Evalueserve's Bhat Urges Workflow Overhauls as 95% of AI Pilots Fail

The ROI Illusion: Why Most Enterprise AI Isn't Paying Off With Evalueserve — Ai4

Ai4

Enterprises miss AI returns because they attach models to isolated edge cases instead of embedding cross-functional squads to rebuild core business workflows around measurable financial metrics.

AI-assisted summary of The ROI Illusion: Why Most Enterprise AI Isn't Paying Off With Evalueserve — original source. Check the original for full context.

Source published: 2026-08-20

Detailed notes

## Background Published by Ai4 on YouTube on August 20, 2026, this presentation features Guru Bhat, EVP and Head of Data & AI Business at Evalueserve. The session breaks down why most enterprise generative AI investments stall before hitting the P&L and outlines practical frameworks to close the gap between AI adoption and measurable business value.

## TL;DR Guru Bhat explains that 95% of enterprise generative AI pilots fail because organizations apply AI to isolated edge use cases rather than re-engineering fundamental business workflows alongside domain experts. To realize real ROI, enterprises must evaluate multi-dimensional value streams—including efficiency, velocity, revenue uplift, and risk compliance—using a three-tiered maturity framework that moves beyond superficial vanity metrics. Bhat advocates for multi-model orchestration grounded in enterprise knowledge graphs, human-in-the-loop validation, and deploying multidisciplinary 'Agentic Squads' composed of forward deployed engineers, domain experts, data specialists, and adoption consultants to overcome end-user inertia and drive sustained P&L outcomes.

## Atmosphere & style The presentation carries an authoritative, pragmatic, and polished keynote tone that cuts through generative AI hype to focus on organizational realities. Bhat speaks with measured confidence, combining light personal humor with deep enterprise credibility drawn from decades of experience at Google Cloud and Wipro. The narrative feels like an executive masterclass, steadily guiding enterprise leaders from broad macroeconomic expenditure projections down to the granular, psychological challenges of convincing an accountant or attorney to trust AI over familiar spreadsheets.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) Roughly 95% of generative AI pilots fail because organizations start with isolated edge use cases and paste AI onto legacy processes instead of re-engineering core workflows. - [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) Enterprise AI spending is projected to reach $2.5 trillion by 2026, making a structured approach to value realization critical. - [04:31](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=271s) Unlike traditional IT ROI focused primarily on cost cutting, AI ROI operates across four simultaneous value streams: efficiency, velocity, revenue uplift, and risk/compliance. - [04:31](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=271s) AI ROI measurement must be treated as an organizational discipline embedded from day one with dedicated teams establishing baseline metrics. - [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) Superficial vanity metrics like login counts and raw token consumption are dangerous and misleading proxies for AI success. - [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) Enterprises should evaluate AI adoption across a three-tiered model: Tier 1 (Access & Usage), Tier 2 (Proficiency & Velocity), and Tier 3 (P&L Financial Outcomes). - [09:33](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=573s) Optimizing tokenomics requires multi-model orchestration to route tasks to cost-effective open-weight or specialized models rather than defaulting to expensive frontier models. - [09:33](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=573s) Data lakes and warehouses must be structured into knowledge graphs to provide the relational context required for reliable agentic AI grounding. - [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) Mission-critical enterprise workflows must integrate human-in-the-loop (HITL) checkpoints to ensure outputs conform to organizational guardrails. - [13:25](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=805s) Technical deployment requires cross-functional 'Agentic Squads' uniting domain experts, forward deployed engineers, data specialists, and adoption consultants. - [13:25](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=805s) Generic top-down change management fails; adoption consulting must speak the daily operational language of specific micro-workflows to prevent users from reverting to legacy tools like spreadsheets. - [15:30](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=930s) Value realization timelines vary by workflow complexity, with measurable returns maturing anywhere from 4 weeks to 3 to 6 months.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) Workflow-First AI vs. Use-Case Traps

Gururaj "Guru" Bhat introduces himself, lightly joking about his name being abbreviated because "it's difficult to pronounce," before outlining his background in enterprise data and AI. Prior to joining Evalueserve, he spent five and a half years at Google Cloud leading go-to-market data and AI practice for North America, and earlier headed the North American data analytics and AI business at Wipro. He highlights that Evalueserve brings over 20 years of domain experience driving mission-critical process workflows across major industry verticals. This background is pivotal: knowing an existing workflow inside out allows an organization to truly re-engineer and incubate AI into that process, rather than simply attempting to paste AI superficially on top of existing steps.

```mermaid flowchart TD accTitle: Pitfall vs Success Path in Enterprise AI subgraph FailingPath["The Standard Failing Path (95% of Pilots)"] A["Isolate an 'Edge' Use Case"] --> B["Apply AI Directly on Legacy Process"] B --> C["Pilot Fails to Move Business Metrics"] end

subgraph SuccessfulPath["The Workflow-First Approach (Top 6%)"] D["Select Critical Business Workflow"] --> E["Re-engineer Workflow with Domain Experts"] E --> F["Pick Measurable, High-Impact Use Cases"] F --> G["Deliver Measurable ROI to Production"] end ```

Addressing the broader market context, Bhat observes that enterprise investment in AI is surging toward an analyst-projected $2.5 trillion by 2026. Despite massive spending, roughly 95% of generative AI pilots fail to deliver business value. The primary culprit is where teams choose to start. Conventional advice urges teams to "pick a right use case" or focus on an edge scenario, but Bhat argues this is fundamentally flawed. Organizations must start with the **workflow**, not isolated use cases. A single workflow typically contains multiple use cases; starting with the overall workflow enables leaders to target the specific sub-cases that actually move the needle for executive leadership and stakeholders.

- **Analyst Projection**: AI spending is estimated to reach $2.5 trillion by 2026. - **Pilot Failure Rate**: 95% of generative AI pilots fail, largely due to selecting the wrong starting workflows or isolated edge use cases. - **Positive Trend**: An increasing number of enterprises are beginning to successfully graduate pilots into full production.

---

[04:31](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=271s) Redefining AI ROI as an Organizational Discipline

When analyzing the top 6% of enterprises that succeed, their primary differentiator is that they begin with deep workflow knowledge and bridge technical leadership with domain understanding. Bhat emphasizes that measuring Return on Investment (ROI) is not new—enterprises have evaluated ROI across three decades of digital, cloud, and mobile transformations. However, AI ROI operates along completely new dimensions that legacy ROI frameworks cannot capture.

Traditional ROI centers almost exclusively on efficiency and cost reduction. In contrast, AI alters four distinct value streams simultaneously: 1. **Efficiency**: Streamlining operational baseline efforts. 2. **Velocity**: Radically accelerating the speed and throughput of work execution. 3. **Revenue Uplift**: Creating new commercial opportunities and top-line expansion. 4. **Risk and Compliance**: Enhancing governance, auditability, and guardrails across business operations.

Because these multidimensional value streams did not exist in prior transformations, Bhat insists that AI ROI cannot be treated as an afterthought. It must be directly embedded into enterprise organizational design. When setting up an organizational structure for AI or agentic transformation, companies must establish dedicated teams responsible for establishing baseline metrics, defining measurement protocols, and tracking outcomes from day one. This structural discipline ensures continued leadership backing and justifiable capital allocation.

---

[06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) Beyond Vanity Metrics: A Three-Tiered Adoption Framework

Many enterprises currently rely on superficial vanity metrics to justify their AI programs—tracking how many employees log into tools, how many pilots are underway, or how many tokens are consumed. Bhat cautions that with surging token costs, raw token consumption is a dangerous and misleading proxy for success. Instead, enterprises must track adoption through a structured, three-tiered maturity model:

Measuring across these three layers replaces vanity volume metrics with outcome-based validation, giving CFOs the clear financial evidence required to release sustained funding.

---

[09:33](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=573s) Tokenomics, Multi-Model Orchestration, and Knowledge Graphs

Addressing the economics of AI ("tokenomics"), Bhat points out the rapid proliferation of open-weight models, reasoning models, and global alternatives from Google, Chinese labs, and open-source research hubs. Organizations do not need to route every single prompt or routine task to expensive frontier models. Instead, enterprises require intelligent model orchestration to route specific tasks to the most cost-effective and task-appropriate model.

```mermaid flowchart TD accTitle: Enterprise AI Orchestration and Context Architecture subgraph OrchestrationLayer["Model Control Plane & Orchestration"] Router["Task & Prompt Router"] M1["Frontier Models"] M2["Open-Weight / Smaller Models"] Router --> M1 Router --> M2 end

subgraph DataArchitecture["Enterprise Context Layer"] KG["Knowledge Graphs (Agentic Grounding)"] DW["Data Warehouses & Data Lakes"] DW --> KG end

KG --> Router ```

Building this multi-model capability requires a robust model control plane backed by enterprise context. While companies frequently claim they already have strong data foundations after decades of building data lakes and data warehouses, agentic AI demands more. For autonomous agents to act reliably, enterprise data must be structured into **knowledge graphs**. Knowledge graphs organize relational context in a format that agentic systems can interpret natively, grounding model actions firmly within enterprise boundaries and preventing expensive, ungrounded failures.

---

[11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) Measurable Use-Case Selection and Human-in-the-Loop Scaling

Bhat reiterates the golden rule for getting started: begin with the workflow, identify the constituent use cases, and prioritize based on measurable business impact. A use case might seem intellectually appealing, but if its business outcome cannot be baseline-measured and validated, it should never be the starting point.

For mission-critical workflows, automated systems must incorporate robust human-in-the-loop (HITL) checkpoints. While not every low-stakes task requires human oversight, core enterprise processes demand human validation to guarantee that outputs conform to strict organizational guardrails and performance metrics.

To bridge the gap between initial pilot validation and full-scale enterprise rollout, Bhat introduces the necessity of **forward deployed engineering**—pairing technical engineers side-by-side with domain practitioners directly inside operational business units to rebuild and scale the workflows from the ground up.

[13:25](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=805s) The Agentic Squad: Bridging Technology, Domain, and Adoption

While Forward Deployed Engineers (FDEs) are exceptional technologists capable of building and applying sophisticated models, they often hit a wall because they lack deep understanding of specific industry process workflows. Every industry operates with distinct operational nuances, compliance standards, risk constraints, and validation protocols. To bridge this gap and deploy AI effectively, enterprise transformation cannot rely on technologists alone; it requires a cross-functional assembly of specialized roles working in unison, which Evalueserve structures as an "Agentic Squad."

```mermaid flowchart TD accTitle: The Agentic Squad Structure subgraph SQUAD["The Agentic Squad"] DE["Domain Expert (Industry & Workflow Nuances)"] FDE["Forward Deployed Engineer (AI & Tech Implementation)"] DP["Data Expert (Context & Data Architecture)"] AC["Adoption Consultant (Workflow-Level Change Mgmt)"] end

subgraph WORKFLOW["Target Business Workflow"] EXEC["Legal Summarization / Monthly Financial Close / Operations"] end

SQUAD --> WORKFLOW ```

This multi-disciplinary squad unites a domain expert who intimately understands the operational process, a forward-deployed engineer who handles the technical implementation, and a data expert responsible for providing the necessary data context. Crucially, the fourth pillar is the adoption consultant. A recurring failure pattern across enterprises is rolling out centralized AI tooling across standard workflows only to find business users entirely ignoring it. Generic, organization-wide change management fails because it does not communicate in the everyday operational language of the end users.

When change management fails to speak to the specific realities of an attorney summarizing legal contracts or an accountant executing the monthly financial close, users inevitably retreat to legacy habits, thinking: *"I'm pretty happy with the Excel I work with every single month. I would not move to this AI new tool that you're bringing to me."* Real adoption requires driving change management directly inside the micro-workflow rather than through top-down mandates.

- **Domain Expert:** Embeds industry-specific compliance, regulatory checks, and workflow intricacies into the solution. - **Forward Deployed Engineer (FDE):** Builds and integrates the applied AI tech stack directly against the business problem. - **Data Specialist:** Curates, cleans, and surfaces the contextual internal data required for reliable model performance. - **Adoption Consultant:** Drives targeted, workflow-level change management so end users abandon legacy spreadsheets for AI-augmented tooling.

---

[15:30](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=930s) Engineering to Outcome and Value Realization

To convert AI roadmaps into actual P&L impact, organizations must shift from tracking technical activity to engineering directly for business outcomes. Value realization frameworks must be tailored around the realization velocity of different workflows. AI impact is not uniform: certain automated workflows may yield measurable returns in as little as 4 weeks, while more complex or deeply integrated enterprise processes may require 3 to 6 months to mature and demonstrate clear financial returns.

Evalueserve provides specialized frameworks and accelerator tooling designed to establish continuous, outcome-based measurement models across all deployed workflows. Closing out the session, Guru invites enterprise leaders to collaborate on designing structured AI value frameworks, testing measurable workflow benchmarks, and systematically scaling AI applications across the broader enterprise portfolio before opening the floor for audience questions.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) knowing an existing workflow inside out allows an organization to truly re-engineer and incubate AI into that process, rather than simply attempting to paste AI superficially on top of existing steps. — Gururaj "Guru" Bhat

> [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) Conventional advice urges teams to "pick a right use case" or focus on an edge scenario, but Bhat argues this is fundamentally flawed. Organizations must start with the workflow, not isolated use cases. — Gururaj "Guru" Bhat

> [04:31](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=271s) Traditional ROI centers almost exclusively on efficiency and cost reduction. In contrast, AI alters four distinct value streams simultaneously: Efficiency, Velocity, Revenue Uplift, and Risk and Compliance. — Gururaj "Guru" Bhat

> [04:31](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=271s) AI ROI cannot be treated as an afterthought. It must be directly embedded into enterprise organizational design. — Gururaj "Guru" Bhat

> [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) With surging token costs, raw token consumption is a dangerous and misleading proxy for success. — Gururaj "Guru" Bhat

> [09:33](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=573s) Organizations do not need to route every single prompt or routine task to expensive frontier models. Instead, enterprises require intelligent model orchestration to route specific tasks to the most cost-effective and task-appropriate model. — Gururaj "Guru" Bhat

> [09:33](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=573s) While companies frequently claim they already have strong data foundations after decades of building data lakes and data warehouses, agentic AI demands more. For autonomous agents to act reliably, enterprise data must be structured into knowledge graphs. — Gururaj "Guru" Bhat

> [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) A use case might seem intellectually appealing, but if its business outcome cannot be baseline-measured and validated, it should never be the starting point. — Gururaj "Guru" Bhat

> [13:25](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=805s) I'm pretty happy with the Excel I work with every single month. I would not move to this AI new tool that you're bringing to me. — Gururaj "Guru" Bhat

> [13:25](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=805s) Real adoption requires driving change management directly inside the micro-workflow rather than through top-down mandates. — Gururaj "Guru" Bhat

> [15:30](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=930s) AI impact is not uniform: certain automated workflows may yield measurable returns in as little as 4 weeks, while more complex or deeply integrated enterprise processes may require 3 to 6 months to mature and demonstrate clear financial returns. — Gururaj "Guru" Bhat

## Mentioned Gururaj "Guru" Bhat, Evalueserve, Google Cloud, Wipro, Ai4, Agentic Squad, Excel

## Video description Guru Bhat - EVP & Head of Data & AI Business, Evalueserve

Most companies aren't failing at AI because the technology is weak. They're failing because they're automating workflows that were never designed for AI in the first place, and calling it transformation.

Gururaj Bhat spent over 25 years inside Google Cloud and Wipro helping the world's largest enterprises put data and AI to work at scale, before an early Google Cloud lunch conversation still shapes how he leads today: real ROI doesn't come from adding AI on top of a broken process. It comes from putting engineers and domain experts in the same room and rebuilding the process itself.

Now leading Data & AI at Evalueserve, Guru will break down why so many AI investments stall before they hit the P&L, what separates companies that see real returns from those stuck in pilot purgatory, and the practical framework he uses to close the gap between AI adoption and AI value.

If your AI roadmap looks impressive in a slide but hasn't moved a single business metric, this session is for you.

[Source](https://www.youtube.com/watch?v=IJyJgsxZCo0)

中文简报

95%的AI试点宣告失败，Evalueserve高管Bhat呼吁重构工作流程企业错失AI回报，是因为它们将模型套用于孤立的边缘案例，而非嵌入跨职能团队围绕可衡量的财务指标重构核心业务流程。 Guru Bhat解释称，95%的企业生成式AI试点项目之所以失败，是因为企业只是将AI应用于孤立的边缘用例，而非与领域专家携手重构根本的业务流程。为了实现切实的投资回报率（ROI），企业必须跳出流于表面的虚荣指标，运用三级成熟度框架来评估涵盖效率、速度、收入增长及风险合规在内的多维价值流。Bhat主张构建基于企业知识图谱的多模型编排机制、推行人机协同验证，并部署由前沿部署工程师、领域专家、数据专员与落地推广顾问组成的多学科“Agentic Squads”，以打破最终用户的惯性阻力，推动可持续的损益成果。约95%的生成式AI试点项目以失败告终，主要症结在于企业往往孤立选择边缘用例并将AI强行拼接至传统流程，而非重构核心工作流。 (0:04) 与侧重削减成本的传统IT投资回报不同，AI ROI同时涵盖效率提升、工作流提速、营收增长以及风险与合规管控四大价值流，企业必须从第一天起将其作为组织级纪律建立基线并持续度量。 (4:31) 企业应摒弃登录量与Token消耗等虚荣指标，通过访问与使用（第一层）、熟练度与流程提速（第二层）以及损益表财务结果（第三层）组成的三层模型来系统评估AI落地价值。 (6:34) 优化Token成本需要多模型编排机制将任务路由至高性价比的模型，同时需将传统数据湖和数据仓库重构为知识图谱，为Agent系统提供可靠的关系型上下文基底。 (9:33) 为弥合AI试点与规模化落地的鸿沟，企业应组建涵盖领域专家、前线部署工程师（FDE）、数据专家与落地咨询顾问的跨职能Agent团队，针对具体微观工作流推动变革与价值实现。 (13:25)

Reading summary

Gururaj Bhat of Evalueserve argues that 95% of enterprise generative AI pilots fail because organizations target isolated use cases rather than re-engineering core workflows. Realizing business value requires evaluating AI across four operational dimensions, supported by intelligent model orchestration and cross-functional deployment squads. Re-engineering Workflows Over Isolated Use Cases (0:04) Enterprise AI spending is projected by analysts to reach $2.5 trillion by 2026, yet Bhat observes that roughly 95% of generative AI pilots fail to deliver business value. This breakdown stems from the common practice of targeting isolated edge use cases or superficially applying AI over legacy processes without modifying how the underlying work is performed. In contrast, the 6% of enterprises that succeed adopt a workflow-first methodology. Because a single business workflow contains multiple component use cases, understanding the entire process allows technical and domain leaders to identify and re-engineer specific high-impact steps that tangibly influence executive-level metrics. Multi-Stream ROI and Three-Tiered Adoption Metrics (4:31) Traditional enterprise transformations measured return on investment primarily through cost reduction and operational efficiency. Bhat asserts that AI alters four value streams simultaneously: baseline efficiency, task execution velocity, top-line revenue uplift, and risk or compliance governance. Capturing these multidimensional returns requires organizations to formalize dedicated teams and baseline measurement protocols from day one. Bhat cautions against tracking vanity metrics such as raw login numbers, total active pilots, or token consumption, noting that high token volume can simply represent inefficient spending. Instead, organizations should evaluate adoption through a three-tiered model: basic access and usage, velocity and workflow shifts in how tasks are executed, and measurable bottom-line or top-line P&L impact. Managing Tokenomics With Orchestration and Knowledge Graphs (9:33) To control the operational economics of AI, enterprises must move away from routing every task to expensive frontier models. Given the rise of open-weight, reasoning, and specialized models, organizations need a model control plane that intelligently routes specific prompts to the most cost-effective, task-appropriate architecture. Additionally, standard data warehouses and data lakes do not provide adequate context for autonomous systems. Bhat explains that enterprise data must be structured into knowledge graphs, which organize relational context natively so agentic models remain grounded within company boundaries. For mission-critical workflows, these automated architectures must be paired with human-in-the-loop checkpoints to enforce compliance. Cross-Functional Agentic Squads and Workflow-Level Change (13:25) Technical specialists such as Forward Deployed Engineers frequently struggle when working alone because they lack deep familiarity with vertical-specific regulatory constraints and operational nuances. To bridge this divide, Bhat details Evalueserve's four-pillar "Agentic Squad," which integrates a forward deployed engineer, a domain expert, a data specialist, and an adoption consultant directly into target workflows. The adoption consultant specifically addresses the tendency of business users to ignore centralized AI rollouts in favor of legacy tools like spreadsheets. By conducting change management directly inside specific micro-workflows—such as legal contract summarization or monthly financial accounting closes—the squad speaks to the end user's daily operational realities rather than relying on top-down corporate mandates. Bhat emphasizes that value realization timelines vary significantly, with simple workflows showing returns within four weeks while complex enterprise processes require three to six months. However, the framework relies heavily on establishing rigorous baseline benchmarks early, leaving open the challenge of how organizations can effectively quantify impact when legacy workflows lack structured historical performance data.

精炼阅读摘要

Evalueserve的Gururaj Bhat指出，尽管分析机构预测到2026年企业人工智能支出将达到2.5万亿美元，但约95%的生成式AI试点未能产生实际业务价值。他认为，破除“AI投资回报率（ROI）幻觉”的关键在于摒弃孤立的用例测试，转为从核心业务工作流切入，并通过组织机制、知识图谱与跨职能团队实现可量化的财务回报。重塑工作流：破除孤立用例陷阱 (0:04) 结合分析师关于2026年全球企业AI支出将攀升至2.5万亿美元的预测，Bhat强调当前约95%的生成式AI试点陷入停滞，其根本原因在于企业普遍采取了错误的切入路径——往往孤立地选择边缘用例，并生搬硬套在旧有业务流程之上。这种修补式的做法无法实质性拉动业务指标。与此相对，排名前6%能够成功将试点转入全面生产环境的企业，其核心差异在于从一开始就依托深厚的领域认知，选定关键业务工作流，并由业务专家协同对工作流进行端到端的重新工程化设计。 Bhat指出，单一业务工作流内部通常嵌套着多个具体用例。若脱离整体工作流而盲目追求个别细分点，往往无法对管理层关注的战略目标产生实质影响。企业必须立足于全局工作流，识别出真正能够产生业务杠杆的关键用例，再进行针对性孵化与落地，从而打通从技术概念验证向规模化生产交付的演进路径。重塑多维ROI并建立三层成熟度指标 (4:31) Bhat主张将AI ROI视为一项贯穿企业组织设计的制度纪律，而非项目交付后的补充评估。过去三十年的数字化、云计算与移动转型中，ROI多聚焦于运营效率与成本削减。然而，AI会同时改变四个维度的价值流：提升基础运营效率（Efficiency）、成倍提高任务流转速度与吞吐量（Velocity）、开辟新的商业机会带来收入增量（Revenue Uplift），以及强化治理与合规风控（Risk and Compliance）。因此，企业必须在推行AI或智能体转型首日便设立专属团队，负责基准指标设定与全程结果追踪。同时，企业亟须摆脱以登录人数、试点项目数或Token消耗量为代表的“虚荣指标”（Vanity Metrics）。在Token采购成本攀升的背景下，原始消耗量完全无法代表业务成效。Bhat提出了三层成熟度评估框架：第一层考核访问与采纳率，验证目标用户的实际使用程度；第二层考核熟练度与流程重构，评估AI是否实质提升了流转速度并改变了任务执行模式；第三层直击财务成果，明确技术落地在损益表（P&L）上带来的底线成本节约或顶线收入增长，为财务管理层持续注资提供客观证据。模型智能编排与知识图谱上下文架构 (9:33) 在探讨AI运行经济学（Tokenomics）时，Bhat指出，随着开源权重模型、具备逻辑推理能力的模型以及来自Google、中国实验室等多方替代方案的涌现，企业完全没有必要将所有常规任务均交由成本昂贵的前沿模型（Frontier Models）处理。企业需要构建智能模型编排系统与控制平面，依照具体任务属性与成本收益比，灵活将请求分发至最具性价比的模型端口。实现高效多模型调度的先决条件是深厚的企业上下文架构。Bhat提醒道，尽管许多企业拥有沉淀数十年建立的数据仓库与数据湖，但这并不等同于具备了支撑自主智能体（Agentic AI）的数据底座。为了确保智能体行动的准确与可靠，企业内部的数据必须重构为知识图谱（Knowledge Graphs）。知识图谱能够以智能体可原生解析的方式组织实体与关系网络，将模型生成边界严格锚定在企业内部真实语境中，杜绝因脱离上下文而导致的昂贵失效。组建智能体战队与微观流程变革管理 (13:25) 对于核心关键流程，系统必须内嵌人在回路（HITL）验证机制，由人工进行关键把关以确保输出符合合规与质量红线。为了跨越从试点到规模化推广的鸿沟，Bhat主张采用前沿部署工程师（FDE）模式。然而，由于FDE等纯技术专家往往缺乏对特定行业在法规、风控和细致操作层面的深入理解，企业必须组建名为“智能体战队”（Agentic Squad）的跨职能团队，协同领域专家（Domain Expert）、FDE、数据专家（Data Specialist）与采纳咨询顾问（Adoption Consultant）四类关键角色共同推进。 Bhat特别指出，AI推广常因忽视微观层面的变革管理而折戟。自上而下的通用培训无法与一线员工建立共鸣；面对合同审查律师或月度财务结账会计师等专业群体，若AI工具未能在其日常工作流的语境下解决具体痛点，员工通常会退守回熟悉的Excel表格工具。变革管理必须深入具体的微观工作流内部推进。此外，AI价值释放的节奏因流程复杂度而异，标准化自动化流程可能在4周内展现回报，而高深度集成的复杂流程通常需要3至6个月的孕育周期才能形成可验证的财务收益。 Bhat总结强调，企业需从追踪技术开发动作转向以业务结果为导向的工程实践，并针对不同工作流建立动态价值评估体系。需要指出的是，该观点主要建立在Evalueserve的服务实践与方法论框架之上，如何在内部数据割裂严重、部门利益交错的传统大型企业中顺利推行跨职能智能体战队与知识图谱改造，仍面临长期的组织协同与实施复杂性挑战。

Detailed notes — zh

## Background 该内容由频道 Ai4 发布于 YouTube 平台，主讲人为 Evalueserve 数据与 AI 业务执行副总裁 Guru Bhat（曾任职于 Google Cloud 及 Wipro）。视频简介指出，大多数企业 AI 投资之所以受挫，是因为盲目在未针对 AI 设计的传统流程上做自动化，本场演讲旨在深入剖析 AI 项目深陷“试点炼狱”的病灶，并提供将技术采纳切实转化为损益表商业价值的实战框架。 ## TL;DR 演讲指出高达 95% 的生成式 AI 试点夭折，根本原因在于企业误从孤立的边缘用例切入，而非重构端到端的业务工作流。Guru 强调 AI 的 ROI 度量必须跳出单一效率思维，扩展到工作流转速度、收入提振与合规控制等维度，并建立采纳度、熟练度与财务成果三层评估模型。为了应对 Token 成本与落地阻碍，企业应结合知识图谱构建统一模型控制平面，并通过由领域专家、前瞻工程师、数据工程师与采纳顾问组成的“智能体特战队”，深入具体岗位语境推动人机协同与组织变革。 ## Atmosphere & style 演讲整体洋溢着务实、辛辣且直击痛点的企业实战氛围。Guru 以自嘲轻松开场，随后以冷峻的数据对比打破行业对孤立用例与虚荣指标的盲目狂热。其表达深入浅出、逻辑严密，既有高屋建瓴的架构与商业洞察，又生动复现了一线业务人员抗拒死板变革的真实心理，带给听众一种拨开炒作迷雾、直面落地现实的沉浸式启发感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) 95% 的生成式 AI 试点失败的根本原因在于从孤立的边缘用例切入，而非从重构端到端的关键业务工作流着手。 - [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) 成功将 AI 落地生产环境的 6% 少数派团队，其核心共同点在于将技术架构与深入的垂直业务流程认知紧密结合。 - [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) AI 重塑了投资回报率（ROI）的维度，除了传统效率之外，还引入了工作流转速度、收入提振与风险合规控制等全新价值流。 - [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) AI 的价值衡量与基线指标必须在项目启动的第一天就融入组织设计（Org Design），而非作为事后补救指标。 - [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 企业需摒弃用户数与 Token 消耗量等虚荣指标，转而构建包含“使用采纳度、工作熟练度、财务成果”的三层评估模型。 - [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 面对 Token 经济学（Tokenomics），企业应建立统一模型控制平面与动态路由机制，避免在所有场景无差别调用昂贵的前沿模型。 - [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 传统的数仓与数据湖底座已不足以支撑 Agentic AI，必须在其之上构建知识图谱，以提供准确的实体关系与业务上下文。 - [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 首发用例必须挑选商业产出最高且结果可精确衡量的场景，并在核心工作流中植入人机协同（HITL）防护机制。 - [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 规模化落地需依靠四位一体的“智能体特战队”（领域专家、前瞻部署工程师、数据工程师、采纳顾问）协同作战。 - [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 变革管理必须深入具体岗位的业务语言语境，避免假大空的宏观技术宣讲，才能真正驱动基层业务人员放弃传统工具。 - [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 业务工作流的价值兑现周期各不相同（4 周至 6 个月不等），企业必须结合工作流生命周期规划度量与收益验证节奏。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) 企业 AI 投资狂潮与“试点炼狱” Guru Bhat 登台后以自己的名字自嘲开场——全名 Gururaj Bhat 发音太拗口，所以大家通常只叫他 Guru。在加入 Evalueserve 担任数据与 AI 业务执行副总裁（EVP）之前，他在 Google Cloud 负责北美区数据与 AI 的市场拓展（Go-to-Market）长达五年半；再往前，他曾领导全球 IT 系统集成商 Wipro 的北美数据分析与 AI 业务。他指出，Evalueserve 过去二十年来深耕各行业的关键业务流程工作流（Business Process Workflows）。这种积累带来的最大差异在于：很多机构仅仅思考“如何将 AI 叠加到现有流程上”，而 Evalueserve 深刻理解业务流程的底细，知道如何对流程进行重构（Re-engineer），进而将 AI 原生孵化到这些工作流中。放眼当前的企业界，关于 AI 的宏观数据铺天盖地。分析机构预测到 2026 年企业在 AI 上的投资将达到 2.5 万亿美元，而且这一数字只增不减。然而与激进投资形成刺眼对比的是，**95% 的生成式 AI（GenAI）试点项目最终都以失败告终**。Guru 直言，失败的根本诱因在于企业“起步的方式完全错了”。业界主流论调一直在劝说企业去挑选所谓的“绝佳用例”（Right use case）或是“前沿用例”（Edge use case），但这恰恰是一个错误的起点。企业真正应该锁定的不是孤立的用例，而是**正确的业务工作流**。 - 分析机构预估到 2026 年全球企业 AI 投资将达 2.5 万亿美元以上。 - 高达 95% 的生成式 AI 试点无法走出试验阶段。 - 绝大多数企业从孤立用例（Use Case）切入，而非从端到端业务工作流（Workflow）切入，导致技术无法兑现业务价值。 ```mermaid flowchart TD accTitle: 传统失败模式与业务工作流驱动模式对比 subgraph WrongWay["95% 企业的失败路径（用例驱动）"] W1["盲目挑选边缘/孤立用例"] --> W2["强行叠加 AI 模型"] W2 --> W3["无法融入实际业务流程"] W3 --> W4["试点失败 / 陷入试点炼狱"] end subgraph RightWay["成功团队的路径（工作流驱动）"] R1["锁定高价值业务工作流"] --> R2["解构并重构业务流程"] R2 --> R3["筛选高可衡量性的子用例"] R3 --> R4["交付可验证的商业 ROI"] end ``` --- ### [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) 成功的少数派：重塑 AI 投资回报率（ROI）的衡量维度尽管整体失败率触目惊心，行业情绪也正在出现积极转变：将 AI 试点成功推进至生产环境的企业数量正在逐日攀升。在成功实现落地的案例中，进一步深挖数据会发现，仅有大约 6% 的顶尖团队真正把事情做对了。这 6% 企业的共同特质就是**始终从工作流出发**。要从工作流切入，团队必须具备扎实的业务流程知识。作为技术人员和 IT 领导者，单懂技术架构远远不够，必须与深谙业务流程的专家结成伙伴，才能将 AI 与真正的业务成果紧密挂钩。谈及 ROI，Guru 指出投资回报率本身并不新鲜，过去三十年无论是数字化转型、云转型还是移动化转型，企业都在做 ROI 测算。但 AI 的独特性在于，它彻底改写了 ROI 的度量维度。传统的转型往往聚焦于“效率”（Efficiency）单一维度，但 AI 不仅重塑效率，还带来了另外三个以往从未测度过的全新价值流： 1. **工作流转速度（Velocity of work）**：彻底改变员工的工作方式与交付周期。 2. **收入提振（Revenue uplift）**：为业务端直接开辟新的增量空间。 3. **风险与合规控制（Risk and compliance）**：在自动化流程中建立合规护栏。正因这几条全新价值流的存在，AI ROI 绝不能成为事后补救的被动指标（Afterthought），而必须在最初就融入**组织设计（Org Design）**之中。在搭建 AI 或智能体（Agentic）转型架构的第一天，就应当设立专职团队来制定度量基线指标（Baseline metrics），明确输出价值的衡量方式，唯有如此才能确保管理层持续注资。 --- ### [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 告别虚荣指标：构建三层采纳度模型与应对 Tokenomics 现实许多企业在评估 AI 进展时仍深陷在错误的虚荣指标（Vanity Metrics）泥潭中。管理层往往只盯着“今天有多少用户用上了 AI”、“我们跑了多少个试点项目”以及“消耗了多少 Token”。但在当下 Token 成本剧烈波动的现实下，盲目追求 Token 消耗量无异于灾难。真正健康的评估体系必须分为三个清晰的层级： | 评估层级 | 核心度量关注点 | 衡量标准与意义 | |:--- |:--- |:--- | | **第一层：使用采纳度 (Adoption)** | 基础工具覆盖率 | 统计组织内部究竟有多少员工真正开始使用赋能的 AI 工具与技术。 | | **第二层：工作熟练度 (Proficiency)** | 流程与速度质变 | 评估员工使用 AI 后是否提升了**工作流转速度**，是否从根本上改变了传统的作业模式。 | | **第三层：财务成果 (Financial Outcome)** | 商业回报与底线效益 | 检验熟练度的提升最终兑现成了多少净收益、节省了多少硬性成本，让 CFO 愿意继续签字打款。 | 与指标设计紧密相连的另一个技术现实是 **Token 经济学（Tokenomics）**。随着开源社区与前沿研究室的井喷式发展——无论是来自中国的高质量模型，还是 Google 及其他前沿实验室开源的开放权重模型（Open-weight models）——企业必须认清一个事实：**绝无必要在内部的所有任务或每一次 Prompt 交互中都调用昂贵的前沿模型（Frontier Models）**。企业亟需构建一套动态模型路由与编排机制，把匹配的模型分配给对应的任务，以实现 Token 效益的最大化。 ```mermaid flowchart LR accTitle: 企业智能体核心控制与数据分层架构 subgraph DataLayer["统一数据与上下文层"] DL1["传统数据湖 / 数据仓库"] --> DL2["企业级知识图谱 (Knowledge Graphs)"] end subgraph OrchestrationLayer["模型控制平面 (Model Control Plane)"] DL2 --> O1["统一编排与路由引擎"] O1 --> M1["前沿模型 (Frontier Models)"] O1 --> M2["垂直/轻量开源权重模型"] end subgraph TaskExecution["分级任务执行"] M1 --> T1["复杂决策 / 逻辑推理"] M2 --> T2["标准化高频任务"] end ``` 企业不应将技术绑定在单一模型供应商上，而是需要建立跨业务单元的**统一模型控制平面（Model Control Plane）**。过去数十年间，IT 团队一直在建设数据仓库和数据湖，并自认为已经打好了数据底座；但对于智能体 AI（Agentic AI）而言，单纯的数仓数湖已经不够用了。为了让智能体能够准确理解上下文并采取接地气的行动，企业必须在原有数据底座之上构建**知识图谱（Knowledge Graphs）**。唯有通过知识图谱将业务实体关系化、情境化，智能体才能获取正确的 Context，从而配合最经济的模型打出理想的 Tokenomics 组合拳。 --- ### [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 落地实操：人机协同（HITL）与“智能体特战队”（Agentic Squad）当企业从工作流着手时，如何选出首发用例？Guru 给出的标准是：**选择能带来最大商业产出、且结果必须可被精确衡量的用例**。如果一个用例看似华丽却无法量化其商业产出，它就绝不应该作为首选启动点。与此同时，在最核心、高风险的关键工作流中，必须植入**人机协同（Human-in-the-Loop, HITL）**机制，以严密的 Guardrails 兜底，确保 AI 的输出指标始终在可控范围之内。当试点要进一步扩展到全局规模时，传统模式便会碰壁。当前硅谷和咨询行业非常流行前瞻部署工程师（Forward Deployed Engineer, FDE）模式。FDE 虽然擅长实现前沿技术，却往往缺乏各垂直行业的具体流程认知，不懂特定行业的合规审计、风险校验与潜规则细节。为此，Evalueserve 提出了名为**智能体特战队（Agentic Squad）**的四位一体跨职能团队架构： - **领域业务专家（Domain Expert）**：透彻理解特定行业端到端的工作流痛点、行业规范与操作细节。 - **前瞻部署工程师（FDE / Technology）**：负责技术工程化落地、系统搭建与模型集成。 - **数据工程师（Data Specialist）**：负责将多源数据结构化，梳理企业知识图谱并为智能体注入精准上下文。 - **采纳顾问（Adoption Consultant）**：攻坚最难落地的最后一公里，在具体工作流层面推行组织变革管理。许多管理层困惑于“工具和算力都给到位了，为什么业务人员就是不用？”Guru 指出，病因在于过去的企业变革管理总是停留在宏观组织层面，说的全是假大空的技术套话，根本没有深入到具体的岗位语境。面向法务合规人员，必须用合同条款摘要审查的具体语言去沟通；面对财务人员，必须站在月末结账平账（Financial closure of books）的操作痛点上去对话。如果不能走进具体工作流，“终端业务人员只会冷笑一声说：‘我用我的 Excel 宏用得好好的，才懒得碰你这套新 AI 工具。’” 在演讲尾声，Guru 强调不同业务工作流的价值兑现周期截然不同：有些轻量工作流在 4 周内就能看到明显的 ROI，有些则需要 3 个月甚至 6 个月。企业必须按工作流的生命周期来规划度量节奏，并欢迎参会者引入经实践验证的评估框架与工具链，避开 ROI 幻觉，让 AI 真正反映在损益表（P&L）的净利润上。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) 95% 的生成式 AI（GenAI）试点项目最终都以失败告终 — Guru Bhat > [00:04](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=4s) 企业真正应该锁定的不是孤立的用例，而是正确的业务工作流 — Guru Bhat > [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) 作为技术人员和 IT 领导者，单懂技术架构远远不够，必须与深谙业务流程的专家结成伙伴，才能将 AI 与真正的业务成果紧密挂钩 — Guru Bhat > [03:19](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=199s) AI ROI 绝不能成为事后补救的被动指标（Afterthought），而必须在最初就融入组织设计（Org Design）之中 — Guru Bhat > [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 绝无必要在内部的所有任务或每一次 Prompt 交互中都调用昂贵的前沿模型（Frontier Models） — Guru Bhat > [06:34](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=394s) 为了让智能体能够准确理解上下文并采取接地气的行动，企业必须在原有数据底座之上构建知识图谱（Knowledge Graphs） — Guru Bhat > [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 选择能带来最大商业产出、且结果必须可被精确衡量的用例。如果一个用例看似华丽却无法量化其商业产出，它就绝不应该作为首选启动点 — Guru Bhat > [11:44](https://www.youtube.com/watch?v=IJyJgsxZCo0&t=704s) 终端业务人员只会冷笑一声说：‘我用我的 Excel 宏用得好好的，才懒得碰你这套新 AI 工具。’ — Guru Bhat ## Mentioned Guru Bhat, Gururaj Bhat, Evalueserve, Google Cloud, Wipro, Ai4, GenAI, Knowledge Graphs, Frontier Models, Tokenomics, Forward Deployed Engineer (FDE), Human-in-the-Loop (HITL), Agentic Squad, Excel ## Video description Guru Bhat - EVP & Head of Data & AI Business, Evalueserve Most companies aren't failing at AI because the technology is weak. They're failing because they're automating workflows that were never designed for AI in the first place, and calling it transformation. Gururaj Bhat spent over 25 years inside Google Cloud and Wipro helping the world's largest enterprises put data and AI to work at scale, before an early Google Cloud lunch conversation still shapes how he leads today: real ROI doesn't come from adding AI on top of a broken process. It comes from putting engineers and domain experts in the same room and rebuilding the process itself. Now leading Data & AI at Evalueserve, Guru will break down why so many AI investments stall before they hit the P&L, what separates companies that see real returns from those stuck in pilot purgatory, and the practical framework he uses to close the gap between AI adoption and AI value. If your AI roadmap looks impressive in a slide but hasn't moved a single business metric, this session is for you. [Source](https://www.youtube.com/watch?v=IJyJgsxZCo0)

Source quotes

[0:04] Guru Bhat: 95% 的生成式 AI（GenAI）试点项目最终都以失败告终 [0:04] Guru Bhat: 企业真正应该锁定的不是孤立的用例，而是正确的业务工作流 [3:19] Guru Bhat: 作为技术人员和 IT 领导者，单懂技术架构远远不够，必须与深谙业务流程的专家结成伙伴，才能将 AI 与真正的业务成果紧密挂钩 [3:19] Guru Bhat: AI ROI 绝不能成为事后补救的被动指标（Afterthought），而必须在最初就融入组织设计（Org Design）之中 [6:34] Guru Bhat: 绝无必要在内部的所有任务或每一次 Prompt 交互中都调用昂贵的前沿模型（Frontier Models） [6:34] Guru Bhat: 为了让智能体能够准确理解上下文并采取接地气的行动，企业必须在原有数据底座之上构建知识图谱（Knowledge Graphs） [11:44] Guru Bhat: 选择能带来最大商业产出、且结果必须可被精确衡量的用例。如果一个用例看似华丽却无法量化其商业产出，它就绝不应该作为首选启动点 [11:44] Guru Bhat: 终端业务人员只会冷笑一声说：‘我用我的 Excel 宏用得好好的，才懒得碰你这套新 AI 工具。’

Source quotes

[0:04] Gururaj "Guru" Bhat: knowing an existing workflow inside out allows an organization to truly re-engineer and incubate AI into that process, rather than simply attempting to paste AI superficially on top of existing steps. [0:04] Gururaj "Guru" Bhat: Conventional advice urges teams to "pick a right use case" or focus on an edge scenario, but Bhat argues this is fundamentally flawed. Organizations must start with the workflow, not isolated use cases. [4:31] Gururaj "Guru" Bhat: Traditional ROI centers almost exclusively on efficiency and cost reduction. In contrast, AI alters four distinct value streams simultaneously: Efficiency, Velocity, Revenue Uplift, and Risk and Compliance. [4:31] Gururaj "Guru" Bhat: AI ROI cannot be treated as an afterthought. It must be directly embedded into enterprise organizational design. [6:34] Gururaj "Guru" Bhat: With surging token costs, raw token consumption is a dangerous and misleading proxy for success. [9:33] Gururaj "Guru" Bhat: Organizations do not need to route every single prompt or routine task to expensive frontier models. Instead, enterprises require intelligent model orchestration to route specific tasks to the most cost-effective and task-appropriate model. [9:33] Gururaj "Guru" Bhat: While companies frequently claim they already have strong data foundations after decades of building data lakes and data warehouses, agentic AI demands more. For autonomous agents to act reliably, enterprise data must be structured into knowledge graphs. [11:44] Gururaj "Guru" Bhat: A use case might seem intellectually appealing, but if its business outcome cannot be baseline-measured and validated, it should never be the starting point. [13:25] Gururaj "Guru" Bhat: I'm pretty happy with the Excel I work with every single month. I would not move to this AI new tool that you're bringing to me. [13:25] Gururaj "Guru" Bhat: Real adoption requires driving change management directly inside the micro-workflow rather than through top-down mandates. [15:30] Gururaj "Guru" Bhat: AI impact is not uniform: certain automated workflows may yield measurable returns in as little as 4 weeks, while more complex or deeply integrated enterprise processes may require 3 to 6 months to mature and demonstrate clear financial returns.

Original transcript

[0:04] All right, just as a brief introduction, uh my name is Guru Rajbad and you can see why they made [0:09] it short because it's difficult to pronounce. Um before joining evalu last month I was part of the [0:17] Google cloud practice go to market for the North America business on data and AI where I spent five [0:24] and a half years and before that I was part of uh a global system integrator called Vipro where [0:31] I used to head the data analytics and AI business for North America business right so evalu I don't [0:38] know if how many of you know evalu So, Evaluerve has been in the business for about 20 years [0:45] driving significant critical business process workflows across many industry segments. Now, [0:53] keep that in mind. What does that mean? That means evalu exactly knows what does the process [1:00] looks like and what it takes to re-engineer that process and enable and incubate AI into those [1:09] process workflows which is very different from hey how do I just apply AI to the process workflows [1:16] that is what makes it unique with you know evalu [clears throat] of course you guys are seeing [1:23] probably so many stats in every session that you are sitting today right but the Uber message what [1:29] we are seeing is there is a significant spend happening or investment happening I would not [1:34] say spend but more investment happening across various enterprises and of course the analysts [1:41] talk about 2.5 trillion in 2026 I'm pretty sure it's just going to go higher uh as well and 95% [1:52] of geni pilot it's fair and I'm pretty sure many of you know reasons but the primary reason is it [2:01] all starts about where do you get started as part of the process right you have heard that hey we [2:09] should pick a right use case we should pick a use case that is edge and I think that that is the [2:16] wrong starting point it is not the use cases that you should pick you should pick the right workflow [2:26] which is the right workflow that you want to enable AI with as a starting point within your [2:33] enterprise and that workflow may have multiple use cases as part of uh you know incubation of AI and [2:41] at that point in time it becomes very critical you pick the right use case within that workflow that [2:47] can deliver value to your business stakeholders or to your leadership or to your enterprise. [2:54] And that's where the most of the pilots fail because you're picking the wrong workflow or [3:00] you're picking the wrong use case to even get started with and hence the application of AI [3:07] to that use case does not deliver the value that you are looking for or your leadership is looking [3:12] for and hence they tend to be failure and but the interesting or the positive sentiment is we [3:19] have started to see a change in direction. More and more enterprises are started to move those [3:26] pilots that they got started into production and that number is started to increase on a day-to-day [3:33] basis, right? Which is the positive sentiment. Now when [clears throat] you look at that 9% who [3:39] have been successful and you double click on hey what is happening in those how many of people are [3:46] getting it right how many people are not getting it right and what are those people the six% who [3:53] are doing it right is some of the points that you know I'm talking about because they always start [4:00] with the workflow they don't start with the use cases and for you to start a workflow you [4:06] need to have the knowledge of the workflow. So as technologists of course we all are technologists [4:11] and IT leaders having the understanding of the workflow and the process becomes critical and when [4:19] you partner with someone who knows that process workflow you are positioning yourself for a higher [4:24] success than otherwise and hence that becomes very critical and how do you tie the AI to business [4:31] outcomes. ROI is not a new thing. We all have been doing probably ROI for last three decades based on [4:39] when we started from digital transformation to cloud transformation to mobile transformation. [4:44] But what is unique with AI is the way you measure the AI outcomes are different than how we have [4:51] measured the ROI previously because it's not only changing the dimension of efficiency which many [4:59] enterprises get started but it's also changing the velocity of the work which you never measured [5:08] previously. And the third one that it is changing is the revenue uplift that you can bring into your [5:14] enterprises. Right? And then of course the last but not the least is the risk and compliance. [5:22] So these value streams in which you measure ROI did not exist previously when you underwent these [5:30] huge transformations in the last few decades. And that's what makes it AI ROI very different. [5:36] And hence AI ROI should not be a afterthought. It should be part of your or design. So as and when [5:45] you define your organization structure to drive AI transformation or agentic transformation, ROI [5:52] should be part of the org design where you have a team to help you define the baseline metrics what [5:59] you are going to measure and determine how you are going to measure the outcomes. That will ensure [6:07] that you get continued investments as you drive transformation. that will ensure you are able to [6:13] pick the right workflows. That will ensure you are able to pick the right use cases for the outcomes [6:19] that deliver more value to you. And that is what these 6% people are doing. Right? And hence they [6:26] have been more successful than others who start on a wrong foot. Right? Now when you look at the [6:34] measurement, you know, if you see many of the enterprises, the measurement they are doing is [6:40] very different. Right? They are saying hey how many number of users are using AI today how many [6:47] pilots we are executing how many tokens they are consuming right that's a wrong measurement right [6:55] I mean I'm pretty sure everyone is aware of the price increase in tokens the tokconomics that [7:03] you know is bombarding the enterprises hence the measurement needs to be different when you look [7:09] at the adoption the way it gets started is yes you will measure how many number of users are [7:16] adopting my uh technology that I am enabling or empowering them with but then the next one that [7:23] you will need to measure is the proficiency now what does proficiency mean proficiency means by [7:30] using that AI am I able to increase the velocity of work or am I able to change the way I work that [7:41] is something you need to measure and then third once you measure the proficiency you need to look [7:48] at what is the financial outcome I'm getting it by doing that so you need to measure the adoption [7:55] in those three layers not just from more users more pilots more tokens but more driven towards [8:03] the workflow more driven towards the proficiency and then the financial outcome that can you can [8:10] achieve as part of it that will ensure that you are measuring the right things to deliver the [8:17] business outcomes that you are expected and that's will make of course many of your CFOs happy and [8:24] he will provide more money to you to invest more right now tokconomics is a very interesting thing [8:31] right it's emerging very different of course many people have uh seeing the progress of the open [8:38] weight models um the models that are coming from China, the models that are being open-sourced by [8:46] thinking labs um by Google and others and many customers are looking at do I need to use the [8:54] frontier models for every single task or prompt that I execute within my organization right and [9:04] you don't require a frontier model for every single task and that means you need to identify [9:11] or build a mechanism which will help you determine what is the right model to use for the right task [9:19] which will ensure that you are not only using the right model but you are using it efficiently and [9:26] hence your tokconomics is much more impactful and that will also ensure that you're building [9:33] a common orchestration layer within your organization because I'm pretty sure you are not [9:38] going to use a technology from just one provider. You are going to use technology from multiple [9:43] providers based on either the business unit, based on the task you're trying to achieve and [9:48] everything. So it requires a single orchestration layer which can bring in the context from your [9:56] enterprise data which means you need to have a unified data layer right and that will ensure [10:04] that you are building the right model control plane as part of the model orchestration layer [10:09] to ensure you are using the right tokconomics. So you have heard lot of people saying hey for [10:15] AI to be working effectively you need to have the right data foundations and you are saying [10:20] hey I've been building data warehouses data lakes for number of years I already have the right data [10:26] foundation but for agentic AI to be working effectively you need to move beyond the right [10:35] data foundation in the form of data links and data warehouses to what we call as knowledge graphs. [10:42] because that is the mechanism which agent AI understands very effectively. So you can ensure [10:48] that the actions taken by your agents are based on the data that is grounded in your enterprise that [10:58] is organized in a knowledge graph fashion right that is what makes being making AI ready means [11:08] so you have a additional layer above and beyond the core data foundations that you may have in [11:13] place now to be estab established for agent AI to work effectively and that drives the context what [11:22] you when you see the context there that's what the context means in order to have the right [11:26] context and hence the right model for doing the right task ensures you are delivering the right [11:32] outcomes and the right toconomic outcomes that is what is important. So now how do you start right [11:44] we said we should start with the workflow not the use case not the model that is the starting point [11:50] and now what else is required as part of that right we need to understand that you know what [11:57] is the use case we need to pick and how do you pick a use case you pick a use case based upon [12:03] where you get the maximum outcome but also that are you able to measure it. If you're not able to [12:12] measure the outcome, it might be the great use case, but that is not where the starting point [12:18] should be. And hence, it's very critical to ensure that [clears throat] you pick the right workflow, [12:24] the right use case, and ensure you have the human in the loop process as part of that, right? [12:32] Because not for every single workflow, not for every single task, but for the critical workflows [12:37] and the critical task, human in the loop is very important to ensure that the outcome that AI is [12:44] delivering is up to the metrics that you are measuring and up to what uh guard rails that [12:52] you want to measure with. Right? So that is how you know we suggest as evalu the approach to how [13:00] you should get started and uh move forward and how do we now engineer to the outcome in terms [13:07] of scale yes you started a pilot you started with the workflow you started with the use case now how [13:12] do we drive it at scale and what is required to drive it at scale now I'm pretty sure the [13:18] industry term of forward deployed engineering is something which all of you have been hearing And [13:25] but where the forward deploy engineering is taking a pause is while they are great technologies to [13:30] apply technology they don't understand the process workflow of different industries because every [13:36] industry has a different process workflow its own nuances its own uh you know validations risk [13:43] compliance and other things. So you need to have a combination of few different role types in order [13:50] to make a effective application of AI and that is what we are calling out here and what we call as [14:00] you know agentic squad that is what you know evalu brings to the table where you combine a domain [14:06] person who understands the workflow associate him with the technology person like a FDE [14:14] support it with the data part who understands the data that is required to bring in the context and [14:21] then the last but not the least which is a very critical element is the adoption consultant. [14:28] Now we are seeing that you know as you enable technologies while at a global level or at a [14:36] federal you know standard level you are applying AI to every workflow but you say hey the business [14:41] users are not adopting it. I already you know enabled them with tools I have enabled them [14:46] with the technology but I'm not seeing it getting adopted and the reason for that is [14:52] because we are not driving the change management and the adoption at the workflow level. We are [14:58] driving it at a organization level which is a generic change management. We are not talking the [15:03] language of those end users who work probably in a summarization of legal documents in his language [15:10] or we are not talking to a person who who is doing a financial closure of books at every month in his [15:16] language. When you don't do that the end user will say hey I'm pretty happy with the Excel I [15:24] work with every single month. I would not move to this AI new tool that you're bringing to me and [15:30] that is why it's important to have these elements as you think through how you want to engineer [15:36] to outcome right so what we are doing from evalu is we can help you [clears throat] to understand [15:47] how a AI value realization should look like right what are the frameworks that you can use [15:55] We have built certain tools that can help you get started quick enough so that you are measuring the [16:01] AI in a right fashion from a outcome standpoint and it's not a one-time because when you start [16:10] building these workflows some of the workflows will give you the outcomes in 4 weeks some of [16:16] them will be in 3 months some of them will be in 6 months so you will need to design for the [16:22] workflows for which you are applying AI and then that's how you measure the ROI and the outcomes [16:27] of it. So please reach out to us if you require help in this area. We'll be able to share you with [16:32] you the frameworks, the tools that we have built that can help you to measure AI effectively and [16:39] then you can apply AI across the board. If there are any questions, let me know. But otherwise, [16:45] thank you for coming and joining us. Hope this was useful. Thanks everyone.