Sage IT Pushes Firms to Replace SaaS Subscriptions With Internal AI

Compounding ROI with AI: From SaaS Disruption to the Fully Agentic Enterprise With Sage IT — Ai4

Ai4

Reinvesting engineering hours saved by AI coding tools into dismantling commercial software contracts generates the direct cost savings needed to fund enterprise-wide agent deployment.

AI-assisted summary of Compounding ROI with AI: From SaaS Disruption to the Fully Agentic Enterprise With Sage IT — original source. Check the original for full context.

Source published: 2026-08-19

Detailed notes

## Background Published by Ai4 on YouTube on August 19, 2026, this presentation features Srikanth Chitipotu, Vice President of Engineering at Sage IT, outlining practical methods for achieving compounding ROI on AI investments. The session description highlights a phased approach where enterprises bypass speculative science projects by capturing quick, low-risk wins—specifically disrupting bloated SaaS subscriptions—to fund the sequential journey toward a fully agentic enterprise.

## TL;DR Srikanth Chitipotu details a four-stage framework for turning AI initiatives into a self-funding, compounding economic flywheel. He begins by exposing the financial pitfalls of runaway token costs and flawed vanity metrics, proposing a comprehensive ROI equation centered on workflow effectiveness and outcome observability. The strategy guides organizations to first unlock developer capacity through agentic coding tools, redeploy that reclaimed bandwidth to dismantle expensive commercial SaaS platforms, establish an autonomous agentic workforce using frugal model architectures, and ultimately transform entire business processes from the ground up.

## Atmosphere & style The talk delivers a pragmatic, fast-paced, and authoritative masterclass that cuts through standard AI hype with sharp operational realism. Chitipotu maintains an earnest, engineering-first delivery, using real-world case studies and concrete architecture diagrams to demystify complex deployment risks. The atmosphere feels intensely tactical and focused, leaving the audience with an empowering, structured blueprint for immediate organizational execution.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) AI investments should be structured as a self-funding compounding flywheel where early wins directly finance successive transformations, rather than being treated as isolated science projects. - [01:00](https://www.youtube.com/watch?v=349T6kz-VA8&t=60s) True operational AI costs often climb despite dropping per-token prices due to shifting tokenizers, complex multi-step tool chains, and steep surges in user adoption. - [02:30](https://www.youtube.com/watch?v=349T6kz-VA8&t=150s) Prematurely adopting flagship frontier models can rapidly burn enterprise budgets, underscoring the need to pace model upgrades to actual realized ROI. - [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) Measuring speed over resolution quality leads to operational rollbacks, as shown by Klarna's need to re-engage human staff after initially replacing them with customer service bots. - [05:30](https://www.youtube.com/watch?v=349T6kz-VA8&t=330s) Accurate AI valuation requires capturing baseline telemetry before production deployment and accounting for hidden costs like 'bot sitting' human oversight. - [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) Developer productivity boosts from AI copilots only generate real business ROI if reclaimed engineering bandwidth is deliberately redeployed into high-value capabilities. - [10:10](https://www.youtube.com/watch?v=349T6kz-VA8&t=610s) A 'conductor' architecture maximizes developer leverage by using a frontier model for planning and code review while delegating discrete implementation tasks to cost-effective specialized agents. - [12:20](https://www.youtube.com/watch?v=349T6kz-VA8&t=740s) Agentic coding fundamentally shifts the build-versus-buy equation, enabling lean engineering teams to dismantle and replace bloated, expensive commercial SaaS tools with custom in-house systems. - [13:55](https://www.youtube.com/watch?v=349T6kz-VA8&t=835s) SaaS replacement should follow a low-risk phased playbook: begin with low blast radius tools, eliminate niche utilities, audit upcoming 12-month renewals, and rebuild around proprietary business logic. - [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) Scaling an operational agentic workforce requires 'frugal architectures' using domain-tuned Small Language Models (SLMs) instead of defaulting to frontier models for routine deterministic tasks. - [17:30](https://www.youtube.com/watch?v=349T6kz-VA8&t=1050s) True agentic transformation reinvents processes from the inside out, elevating human employees from manual drafters into strategic supervisors and auditors of AI-generated work. - [18:54](https://www.youtube.com/watch?v=349T6kz-VA8&t=1134s) The AI compounding loop is sequential: agentic development creates bandwidth, SaaS disruption captures hard dollar savings, and an agentic workforce drives full enterprise transformation.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) Introduction and the Contrarian ROI Reality

Srikanth Chitipotu, Vice President of Engineering at Sage IT, opens his talk by setting the stage for a practical, step-by-step playbook designed to achieve compounding returns on AI investments. Rather than viewing AI as isolated science experiments or merely taking incremental cost savings from one initiative to another, the focus is on creating a self-funding flywheel where initial wins compound into larger organizational transformations. While there is plenty of contrarian skepticism surrounding AI disruption, Sage IT approaches the technology from a position of conviction: AI will drive immense long-term value, provided enterprises approach its implementation responsibly and measure its business impact accurately.

Before laying out the compounding stages, Chitipotu addresses the hard financial realities often overlooked by enterprise teams. Although base per-token inference prices appear to trend downward on surface level, the real operational cost of running AI use cases is frequently climbing. Underlying tokenizers are shifting, and increasingly complex agentic architectures trigger continuous chains of tool calls. Chitipotu highlights frontier model pricing dynamics—such as newer flagship releases costing significantly more per million tokens than prior versions—as a reason not to prematurely jump to the largest model. Teams must balance development and runtime expenditures, bearing in mind that high-performing AI features naturally drive five-to-six-fold spikes in user adoption, compounding infrastructure bills rapidly.

- **Underlying Cost Inflation**: Evolving tokenizers, multi-step tool calls, and high token volume can drive up real workload costs even when nominal list prices appear low. - **Run-Time Demand Surges**: As internal and external users realize the power of an AI feature, usage volume expands dramatically (e.g., from 1–3 runs to 5–6 runs per workflow), causing runtime budgets to balloon. - **Premature Frontier Adoption**: Enterprises frequently exhaust token budgets prematurely (citing real-world examples like Uber burning planned allocations in 3–6 months), underscoring the need to pace model upgrades based on realized ROI.

---

[04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) Where ROI Calculations Break and How to Measure Value

ROI calculations frequently fail due to three core pitfalls: measuring vanity metrics rather than true business outcomes, neglecting pre-deployment instrumentation, and failing to plan for regulatory or vendor constraints. Citing the widely discussed Klarna case study from 2024, Chitipotu points out that replacing 700 human workers with automated bots initially looked like a major efficiency win when measuring speed of resolution. However, the organization subsequently had to pivot back and re-engage human staff because the metric ignored quality of resolution. Furthermore, without baseline operational telemetry in place before an AI deployment goes live, organizations lose the ability to accurately benchmark performance after the fact. Regulatory shifts or open-weights compliance requirements can also trigger expensive rewrites if architectural flexibility is not planned from day one.

To evaluate AI investments accurately, Chitipotu presents a five-part ROI framework combining capacity released, hard costs saved, time saved, risk averted, and total operational cost—evaluated against an overarching fifth pillar: **effectiveness**.

``` (Capacity Released + Cost Saved + Time Saved + Risk Averted) Overall ROI = ------------------------------------------------------------- Total Cost * [ Effectiveness Multiplier ] ```

Total cost encompasses token usage, underlying compute, build overhead, tooling licenses, maintenance, and "bot sitting"—the ongoing human-in-the-loop engineering effort required to monitor and guide models. To measure effectiveness, Chitipotu advocates moving beyond low-level LLM tracing (e.g., isolated latency or prompt tokens) to full "outcome observability", tracking whether end-to-end business workflows actually complete correctly without unintended negative side effects.

- **Resolution Quality vs. Speed**: Tracking fast closure times without monitoring resolution accuracy leads to high customer churn and operational rollbacks. - **Pre-Deployment Telemetry**: Comprehensive baseline measurements must be captured before deploying AI into production, as production workflows cannot easily be rolled back later for benchmarking. - **"Bot Sitting" Overhead**: True total cost must include human intervention time, reviewer cycles, and engineer oversight.

---

[08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) Stage 1: Agentic Build and Engineering Redeployment

The first foundational stage of compounding ROI is **Agentic Build**, which focuses on supercharging internal development teams with AI tooling. Chitipotu considers deploying copilots and AI coding assistants table stakes for modern engineering organizations. However, simply freeing up developer hours does not create ROI on paper unless that recovered velocity is deliberately captured and redeployed onto high-value business initiatives. Without strategic reallocation, developer time savings simply evaporate into routine slack by Monday morning.

In practice, a structured rollout across an engineering team can yield up to a 60% boost in developer productivity. In an example deployment involving 15 software engineers, reclaiming 60% of their bandwidth produced significant capacity gains while noticeably reducing code defects in the final delivered product.

``` [ Step 1: Copilots ] [ Step 2: Agent Execution ] Individual Dev Tools ---> 80-90% Autonomous Tasks (Claude Code, etc.) (Human reviews outcome) | v [ Step 4: Full Autonomy ] [ Step 3: Conductor Pattern ] Hands-off execution <--- Frontier Model Conductor + (Future Horizon) Cheaper Specialized Agents ```

Chitipotu details a practical four-step progression for software development: starting with basic developer copilots, advancing to task-oriented agents handling 80% to 90% of routine coding, moving into an orchestration layer using a "conductor" architecture, and eventually reaching autonomous execution. In the conductor model, an intelligent frontier model breaks complex tasks into discrete tickets, passes them to cost-effective specialized agents, and verifies the consolidated code for accuracy—amplifying developer productivity by 3x to 4x.

- **Redeployment Imperative**: Time saved by developers must be explicitly budgeted into new capability builds rather than left unmanaged. - **Conductor Architecture**: Offloads modular development sub-tasks to inexpensive smaller models while preserving frontier model intelligence strictly for orchestration and final verification. - **Measurable Quality Gains**: In addition to raw velocity, codebases developed with structured agentic workflows exhibit measurably higher defect-free release rates.

---

[12:20](https://www.youtube.com/watch?v=349T6kz-VA8&t=740s) Stage 2: Disrupting Bloated SaaS and Reclaiming Business Logic

Once developer capacity is unlocked in Stage 1, organizations must avoid sinking those hours into vague research bets and instead direct them toward **SaaS Disruption**. Chitipotu highlights how enterprise software portfolios have accumulated massive bloat over decades. The historical boundary between "build vs. buy" has radically shifted: while enterprises previously bought external software licenses to avoid maintenance overhead and delivery risk, modern agentic tooling allows small engineering teams to rapidly build, customize, and maintain core business tools in-house.

Instead of paying recurring seat licenses for bloated commercial suites, teams can target legacy software and replace it with custom-tailored internal platforms. In one deployment, an organization successfully dismantled an expensive commercial CRM platform, replacing it with a focused internal system tailored to their specific operational workflows without requiring manual fallbacks.

``` PHASE 1: Low Blast Radius --> Target isolated systems with few dependencies PHASE 2: Esoteric / Niche --> Replace single-purpose tools (FHIR, Doc Mgmt) PHASE 3: Core Logic --> Build proprietary workflows, not vendor opinions PHASE 4: 12-Month Renewals --> Audit upcoming license renewals for teardown ```

To execute this transition safely, Chitipotu outlines an incremental teardown strategy: 1. **Low Blast Radius Platforms**: Begin with peripheral tools that have minimal ecosystem integrations to prove the build pipeline with near-zero operational risk. 2. **Esoteric & Niche Tools**: Target domain-specific utilities, such as specialized healthcare FHIR parsers or niche document management software, reclaiming licensing capital for core innovation. 3. **Core Business Functionality**: Rebuild primary operational workflows around proprietary business processes rather than forcing staff to conform to an external vendor's opinionated product design. 4. **Upcoming Contract Renewals**: Systematically audit software renewals scheduled over the next 12 months and pre-emptively build custom replacements.

---

[15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) Stage 3 & 4: Agentic Workforce and Full Enterprise Transformation

The compounding journey advances into **Agentic Workforce** (Stage 3), where AI agents transition from developer tools into first-class operational participants. In this phase, autonomous agents are provisioned with enterprise identities, access controls, security boundaries, and governance protocols. Rather than defaulting to expensive frontier models for every workflow step, organizations implement "frugal architectures" using focused Small Language Models (SLMs) and domain-tuned models optimized on proprietary internal data. Chitipotu cautions against placing non-deterministic LLMs in front of standard deterministic business logic, advising teams to adopt a "storming before forming" mindset—rapidly testing five agentic concepts and scaling the single winning design. In real-world customer service and operations environments, agentic workforce deployments have recovered two full business days per service request, driving $1.5 million in bottom-line recovery within the first three months.

The final evolution is **Agentic Transformation** (Stage 4), where organizations fundamentally reinvent business processes from the ground up rather than automating fragmented legacy steps. Employees shift from manual drafters and executors into reviewers and strategic supervisors—such as moving from writing complex legal contracts by hand to auditing AI-synthesized agreements. Chitipotu emphasizes that genuine enterprise transformation must emerge internally from frontline domain experts rather than being dictated by external consultants, turning AI into a sustainable, compounding competitive advantage.

- **SLMs over Frontier Models**: Deploy small, specialized models tuned on enterprise data to minimize inference costs and maximize ROI in the denominator. - **Frugal Architecture**: Keep deterministic rules out of LLM prompts; only deploy probabilistic agentic layers where genuine reasoning or linguistic interpretation is required. - **Operational Payoff**: Streamlining multi-agent operational workflows can reduce service request cycle times by up to two days, yielding measurable multi-million dollar returns within the first quarter of deployment.

[18:54](https://www.youtube.com/watch?v=349T6kz-VA8&t=1134s) The Compounding Cycle of Agentic Transformation and Closing Remarks

The culmination of the strategy lies in how each phase naturally feeds into and funds the next. The operational capacity gained from initial agentic development is strategically repurposed to disrupt existing, expensive SaaS tools. This disruption yields direct, flat cost savings that can be immediately reinvested into building out a dedicated agentic workforce. Through this sequential progression, organizations gain deep, hands-on learning about how autonomous agents interact with core workflows, providing the foundation needed to fundamentally transform the entire enterprise.

With his presentation time winding down to the final twenty seconds, the speaker concludes by inviting attendees to continue the conversation in person or visit the Sage IT team at booth 316 for questions and deeper discussions on executing this strategy.

- **The Compounding Loop:** Agentic Build Capacity $\rightarrow$ SaaS Disruption & Flat Cost Savings $\rightarrow$ Agentic Workforce Investment $\rightarrow$ Business Process Transformation. - **Connect:** Sage IT booth 316.

## Notable quotes > The edge isn't whether you build; it's whether the returns compound. So we flip it: realize ROI now, and let each win fund the next. — Srikanth Chitipotu

> It starts where returns are fastest and lowest-risk - disrupting the bloated SaaS you're already overpaying for - then compounds from there, all the way to a fully agentic enterprise. No big bets, no science projects, just real value one win at a time. — Srikanth Chitipotu

> [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) AI will drive immense long-term value, provided enterprises approach its implementation responsibly and measure its business impact accurately. — Srikanth Chitipotu

> [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) Without baseline operational telemetry in place before an AI deployment goes live, organizations lose the ability to accurately benchmark performance after the fact. — Srikanth Chitipotu

> [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) Simply freeing up developer hours does not create ROI on paper unless that recovered velocity is deliberately captured and redeployed onto high-value business initiatives. — Srikanth Chitipotu

> [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) Without strategic reallocation, developer time savings simply evaporate into routine slack by Monday morning. — Srikanth Chitipotu

> [12:20](https://www.youtube.com/watch?v=349T6kz-VA8&t=740s) The historical boundary between 'build vs. buy' has radically shifted: while enterprises previously bought external software licenses to avoid maintenance overhead and delivery risk, modern agentic tooling allows small engineering teams to rapidly build, customize, and maintain core business tools in-house. — Srikanth Chitipotu

> [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) Keep deterministic rules out of LLM prompts; only deploy probabilistic agentic layers where genuine reasoning or linguistic interpretation is required. — Srikanth Chitipotu

> [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) Genuine enterprise transformation must emerge internally from frontline domain experts rather than being dictated by external consultants, turning AI into a sustainable, compounding competitive advantage. — Srikanth Chitipotu

> [18:54](https://www.youtube.com/watch?v=349T6kz-VA8&t=1134s) The operational capacity gained from initial agentic development is strategically repurposed to disrupt existing, expensive SaaS tools. This disruption yields direct, flat cost savings that can be immediately reinvested into building out a dedicated agentic workforce. — Srikanth Chitipotu

## Mentioned Srikanth Chitipotu, Sage IT, Ai4, Klarna, Uber, Claude Code, FHIR

## Video description Srikanth Chitipotu - VP Engineering, Sage IT

Doing AI is table stakes now. Everyone has multiple programs in production and many more in flight. The edge isn't whether you build; it's whether the returns compound. So we flip it: realize ROI now, and let each win fund the next. It starts where returns are fastest and lowest-risk - disrupting the bloated SaaS you're already overpaying for - then compounds from there, all the way to a fully agentic enterprise. No big bets, no science projects, just real value one win at a time. We've done it across multiple domains. Join us, and we'll show you how to start in yours.

[Source](https://www.youtube.com/watch?v=349T6kz-VA8)

中文简报

Sage IT 推动企业用内部 AI 替代 SaaS 订阅将 AI 编程工具节省的工程工时重新投入到解约商业软件合同中，所带来的直接成本削减可为企业级 Agent 部署提供资金支持。 Srikanth Chitipotu 详细介绍了一套四阶段框架，旨在将 AI 举措转化为自我造血、持续复利的经济飞轮。他首先揭示了失控的 Token 成本与虚荣指标所带来的财务陷阱，并提出了一套以工作流效能和结果可观测性为核心的综合 ROI 方程。该战略引导企业首先利用 Agent 编程工具释放研发产能，再将节省出的人力用于替代昂贵的商业 SaaS 平台，通过轻量节约型模型架构建立自主 Agent 员工团队，并最终自下而上彻底重塑整个业务流程。 Sage IT 指出，企业应将 AI 投资设计为自我造血的复利飞轮，通过早期快速见效的成果直接资助后续转型，而非停留在孤立的科学实验。 (0:04) 尽管名义上的每 token 标价呈下降趋势，但底层分词器演变、复杂的长链工具调用以及用户使用量的激增，常导致企业实际的 AI 运行时账单大幅攀升。 (0:04) 仅衡量解决速度而忽视交付质量会导致业务回退（如 Klarna 的反思案例），精准评估 AI 价值需在投产前采集基线遥测，并算入人工“看护机器人”（bot sitting）等隐性运维成本。 (4:13) AI 辅助工具可提升高达 60% 的开发产能，但该时间红利必须被主动重定向至核心业务；通过“指挥官”架构，前沿模型负责规划与审查，并将离散任务交由低成本专用 Agent 完成。 (8:22) Agentic 编程彻底颠覆了“自建与外购”的权衡，精益团队能够自建系统拆换昂贵臃肿的商用 SaaS，并将节省下的巨额软件许可费用直接注资于构建企业级 Agent 员工。 (12:20)

Reading summary

Srikanth Chitipotu of Sage IT outlines a framework for compounding returns on enterprise AI, arguing that companies must capture developer productivity gains to dismantle bloated commercial software portfolios and fund an autonomous agentic workforce. Chitipotu contends that sustainable ROI requires tracking business effectiveness and managing operational expenses rather than relying on vanity metrics or premature upgrades to frontier models. Hidden Costs and Flawed ROI Metrics (0:04) Chitipotu warns that nominal declines in per-token inference prices obscure rising operational costs, which are driven by changing tokenizers, complex tool-calling chains, and post-deployment adoption spikes that routinely expand workflow runs five- to six-fold. Organizations that prematurely deploy expensive flagship models risk rapid budget depletion, an outcome Chitipotu illustrates by citing examples such as Uber exhausting planned token allocations within three to six months. Traditional ROI calculations frequently fail by measuring velocity rather than operational quality and by ignoring the costs of "bot sitting"—the ongoing human-in-the-loop engineering required to supervise models. Citing a 2024 case where Klarna replaced 700 workers with automated bots to optimize resolution speed but subsequently had to re-engage human staff when resolution quality suffered, Chitipotu argues that speed without accuracy leads to rollbacks. To track true returns, he proposes an ROI model that balances capacity released, direct cost savings, time saved, and risk averted against total operational costs, scaled by an "effectiveness multiplier" reliant on pre-deployment baseline telemetry. Capturing Developer Bandwidth Through Agentic Build (8:22) The first operational phase, Agentic Build, focuses on accelerating internal software engineering. In practice, structured rollouts of AI coding tools can yield up to a 60% boost in developer productivity while reducing defect rates in production codebases, as demonstrated in a 15-engineer deployment cited by Chitipotu. However, he cautions that developer time savings do not generate real ROI if the reclaimed hours simply dissipate into routine slack; organizations must deliberately budget and redeploy this recovered capacity toward strategic business projects. Chitipotu outlines a four-step progression for engineering teams: deploying individual copilots, moving to task-oriented agents handling 80% to 90% of routine coding, adopting a "conductor" architecture, and ultimately achieving full autonomous execution. Under the conductor pattern, an advanced frontier model translates high-level tickets into modular development tasks, delegates them to lower-cost specialized agents, and verifies the merged code, which Chitipotu claims can amplify developer velocity by three to four times. Dismantling Bloated Enterprise SaaS Portfolios (12:20) Rather than absorbing recovered developer velocity into open-ended research, Chitipotu argues that enterprises should direct this capacity toward "SaaS Disruption." The historical build-versus-buy equation has shifted: where companies once purchased off-the-shelf software to avoid maintenance burdens, agentic development tools now allow small engineering teams to rapidly build, customize, and maintain core operational platforms internally, eliminating costly recurring seat licenses. To execute this without operational disruption, Chitipotu details a four-phase teardown strategy: start with low-blast-radius peripheral platforms with few integrations; replace niche or esoteric utilities, such as specialized healthcare FHIR parsers or single-purpose document management tools; rebuild core proprietary workflows, such as replacing a commercial CRM with an internal system tailored to existing business logic; and systematically audit commercial software contracts due for renewal over the next 12 months. Deploying Frugal Workforces and Process Reinvention (15:38) The compounding model advances into an "Agentic Workforce," where autonomous agents receive enterprise identities, permissions, and security boundaries. Chitipotu advocates for "frugal architectures" that deploy Small Language Models (SLMs) and domain-tuned models on proprietary data, reserving non-deterministic models for reasoning while keeping standard deterministic business rules out of LLM prompts. In one operational customer service implementation, Chitipotu reports that multi-agent automation cut service request cycle times by two business days, generating $1.5 million in bottom-line recovery within three months. In the final stage, "Agentic Transformation," organizations redesign entire workflows from the ground up rather than automating fragmented legacy steps. Workers transition from manual execution, such as drafting complex contracts, to auditing and supervising AI-generated deliverables. Chitipotu emphasizes that genuine enterprise reinvention must be directed internally by frontline domain experts rather than external consultants to establish a defensible competitive advantage. Sage IT's framework posits that enterprise AI generates compounding returns only when managed as a self-funding loop: engineering gains dismantle external software licensing costs, freeing up capital to build a domain-specialized agentic workforce. The viability of this strategy hinges on an organization's ability to measure resolution effectiveness rather than vanity speeds, maintain custom software in-house over the long term, and avoid the cost inflations associated with over-relying on flagship frontier models.

精炼阅读摘要

Sage IT工程副总裁Srikanth Chitipotu提出，企业实现AI复利投资回报率（ROI）的关键在于构建自我供资的闭环飞轮：利用智能体工程释放的开发产能自主研发替代臃肿的商业SaaS，再将节约的许可费用再投资于智能体劳动力与业务流程重构，同时必须摒弃虚荣指标以防范实际运行成本的失控膨胀。真实ROI评估模型与隐藏运营成本 (4:13) Chitipotu指出，尽管名义上的Token单价呈下降趋势，但分词器演变、多步工具调用链以及用户使用频次激增（例如工作流调用量激增5至6倍），常导致企业实际运营成本大幅攀升，部分企业（如Uber）甚至在3至6个月内耗尽年度Token预算。此外，仅衡量处理速度而忽略解决质量会引发严重业务倒退，例如Klarna在2024年以自动化Bot替代700名员工后，因未兼顾解决质量最终不得不重新引入人工服务。为此，Sage IT提出了五要素综合ROI模型，将“释放产能、节约硬性成本、节约时间与规避风险”之和除以“总成本”，并乘以核心支柱“有效性乘数”。总成本必须涵盖算力、工具许可、系统维护以及工程师持续监控干预的“看护成本”（bot sitting）。Chitipotu强调，企业必须在系统上线前建立基线遥测指标，并转向追踪端到端业务能否正确闭环的“结果可观测性”，避免将不可回滚的工作流盲目推向生产环境。工程自主构建与开发产能重分配 (8:22) 在ROI复利演进的第一阶段（Agentic Build），企业通过部署AI辅助编程工具提升研发效能。Chitipotu认为配置Copilot已是现代研发团队的底线，但单纯节省开发时间若不主动进行战略性重分配，释放的工时便会在日常松懈中迅速蒸发。在实际案例中，对15名软件工程师开展结构化智能体部署，不仅释放了高达60%的开发产能，还显著降低了交付成果中的代码缺陷率。该阶段的技术演进包含四个步骤：从单兵Copilot工具，发展到接管80%至90%常规编码的任务智能体，再迈向“指挥官架构”（Conductor Pattern），最终抵达全自主执行。在指挥官模式中，高智能的前沿模型仅负责任务拆解与最终代码校验，并将具体开发子模块分派给低成本的专业化小模型协同执行，从而实现3至4倍的工程生产力放大。渐进式拆解SaaS以收回业务逻辑 (12:20) 当第一阶段释放出工程产能后，企业应避免将资源投入模糊的研究实验，而应转向第二阶段“颠覆SaaS”（SaaS Disruption）。长期以来企业软件采购充斥着冗余功能与持续的坐席授权费，而现代智能体工具的成熟使得“自研与外购”（build vs. buy）的权衡发生逆转，小型工程团队已具备低风险快速自研内部工具的能力。在实践中，已有企业成功拆解并替换了昂贵且臃肿的商业CRM平台，重构出紧密贴合自身工作流的内部系统。为确保平稳过渡，Chitipotu给出了四步拆解路径：首先替换生态依赖极少的“低爆炸半径”边缘工具；其次攻关特定垂直领域的单用途组件（如医疗FHIR解析器或文档管理系统）；第三步重构核心业务逻辑，使系统契合企业专有流程而非供应商的设计主张；最后系统排查未来12个月内面临续约的商业软件合同，提前开展定制替代。节俭架构与企业智能体劳动力转型 (15:38) 在第三阶段“智能体劳动力”（Agentic Workforce）中，AI智能体被赋予企业身份凭据、访问权限控制和安全边界，直接参与业务运营。Chitipotu建议采用“节俭架构”（frugal architectures），优先使用基于企业专有数据微调的小语言模型（SLMs），严禁将非确定性的LLM强加在现有的确定性业务规则之上，并采取“先激荡后成型”策略（快速测试5个智能体概念，仅扩展1个胜出方案）。在客服与运营场景中，多智能体工作流将单个服务请求耗时缩短了整整两个工作日，在部署前三个月内实现了150万美元的净收益回收。最终的第四阶段是“全企业智能体转型”（Agentic Transformation），即从零开始彻底重构业务流程，而非修补传统碎片化环节。员工职责由此发生结构性转变，从手工起草者变为监督与审计人员（例如从手写复杂法律合同转为审核由AI生成的协议）。Chitipotu强调，真正的企业转型必须由一线业务专家自下而上驱动，而非依赖外部咨询机构灌输，以此形成可持续的长期竞争壁垒。 Sage IT总结强调，通过智能体开发释放的工程产能颠覆昂贵SaaS，SaaS节省的硬性资金再反哺智能体劳动力，从而形成了层层递进的ROI复利飞轮。然而，这一路径的成功依赖于企业在起跑阶段对总拥有成本与有效性的严谨度量，若忽视底层Token膨胀与真实业务完成质量，过早引入前沿大模型将可能迅速侵蚀预期回报。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 19 日在 YouTube 平台发布，主讲人为 Sage IT 工程副总裁 Srikanth Chitipotu。视频围绕企业如何摆脱盲目搞 AI“科研项目”的误区展开，重点阐述从颠覆臃肿高昂的传统 SaaS 软件切入，以低风险、快速回报逐步演进至全方位智能体企业的复合 ROI 实现路径。 ## TL;DR 演讲指出企业在 AI 落地中普遍面临指标错位、基线缺失以及推理成本失控等严峻挑战，并提出了涵盖产能释放、直接节约、时间价值与风险规避的五维度量公式与“结果可观测性”理念。随后，讲者系统化拆解了四阶段演进战术：首先通过智能构建（Agentic Build）释放研发产能；进而利用多余产能自研替代低性价比的臃肿 SaaS；接着引入赋予明确权限与安全边界的“智能体劳动力”并推行精益架构；最终由内而外重构业务流程，形成上一阶段收益反哺下一阶段创新的复合 ROI 飞轮。 ## Atmosphere & style 整场演讲风格务实笃定、直击痛点，兼具资深工程管理者的严谨理性与战略前瞻性。主讲人毫不避讳地拆解了诸如 Klarna 客服裁员反弹与算力账单失控等行业真实翻车案例，以层层递进的架构逻辑和清醒的成本账本打破了盲目追逐前沿模型的浮躁叙事。观众能清晰感受到一种高度崇尚“用数字说话、以战养战”的实战派工程交付氛围，沉稳且极具实操说服力。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) AI 落地核心在于实现 ROI 的复利增长，即用早期低风险、快速回报的收益持续反哺后续高阶智能化建设。 - [01:39](https://www.youtube.com/watch?v=349T6kz-VA8&t=99s) 大模型单价下调并不等于账单减少，由于分词器变化和 Agent 工具调用激增，端到端任务消耗的 Token 总量与成本反而显著攀升。 - [01:39](https://www.youtube.com/watch?v=349T6kz-VA8&t=99s) 构建与运行成本极易双重失控，AI 体验提升会导致调用频次激增数倍，若未算清账目可能出现推理成本反噬人力节约效益的局面。 - [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) 企业 AI 算不清账的核心原因包括指标严重错位（过度看重解决速度而忽视质量，如 Klarna 案例）、上线前缺少基线插桩以及合规风险导致的架构重写。 - [06:26](https://www.youtube.com/watch?v=349T6kz-VA8&t=386s) Sage IT 提出多维 ROI 公式，将产能释放、直接节约、时间价值与风险规避除以包含 Token、基础设施及人工看护（Bot Sitting）在内的全口径成本。 - [06:26](https://www.youtube.com/watch?v=349T6kz-VA8&t=386s) 衡量 AI 效果应从单纯的 LLM 链路追踪转向全局“结果可观测性（Outcome Observability）”，关注交付物在真实业务环境中的达成度与良品率。 - [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) 第一阶段智能构建（Agentic Build）的关键在于将释放的研发速率明确重新部署到战略业务中，避免节约的时间被日常琐事稀释。 - [11:17](https://www.youtube.com/watch?v=349T6kz-VA8&t=677s) 工程智能化演进正从单纯的代码补全向指挥官（Conductor）多智能体编排推进，通过“顶级前沿调度 + 廉价小模型执行”大幅倍增工程师产能。 - [12:20](https://www.youtube.com/watch?v=349T6kz-VA8&t=740s) 随着 AI 开发门槛大幅降低，自研与采购的平衡线转移，企业可主动替换昂贵、低效且臃肿的传统商业 SaaS。 - [13:46](https://www.youtube.com/watch?v=349T6kz-VA8&t=826s) SaaS 替代应遵循低爆炸半径、攻克垂直小众场景、系统性重构核心业务逻辑并锚定即将到期续约清单的四步战术。 - [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) 打造智能体劳动力需将 Agent 视作一等公民，赋予身份、权限与治理规范，并坚持聚焦专有智能、推行精益架构。 - [17:18](https://www.youtube.com/watch?v=349T6kz-VA8&t=1038s) AI 转型的终极形态是业务重构，通过智能构建释放产能、颠覆 SaaS 节省开销、打造智能体劳动力并最终实现业务闭环的复利飞轮。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) 开场与议程：如何实现 AI 投资回报率（ROI）的复利增长演讲由 Sage IT 工程副总裁 Srikanth Chitipotu 开启。他开门见山地指出，当今企业面临的核心挑战不是“要不要做 AI”，而是如何让 AI 的投资回报率产生“复利效应（Compounding ROI）”。很多企业以为做 AI 只是把一个项目的收益搬到另一个项目上，但真正的逻辑应该是：用当前阶段验证过的低风险、快速回报收益，去反哺和投资下一阶段更复杂的智能化系统，形成自给自足的复利飞轮。在正式展示实操指南（Playbook）之前，Srikanth 梳理了当天的分享框架： 1. **认知校准（Level Setting）**：拆解企业在落地 AI 时最容易在哪些地方踩坑、计算逻辑为何常常失效； 2. **ROI 测量体系**：介绍 Sage IT 跨多家客户实践验证的 ROI 精确度量公式； 3. **分阶段实操指南（Playbook）**：按阶段推进智能化建设； 4. **全景复利演进图**：展示各个阶段之间如何相互支撑、层层放大企业价值。针对当下业界对 AI 的态度分歧，Srikanth 明确了 Sage IT 的立场：虽然市场上存在许多对于 AI 颠覆性影响的悲观或对抗性观点，但他们坚定相信 AI 从长远看能够真正赋能业务，而技术的驾驭者必须以极度负责且重实效的态度来推进这一进程。 --- ### [01:39](https://www.youtube.com/watch?v=349T6kz-VA8&t=99s) 成本黑洞与隐形陷阱：模型升级与推理调用的隐藏代价很多人直觉上认为大模型的调用单价在不断走低，但实际账单却往往节节攀升。Srikanth 揭示了背后的隐形机制：虽然单个 Token 的标价看似下调，但底层分词器（Tokenizer）的设计在变，更重要的是，随着系统复杂度上升，Agent 调用的外部工具（Tool Calls）次数呈几何级数增加。这意味着完成同一个端到端任务所消耗的 Token 总量激增，最终的实际业务成本反而显著上涨。以市场上高端前沿模型为例，例如 GPT-4.5（发布于 2025 年初，每百万 Token 约 5 美元 / 30 美元），其成本相比此前版本直接翻倍。这表明企业绝不能盲目追赶最新的前沿大模型，必须依据实际业务收益精准评估其边际价值。除了底层模型定价，构建成本（Build Cost）与运行时成本（Run-time Cost）的双重失控也是常见的算力黑洞： - **构建成本失控**：项目在前期探索和原型搭建时极易耗尽预算，如 Uber 曾公开表示在短短 3 到 6 个月内就烧光了大量 Token 配额，类似现象在诸多企业客户中屡见不鲜； - **运行时成本激增**：当一个 AI 应用体验极佳时，其使用频次会呈指数级暴涨。过去人工流程用户可能每周只使用 3 次，引入体验优秀的 AI 后，使用频率可能直接飙升至 5 到 6 倍，推理成本随之水涨船高； - **客服退潮案例**：在某知名客服自动化部署案例中，由于前期没有算清调用与复购账目，模型推理费用直接吞噬了人力节约收益，最终团队不得不将 AI 客服系统紧急下线。 --- ### [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) 算盘为何打不响：指标错位、基线缺失与合规风险为什么众多雄心勃勃的 AI 计划最终无法算平账目？Srikanth 指出，ROI 计算往往在三个关键环节彻底崩溃：首先是**度量指标严重错位（Wrong Metric）**。最典型的例证便是金融科技公司 Klarna。Klarna 曾在 2024 年 2 月高调宣布借助与 OpenAI 的合作裁撤或替代了 700 名客服员工，然而仅一年之后，公司便不得不重新调整策略、将人类员工召回。正如其 CEO 后续反思的那样，他们在一开始就锚定了错误的考核指标——团队过度关注“单次问题解决的速度（Speed of Resolution）”，却牺牲并忽视了最为核心的“问题解决质量（Quality of Resolution）”，最终导致客户满意度受损与隐性维护成本飙升。 ```mermaid flowchart TD accTitle: AI ROI Failure Causes subgraph S1["指标与度量层"] A["指标错位 (关注速度忽视质量，如Klarna)"] B["缺乏事前插桩 (缺少Baseline，效果无法归因)"] end subgraph S2["外部约束层"] C["监管与数据合规限制"] D["模型被迫重写/替换为开源权重"] end subgraph S3["最终财务结果"] E["算力账单失控"] F["业务价值缩水，系统被迫下线"] end A --> F B --> E C --> D D --> E E --> F ``` 其次是**上线前未做埋点插桩（No Instrumentation Before Deployment）**。如果没有在 AI 系统接入之前建立起严格、客观的业务基线（Baseline），上线之后任何数据归因都会沦为主观臆断。AI 部署在真实用户面前后，企业很难再将其完全撤下以做“空白对照”。没有事前对比数据，真实的 ROI 将永远无法被量化。最后是**监管合规与外部限制（Regulations & Lock-in）**。强监管行业通常需要耗费数年时间满足合规审查。一旦合规政策或安全禁令叫停特定闭源模型，企业将被迫将整个架构迁移至开源权重（Open-weights）模型，这通常意味着整个业务工程栈必须彻底推倒重写，前期投入将遭遇毁灭性的折旧。 --- ### [06:26](https://www.youtube.com/watch?v=349T6kz-VA8&t=386s) ROI 度量黄金公式：五项指标与“结果可观测性” 为了摆脱传统的糊涂账，Sage IT 建立了一套贯穿全生命周期的五维度量体系。核心公式由四大显性指标构成，并由第五个核心隐性指标——“有效性（Effectiveness）”作为统摄： $$ \text{ROI} = \frac{\text{释放的产能 (Capacity Released)} + \text{节约的直接成本 (Cost Saved)} + \text{节约的时间价值 (Time Saved)} + \text{风险规避收益 (Risk Averted)}}{\text{总拥有成本 (Total Cost: Token + 基础设施 + 构建 + 辅导与看护人力 + 运维许可)}} $$ - **分子：多元业务收益** - **释放的产能（Capacity Released）**：人员被 AI 释放出来从事高附加值工作的时间份额； - **成本与时间节约（Cost & Time Saved）**：基础设施精简、外包减少以及全流程交付周期的压缩； - **风险规避（Risk Averted）**：通过前置拦截缺陷、合规审查带来的故障与诉讼风险规避。 - **分母：全口径总成本（Total Cost）** - 除了 Token 调用账单、底层算力基础设施构建与维护成本外，必须严格计入**人机协同/看护成本（"Bot Sitting"）**，即工程师作为 Human-in-the-loop 在旁修正、微调和指导 AI 所消耗的高昂工时。 - **核心控制维度：结果可观测性（Outcome Observability）** - “有效性（Effectiveness）”极难被单纯的标量捕捉。传统方法仅聚焦于 LLM 调用的链路追踪（LLM Traces）或单次交互质量，而 Sage IT 主张推行**结果可观测性（Outcome Observability）**——从整个技术栈与端到端业务场景的全局视野出发，监测最终交付物在真实业务环境中的良品率与业务达成度。 --- ### [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) Playbook 阶段一：智能构建（Agentic Build）与工程产能释放在实操演进指南（Playbook）中，第一阶段是每个研发团队的基础门槛：**智能构建（Agentic Build）**。Srikanth 直言，如果团队目前还没有在日常开发中全面普及 AI 编码辅助工具，那就属于严重滞后。然而，仅仅让工程师装上插件是远远不够的，这一阶段的核心成败取决于**“释放出的研发速率（Velocity）是否被真正重新部署（Redeployed）”**。如果工程师因为 AI 节省了 30% 的时间，但这些时间没有被明确规划投向新的战略性功能研发，那么所节省的产能“到周一早上就会被日常琐事彻底稀释”，在财务报表上完全隐形。在一个包含 15 名软件工程师的实际客户交付案例中： - 团队通过系统化引入 AI 工具链，实现了 **60% 的综合生产力提升**； - 这一提升直接折算为相当于释放了近 9 名全职工程师的产能，并在报表中转化为可度量的 ROI 收益； - 从“结果可观测性”的角度评估，最终交付的软件版本缺陷率显著下降，交付质量迎来了质的飞跃。 ```mermaid flowchart LR accTitle: Evolution of Agentic Engineering A["步骤 1: 辅助赋能 (Copilot / Claude Code)"] --> B["步骤 2: Agent 主导执行 (80-90% 工作自主完成)"] B --> C["步骤 3: 编排指挥官 (Conductor 指挥专业 Agent)"] C --> D["步骤 4: 全自动自治工程系统"] ``` 在落地演进路径上，团队归纳了工程智能化的演进阶梯： - **步骤 1：Copilot 辅助赋能**：在 IDE 内配置 GitHub Copilot、Claude Code 等交互式编程插件，完成基础代码补全（绝大多数成熟团队已迈过此阶段）； - **步骤 2：Agent 承担主体开发**：让自治编码 Agent 独立承担 80% 到 90% 的工程实现任务，人类工程师退居二线担任审查者（Reviewer），确保交付物符合规范； - **步骤 3：指挥官（Conductor）多智能体编排**：当前顶尖团队正在试验的方向。系统设立一个利用顶级前沿模型驱动的核心“指挥官（Conductor）”，由其将复杂业务需求层层拆解，调度各个专业小模型或领域 Agent 协同作业。该架构已在多家合作客户中进入试点阶段，并展现出极具爆发力的效能表现。 ### [11:17](https://www.youtube.com/watch?v=349T6kz-VA8&t=677s) 分布式智能与工程师产能的倍增释放通过合理的架构设计，企业可以将复杂的工程任务拆解分配给三到四个轻量级智能体（Sub-agents）。这些智能体不需要依赖昂贵的前沿模型（Frontier Models），只需采用成本低廉的模型即可完成基础且琐碎的工作；最后再由前沿模型担任“指挥家”（Frontier Conductor）进行结果评估与把关，确保交付质量的精准度。这种“前沿调度 + 廉价执行”的组合拳，能够让每位工程师的工作效率直接实现三到四倍的飞跃。尽管完全自主的智能体（Agent Autonomy）目前看来仍有一定距离，但部分走在前沿的企业和顶尖产品公司已经开始在此方向积极探索。通过引入智能体驱动构建（Agentic Build）所带来的直接收益，就是为企业彻底释放出宝贵的工程研发产能。而如何重新分配并利用好这些被释放的产能，正是迈向下一阶段的核心关键。 --- ### [12:20](https://www.youtube.com/watch?v=349T6kz-VA8&t=740s) 颠覆臃肿 SaaS：重塑“自研与采购”的边界当工程团队释放出多余的研发产能后，企业应立即着手审视当前正在付费的各类 SaaS 产品。长期以来，这一领域一直被严重忽视，但其实蕴藏着巨大的降本潜力——许多组织内部充斥着多年前引入、如今却食之无味弃之可惜的沉重 SaaS 工具。过去，企业在“自研还是采购”（Build vs. Buy）的抉择中往往倾向于采购，核心原因在于规避自行构建和后期维护的技术风险。然而，随着 AI 赋能开发，自研与采购的平衡线已经发生了根本性转移。自研门槛的大幅降低，促使越来越多的组织选择主动替换掉市场上那些价值持续缩水、性价比极低的传统 SaaS 软件。例如在客户关系管理（CRM）系统的替换实践中，不仅实现了预期的投资回报率（ROI），更通过智能自动化彻底消除了后期人工兜底（Manual Fallback）的维护负担。 ```mermaid flowchart LR accTitle: AI 时代自研与采购决策的转变逻辑 subgraph Past["传统决策模式"] A["高昂自研与维护成本"] --> B["倾向采购商业 SaaS"] B --> C["承担长期冗余订阅费用"] end subgraph Present["AI 智能体时代"] D["智能体构建释放研发产能"] --> E["自研门槛与成本断崖式下降"] E --> F["重构/替代传统 SaaS"] F --> G["消除维护负担与软件溢价"] end ``` --- ### [13:46](https://www.youtube.com/watch?v=349T6kz-VA8&t=826s) SaaS 替代战术手册：从低影响领域到核心业务掌控为了稳步推进 SaaS 替代并最大化投资回报，企业应当遵循一套结构清晰的落地手册（Playbook）： ```mermaid flowchart TD accTitle: SaaS 替代与重构的四步战术手册 subgraph S1["第一步：低爆炸半径"] A["锁定与外部系统集成极少、风险可控的产品"] end subgraph S2["第二步：小众与垂直场景"] B["替换特定行业高溢价单点工具 (如医疗 FHIR、专用文档管理)"] end subgraph S3["第三步：夺回业务主导权"] C["按自身业务逻辑重构核心功能 (摆脱外部厂商的预设框架)"] end subgraph S4["第四步：滚动替代周期"] D["锚定未来 12 个月内即将到期的续费清单"] end S1 --> S2 --> S3 --> S4 ``` * **瞄准低爆炸半径（Low Blast Radius）产品**：优先向那些在整个 IT 生态中集成极少、影响面可控的边缘 SaaS 开刀，以最低风险验证替代可行性。 * **攻克垂直小众（Esoteric）场景**：针对医疗行业的 FHIR 标准工具或特定文档管理等非常小众且昂贵的软件进行定制替代。这并非因为这些产品本身劣质，而是为了将昂贵的订阅预算省下来用于更高回报的创新。 * **摆脱供应商束缚，重塑业务逻辑**：在完成前两步积累后，直接接管核心业务功能并按照自己的方式进行自研。正如讲者所强调的：“按照你自己的方式构建业务功能，而不是遵循供应商认为你应该如何运营业务的观点（Not a vendor's opinion of how you should run your business）”，这是拉开与竞争对手差距的关键分水岭。 * **跟进即将到期的续约清单**：系统性排查未来 12 个月内面临续约的所有 SaaS 合同，将其列入自研替换的攻坚计划中。 --- ### [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) 打造智能体劳动力与推行精益架构进入“智能体劳动力”（Agentic Workforce）阶段后，企业需要将智能体视作系统中的“一等公民”（First-class Citizens）。这意味着必须为智能体赋予明确的身份标识（Identity）、操作权限（Authority），并建立完善的安全防护与治理规范。在实际落地中，智能体工作流的引入使得每个服务请求的处理时间缩短了整整两天，并在部署后的前三个月内直接挽回了 150 万美元的经济效益。在架构推行层面，讲者提出了三条核心实践准则： * **聚焦专有智能而非盲目追求前沿大模型（Focused Intelligence over Frontier Intelligence）**：充分利用企业自身的专有数据，去优化参数量更小的轻量化小模型（SLM）或垂直领域 LLM，而不是在所有场景前无脑堆砌最昂贵的顶级模型。降低底层成本分母，ROI 自然会水涨船高。 * **精益架构（Frugal Architecture）**：坚持非必要不接入。确定性的业务逻辑绝不应该交给具有不确定性的 LLM 或智能体来处理。 * **先头脑风暴再定型（Storming before Forming）**：在方案最终固化前充分试验多种路径，“尝试五个点子，挑选最终胜出的那一个（Try five ideas, pick the one that wins）”。 --- ### [17:18](https://www.youtube.com/watch?v=349T6kz-VA8&t=1038s) 业务重构与复合 ROI 的飞轮闭环 AI 转型的终极目标是实现全方位的“智能体业务变革”（Agentic Transformation），即重新思考并重构企业的运营本质。以合同管理为例，员工的角色发生了根本转变——人们不再需要耗费大量时间从零撰写合同，核心精力转向了对 AI 生成内容的审核与确认。这种深层次的变革必须让全员与一线最终用户深度参与，“变革来自于内部（Change comes from within）”，绝非仅靠外部顾问的一面之词。尽管这一阶段在时间节省、速度提升和风险规避上的收益往往难以用单一数字直接量化，但它构成了企业的决定性时刻。这种复合 ROI 的演进飞轮呈现出层层递进的良性循环： ```mermaid flowchart LR accTitle: 复合 ROI 四阶段演进飞轮 A["阶段一：智能体构建 (Agentic Build)"] -->|"释放工程师产能"| B["阶段二：颠覆传统 SaaS (Disrupting SaaS)"] B -->|"节省固定软件开销"| C["阶段三：智能体劳动力 (Agentic Workforce)"] C -->|"沉淀智能体与流程认知"| D["阶段四：重塑业务模式 (Transformation)"] D -.->|"持续反哺与放大收益"| A ``` 在演讲尾声，讲者总结道，通过自研替代 SaaS 获得的直接资金节约，可以反哺投入到智能体劳动力的建设中，进而提炼出对业务和流程的全新认知，最终推动企业的全方位智能化重塑。讲者最后邀请参会者前往 316 号展位进一步深入交流。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=349T6kz-VA8&t=4s) 用当前阶段验证过的低风险、快速回报收益，去反哺和投资下一阶段更复杂的智能化系统，形成自给自足的复利飞轮。 — Srikanth Chitipotu > [01:39](https://www.youtube.com/watch?v=349T6kz-VA8&t=99s) 虽然单个 Token 的标价看似下调，但底层分词器（Tokenizer）的设计在变，更重要的是，随着系统复杂度上升，Agent 调用的外部工具（Tool Calls）次数呈几何级数增加。 — Srikanth Chitipotu > [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) 团队过度关注“单次问题解决的速度（Speed of Resolution）”，却牺牲并忽视了最为核心的“问题解决质量（Quality of Resolution）”，最终导致客户满意度受损与隐性维护成本飙升。 — Srikanth Chitipotu > [04:13](https://www.youtube.com/watch?v=349T6kz-VA8&t=253s) 如果没有在 AI 系统接入之前建立起严格、客观的业务基线（Baseline），上线之后任何数据归因都会沦为主观臆断。 — Srikanth Chitipotu > [08:22](https://www.youtube.com/watch?v=349T6kz-VA8&t=502s) 释放出的研发速率（Velocity）是否被真正重新部署（Redeployed）。如果工程师因为 AI 节省了 30% 的时间，但这些时间没有被明确规划投向新的战略性功能研发，那么所节省的产能“到周一早上就会被日常琐事彻底稀释”，在财务报表上完全隐形。 — Srikanth Chitipotu > [13:46](https://www.youtube.com/watch?v=349T6kz-VA8&t=826s) 按照你自己的方式构建业务功能，而不是遵循供应商认为你应该如何运营业务的观点（Not a vendor's opinion of how you should run your business） — Srikanth Chitipotu > [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) 聚焦专有智能而非盲目追求前沿大模型（Focused Intelligence over Frontier Intelligence） — Srikanth Chitipotu > [15:38](https://www.youtube.com/watch?v=349T6kz-VA8&t=938s) 尝试五个点子，挑选最终胜出的那一个（Try five ideas, pick the one that wins） — Srikanth Chitipotu > [17:18](https://www.youtube.com/watch?v=349T6kz-VA8&t=1038s) 变革来自于内部（Change comes from within） — Srikanth Chitipotu ## Mentioned Srikanth Chitipotu, Sage IT, Ai4, GPT-4.5, OpenAI, Uber, Klarna, GitHub Copilot, Claude Code, FHIR ## Video description Srikanth Chitipotu - VP Engineering, Sage IT Doing AI is table stakes now. Everyone has multiple programs in production and many more in flight. The edge isn't whether you build; it's whether the returns compound. So we flip it: realize ROI now, and let each win fund the next. It starts where returns are fastest and lowest-risk - disrupting the bloated SaaS you're already overpaying for - then compounds from there, all the way to a fully agentic enterprise. No big bets, no science projects, just real value one win at a time. We've done it across multiple domains. Join us, and we'll show you how to start in yours. [Source](https://www.youtube.com/watch?v=349T6kz-VA8)

Source quotes

[0:04] Srikanth Chitipotu: 用当前阶段验证过的低风险、快速回报收益，去反哺和投资下一阶段更复杂的智能化系统，形成自给自足的复利飞轮。 [1:39] Srikanth Chitipotu: 虽然单个 Token 的标价看似下调，但底层分词器（Tokenizer）的设计在变，更重要的是，随着系统复杂度上升，Agent 调用的外部工具（Tool Calls）次数呈几何级数增加。 [4:13] Srikanth Chitipotu: 团队过度关注“单次问题解决的速度（Speed of Resolution）”，却牺牲并忽视了最为核心的“问题解决质量（Quality of Resolution）”，最终导致客户满意度受损与隐性维护成本飙升。 [4:13] Srikanth Chitipotu: 如果没有在 AI 系统接入之前建立起严格、客观的业务基线（Baseline），上线之后任何数据归因都会沦为主观臆断。 [8:22] Srikanth Chitipotu: 释放出的研发速率（Velocity）是否被真正重新部署（Redeployed）。如果工程师因为 AI 节省了 30% 的时间，但这些时间没有被明确规划投向新的战略性功能研发，那么所节省的产能“到周一早上就会被日常琐事彻底稀释”，在财务报表上完全隐形。 [13:46] Srikanth Chitipotu: 按照你自己的方式构建业务功能，而不是遵循供应商认为你应该如何运营业务的观点（Not a vendor's opinion of how you should run your business） [15:38] Srikanth Chitipotu: 聚焦专有智能而非盲目追求前沿大模型（Focused Intelligence over Frontier Intelligence） [15:38] Srikanth Chitipotu: 尝试五个点子，挑选最终胜出的那一个（Try five ideas, pick the one that wins） [17:18] Srikanth Chitipotu: 变革来自于内部（Change comes from within）

Source quotes

Srikanth Chitipotu: The edge isn't whether you build; it's whether the returns compound. So we flip it: realize ROI now, and let each win fund the next. Srikanth Chitipotu: It starts where returns are fastest and lowest-risk - disrupting the bloated SaaS you're already overpaying for - then compounds from there, all the way to a fully agentic enterprise. No big bets, no science projects, just real value one win at a time. [0:04] Srikanth Chitipotu: AI will drive immense long-term value, provided enterprises approach its implementation responsibly and measure its business impact accurately. [4:13] Srikanth Chitipotu: Without baseline operational telemetry in place before an AI deployment goes live, organizations lose the ability to accurately benchmark performance after the fact. [8:22] Srikanth Chitipotu: Simply freeing up developer hours does not create ROI on paper unless that recovered velocity is deliberately captured and redeployed onto high-value business initiatives. [8:22] Srikanth Chitipotu: Without strategic reallocation, developer time savings simply evaporate into routine slack by Monday morning. [12:20] Srikanth Chitipotu: The historical boundary between 'build vs. buy' has radically shifted: while enterprises previously bought external software licenses to avoid maintenance overhead and delivery risk, modern agentic tooling allows small engineering teams to rapidly build, customize, and maintain core business tools in-house. [15:38] Srikanth Chitipotu: Keep deterministic rules out of LLM prompts; only deploy probabilistic agentic layers where genuine reasoning or linguistic interpretation is required. [15:38] Srikanth Chitipotu: Genuine enterprise transformation must emerge internally from frontline domain experts rather than being dictated by external consultants, turning AI into a sustainable, compounding competitive advantage. [18:54] Srikanth Chitipotu: The operational capacity gained from initial agentic development is strategically repurposed to disrupt existing, expensive SaaS tools. This disruption yields direct, flat cost savings that can be immediately reinvested into building out a dedicated agentic workforce.

Original transcript

[0:04] I'm going to talk about compounding ROI with AI. Um so uh this is uh we'll we'll go through this [0:11] and there's at the end there's a playbook which you can apply uh going back and taking back. Uh so [0:18] let's get started. So here is what uh the agenda looks like. Uh before we talk the playbook, [0:23] before we talk about uh how it can be applied, so I want to level set a little bit of what [0:29] usually breaks and you know what what happens, how the how the calculation should be made, [0:34] uh what is the way that we have used and we have used at multiple organizations to measure this [0:39] and uh then you know we'll we'll run through the playbook after that. And I want to leave you all [0:44] with a slide that shows you uh how how it would uh how one stage uh you know kind of compounds [0:50] on the other. So it's not really just taking um ROI from one to another and applying it, [0:56] it actually does compound. So I want to show you all that. Uh so before we get started, I just want [1:01] to quickly introduce myself. So vice president of engineering, uh I've been with Sage ID for [1:05] a few years now. I've got career across multiple places before that. Uh fully in the technology, [1:12] um you know we I want to add that uh last line there just to um you know highlight that Sage ID [1:20] as a company is a believer. I know there is a there's contrarian view of AI as a disruption. [1:27] Uh it's impacting but but we are we strongly believe AI is something that is going to help um [1:33] in the long run. Uh and and we all should be responsible for that. Um so yeah, let's let's talk [1:39] ROI um first thing. So uh you know the the the first is this is a little bit controversial, so [1:48] uh we've all seen the model costs go down uh to some degree, but what what we are not realizing [1:54] it is the tokenizers behind are changing. The number of tool calls are changing. Like which [2:00] ultimately means that your token cost, your cost to run the same use case is increasing. So you got [2:08] to factor this in. So that's one of the key thing and you know GPT-5.1 5 released 23rd April $5 [2:17] $30 per million tokens. That's double 5.4. So that's just an example not not to say anything [2:23] bad about them but it's it's just like it's underlying cost that you wouldn't see so you [2:28] shouldn't just jump right into the newer frontier models. So we should pace ourselves depending on [2:33] you know what what what is the benefit that you'd get with that. So build and run cost so this is [2:37] very key thing that we've noticed. So build time kind of you know gets out of hand. We've seen [2:44] good examples of this. I think Uber is the one that announced that they've run through their [2:49] tokens in 3 months 6 months. So we've seen that at some of the organizations that we've worked with. [2:53] The run time cost is another thing that is hidden. So you have to factor it in. The more [3:00] powerful your use case is the more people are going to start using it. So you'll have to factor [3:06] that in. Like it's it's not just a matter of oh people used to use it three times but now once [3:10] with AI they will see better use of it which means that you should expect the five times six time [3:15] usage of it. So that's that needs to be factored in as well. So we've seen that common bites of [3:21] customer service example I think which is very well known in in one of our deployments we've [3:27] had to take that back down because you know it it it did not make economic sense. The ROI [3:33] didn't add up properly. So that is one. The next one is and this is a little bit tough to measure. [3:39] So effectiveness and and what what we mean by that is how well the AI is actually going to be [3:46] effective in in coming up with the use case and I I there are a few things that are specifically [3:52] called out like you know we we've seen this with um uh any advertised use case out there. [4:00] AI spend and how much it costs to build AI is clearly advertised, but what did really [4:07] uh you know improve at the end of the day? That's a little bit tough to measure. So, so I want to [4:13] uh move on to how or where calculations break. So, exactly the effectiveness uh that we were looking [4:19] at. So, that's the wrong metric. That's that's not measured properly. If not measured properly, [4:24] you're measuring the wrong thing. So, this is a very well-known example as well. So, Klarna [4:29] um is is the company for this. I don't want to call that out on the slides here, but [4:34] uh you know they've you know famously advertised in 2024 uh February saying [4:39] that okay they're replacing 700 of the people with uh you know the the OpenAI collaboration, [4:44] but a year after they had to pivot back and bring uh the humans back. And that's because and and [4:51] this is quoted by their CEO as well. They're measuring the wrong thing. They're measuring uh [4:56] how quickly they're achieving the resolution, but not how uh what is the quality of resolution. So, [5:02] that's that's important to measure as well. Um and then the next one is uh no instrumentation before [5:09] deployment. This is again another famous example. Uh if you if we don't have instrumentation done [5:14] before um we it's going to be very tough to quantify because it's all it's always going to [5:20] look uh very positive if if that is not properly factored in. So, we've we've seen that happen. [5:26] Uh you cannot go back like once once you have deployed AI uh in production in front of the use [5:31] cases in front of your customers, you cannot go back and take it out and and then measure [5:36] what it would look like once you have AI. So, you need to factor this in as well. Um [5:42] and then finally I think this was brought up by uh so some of the uh you know leaders when they were [5:47] talking the keynote as well. So, restrictions, regulations that is outside of our control. So, [5:51] we have to figure out, okay, what does that mean if someone stops me using from whether [5:56] AI in this whole use case or whether this model in this use case. So, so we need to definitely factor [6:02] that in. It's a couple of I've just listed couple of examples. I mean, I'm sure there are others. [6:06] Um so, yeah, this usually takes years if if it is in a regulated industry, uh but if you are [6:12] if you need to change your model altogether to something open weights, for example, [6:17] then that's a rewrite of your use case. So, so you need to factor that in. So, [6:21] uh just, you know, that kind of level sets us where they would break. All right. Now, [6:26] let's look at how do we measure ROI and and this is something that we've done for for majority of [6:32] our use cases. So, I I want to just um use this as a way to guide uh how the playbook would work. [6:40] Um so, there is five metrics here. Although in the equation, there there is only four. [6:45] Uh the last one is effectiveness, which we'll talk about in a minute. Uh the the [6:50] the few that are there, capacity release, cost save, and time save. So, those are your benefits. [6:55] So, what does what does AI give you? Whether whether it can give you capacity, whether it [6:58] can give you cost, or it will save you time doing whatever, right? And this doesn't point only to [7:06] uh AI for engineers. This this is general uh AI when it is applied, it is going to save you money [7:12] with with doing one of these. And there is risk aversion as well that is factored in the last [7:16] metric. Um then then the denominator is all total cost. So, what does that mean? So, it is cost of [7:23] tokens, infrastructure, build, and human as well. So, we also have to account for [7:29] bot sitting, which is what the term is called now. So, that that that would mean human in the loop, [7:34] that would mean your engineer spending time to actually uh guide the AI to achieve the outcome. [7:41] So, that is needed as well, and then obviously maintenance and licensing licensing and all that. [7:45] Um so, now the fifth and important metric, which is effectiveness, [7:49] that is very tough to measure. Um what we've done is we've started measuring it [7:56] for as an outcome observability is what we call it. So, instead of looking at LLM traces, instead [8:01] of looking at what is happening in a specific interaction, we kind of started looking at [8:07] a little bit superficial and looking at use cases well. So, for a given use case, [8:12] what is exactly happening through the stack, right? So, so that's that's really what what we [8:17] um try to point out. So, that's that's what we call as a outcome observability. [8:22] Now, let's get to the playbook. Um so, the main thing here is there is it's a it's a [8:31] um incremental stage incremental stages that we go through. So, this is stage one. So, agentic build, [8:38] I think we've all done this uh I I and I think majority of the organizations are already here. [8:43] It's stable stakes now. If not, I think we should talk why what what's stopping you. Uh so, this is [8:50] letting your engineers letting your whole whole dev dev team uh start applying copilots and and [8:56] all that. So, we also have to make sure that velocity is released and redeployed, and that's [9:02] the key thing. So, without redeployment, it's it's a it's a loss. Like, it's it's it won't be [9:07] visible. It will just get absorbed by you know, by Monday morning. So, you have to make sure [9:12] it gets redeployed. So, that is the key thing. So, how did it look like in practice and what is [9:17] the ROI look like? So, I've included an example. Obviously, I I I didn't quote the customer, so we [9:22] can talk in the booth if if there is an interest of what it is. Uh but, we've we're able to achieve [9:29] 60% of productivity, and that translated to the ROI uh math that is shown here. So, it released [9:35] capacity. Uh this is 60% of 15 engineers is what we are quoting here. I mean, this is just one [9:42] deployment. This is multiple others that we've done where it changed quite quite a quite a bit. [9:48] So, what is the effectiveness here, right? So, as I said, this is not that easy to measure. But what [9:53] we ultimately observed is that the the product that is released is is much more defect free, [10:00] right? So, that's one of the key metric that we've encountered. Now, moving on to the next So, before [10:07] moving on to the next one, I want to get something to show how it would how you would approach this [10:13] and what are the different steps to follow while while you apply this. I am I think there is [10:19] four steps here. I think it can be debated whether this is four or three, but regardless, [10:24] like first one is we enable Copilots, Claude codes, whatever is that you prefer to use. I think [10:31] that needs to be enabled. So, that's that's stable stakes. That's done already, I'm assuming. Now, [10:36] now comes the agents doing majority of the work. And this is where we let agents do 80 90% of it [10:44] and we are kind of reviewing, kind of making sure that it is producing the right outcome. [10:49] And majority of the organizations are already here. And and I think some are moving past this. [10:55] So, the next comes the agents orchestrating it. So, what that means is and we've started [11:01] piloting with couple of our organizations that we work very closely with and we're seeing some [11:05] very good traction. So, I was talking to a few folks here as well. So, what we do is [11:10] we call it as a conductor, for example, which uses the frontier intelligence and it divides it [11:17] the work into three or four agents, which does not need frontier models, which would need like a [11:24] cheaper model cheaper versions of the models to do minimal work. And then they come back to the [11:28] frontier conductor, which does the evaluation, which makes sure that the work is accurate and all [11:34] that. So, that is how agents like you you can do three times more productive even the productivity [11:40] that is gained by enabling agents and co-pilots is one thing, but if you are able to do this, you are [11:47] getting the three times four times more productivity from each of your engineers. [11:52] And the last step is agents autonomy. I think it's a little bit far away and some of the probably [11:58] very well advanced organizations might be doing it already and and there is good [12:04] product companies that are attempting this as well that we we talked we kind of talked to all the [12:08] time. So, so that's kind of it. Like if you are looking at this agentic build as the first step. [12:14] So, what what this gives you is it gives you engineering capacity that you need to redeploy. [12:20] So, that's the key takeaway here. And [snorts] how do we redeploy it? And this is this is stage two. [12:25] So, once the engineering capacity is recovered like we have to look at [12:30] what are the SaaS products that are being that you are paying today? And and I think this is [12:37] very well underserved, but we are seeing huge traction with a lot of conversations that we have [12:44] with all organizations that we work with. So, there is a lot of potential in terms of [12:49] SaaS products that you don't really need, but you have been carrying over the years because [12:55] that is something that was added into your organization a while back. So, that is where [13:01] that that is what this means. So, the build and buy line has kind of moved a little bit. So, [13:06] earlier if an organization is debating okay, whether I should buy or whether I should build, [13:11] they usually tend towards buying it. That's that's because the risk aversion is there [13:16] and you don't want to take that risk of okay, building this and then maintaining it. But now [13:21] since the lines have moved, so we are seeing more and more that organizations are attempting [13:25] to build and and talk about how we need to do that. So, what did this just quoted few example [13:34] SaaS products out there which are seeing you know down down trend in terms of losing their value [13:40] and and kind of is reflective of that. What did it mean in practice and what is ROI we looked at? So, [13:46] we replaced the CRM and the map there is kind of showing you what what it achieved. And [13:54] the key takeaway again the effectiveness is the measurement of it's it's not just replacing that [14:04] and and leave leaving at that. So, we also make sure that there is no manual fallback on that. [14:12] So, here is a playbook that looks like how to approach this. So, we start off with the low [14:17] blast radius products. The products that have very minimal integrations in rest of your ecosystem. [14:24] So, that is first like go after them first. Then esoteric. So, there is examples like you know fire [14:32] in the health care industry for example. There are examples for document management which are [14:37] very very niche, very very peculiar in their use case. You you need to go after them. Not because [14:43] they are bad but because you better you utilize that money for something else. So, that's another. [14:50] And then the third is I think once you've done these two steps you're ready for [14:54] taking the business functionality and building it your own way. Not a vendor's opinion of how [15:01] you should run your business. And this is key. This is going to put you on the track to kind of [15:07] be ahead of your competition. Right. And then lastly, I mean once you've [15:11] done this obviously you'd be looking at okay what else can I replace. Right. So, [15:14] I really don't have to or we really don't have to tell you but you'd be looking at [15:18] okay what is coming up for renewal in the next 12 months. Let let me go after that. Right, [15:22] so this kind of is what we have seen and and there is organizations already doing this. As I said, [15:30] really very underserved. Like we should definitely look at this. Right, now let let's cover [15:38] the next stage which is agentic workforce. This is agents as as a as first class citizens. So we [15:45] should be looking at how do you make agents run in your organization? So give them an identity, [15:51] authority, make sure they are secure, make sure they are governed and all that, but make sure the [15:55] agentic workflows is deployed. And you know, in in practice and ROI, we are able to deploy it and [16:02] recover two days per service request. And that translated to 1.5 million [16:07] recovered in the first three months of deployment. And then the [16:11] again, effectiveness is mentioned here, so I'm going to skip ahead. So in terms of applying this [16:16] playbook, the stages, I think this is one of the key thing that I want to give out here is a [16:22] focused intelligence over frontier intelligence. So if if you look at SLMs, domain specific LLMs, [16:28] I think this is where you need to take organizational data that you have and make sure [16:33] you optimize a lower weight model, lower parameter model than just putting it your the the best model [16:41] out there in front of everything. So that is you know, that that that needs to be factored into [16:45] your ROI. So the the bottom denominator, if you reduce your cost, your ROI automatically goes up, [16:51] right? And then frugal architecture. So you should not deploy agents unless you know that there is [16:56] a need for it. Like your deterministic logic shouldn't live in with an agent, with an LLM. [17:01] Storming before forming and again, this is one thing that we kind of brainstorm, make sure [17:06] all of our organizations do this. So try five ideas, pick the one that wins. [17:12] And ultimately creating for the business. So the last stage here is agentic transformation. So, [17:18] this is rethinking how your business is run. Since we've done enough on building agents and adding [17:25] agentic layers, uh so now we are able to take it to the next level, kind of reimagine the process. [17:31] So, what for one example is you are no longer looking at contracts or you are no longer working [17:36] on writing contracts, for example. Your main focus and main work now is reading through the one that [17:43] AI has generated. Right? So, that's an example. So, we we've done this. So, again, the make [17:48] sure you are including your um employees and and users uh and your um people in the transformation. [17:56] Uh it's not someone like me that comes in and tells you how to do your business. So, so again, [18:00] the there is question marks in there in in the ROI math here. And the reason for that is it's [18:05] not quantifiable. It's not so easy to quantify how much uh you know, what what's the time saved, [18:11] what is the uh speed gained, or what is the risk averse. So, that is something that you would know [18:16] for your organizations. You should factor that in and calculate it. So, how do you get through this [18:20] transformation is these four stages. So, this is I I I would say consider this your defining [18:26] moment. So, if your competitor is working on something that that you are in, what do they [18:32] don't want to build or what do they want to bypass altogether? Right? So, that that is that is the [18:37] uh for one thing you to you need to look at. Right? And think bigger, and then last, [18:42] don't leave behind. It should change comes from within is is what we call it. Like transformation [18:48] um should al- always be part of, you know, um your your organization, your people. So, they should be [18:54] included. And here is how the stages compound. Um as I said, you know, the the the the capacity [19:00] gained from agentic build is uh repurposed for disrupting your SaaS, which will give you flat [19:08] um cost savings, which can be applied to building agentic workforce. And then you will learn about [19:13] how the agents and how the processes are, which can help you transform your business. [19:19] So, I think I only have 20 seconds left, but I'm here and we are also in booth 316. If you have any [19:26] questions, if you want to talk to me, please do. So, that's all. Thank you. Thank you. [applause]