220,000 annually, non-labor prototype expenses averaging $85,000 per use case combine with administrative friction to lock up nearly $8 million of engineering payroll. Delays averaging 4.5 months across eight planned initiatives forfeit an estimated $4 million in delayed and unrealized business value.
Conversely, a collaborative platform model shifts engineers away from minor prompt adjustments toward shared infrastructure and reusable components. According to Tooloee's model, this architecture can reduce per-application development and iteration costs by 70% to 80% while enabling engineers to support 5x to 10x more initiatives. Reclaiming just 30% of engineering efficiency would recover roughly $4 million in direct engineering capacity alongside $3.6 million in value from eliminated deployment delays.
Four Operational Pillars for Compound Returns (9:05)
To reverse value destruction, Tooloee outlines four structural pillars. First, organizations must develop exponential engineering scalability by redirecting scarce developers away from single-use tools toward composable architecture, including standardized integration patterns and secure data pipelines. Second, enterprises must empower domain specialists—such as customer support, finance, and supply chain operators—to optimize prompts and evaluations directly through governed guardrails. Direct iteration prevents institutional context from being lost in engineering backlogs.
The third and fourth pillars expand organizational capabilities into an enduring operational moat. Pillar three centers on workforce multiplication, augmenting proprietary institutional knowledge rather than prioritizing pure headcount reduction via vertical agents. Pillar four replaces quarterly development cycles with same-day operational optimization, allowing enterprises to adapt models dynamically to shifting market conditions and customer inputs.
Platform Foundations and Phased Engineering Execution (15:01)
Translating this operating model into production requires three non-negotiable platform capabilities: secure infrastructure featuring bi-directional synchronization between code and visual interfaces, real-time testing and evaluation environments for non-technical domain specialists, and deep operational workflow integration across the enterprise.
Implementation follows a traditional "crawl, walk, run" trajectory. Tooloee explains that Phase One focuses on engineering foundations, where developers construct central platform integrations, connect enterprise data pipelines, and assemble custom, reusable modules. These components form the basis for initial workflows, minimum viable products, and cross-departmental field trials, establishing the infrastructure needed for subsequent deployments to operate at substantially lower marginal costs.
Tooloee concludes that enterprise AI returns depend on organizational delivery models and accessible platform infrastructure rather than raw underlying model capabilities. However, because session time constraints curtailed the presentation, detailed operational frameworks for the subsequent "walk" and "run" implementation phases were left unaddressed.
## Background
该内容由 Ai4 频道于 2025 年 8 月 27 日发布在 YouTube 平台上，演讲者为来自协同 AI 平台 Vellum 的市场拓展（GTM）负责人 Rasam Tooloee。视频围绕企业级 AI 投资回报率（ROI）的困境展开，深入剖析了为何高达 85% 的企业 AI 项目未达预期，并提出了如何通过协同平台打通工程与业务壁垒，将线性成本浪费转化为指数级复合回报的战略路径。
## TL;DR
演讲指出，当前企业 AI 投资效率低下的根源并非模型与算法等底层技术不足，而是传统的线性软件工程模式无法适配非确定性 AI 开发所带来的“协作鸿沟”。通过构建支持工程与业务双向同步的协同 AI 平台，工程师可从繁琐的 Prompt 调优中抽身去沉淀高复用性底层资产，而一线业务专家则能直接进行同日迭代与优化。这种平台经济学模式不仅能降低 70% 至 80% 的单应用开发成本并大幅减少项目延期损失，还能通过全员赋能构筑起企业的专有数据与专业知识护城河。
## Atmosphere & style
整场演讲节奏紧凑、干货密集且极具务实说服力。开场时讲台屏幕发生闪烁，演讲者机智地借此调侃企业不稳定的 AI ROI，迅速拉近了与观众的距离并化解了尴尬；在随后的分享中，他结合详实的财务测算数据与生动的架构对比，层层推进论点。结尾处虽然遭遇主持人严苛的“30 秒倒计时”催促，演讲者依然在紧迫中保持从容与幽默，用一记直击核心的战略总结给全场留下了深刻印象。
## Key takeaways
- [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) 企业在 AI 领域普遍面临投资回报率低下的困境，决定能否实现复合回报的核心在于速度优势、市场响应能力以及持续创新迭代速率，而非单纯拥有顶尖技术。
- [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) 全球企业在 AI 领域投入巨大，但高达 85% 的 AI 项目从未成功落地生产环境，其根源在于用传统的确定性软件开发流程应对非确定性 LLM 所造成的“协作鸿沟”。
- [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 在传统的工程主导模式下，AI 工程师 60% 至 70% 的时间被消耗在跨部门琐碎沟通与系统维护中，导致极高的单用例成本与严重的技术交付瓶颈。
- [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 协同平台模式通过沉淀可复用组件，可使工程团队在不按比例扩招的情况下支撑 5 到 10 倍的 AI 项目，并将单应用开发成本大幅降低 70% 至 80%。
- [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 通过消除平均 4.5 个月的项目交付延误并释放工程产能，协同平台模式能为典型中型企业挽回数百万美元的机会收益并盘活工程价值。
- [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 实现变革的支柱之一在于打造工程杠杆（Engineering Leverage），让工程师专注于构建像金融资产一样随时间持续升值的底层集成模式与组件库。
- [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 变革的第二支柱是在安全护栏保障下，赋予客服、财务等业务领域专家独立调优 AI 的能力，避免知识在跨部门传递中失真，沉淀出“复合智能”。
- [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) AI 变革的目标不是单纯裁汰员工，而是推动全员能力转型，将专有业务经验与内生数据通过 AI 进行指数级放大，筑牢竞争壁垒。
- [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 协同模式能够以“同日发现机会并完成优化上线”的极致响应速度，颠覆传统的季度交付周期，在市场变化中夺得先发优势。
- [15:01](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=901s) 协同 AI 落地依赖三大基础设施基石：支持代码与可视化界面双向同步的安全底座、面向业务专家的交互环境，以及全组织维度的运营融入。
- [16:16](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=976s) 实施协同平台应遵循“爬、走、跑”三阶段演进策略，在第一阶段重点打通数据管道与系统集成，通过 MVP 快速确立平台经济学模型。
- [17:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=1028s) 在全组织范围内打造“全员皆可访问”的 AI 部署底座，是建立高交付速度并从根本上扭转企业 AI 投资回报率的关键。
## Detailed walkthrough
### [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) AI 投资回报率的困境与三大核心价值驱动力
演讲开场时，讲台上的投影屏幕出现了一些闪烁，来自 Vellum 的 Rasam Tooloee 幽默地借此自嘲：“刚才屏幕有点闪烁，就像企业当前的 AI ROI 一样不稳定。”不过设备很快恢复了正常，他随即切入正题。当前不论是中型企业还是大型跨国集团，都在向人工智能领域砸入数以百万计甚至千万计的资金，但残酷的现实是，绝大多数企业获得的投资回报都微乎其微。Rasam 指出，这一问题必须尽早解决，因为随着时间推移，低效的模式会产生对企业极为不利的“负向复利”；如果不能赶在竞争对手之前摸清门道，差距将会被迅速拉大。
Rasam 强调，在 AI 领域获取高额 ROI 的核心从来不只是“拥有最顶尖的技术”。技术固然是基础，但决定企业能否实现复合回报的，是以下三大关键能力：
- **速度优势（Speed Advantage）**：能够以极高的效率将 AI 应用快速推向生产环境。
- **市场响应能力（Market Response）**：在 AI 部署上线后，能够根据瞬息万变的市场反馈即时调整与迭代。
- **创新迭代速率（Innovation Velocity）**：在应用进入生产阶段后，拥有持续优化、打磨和升级系统的产能与机制。
现实中许多企业在早期阶段做出了几个原型，但一旦业务规模扩大，整个团队就陷入了疲于应付的困境。企业追求的回报不应是线性的微弱增长，而应该通过这三大关键因素形成长期的“指数级复合效应”（Collaborative Compound Effects）。
---
### [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) 现实差距：85% 的 AI 项目为何折戟与“协作鸿沟”
从宏观数据来看，企业对 AI 的投入与实际产出之间存在巨大的鸿沟。2024 年，全球企业在 AI 上的支出接近 140 亿美元（较前一年增长了 6 倍），但其中高达 85% 的 AI 项目从未成功进入生产环境。换算下来，企业在 AI 上的浪费高达 110 多亿美元。更令人触目惊心的是，仅有 1% 的企业高管认为自己公司的 AI 部署达到了“成熟”水平。绝大部分项目依然被困在早期的 PoC（概念验证）或原型开发阶段。
这一困境的根源并不在底层技术。底层大语言模型（LLM）正变得越来越强，算法与开发框架也日趋精密，真正滞后的是**协作模式（Collaboration Model）**。许多企业试图套用传统的软件工程流程——那些周期长达 4 至 6 周甚至 4 至 6 个月的流水线去开发 AI。然而，传统工程流程适用于确定性软件，无法应对 LLM 这类非确定性（Non-deterministic）模型的调优与上线需求。这种“协作鸿沟”正是导致 AI 经济模型崩溃、成本居高不下的头号元凶。
```mermaid
flowchart LR
accTitle: The Collaboration Gap in Enterprise AI
subgraph S1["传统线性模式 (Legacy Pipeline)"]
A1["业务提需求"] --> B1["工程排期编码"]
B1 --> C1["4-6 个月交付上线"]
C1 --> D1["市场变化/维护瓶颈"]
end
subgraph S2["协同复合模式 (Collaborative Platform)"]
A2["工程沉淀可复用组件"] --> B2["统一协作平台"]
B2 --> C2["业务专家同日迭代优化"]
C2 --> D2["指数级复利回报 (ROI)"]
end
```
---
### [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 平台经济学：工程驱动模式 vs 协同平台模式的成本与效益剖析
如果继续沿用纯工程主导的传统软件开发范式，企业的成本结构将是“单用例高额投入 + 持续年度维护”。在这种架构下，工程师需要全程包揽需求对接、编写代码、提示词微调、收集反馈和后期维护。统计显示，企业 AI 工程师 60% 至 70% 的时间都被消耗在跨部门沟通、琐碎协调和系统维护等低效事务中。在市场上 AI 人才极度稀缺且昂贵的情况下，这种做法直接导致了严重的工程瓶颈与技术债务，完全无法形成规模化优势。
与之相对的“协同平台模式”（Collaborative Approach）则主张通过统一平台打通工程与业务的壁垒：
- **释放工程核心产能**：工程师不再充当“修改 Prompt”或“调换模型”的人工接口，而是专注于构建具有核心竞争壁垒的高价值底层逻辑与可复用组件。
- **5-10 倍的人效提升**：在不按比例扩充技术团队的前提下，每位工程师所能支撑的 AI 业务项目数量提高 5 到 10 倍。
- **70% - 80% 的单应用成本骤降**：随着开发流程标准化与组件高度复用，新应用的边际开发成本将大幅下挫。
为了更直观地呈现两者的财务差距，Rasam 给出了一个典型的中型企业（Mid-Market）测算场景：
| 评估维度 | 传统工程主导模式 (Engineering-Focused) | 协同平台模式 (Collaborative Platform) |
|:--- |:--- |:--- |
| **基础项目与人力成本** | 8 个用例，非人力成本 68 万美元/年；60 名工程师（年均综合成本 22 万美元），其中 60% 时间用于协调与维护，产生近 **800 万美元** 的隐形成本消耗。 | 工程师专注底层资产构建，回收 **30% 的人效**，直接为企业盘活约 **400 万美元** 的工程价值。 |
| **延期机会成本** | 因交付瓶颈平均延误 4.5 个月（按单项目 80 万美元/月潜在业务价值计算），造成近 **400 万美元** 的机会收益损失。 | 敏捷上线抹平交付延迟，挽回并实现约 **360 万美元** 的业务机会价值。 |
| **组织扩展能力** | 严重依赖工程单点，技术债务堆积，全业务线排队等待，扩展性极低。 | 业务线自主迭代，边际协作成本极低，形成指数级滚雪球式的复利价值。 |
---
### [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 成功实现 AI 变革的四大支柱
为了真正摆脱资源黑洞并实现投资回报复利，企业需要依靠四大支柱构建全新的 AI 运作体系：
#### 支柱一：让工程资源产生指数级规模效应（Engineering Leverage）
AI 工程师是企业最宝贵的技术资产。传统模式下一个团队只能死磕一两个独立应用，而在协同模式下，工程师致力于构建可复用的组件库、通用的集成模式以及优化的基础设施。这些底层资产像金融资产一样随时间持续升值，单套平台即可支撑数十个业务部门的上百种 AI 场景。
#### 支柱二：业务领域专家独立驱动 AI 持续优化（Domain Experts Optimization）
“独立”（Independently）是这一支柱的核心。只要平台具备完善的护栏（Safeguards）、合规治理与安全控制，企业就应大胆赋予业务人员直接调优 AI 的权限：
- **客服管理者**：拥有成千上万小时一线对话沉淀的直觉与洞察。
- **财务分析师**：具备多年行业研判经验，能精准校正预测模型。
- **供应链运营专家**：掌握深厚的组织内生经验，懂得如何让流程在实际场景中运转顺畅。
如果每次调优都需要业务人员开会、写需求文档、由工程师在代码中摸索修改，领域知识在传递中就会严重失真。让专家直接介入 AI 优化，系统才能沉淀出真正的“复合智能”（Compound Intelligence）。
#### 支柱三：全员赋能与劳动力转型，构筑企业竞争壁垒（Workforce Transformation）
AI 变革的目标不是单纯地用自动化 Agent 裁汰员工，而是将全体员工的知识、洞见与敏锐度通过 AI 进行指数级放大。当全体员工都深度掌握 AI 工具并与日常业务融为一体时，企业的内生专有数据与专业知识就转化为了难以被复制的护城河（Moat），从而在市场份额与创新能力上建立先发优势。
#### 支柱四：以“同日迭代”对抗“季度周期”，实现极致的市场响应速度（Market Speed Edge）
在传统流程下，从发现业务机会到提需求、排期开发、测试再到部署，往往需要数月甚至按季度计算，彼时市场风口早已过去。而在协作型 AI 原生组织中，业务团队能够做到“当天发现机会、当天完成优化与上线”（Same-day Optimization vs Quarterly Cycles）。更快的响应带来更多业务触点，产生更多真实数据，进而驱动更快速的迭代——当竞争对手还在开会复盘上个季度的变化时，你的团队已经在定义下个季度的行业创新。
---
### [15:01](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=901s) 落地协同 AI 的三大核心基础设施要素
要将上述愿景转变为工程与业务现实，企业必须具备三大基础设施基石：
- **安全、可扩展的底层基础设施（Secure & Scalable Infrastructure）**：具备企业级安全合规体系、生产级监控与追踪能力、可复用的逻辑组件，以及至关重要的**双向同步机制（Bidirectional Sync）**——工程师可以直接在代码库中编程，而非技术背景的业务人员则可以在直观的可视化界面（Visual UI）中查看逻辑、进行调试与优化，两者底层的代码与配置保持完全实时同步。
- **面向领域专家的交互环境（Visual Interface for Domain Experts）**：提供低门槛的可视化工作台，让业务专家能够在安全边界内即时获得调优反馈，直观理解 AI 逻辑并快速进行微调。
- **全组织维度的运营融入（Organizational Support & Embedding）**：明确认识到 AI 不仅是一个技术或 IT 部门的孤立项目，而是一场深刻的业务运营转型，必须将其全方位嵌入到企业的日常经营与核心业务流程中。
### [16:16](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=976s) 落地实施的三阶段框架与工程底座构建
在企业落地协同 AI 平台时，需要遵循循序渐进的实施框架，也就是常说的“爬、走、跑”（crawl, walk, run）三阶段演进策略。第一阶段（Phase 1）的核心目标，是通过卓越的工程能力系统化地构建复合价值（compound value）。在这一阶段，技术团队负责搭建平台基础设施，完成关键的企业级系统集成，确保数据管道（data pipelines）能够顺畅贯通并赋能各类 AI 业务场景。
与此同时，工程团队会开发出一套自定义的可复用组件，并启动第一批工作流，以此支持各业务部门开展最小可行性产品（MVP）验证和早期现场试验（early field trials）。这一奠基性阶段的工作至关重要，正是它正式确立了前面所讨论的平台经济学模型，为后续规模化扩展打下成本与效率的基础。
```mermaid
flowchart TD
accTitle: Phase 1 Implementation Framework
subgraph S1["第一阶段：工程底座建设"]
A["企业级系统集成 & 数据管道"] --> B["开发自定义可复用组件"]
B --> C["启动初始工作流与 MVP 试验"]
end
C --> D["确立平台经济学基础 (Platform Economics)"]
```
- **实施路径**：遵循“爬、走、跑”三阶段渐进式框架。
- **第一阶段核心任务**：技术团队负责底层企业集成、数据流水线打通、构建可复用自定义组件。
- **业务验证**：快速向各业务部门输出 MVP 和早期试点，确立后续带来复合回报的平台经济模型。
---
### [17:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=1028s) 现场交锋收尾与全员可访问性带来的经济学变革
正当演讲者准备深入展开后续阶段时，现场主持人介入打断，提示演讲时间已到。尽管此前因技术故障已额外延长了几分钟，但时间依然用尽。演讲者幽默地争取道：“我能再要两分钟做个总结吗？”主持人则回应：“只能给你 30 秒。”
在极其紧凑的 30 秒倒计时中，演讲者迅速给出了参会指引并留下了核心结语：如果参会者想要了解未完待续的内容，可以前往展厅中央的 348 号展位（Booth 348），团队成员和驻守现场的同事 Nico 都在那里随时准备深入交流。最后，演讲者向全场留下了最具战略意义的总结：“在整个组织中构建一个全员皆可访问的 AI 部署底座，是建立交付速度（velocity）的真正基石，而这种速度将从根本上改变驱动企业 AI 投资回报的经济学计算方式。”
- **展台信息**：展厅正中央的 348 号展位（Booth 348），现场有团队成员及同事 Nico 驻守交流。
- **核心总结观点**：企业部署 AI 的策略必须建立在“全员可访问（accessible to entire organization）”的基础之上，这才是提升业务迭代速度、彻底扭转 AI 投资经济模型并实现复合回报的关键。
## Notable quotes
> [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) 刚才屏幕有点闪烁，就像企业当前的 AI ROI 一样不稳定。 — Rasam Tooloee
> [00:04](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=4s) 在 AI 领域获取高额 ROI 的核心从来不只是‘拥有最顶尖的技术’。 — Rasam Tooloee
> [02:41](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=161s) 真正滞后的是协作模式（Collaboration Model）。 — Rasam Tooloee
> [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 企业 AI 工程师 60% 至 70% 的时间都被消耗在跨部门沟通、琐碎协调和系统维护等低效事务中。 — Rasam Tooloee
> [04:29](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=269s) 工程师不再充当‘修改 Prompt’或‘调换模型’的人工接口，而是专注于构建具有核心竞争壁垒的高价值底层逻辑与可复用组件。 — Rasam Tooloee
> [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 让专家直接介入 AI 优化，系统才能沉淀出真正的‘复合智能’（Compound Intelligence）。 — Rasam Tooloee
> [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) AI 变革的目标不是单纯地用自动化 Agent 裁汰员工，而是将全体员工的知识、洞见与敏锐度通过 AI 进行指数级放大。 — Rasam Tooloee
> [09:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=548s) 业务团队能够做到‘当天发现机会、当天完成优化与上线’（Same-day Optimization vs Quarterly Cycles）。 — Rasam Tooloee
> [15:01](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=901s) 工程师可以直接在代码库中编程，而非技术背景的业务人员则可以在直观的可视化界面（Visual UI）中查看逻辑、进行调试与优化，两者底层的代码与配置保持完全实时同步。 — Rasam Tooloee
> [17:08](https://www.youtube.com/watch?v=j1PYD4N0cCM&t=1028s) 在整个组织中构建一个全员皆可访问的 AI 部署底座，是建立交付速度（velocity）的真正基石，而这种速度将从根本上改变驱动企业 AI 投资回报的经济学计算方式。 — Rasam Tooloee
## Mentioned
Rasam Tooloee, Vellum, Ai4, Nico, LLM (大语言模型)
## Video description
Rasam Tooloee, GTM, Vellum
Abstracts
With 85% of AI projects failing to reach production, enterprises are wasting $11 billion of their $13.8 billion annual AI spend. The root cause isn't technology—it's the collaboration gap that creates linear costs instead of exponential value. Discover how collaborative AI platforms transform this waste into compound returns: engineering teams support 5-10x more AI initiatives without proportional headcount increases, per-application costs drop 70-80%, and business teams optimize AI in hours rather than weeks—multiplying ROI with each deployment.
[Source](https://www.youtube.com/watch?v=j1PYD4N0cCM)
Source quotes
[0:04] Rasam Tooloee: 刚才屏幕有点闪烁，就像企业当前的 AI ROI 一样不稳定。
[0:04] Rasam Tooloee: 在 AI 领域获取高额 ROI 的核心从来不只是‘拥有最顶尖的技术’。
[2:41] Rasam Tooloee: 真正滞后的是协作模式（Collaboration Model）。
[4:29] Rasam Tooloee: 企业 AI 工程师 60% 至 70% 的时间都被消耗在跨部门沟通、琐碎协调和系统维护等低效事务中。
[4:29] Rasam Tooloee: 工程师不再充当‘修改 Prompt’或‘调换模型’的人工接口，而是专注于构建具有核心竞争壁垒的高价值底层逻辑与可复用组件。
[9:08] Rasam Tooloee: 让专家直接介入 AI 优化，系统才能沉淀出真正的‘复合智能’（Compound Intelligence）。
[9:08] Rasam Tooloee: AI 变革的目标不是单纯地用自动化 Agent 裁汰员工，而是将全体员工的知识、洞见与敏锐度通过 AI 进行指数级放大。
[9:08] Rasam Tooloee: 业务团队能够做到‘当天发现机会、当天完成优化与上线’（Same-day Optimization vs Quarterly Cycles）。
[15:01] Rasam Tooloee: 工程师可以直接在代码库中编程，而非技术背景的业务人员则可以在直观的可视化界面（Visual UI）中查看逻辑、进行调试与优化，两者底层的代码与配置保持完全实时同步。
[17:08] Rasam Tooloee: 在整个组织中构建一个全员皆可访问的 AI 部署底座，是建立交付速度（velocity）的真正基石，而这种速度将从根本上改变驱动企业 AI 投资回报的经济学计算方式。
Source quotes
Rasam Tooloee: The root cause isn't technology—it's the collaboration gap that creates linear costs instead of exponential value.
[2:41] Rasam Tooloee: collaboration gap
[9:05] Rasam Tooloee: compound intelligence
[12:18] Rasam Tooloee: same-day optimization versus quarterly cycles
[16:16] Rasam Tooloee: crawl, walk, run
[17:08] Moderator: Unfortunately, that is time.
[17:08] Rasam Tooloee: Can I just have two minutes to wrap?
[17:08] Moderator: I can give you 30 seconds.
Original transcript
[0:04] Thank you for being here. Uh really
[0:06] appreciate it. Uh I uh we're have we
[0:09] were having a bit of an issue with
[0:10] flickering screens just like um AI ROI.
[0:14] It was inconsistent, but uh looks like
[0:16] we've stabilized. So hopefully this will
[0:17] work uh well and we can cover the topic.
[0:20] Um we'll be talking about AI ROI today.
[0:22] I I suspect if you're in this room and
[0:25] you're curious about this topic, you're
[0:27] either in the middle of your AI journey
[0:28] and and um not necessarily getting the
[0:31] level of return on your investments that
[0:32] you had hoped or you've heard a lot
[0:34] about some of the challenges that
[0:36] organizations are having uh with getting
[0:38] ROI. Um a lot of organizations uh
[0:41] midsize large enterprises are are
[0:44] investing millions of dollars into AI
[0:46] and uh the reality is they're seeing u
[0:49] nominal returns on their investments.
[0:51] So, we want to take an interesting look
[0:52] today on why that's happening. A unique
[0:54] perspective that's hopefully
[0:55] thoughtprovoking for you. Um, really the
[0:58] news uh or the the core message I want
[1:00] to convey to you is this is something
[1:02] you want to get ahead of early. The
[1:03] longer you wait, the more it compounds
[1:05] against you. Um, and uh you know, you
[1:08] want to figure this out before your
[1:10] competitors do. That's uh that's pretty
[1:11] key. I'm actually going to click in two
[1:14] places, so bear with me if uh there's a
[1:16] bit of a delay. Um so really um we
[1:21] believe uh I should mention that uh I'm
[1:24] with an organization called Vellum. We
[1:25] build an AI development platform. We do
[1:27] have a booth here if you want to come
[1:28] see us um after the talk. We'd love to
[1:31] answer any questions that you may have.
[1:33] Um really u ROI when it comes to AI is
[1:36] not necessarily about having the best
[1:38] technology. Obviously that's important
[1:40] but there are other critical factors in
[1:42] our mind these three critical factors
[1:44] that contribute to your ROI. First is
[1:46] having a speed advantage and what that
[1:48] really means is the ability to deploy AI
[1:51] rapidly quickly very important. Uh
[1:54] market response means once you deploy it
[1:56] you can quickly adapt your AI to changes
[1:58] that happen in the market. Uh and third
[2:01] is your innovation velocity. So that's
[2:03] the ability and the capacity to
[2:05] continuously improve on the AI once it's
[2:08] in production. These are critical
[2:09] pieces. We see organizations all the
[2:11] time that build early prototypes, deploy
[2:14] them, and as usage scales, they
[2:16] challenge they're challenged with being
[2:18] able to keep up. So the idea is the
[2:21] returns shouldn't just be linear. They
[2:23] should have exponential value. They
[2:24] should compound over time. And that's
[2:26] what these these three critical factors
[2:28] get you. Um, so we have this business
[2:31] value framework in mind as we talk about
[2:33] this. And the idea is this collaborative
[2:35] compound effects that happen over time.
[2:37] So let me show you specifically what
[2:38] that means.
[2:41] Um so again the reality gap is the fact
[2:44] that in 2024 organizations spend almost
[2:47] 14 billion on AI. So a six-fold increase
[2:50] from the prior year. Um but 85%
[2:55] failed to reach uh production which is
[2:58] if you do the math that's over 11
[3:00] billion in wasted investment by large
[3:02] enterprises or by enterprises mid and
[3:04] large. Um it's uh it's a challenge that
[3:08] executives have to face every day. 1% of
[3:10] executives that describe their AI
[3:12] deployments as actually being mature. Uh
[3:14] which is a pretty small number if
[3:15] considering the massive budgets that are
[3:17] going into building and deploying AI.
[3:19] And many of the uh projects are still
[3:21] stuck in early stage uh prototyping and
[3:24] early stage deployments. So um what's
[3:28] also interesting is you c you um
[3:32] um you know look at the fact that the
[3:34] techn is fundamentally getting a lot
[3:36] better. Uh LLMs are more powerful. The
[3:39] models keep getting better. The
[3:40] frameworks are getting more
[3:41] sophisticated. But what hasn't kept up
[3:44] is the collaboration model that's
[3:46] required for AI to scale effectively.
[3:48] Um, and that's really at the core of
[3:50] this compound effect that we're talking
[3:52] about that will change the dynamics, the
[3:54] platform economics of what you're
[3:56] getting from your AI. Um, so um, really
[4:00] what I would net it out at is net it out
[4:02] as is applying legacy development
[4:05] platforms uh, with four to six week four
[4:08] to six month product life cycles and and
[4:11] pipelines.
[4:12] um they are proven but they don't
[4:15] necessarily apply to AI and these
[4:17] non-deterministic models and what it
[4:19] takes to prime them for production. So
[4:22] the collaboration gap, keep this in
[4:23] mind, is the number one reason from our
[4:25] perspective the economics fail over
[4:26] time.
[4:29] So um let's let's take a deeper look at
[4:32] this collaborative platform economics
[4:34] that we're talking about. Um when you
[4:36] look at engineering focused approaches
[4:38] based on these legacy paradigms, what
[4:40] does that look like from a cost
[4:41] structure perspective? Typically what
[4:42] that means is investment per use case
[4:45] plus annual maintenance. So engineering
[4:48] owning the process end to end, writing
[4:50] code, gathering requirements, um
[4:53] iterating and refining based on feedback
[4:55] and then really owning the solution. Uh
[4:57] from a resource allocation perspective,
[4:59] it's common to see 60 to 70% of
[5:01] engineering time uh spent on
[5:03] coordination and maintenance of each
[5:06] solution that's built. If you think
[5:08] about that over time, that's a lot of
[5:10] technical debt that the engineering team
[5:12] and if you I don't know what your
[5:13] organization's like, but typically what
[5:14] we see is AI engineers are not a dime a
[5:17] dozen. They're hard to find. It's a
[5:18] difficult skill set to find. Um more and
[5:21] more adjacent skill sets are getting
[5:23] into AI, but still it's a challenge. Um
[5:26] from a scaling advantage perspective,
[5:28] the scaling advantage is low with this
[5:30] approach because each AI initiative
[5:32] requires engineering investment and time
[5:34] and resources. Um and the result is
[5:37] engineering bottlenecks and technical
[5:39] debt. It's a very common pattern and if
[5:42] if your environment is different, I'd
[5:43] love to hear about it after after the
[5:45] talk because I'd be curious to hear what
[5:46] you've done differently. From a
[5:48] collaborative approach perspective, the
[5:50] key is a unified platform. So a unified
[5:53] platform that enables a wide range of
[5:55] applications not not the answer that's
[5:57] not the answer by itself but it's the
[5:58] foundation for a collaborative approach.
[6:00] So your platform choice is key. Um
[6:03] engineering time can then be redirected
[6:05] to strategic work. So instead of
[6:07] engineers being prodded to update a
[6:09] prompt or try a different model or or
[6:12] you know tweak something in the logic of
[6:14] the AI their focus is on building more
[6:16] strategic things that are
[6:17] differentiators for your business. Um
[6:20] what it it should result in is 5 to 10x
[6:23] uh more invest AI initiatives per
[6:25] engineering uh resource and per
[6:27] engineering spend. Um and the idea is 70
[6:31] to 80% reduction in per application
[6:33] development costs over time. So as these
[6:35] things as these advantages compound
[6:37] speed market responsiveness innovation
[6:39] velocity over time you should see
[6:41] significantly different returns.
[6:45] So let's let's dive a little bit deeper
[6:47] and just look at a mid-market scenario.
[6:49] Um so if you think about engineering
[6:51] cost number of projects uh the cost of
[6:55] actually deploying AI uh engineering
[6:57] focused approach in this particular
[6:58] scenario uh project cost non- labor 85k
[7:02] average per use case uh for a total of
[7:04] 680k per year. um the actual engineering
[7:07] cost. Let's let's uh um imagine a
[7:11] scenario where we have 60 engineers,
[7:13] 220,000 per year loaded cost for having
[7:16] these engineers as a part of your staff.
[7:18] Um and 60% of their time is uh in
[7:21] coordination, overhead and maintenance
[7:22] of AI solutions. So that's almost $8
[7:24] million in the scenario of um effort
[7:27] that could be reallocated elsewhere or
[7:30] it could be reclaimed partially to get
[7:32] reallocated elsewhere. From an
[7:34] opportunity cost perspective um if
[7:37] things are move not moving with the
[7:39] speed and the velocity that they could
[7:41] there is um there's an opportunity
[7:44] opportunity cost associated with lost
[7:46] revenue. So in this scenario 4 and a.5
[7:48] months delay eight projects 800k per
[7:52] month potential value almost $4 million
[7:55] u and then the impact is coordination
[7:57] overhead technical debt limited
[7:59] organizational scalability because
[8:00] limited resources trying to deploy more
[8:02] and more solutions. It just it's taking
[8:05] all of your workloads. Envision taking
[8:06] all the workloads in your business and
[8:08] trying to stuff them through your your
[8:10] engineering pipeline. is there's bound
[8:12] to be limitations in how how much
[8:13] agility you have and how fast you can
[8:15] move. So contrast that with this
[8:18] collaborative approach where you have
[8:20] platform infra reusable components and
[8:22] I'll get into what I mean by reusable
[8:23] components in a moment. Um if you can
[8:26] reclaim 30% efficiency uh in your
[8:28] engineering workforce that's just
[8:30] inherently $4 million in value in this
[8:32] scenario. um if you can um reclaim the
[8:36] opportunity cost because you can move
[8:39] more quickly and gain those months back
[8:40] in terms of getting these solutions into
[8:42] production that's 3.6 million in
[8:45] realized value in this case. Um so
[8:47] ultimately what you're striving for is
[8:49] increased velocity plus scalability and
[8:52] small over time small incremental
[8:54] overheads per business unit as you
[8:56] deploy more applications. So that's the
[8:58] compound effect that we're striving for
[9:00] based on those uh those foundational
[9:02] components that I talked about before
[9:03] and based on the pillars that I'll share
[9:05] with you. So speaking of the pillars,
[9:08] there are four pillars in our uh in our
[9:10] mind to successful AI transformation and
[9:12] we'll go through each one. Um the first
[9:14] is and I've alluded to this a bunch of
[9:16] times. Your engineering resources are
[9:18] precious. The work that they do to build
[9:21] your platforms should scale
[9:23] exponentially. Um and what that really
[9:25] means is um traditionally again speaking
[9:28] one engineer or a small team builds one
[9:30] or two AI applications with a
[9:33] collaborative platform the same platform
[9:35] supports dozens of stakeholders because
[9:37] they're building reusable composable
[9:40] components um whatever your organization
[9:42] is the integration patterns you need uh
[9:45] the optimization you need for your
[9:46] infrastructure that can get reused
[9:49] across the various lines of business um
[9:51] and the various uh AI workflows and
[9:53] processes.
[9:54] U so over time engineering effort grows
[9:57] into assets potentially multi-million
[9:59] dollar assets that appreciate over time.
[10:02] Um so the the idea is the platform
[10:05] supports a vast array of businesses or
[10:08] business applications. Uh and each each
[10:11] additional business application that
[10:12] uses it this adds value to the
[10:14] organization. Um and and again this is
[10:17] the core foundation of the compound
[10:18] effect. We're going to build on top of
[10:19] it. But this is foundational. It's
[10:21] fundamental. Obviously it's AI. So we
[10:22] know engineering work is precious.
[10:24] Pillar two focuses on business experts
[10:28] independently and independently is is
[10:30] the key word here driving continuous AI
[10:32] optimization. Now that seems like a
[10:35] stretch that's that's that's a bridge to
[10:37] cross for them to independently do it.
[10:39] So the idea is how do you give them a
[10:41] platform that makes this accessible and
[10:44] usable independently in a way that's
[10:46] safe? It's got the safeguard, the
[10:48] safeguards, the governance, the
[10:49] compliance, the security constraints in
[10:51] place so that you can open that uh that
[10:54] floodgate of trust. Um but if they can
[10:58] independently drive continuous
[11:00] optimization, this result in the
[11:02] compound effect accelerating. Um think
[11:04] about it from a different uh you know
[11:06] set of domain experts perspective. Your
[11:07] customer service leaders have gained um
[11:11] insight about customer interactions from
[11:14] thousands of hours of experience. uh in
[11:17] being in those kinds of situations. Your
[11:18] financial analysts can refine forecasts
[11:21] because they have uh competitive and
[11:23] market intelligence that took years to
[11:25] develop. Your engineers can sit in a
[11:28] room and get feedback on how AI is
[11:30] performing and take that feedback and
[11:32] try to code it. But if your domain
[11:34] experts can directly engage with the AI,
[11:36] that domain expertise is not going to
[11:38] just be transferred through a set of
[11:40] meetings requirements or some data that
[11:42] the engineering team reviews. It's it's
[11:44] deeper than that. It's more substantive
[11:46] than that. Same thing for operational
[11:48] leaders. Your operational leaders have
[11:50] institutional knowledge. They can take
[11:52] supply chain um and make AI that
[11:55] actually works for supply chain,
[11:56] optimize and refine AI that actually
[11:58] works. So in a nutshell, your experts
[12:00] can directly optimize your AI systems
[12:02] and through that you get greater
[12:04] compound returns. You get compound
[12:06] intelligence. So every optimization
[12:08] makes the system smarter and the closer
[12:11] your domain experts are to making that
[12:13] optimization, the greater leverage you
[12:14] have as an organization.
[12:18] So pillar three over time results in
[12:21] workforce transformation, right? Because
[12:22] what what you're doing is you're
[12:24] engaging your entire organization closer
[12:27] to AI. Um and that's going to give you a
[12:29] competitive edge. But again, that takes
[12:31] time. Uh and it takes right platform,
[12:34] right sort of access, right sort of
[12:35] constructs. But this isn't about
[12:37] replacing workers. This is about making
[12:39] them exponentially more valuable. Um
[12:41] there are a lot of you there's a lot of
[12:43] talk in the industry, a lot of vendors
[12:44] here. There's definitely a lot of value
[12:46] to be gleaned from AI agents doing
[12:48] vertical type of work. Um so this isn't
[12:51] it's not one or the other but there is a
[12:54] vast opportunity for not replacing
[12:56] workers but actually scaling them
[12:58] scaling their expertise their knowledge
[13:01] um their u their acumen through AI. And
[13:05] that's how you build a substantive mode
[13:07] because you're taking all of that in uh
[13:09] intellectual property and and domain
[13:10] knowledge that your organization has and
[13:12] accelerating it with AI and building
[13:14] that mode. Um so what that can give you
[13:17] is first mover advantage. Um that can
[13:20] that can be measured in all kinds of KPI
[13:22] and metric like greater market share. Um
[13:24] so your your talent gets better through
[13:27] the use of AI. you can drive more
[13:29] innovation and you strengthen your
[13:31] market position hopefully before your
[13:33] competition because that's what's going
[13:35] to give you the competitive mode that
[13:36] you're looking for. And pillar four is
[13:39] how you boost your competitive edge. Uh
[13:42] because now think about it, you're more
[13:44] more AI native organization. You're more
[13:46] AI native workforce. Um your ability to
[13:50] respond at market speed is on a whole
[13:52] different level. You can identify an
[13:54] opportunity. Um think well if you think
[13:58] about it in tra the traditional paradigm
[13:59] of going back into the AI development
[14:01] pipeline um using traditional paradigms
[14:04] if an opportunity exists for
[14:06] optimization with AI uh you have to
[14:08] identify the uh opportunity you have to
[14:11] specify the requirements you have to
[14:12] build a solution you have to deploy it
[14:15] right well by then the market's moved on
[14:18] right the the opportunities moved on
[14:20] with collaborative AI you get same day
[14:22] optimization versus quarterly cycles
[14:25] That's really a key construct quarterly
[14:26] cycles versus same day iteration once
[14:29] you're AI native and have an AI native
[14:31] workforce. Um you can counter
[14:33] immediately you can adapt instantly and
[14:36] that's what results in this compound
[14:38] effect becoming dramatic. Faster
[14:39] response uh results in more market
[14:42] opportunities. It generates more data
[14:44] and insights for your organization which
[14:45] then allows you to have even faster
[14:47] response. So hopefully you end up in a
[14:50] position where your competitors while
[14:51] your competitors are still planning on
[14:53] how to respond to the changes of the
[14:55] last quarter, you're focused on driving
[14:57] the innovation for the next quarter.
[15:01] Um, so from a basic requirements
[15:03] perspective, what does it take to
[15:05] actually pull this off? Like this is all
[15:07] conceptually great, but let's get
[15:08] pragmatic. Like what is what does it
[15:10] actually take to make this happen? So
[15:12] there are three essential elements that
[15:14] you have to have in place. First of all,
[15:15] you need a first you need a secure and
[15:17] scalable infrastructure. So that has to
[15:20] come with enterprise security. It has to
[15:22] come with these reusable components that
[15:24] your your engineers code and and make
[15:26] available to the rest of the
[15:27] organization. You have to have
[15:28] production grade monitoring. Um and uh
[15:32] this idea of a birectional sync. So
[15:34] think about envision an environment
[15:36] where engineers can code and code and
[15:38] non-engineers can interact with that
[15:40] code in a visual UI and it's fully
[15:42] birectionally synced. uh non-engineers
[15:45] can intuitively understand what the AI
[15:47] is doing, how it's executing and they
[15:48] have a user-friendly environment to help
[15:50] be part of the optimization process. The
[15:53] second thing you need is domain experts
[15:55] obviously uh engage with these uh visual
[15:58] interfaces and getting real time
[15:59] feedback on their on their iterations
[16:01] and enhancements. Uh and then third
[16:04] organizational support that embeds AI
[16:06] throughout the operations. So this isn't
[16:08] just a technology project that's really
[16:10] important. Um it's a business
[16:12] transformation. So AI is becoming
[16:14] embedded in how you operate as an
[16:16] organization through a platform that
[16:17] makes that possible.
[16:20] So uh from an implementation framework
[16:23] perspective there are three phases to
[16:25] actually make this a reality. This it's
[16:27] obviously you have to crawl walk run as
[16:29] they say. Um phase one is how you build
[16:32] compound value systematically through
[16:34] engineering excellence. uh technical
[16:37] teams are building the platform um with
[16:40] enterprise integrations that are
[16:41] required for the data your data pipeline
[16:43] and to flow through your AI use cases.
[16:47] Um they're creating again these custom
[16:48] reusable components. They're initiating
[16:51] the initial workflows um that will get
[16:54] will your MVPs your early field trials
[16:57] throughout the early field trials
[16:59] throughout the various departments in
[17:00] your organization. And this foundational
[17:03] phase is what establishes the platform
[17:05] economics that we've been talking about.
[17:08] Um, am I good on time? Yeah, I
[17:10] >> actually do have to interrupt to cut you
[17:12] off. Unfortunately, that is time.
[17:14] >> It is time. Okay.
[17:15] >> That is I give you an extra couple
[17:17] minutes for the technical difficulties.
[17:18] >> Okay. Can I just have two minutes to
[17:20] wrap? All right.
[17:21] >> I can give you 30 seconds.
[17:22] >> 30 seconds. All right. Um, we are in uh
[17:25] if you want to hear more about this cuz
[17:27] I, you know, didn't get to finish the
[17:28] story. We're in booth 348. We're right
[17:30] in the middle of the expo floor. There
[17:32] are a few of us there staffing. I'll
[17:33] still be here after the session is done.
[17:35] Obviously, I have my counterpart Nico
[17:37] here somewhere, too. If you want to
[17:38] chat, he's in the corner. Um and uh you
[17:41] know what what I'll leave you with I
[17:43] guess is um the approach that you take
[17:47] to deploying AI built on a a foundation
[17:50] that's accessable accessible to your
[17:52] entire organization is really the basis
[17:54] that will build the velocity that will
[17:56] change the economic um calculation that
[18:00] drives your AI returns. So uh thank you
[18:02] for being here. Appreciate it and uh
[18:04] have a great rest of your event.
