Freshworks Unveils Mission Control Frame to Calibrate AI Agency

Mission Control - Giving AI Agency (Without Losing Control) With Freshworks — Ai4

Ai4

Autonomous AI handles routine IT tasks to cut ticket resolution times by 76% while preserving human strategic oversight.

AI-assisted summary of Mission Control - Giving AI Agency (Without Losing Control) With Freshworks — 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 Payal Patel, Senior Director of Solution Engineering for North America at Freshworks. According to the page description, the session explores how service operations can adopt a "Mission Control" architecture to grant AI genuine agency for autonomous execution while keeping human leaders in strategic control.

## TL;DR Payal Patel addresses the binary mindset trap that stalls enterprise AI initiatives—either rushing deployment without governance or delaying value out of excessive caution. To bridge this gap, she introduces Freshworks' 3-Layer Mission Control Framework, which ranges from fully autonomous execution of routine tickets (Layer 1) to AI-assisted copilot workflows (Layer 2) and human-led strategic intelligence (Layer 3). Real-world implementations, such as Manchester Metropolitan University, demonstrate how calibrating agency achieves a 65% ticket deflection rate and a 76% faster resolution time without sacrificing operational control.

## Atmosphere & style The presentation opens with a warm, conversational, and energetic vibe as Payal Patel jokes about being the final speaker standing between the audience and evening festivities after winning two thousand dollars at craps. Her delivery is pragmatic, authoritative, and deeply grounded in executive-level enterprise realities, engaging the room through interactive polls that reveal shared industry anxieties around AI autonomy. The narrative arc smoothly transitions from lighthearted banter to structured strategic insights and concrete metrics, offering listeners a reassuring, clear, and actionable path forward.

## Key takeaways - [00:08](https://www.youtube.com/watch?v=kU20tH50Yb8&t=8s) Organizations are shifting away from debating whether to use AI and are now focused on calibrating AI agency and defining boundaries for human intervention. - [01:41](https://www.youtube.com/watch?v=kU20tH50Yb8&t=101s) AI agency is defined as an AI system's capability to act, execute tasks, and make end-to-end decisions without requiring human approval at every step. - [03:47](https://www.youtube.com/watch?v=kU20tH50Yb8&t=227s) Successful AI deployment fails when organizations fall into the binary mindset trap of either uncontrolled speed without governance or excessive caution that stalls pilots. - [03:47](https://www.youtube.com/watch?v=kU20tH50Yb8&t=227s) Overcoming speed-first AI failures requires addressing three key pillars: establishing clear governance and ownership, measuring pre- and post-implementation outcomes, and preventing multi-year rollout complexity. - [07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) Excessive caution and operational friction cost a typical 500-person organization $2.5 million annually, with employees losing 7 hours per week searching across disconnected systems. - [10:01](https://www.youtube.com/watch?v=kU20tH50Yb8&t=601s) The Mission Control Framework structures AI agency into three distinct operational layers: Autonomous AI Foundation, Agency Assistance (Copilot), and Strategic Intelligence. - [16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) Layer 1 (Autonomous AI Foundation) resolves routine, low-risk requests like password resets and policy inquiries end-to-end with zero human intervention. - [16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) Layer 2 (Agency Assistance) aids human agents during complex troubleshooting by summarizing context, drafting responses, and recommending solutions for 1-click execution. - [16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) Layer 3 (Strategic Intelligence) transforms raw operational logs into proactive insights, surfacing performance anomalies and root causes to assist strategic decision-making. - [20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) Calibrating AI agency across all three layers allows organizations to achieve benchmarks such as a 65% ticket deflection rate and 76% faster resolution time. - [21:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1297s) Manchester Metropolitan University successfully implemented the 3-layer Mission Control model to support 22,000 students and staff, balancing automation speed with human control.

## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=kU20tH50Yb8&t=8s) Introduction, Vegas Wins, and the Evolution of the AI Conversation

Payal Patel, Senior Director of Solution Engineering for North America at Freshworks, kicks off the session by acknowledging her slot as the final presentation of the day. She jokes with the audience about holding them back from drinks and gambling, boasting that she won two thousand dollars at the craps table the previous night and inviting attendees to join her there later. Transitioning to the core topic, Patel notes how dramatically the AI conversation has shifted over the past year. Organizations are no longer debating whether AI can help; rather, having experienced its initial potential, they are now grappling with how much AI should actually do, how much value can be extracted from it, and where human intervention must remain in control.

Before diving into the core framework, Patel takes a quick poll of the room to gauge the current state of AI adoption among attendees:

* **Current AI Deployment:** She asks how many in the room already have AI running in production, receiving a strong show of hands. * **Autonomy Comfort Level:** When asking those same individuals to keep their hands up if they feel completely comfortable with the level of autonomy granted to their AI, almost every hand drops.

This drop in confidence highlights the exact problem the session addresses: introducing "Mission Control," a operational framework designed to grant AI sufficient agency to deliver real value while ensuring human leaders retain control over critical decision-making.

[01:41](https://www.youtube.com/watch?v=kU20tH50Yb8&t=101s) The Real Challenge: Calibrating AI Agency vs. Control

Patel introduces herself and her role at Freshworks, leading the solution engineering team across North America. Her team works closely with C-level executives—specifically CIOs and CTOs—to integrate AI across their service operations and broader software portfolios. Through these executive conversations, Patel observes that the primary roadblock is rarely the underlying capability of the technology. Instead, the central challenge is determining where AI should be permitted to act independently versus where it should assist human decision-making so that leaders maintain necessary oversight.

She argues that organizations have been asking the wrong questions. The early AI dialogue centered on whether to use AI at all, but most enterprises have already moved past basic experimentation by embedding copilots and AI workflows into daily operations. The real question today is how much organizational responsibility and authority should be delegated to AI. Patel defines **AI agency** as the system's ability to act, execute tasks, and make decisions without awaiting human intervention. The companies seeing the highest return on investment are those being deliberately intentional about defining these boundaries—using AI autonomously for repetitive, well-defined inquiries so human teams can focus on complex, high-judgment scenarios.

[03:47](https://www.youtube.com/watch?v=kU20tH50Yb8&t=227s) The Binary Mindset Trap: Speed Without Strategy vs. Paralysis by Caution

Patel identifies the current primary barrier to successful AI adoption not as technology, but as a flawed, binary mindset. Organizations tend to fall into one of two extreme traps:

1. **Uncontrolled Speed:** Driven by intense pressure to automate, some organizations deploy AI rapidly without establishing clear governance, ownership, or accountability. This rush results in inconsistent user experiences, fractured trust, and unclear business outcomes. 2. **Excessive Caution:** Fearing security or operational risks, other organizations move too slowly. They keep AI confined to small-scale pilots that never reach enterprise scale, delaying value creation out of a false sense of safety.

``` ┌───────────────────────────────────────────────────────────────────┐ │ THE BINARY MINDSET TRAP │ ├───────────────────────────────────┬───────────────────────────────┤ │ UNCONTROLLED SPEED (RUSH) │ EXCESSIVE CAUTION (DELAY) │ ├───────────────────────────────────┼───────────────────────────────┤ │ • Moves fast without control │ • Moves slow & falls behind │ │ • Unclear governance & ownership │ • Pilots fail to reach scale │ │ • Inconsistent experiences │ • Excessive risk aversion │ └───────────────────────────────────┴───────────────────────────────┘ │ ▼ ┌───────────────────────────────────────────────────────────────────┐ │ MISSION CONTROL FRAMEWORK │ │ Intentionally calibrates where AI acts autonomously │ │ and where humans remain in strategic control. │ └───────────────────────────────────────────────────────────────────┘ ```

Conducting another audience poll, Patel asks how many are actively trying to expand AI into more business processes, receiving widespread agreement. However, when she asks how many have successfully secured broad organizational approval and adoption for those expanded processes, only two hands go up—and notably, not the same hands that indicated comfort with autonomy earlier. Patel stresses that success isn't about how fast a company moves, but how intentionally it drives real adoption. Buying AI software yields no value if the organization fails to adopt it effectively.

- **The Three Pillars of AI Failure in Speed-First Approaches:** - **Governance & Ownership:** Initiatives stalled by a lack of clear operating models and aligned stakeholders. - **Measuring Business Outcomes:** Failure to define and track pre- and post-implementation metrics to verify if the AI delivers its intended value. - **Implementation Complexity:** Multi-year legacy-style rollout cycles that kill momentum; modern AI implementations should never take a full year to execute.

[07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) The Heavy Cost of Complexity and Delay

Turning to the other extreme, Patel illustrates the severe financial and operational costs borne by companies that hesitate to adopt AI. Staying cautious does not mean standing still—it means falling behind in operational speed. In service operations, delaying automation leaves employees drowning in routine, low-value inquiries such as password resets or basic employee benefit questions.

To quantify this, Patel shares concrete benchmark statistics regarding the price of operational friction in enterprise environments:

- **$2.5 Million Annual Cost:** The estimated annual cost of operational complexity for a typical 500-person organization. - **7 Hours Lost Per Week:** The average amount of time an employee wastes every single week searching across disconnected systems, toggling between tools, and navigating manual workflows. - **Opportunity Cost of Fragmentation:** Disjointed systems slow down execution, whereas tool consolidation eliminates friction and enables faster business growth.

Patel reinforces this point by quoting Freshworks CEO Dennis Woodside: "Complexity is a choice to grow slower than you should."

[10:01](https://www.youtube.com/watch?v=kU20tH50Yb8&t=601s) Defining AI Agency and the 3-Layer Mission Control Framework

To establish a path forward, Patel formalizes the definition of **AI Agency**: the ability of an AI system to make decisions and complete entire workflows within predefined boundaries without requesting human approval at every step. This differs from simple automation because agentic systems operate autonomously to finish tasks end-to-end—such as verifying employee PTO balances, submitting time-off requests, and notifying managers automatically.

She breaks down the spectrum of AI involvement across operations into a structured 3-Layer Mission Control Architecture:

``` ┌───────────────────────────────────────────────────────────────────┐ │ LAYER 3: STRATEGIC INTELLIGENCE LAYER │ │ Humans lead strategy; AI surfaces trends, systemic risk, │ │ and proactive operational insights. │ ├───────────────────────────────────────────────────────────────────┤ │ LAYER 2: AGENCY ASSISTANCE LAYER (COPILOT) │ │ AI assists humans; gathers context, drafts responses, │ │ and recommends actions for 1-click execution. │ ├───────────────────────────────────────────────────────────────────┤ │ LAYER 1: AI FOUNDATION LAYER (AUTONOMOUS) │ │ AI acts independently; resolves routine, repeatable, │ │ low-risk requests with zero human intervention. │ └───────────────────────────────────────────────────────────────────┘ ```

- **Layer 1 (AI Foundation):** Fully autonomous action for repeatable, low-risk requests (e.g., password resets, policy Q&A, standard service requests). - **Layer 2 (Agency Assistance):** Human-assisted execution where AI copilots summarize ticket histories, recommend next steps, and draft responses for quick human approval. - **Layer 3 (Strategic Intelligence):** Human-led strategic oversight where AI analyzes underlying system data to surface operational trends, anomalies, and proactive recommendations.

Polling the audience on where their organizations currently sit, Patel finds that the vast majority are at Stage A (AI assisting while humans do most of the work), only one person is at Stage B (autonomous resolution), and none are at Stage C (full agentic maturity). She reassures the audience that jumping to full autonomy overnight is a mistake; the goal is to systematically match specific workflows to their appropriate level of agency.

- **Freshworks Company Benchmarks:** - **Founded:** 2010 - **Scale:** Public company with over 75,000 customers and 4,500+ employees. - **Financial Metrics:** Greater than $810 million in annual revenue guidance. - **Architecture:** Unified platform anchored by a central CMDB (Configuration Management Database) mapping people, assets, and service interdependencies to supply Freddy AI with full context.

[16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) Mission Control in Action: Layers 1, 2, and 3 Detailed

Patel details how the three layers of the Mission Control framework operate in real-world IT and service management scenarios using Freshworks' Freddy AI:

#### Layer 1: Autonomous AI Foundation At this layer, AI agents resolve routine issues end-to-end across channels like Microsoft Teams, Slack, email, or self-service portals with zero human intervention. - *Example 1 (Password Resets):* An employee locked out at 2:00 AM interacts with Freddy AI. The AI authenticates the employee's identity, enforces organizational security policies, and executes the password reset automatically. - *Example 2 (Contextual Knowledge & HR Inquiries):* For requests like parental leave policies, Freddy AI searches across integrated content repositories (Google Drive, Confluence, SharePoint), provides tailored answers, prompts for necessary follow-ups, and cites exact source documentation to maintain trust. - *Multi-lingual Support:* Resolves requests natively in the user's preferred language across global workforces.

#### Layer 2: Agency Assistance (Copilot) Designed for complex troubleshooting where human judgment is required, Layer 2 focuses on agent velocity and guidance inside the service desk UI: - *Instant Context & Drafting:* Upon opening a ticket, Freddy AI digests the history, surfaces relevant knowledge articles, drafts suggested responses, and sets initial categorization and routing. - *Similar Ticket Recommendations:* It identifies historically resolved tickets with similar symptoms, presenting solutions directly to the agent so they don't have to search manually. - *Continuous Feedback Loop:* It analyzes recurring agent actions to suggest new knowledge base additions and process improvements.

#### Layer 3: Strategic Intelligence Layer 3 solves the data clarity problem faced by CIOs by transforming raw operational logs into actionable executive guidance. Rather than requiring leaders to run and analyze complex reports manually, the AI proactively surfaces critical trends and root causes.

Patel displays a live UI example illustrating Layer 3 intelligence: - *Anomalies & Root Cause Detection:* The interface alerts leadership that the average resolution time across the organization has unexpectedly spiked by 200%. Instead of leaving leaders to investigate, Freddy AI automatically identifies that the spike is specifically driven by network and hardware tickets within the IT Operations team, providing an immediate root-cause breakdown.

[20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) IT Operations Root Cause Analysis and Performance Benchmarks

Payal Patel shifts focus to the operational intelligence layer, explaining how leaders can utilize AI-driven insights to investigate root causes behind system anomalies. When IT operations experience sudden disruptions, the framework gives leaders the visibility to interrogate the environment—allowing them to isolate whether a ticket spike stems from a recent infrastructure deployment or a specific underlying systemic fault. This capability ensures that instead of treating symptoms, teams can analyze the precise structural changes driving performance degradation and resolve operational friction at the source.

Patel highlights key performance benchmarks achieved by Freshworks customers who have deployed this three-layered AI architecture. By granting AI calibrated agency, organizations are driving massive gains across key operational metrics while simultaneously freeing up internal teams from repetitive manual workloads.

- **76% Faster Resolution Time:** Significant reduction in mean time to resolve (MTTR) across complex service requests. - **65% Ticket Deflection Rate:** Average autonomous resolution of routine tickets via Layer 1 self-service agents without human intervention. - **Agent Capacity Saved:** Substantial recovery of working hours, allowing support teams to focus on high-value, complex edge cases.

[21:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1297s) Real-World Success: Manchester Metropolitan University

To illustrate the practical impact of the Mission Control model, Patel shares a final case study featuring Manchester Metropolitan University. Tasked with supporting approximately 22,000 students and staff, the university faced rapidly increasing ticket volumes that strained traditional support operations. By operationalizing the three layers of agency, the institution successfully balanced automation speed with human control:

``` +-----------------------------------------------------------------+ | MANCHESTER METROPOLITAN UNIVERSITY | | (22,000 Students & Staff) | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | LAYER 1: Freddy AI Agent | | Drives self-service resolution for students & employees | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | LAYER 2: Freddy Copilot | | Assists agents to resolve escalated tickets faster | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | LAYER 3: AI Insights | | Equips leaders with data to make strategic operational choices | +-----------------------------------------------------------------+ ```

Patel wraps up the presentation by thanking the audience, opening the floor to audience Q&A, and inviting attendees to continue the discussion at the Freshworks booth located in the main expo hall.

## Notable quotes > [07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) Complexity is a choice to grow slower than you should. — Dennis Woodside

## Mentioned Ai4, Payal Patel, Freshworks, NASA, Apollo 13, Dennis Woodside, Freddy AI, Freddy Copilot, Microsoft Teams, Slack, Google Drive, Confluence, SharePoint, Manchester Metropolitan University

## Video description Payal Patel - Senior Director, Solution Engineering, North America, Freshworks

The AI conversation has become uncomfortably binary: either hand everything to AI and hope for the best, or resist automation entirely and fall behind. But the organizations achieving real results are doing something different—they're building Mission Control architectures where AI has genuine agency to act autonomously, while humans maintain strategic oversight. This is the lesson from NASA's Mission Control: autonomous systems handled thousands of micro-decisions during Apollo missions, but flight directors always knew what was happening and could intervene when judgment mattered. The spacecraft didn't ask permission to adjust trajectory for every solar wind fluctuation—that would be absurd.

But when an oxygen tank exploded on Apollo 13, human expertise took over immediately. Service operations need the same architecture. AI agents that deflect 65% of routine tickets without human review. Copilots that suggest responses agents can accept with one click—or override when context demands it. Insights that surface trends proactively but let leaders decide what action to take. This is AI with agency: sophisticated enough to act independently, architected carefully enough that you never lose control. Through real examples and the Mission Control framework, we'll explore how to calibrate agency correctly—knowing when AI should act autonomously (password resets, routine requests, policy questions), when it should assist human decision-making (complex troubleshooting, escalations, edge cases), and when humans must stay in full control (strategic decisions, sensitive situations, novel problems). The result is service operations where AI handles exponentially more work while leaders feel more in control, not less.

Key takeaways are, the agency paradox: why giving AI more autonomy (in the right places) actually increases human control over outcomes. Mission Control architecture: the three layers of agency (autonomous action, assisted decisions, human-led strategy). Calibrating agency correctly: frameworks for deciding what AI handles independently vs. where humans stay in the loop. Real benchmarks from organizations getting agency right: 65% autonomous deflection, 76% faster resolution, zero loss of quality or control. Governance without bureaucracy: how to build oversight that enables speed rather than blocking it. The agentic AI maturity curve: progressing from AI-assisted to AI-autonomous operations while maintaining strategic control

[Source](https://www.youtube.com/watch?v=kU20tH50Yb8)

中文简报

Freshworks推出任务控制框架以校准AI自主权自主AI处理日常IT任务将工单解决时间缩短76%，同时保留人类的战略监督。 Payal Patel指出了阻碍企业AI项目推进的二元思维陷阱——要么在缺乏治理的情况下仓促部署，要么因过度谨慎而延误价值实现。为了弥合这一差距，她介绍了Freshworks的三层任务控制框架（3-Layer Mission Control Framework），涵盖从日常工单的全自动执行（第1层）、AI辅助的副驾驶工作流（第2层）到人类主导的战略情报（第3层）。Manchester Metropolitan University等实际应用案例表明，在不牺牲运营控制权的前提下，校准自主权可实现65%的工单分流率，并将解决时间缩短76%。各机构正从争论是否使用AI转向关注校准AI自主权并界定人类介入的边界。 (0:08) AI自主权被定义为AI系统在无需每一步都获得人类批准的情况下，采取行动、执行任务并做出端到端决策的能力。 (1:41) 当机构陷入二元思维陷阱时，AI部署便难以成功——要么是缺乏治理的失控速度，要么是导致试点停滞的过度谨慎。 (3:47) 克服“速度优先”导致的AI失败需要解决三大支柱问题：建立明确的治理与权责归属、衡量实施前后的成效，以及防止长达数年的推广复杂性。 (3:47) 过度谨慎和运营摩擦每年会给典型的500人规模机构造成250万美元的损失，员工每周在互不相通的系统间进行搜索会耗费7小时。 (7:56)

Reading summary

Freshworks' Payal Patel argues that enterprise AI adoption is hindered not by technological limitations, but by an uncalibrated binary choice between reckless deployment and paralyzing risk aversion. To realize measurable returns, organizations must implement a multi-layered governance model that grants AI autonomous execution over routine tasks while anchoring strategic control with human operators. The False Binary Between Speed and Paralysis (3:47) Patel observes that enterprise leaders frequently succumb to one of two operational extremes when deploying AI. Some organizations accelerate implementations without establishing governance, clear accountability, or pre- and post-deployment success metrics, resulting in fragmented user trust and disjointed workflows. Conversely, risk-averse organizations stall initiatives within protracted pilot phases, fearing operational disruption while failing to reach enterprise scale. Audience polling during the session mirrored this tension: while many enterprises run AI in production, nearly all participants acknowledged discomfort with granting autonomous authority to these systems. Patel warns that hesitation incurs steep financial and productivity penalties. Drawing on enterprise benchmarks, she estimates that operational friction costs a typical 500-person organization $2.5 million annually, with workers losing an average of seven hours per week navigating disconnected tools and legacy manual processes. According to Patel, modern AI rollouts should avoid multi-year implementation cycles, focusing instead on targeted process adoption that directly reduces systemic friction. The Three-Layer Architecture for Calibrated Agency (10:01) To bridge the gap between human oversight and autonomous action, Patel defines AI agency as the capacity of a system to execute entire workflows end-to-end within predefined boundaries without requiring manual step-by-step approval. Rather than treating autonomy as an all-or-nothing switch, Freshworks structures deployment across a three-layer 'Mission Control' operational model anchored in a unified Configuration Management Database (CMDB) that maps dependencies across personnel and technical assets. The framework stratifies responsibilities into distinct operational tiers: Layer 1 functions as an autonomous foundation handling repeatable, low-risk requests independently; Layer 2 serves as an agency assistance copilot that summarizes context and drafts actions for human approval; and Layer 3 acts as strategic intelligence, where AI analyzes system-wide operational data to surface trends and operational anomalies for executive leadership. Operational Workflows from Self-Service to Intelligence (16:14) In practical enterprise IT and service operations, Layer 1 executes routine transactions—such as automated off-hours password resets enforcing security authentication or answering leave inquiries by indexing repositories like Confluence, SharePoint, and Google Drive with strict source citations. Layer 2 accelerates internal service desk capacity by automatically ingesting ticket histories, proposing categorized resolutions, and recommending precedent solutions from previously resolved tickets to reduce manual agent research. At the executive tier, Layer 3 addresses operational opacity by converting raw service logs into diagnostic insights. Rather than requiring administrators to assemble manual reporting, the system autonomously surfaces anomalies and pinpoints root causes. Patel demonstrated this via an operational scenario where a 200% surge in organizational ticket resolution time was automatically diagnosed and attributed directly to underlying hardware and network issues within the IT Operations team. Performance Benchmarks and Real-World Validation (20:54) Patel asserts that deploying this calibrated agency model yields significant operational efficiency gains. Freshworks reports that organizations implementing the three-layered architecture achieve an average ticket deflection rate of 65% through Layer 1 autonomous resolution, alongside a 76% reduction in mean time to resolve (MTTR) across escalated, human-assisted service requests. To demonstrate execution at scale, Patel cited Manchester Metropolitan University, an institution supporting roughly 22,000 students and staff. By routing high-volume, routine requests to autonomous Layer 1 self-service agents and equipping staff with Layer 2 copilots, the university absorbed expanding service demands without sacrificing human oversight on high-judgment, complex cases. Patel emphasizes that enterprises should avoid attempting an instantaneous leap to complete autonomy. Successful adoption requires an intentional, incremental calibration that matches specific operational workflows to their corresponding level of AI agency, ensuring organizations capture efficiency gains without relinquishing executive control.

精炼阅读摘要

Freshworks北美解决方案工程高级总监Payal Patel指出，企业应用人工智能的核心瓶颈已从底层技术能力转向治理校准；企业必须打破盲目求快与过度谨慎的二元对立，通过分层的“任务控制”架构为AI划定明确的自主代理权边界，在保障人工战略控制的同时释放业务价值。摆脱二元陷阱与运营内耗的沉重代价 (1:41) Patel指出，企业界关于人工智能的讨论已从早期的“是否引入”转变为“赋予AI多少自主权”。现场调查显示，尽管多数企业已在生产环境中部署了AI工具，但几乎没有管理者对其系统当前的自主权水平感到完全放心，且仅有极少数人成功在全组织范围内推广扩展后的业务流程。Patel分析，企业通常会陷入两种极端的思维陷阱：一种是在缺乏清晰治理、责任归属和效果衡量机制的情况下盲目追求自动化速度，导致用户体验破碎与信任缺失；另一种则因过度担忧安全与运营风险而裹足不前，使AI长期受限于小规模试点，错失规模化价值。延迟引入自动化同样伴随着高昂的财务与时间损耗。根据Patel引述的基准数据，一家500人规模的典型企业每年因系统与流程复杂性承受约250万美元的间接损失；员工平均每周在分散工具切换与手动信息检索中浪费7小时。正如Freshworks首席执行官Dennis Woodside所总结，“复杂性是企业选择放慢增长步伐的结果”，缺乏治理的盲目求快与停滞不前的过度谨慎都会转化为切实的运营拖累。任务控制架构与AI代理权的分级校准 (10:01) 为破解控制权与效率的冲突，Patel将“AI代理权”（AI Agency）界定为：AI系统在预设安全边界内自主做出决策并端到端完成完整工作流的能力，无需在每一步骤停滞等待人工审批。这与单点触发的自动化具有本质区别，代理型系统能够跨越多个业务系统联动执行操作。为此，Freshworks提出了三层“任务控制”（Mission Control）架构，将AI的参与形态系统性解构为自主基础层、辅助增强层和战略智能层。现场调研反映出显著的应用成熟度断层：绝大多数受访企业的运营仍停留在以人类为主、AI仅做边缘辅助的初始阶段（Stage A），仅有一家实现了部分业务的自主闭环（Stage B），尚无企业达到完全成熟的代理运作状态（Stage C）。Patel强调，企业无需盲目追求全盘自动化，关键在于结合业务风险与标准化程度进行分层匹配。该架构依托Freshworks集成了配置管理数据库（CMDB）的统一平台，为AI引擎Freddy AI映射组织内人员、资产与服务依赖关系的完整上下文。三层控制模型在服务管理中的具体运作 (16:14) 在IT与企业服务管理的实际场景中，三层任务控制架构形成了清晰的分工界限。在第一层“AI自主基础层”，系统针对高频、低风险且高度标准化的常规事务实现零人工介入的闭环处理。例如在非工作时间处理员工密码重置时，AI可自主完成身份验证、策略校验与系统执行；在员工咨询带薪育儿假政策时，AI能够跨Google Drive、SharePoint等文档库提取精确条款，提供多语言解答并附带原始出处链接以建立信任。第二层“辅助增强层（Copilot）”专为需要人类专业判断的复杂疑难工单设计。当技术工程师介入时，系统会自动提炼工单历史、匹配过往类似故障的解决方案，并预先生成回复草稿与分类标签供工程师一键审核，同时将处理模式沉淀为新的知识库建议。第三层“战略智能层”则解决决策层面临的数据孤岛与分析延迟。AI通过持续监控底层系统日志与工单态势主动识别异常；例如当系统监测到整体平均解决时间突发激增200%时，战略智能层会自动向下穿透，明确提示该波动系由IT运营团队处理的网络与硬件故障引起，为管理者省去繁琐的手动报表排查。量化运营基准与曼彻斯特城市大学实践 (20:54) Patel披露的Freshworks客户综合基准数据显示，通过对AI代理权进行分层校准，部署该架构的企业在复杂服务请求上的平均解决时间（MTTR）缩短了76%，同时依托第一层自主渠道实现了65%的常规工单自动拦截分流（Ticket Deflection Rate），显著回收了人工团队的工作时长，使其能够专注于高价值异常场景。曼彻斯特城市大学（Manchester Metropolitan University）的实际部署验证了这套模型的有效性。面对覆盖2.2万名师生的庞大支持需求，该校传统支持渠道曾面临严重负荷。通过将三层框架落地，该校底层利用Freddy AI代理消化日常师生自助请求，中间层依托Copilot加速技术人员处理复杂工单，顶层借助AI洞察为管理团队提供运营决策依据，在保障治理控制与规范的前提下化解了服务瓶颈。 Patel总结指出，企业能否释放AI价值的关键在于审慎界定代理边界与运营模式，而非单纯比拼工具上线速度。然而，演讲中呈现的数据主要来自Freshworks自身客户基准与具体案例，并未详细阐述跨异构遗留系统整合时的技术摩擦，以及在高度监管行业中如何动态调整AI自主决策容错阈值的具体准则。

Detailed notes — zh

## Background 本内容由 Ai4 频道于 2026 年 8 月 20 日发布在 YouTube 平台上，演讲者为 Freshworks 北美解决方案工程高级总监 Payal Patel。页面介绍指出，该演讲探讨了如何摆脱“完全放手”与“抗拒自动化”的二元对立，借鉴 NASA 任务控制中心的理念构建分层架构，使 AI 既能拥有自主行动的代理权，又能让企业管理者牢牢掌控战略大局。 ## TL;DR 演讲围绕企业在应用 AI 时“既想获得自动化价值又担忧失去控制”的核心焦虑展开，深入剖析了缺乏战略的速度与过度复杂性所带来的隐形成本。为此，Payal 提出了由三层构成的 Mission Control 任务控制中心架构：第一层通过 Freddy AI Agents 自主闭环处理密码重置等高频低风险请求；第二层借助 Freddy Copilot 为人工坐席实时推荐回复与梳理上下文；第三层则依托战略智能提供告警聚类与根因分析（RCA），辅助管理层决策。最后，结合曼彻斯特城市大学等实际落地案例，展示了该架构实现 65% 工单拦截率与 76% 解决提速的卓越成效。 ## Atmosphere & style 整场演讲氛围幽默且充满互动感，讲者以在拉斯维加斯赢钱的轻松趣事破冰，随后通过现场举手表决精准切中观众在部署 AI 时的普遍焦虑。叙事逻辑层层递进、干货密集，将企业软件平台与 NASA 任务控制中心的生动类比相融合，既有直击痛点的商业金句，又有清晰的架构拆解与扎实的客户数据支撑。整体节奏明快、务实接地气，让听众在轻松共鸣之余迅速建立起对 AI 代理权落地与治理的全局认知。 ## Key takeaways - [00:08](https://www.youtube.com/watch?v=kU20tH50Yb8&t=8s) 企业在 AI 应用上面临的核心焦虑已从“AI 能否起效”转向“应赋予 AI 多大自主权，以及人类如何在关键节点保持掌控”。 - [01:41](https://www.youtube.com/watch?v=kU20tH50Yb8&t=101s) AI 代理权（Agency）的本质在于 AI 能够在明确界定的边界内独立做决策并完成闭环，无需每一步都等待人工审批。 - [03:47](https://www.youtube.com/watch?v=kU20tH50Yb8&t=227s) 推进 AI 必须摆脱二元对立思维：既不能在缺乏治理与度量的情况下盲目激进，也不能因过度谨慎而局限于极小试点导致掉队。 - [05:49](https://www.youtube.com/watch?v=kU20tH50Yb8&t=349s) 缺乏战略的盲目提速会导致三大陷阱：跨部门治理缺失、业务价值无法量化衡量，以及实施过度复杂沦为遗留包袱。 - [07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) 过度复杂性对企业增长构成隐形侵蚀，500 人的企业每年因系统摩擦和信息查找浪费可产生超 250 万美元的隐形成本。 - [10:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=637s) Mission Control 架构通过三层代理体系平衡自主与控制：底层自主闭环、中间层 Copilot 人机协作、顶层战略智能洞察。 - [13:13](https://www.youtube.com/watch?v=kU20tH50Yb8&t=793s) 企业 AI 成熟度是逐步演进的过程，需依托统一平台和 CMDB 资产拓扑等深厚业务上下文，才能实现安全可靠的操作执行。 - [16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) 在 AI 基础层，智能体可跨 Teams、Slack 等多渠道自主完成密码重置、制度答疑等低风险高频任务并精准溯源知识库。 - [18:00](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1080s) 在代理协作层，Freddy Copilot 能秒级提炼工单上下文、自动推荐分类与回复草稿，并根据高频痛点主动反哺知识库。 - [19:45](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1185s) 在战略智能层，AI 主动进行告警聚类与异常根因分析（RCA），让 IT 管理者在事故爆发前精准掌握态势并做出决策。 - [20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) 任务控制中心架构在实际落地中带来了显著成效：客户平均解决时效提升 76%，实现高达 65% 的常规工单自主拦截率。 - [21:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1297s) 曼彻斯特城市大学通过完整部署三层 AI 代理架构，成功在应对 22,000 名师生激增工单的同时实现精细化掌控。 ## Detailed walkthrough ### [00:08](https://www.youtube.com/watch?v=kU20tH50Yb8&t=8s) 开场破冰：我们是否真的对 AI 的自主权感到安心？作为全天议程的最后一位演讲者，Payal Patel 以轻松幽默的方式开场。她笑称自己成了挡在大家去享受饮品和体验拉斯维加斯娱乐之间的“最后一道关卡”，并分享了自己前一晚在掷骰子赌桌（craps table）赢了 2000 美元的兴奋趣事。随后，她切入正题：在过去一年中，围绕 AI 的行业对话发生了翻天覆地的变化。大家不再争论“AI 到底能不能帮到我们”，因为多数企业已经亲眼见识了 AI 的潜力。现在的核心问题变成了：**我们究竟应该让 AI 承担多少工作？如何从中榨取最大价值？以及人类应该在哪些环节牢牢保持控制权？** 为了摸底现场现状，Payal 与观众进行了一次即兴互动。当她询问台下有多少人已经在生产环境中部署了 AI 时，现场有大量观众举手；但当她紧接着要求“如果对自己赋予 AI 的自主权水平感到完全放心，请继续保持举手”时，几乎所有举起的手都在瞬间放了下去。这生动地揭示了当前企业应用 AI 时的普遍焦虑与困境。为此，Payal 正式引出了本次演讲的核心——“Mission Control（任务控制中心）”架构：一种既能赋予 AI 充分自主权以创造业务价值，又能确保关键决策权始终掌握在人类手中的治理框架。 ### [01:41](https://www.youtube.com/watch?v=kU20tH50Yb8&t=101s) 重新定义议题：从“AI 能做什么”到“AI 代理权（Agency）” Payal 介绍了自己的背景。作为 Freshworks 北美解决方案工程团队的负责人，她长期与众多 CIO 和 CTO 深度探讨如何在整个产品组合中推进 AI 运营与服务运营（Service Operations）。在交流中，她发现企业面临的最大挑战从不是 AI 的技术能力是否达标，而是如何在“让 AI 独立行动”与“在必要时由人类接管控制”之间找到平衡点。过去行业讨论的焦点往往停留在“AI 能做什么”或“我们要不要用 AI”，但绝大多数企业早已跨过了这个初级阶段——大家都在部署各类 Copilot，将 AI 嵌入日常操作。真正的核心命题在于：**企业应该赋予 AI 多大的责任？** 这正是“代理权（Agency）”的本质。所谓代理权，是指 AI 能够在明确界定的边界内，无需等待人类介入或审批，即可自主做决策并完成端到端工作的能力。取得显著成效的企业并非盲目上线工具，而是非常有策略性地规划 AI 的独立行动范围与人类的监督节点。这绝不是为了取代人工，而是让 AI 接管结构化、高重复性的例行事务，从而释放人力去攻克企业中最复杂的突发问题。 ### [03:47](https://www.youtube.com/watch?v=kU20tH50Yb8&t=227s) 摆脱二元对立：盲目狂奔失控 vs 过度保守落后当前企业推进 AI 的最大阻碍不是底层算法，而是非黑即白的二元思维定式： ```mermaid flowchart LR accTitle: AI Adoption Extremes vs Mission Control subgraph S1["极端路线"] A["盲目激进"] -->|"缺乏治理与度量"| B["体验不一致<br/>失去控制与信任"] C["过度谨慎"] -->|"试点停滞不前"| D["流程繁琐<br/>掉队与效率损失"] end subgraph S2["平衡之道"] E["Mission Control 架构"] -->|"校准代理权边界"| F["低风险自主闭环<br/>高复杂度人机协同"] end ``` 一方面，不少企业迫于竞争压力急于全面自动化，但由于缺乏清晰的治理架构，导致服务体验参差不齐，进而引发了对 AI 信任度和责任归属的质疑；另一方面，另一些企业因过度担忧安全与合规而步履蹒跚，将 AI 局限在极小的试点范围内，迟迟无法实现规模化落地。前者是在失控中盲目狂奔，后者则因慢条斯理而错失良机。现场调查再次印证了这种割裂：当被问及有多少人正积极将 AI 扩展到更多业务流程中时，不少人举手；但当问到有多少人已经成功推进并获得了全公司自上而下的广泛认可与采纳时，全场仅有寥寥两人举手。这说明“买工具”很容易，但要在组织内部建立起有策略的采纳机制并释放价值，才是真正的难点所在。 ### [05:49](https://www.youtube.com/watch?v=kU20tH50Yb8&t=349s) 缺乏战略的速度陷阱：治理、度量与实施复杂度许多企业在 AI 浪潮中快速启动试点、采购工具、能自动化的地方就强行自动化，然而问题不在于速度本身，而在于“缺乏战略的速度（speed without strategy）”。盲目提速往往会撞上三堵大墙： - **治理缺失（Governance）**：缺乏清晰的所有权界定、问责机制、运营模型以及跨部门利益相关者的有效对齐，导致 AI 举措无法发挥全部潜能。 - **业务价值无法度量（Business Value）**：很多团队在部署前根本没有明确基准指标，上线后也无法衡量 AI 是否达成了既定结果。AI 绝不是“一次性打开开关即可”的工具，而需要持续监控、校准和迭代优化。 - **实施过度复杂（Complexity）**：漫长的实施周期和大量的深度定制，让 AI 项目沦为另一个沉重的传统 IT 遗留包袱。如果引入一套 AI 软件需要耗费整整一年才能上线见效，那就彻底失去了采用现代 AI 的敏捷意义。 ### [07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) 迟缓与复杂的隐形代价：从 CEO 警句看运营摩擦与盲目狂奔相对，过度谨慎和拖延采用 AI 同样会带来沉重代价。在服务运营场景中，AI 本可以迅速消化密码重置、员工福利咨询等内部高频日常请求，并自动调取正确的知识库内容，从而省去员工跨系统翻找信息的时间。 Payal 列举了一组极具冲击力的数据来揭示复杂性对企业的侵蚀： - **巨额隐形成本**：对于一家 500 人的中型企业，运营复杂性带来的隐形损失每年可超过 **250 万美元**。 - **严重的工时浪费**：员工平均每周要耗费近 **7 个小时**在不同工具间频繁切换、手动翻找信息和应付繁琐的非自动化流程上。 - **居高不下的机会成本**：工具泛滥和多平台割裂极大地拖慢了企业的响应速度。她援引了 Freshworks CEO Dennis Woodside 的名言：“**复杂度是一种让人增长放缓的主动选择（Complexity is a choice to grow slower than you should）**。”如果企业希望加速增长，就必须主动消除系统中的摩擦与多余复杂度。 ### [10:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=637s) Mission Control 三层代理架构与任务校准在明确了任务性质决定 AI 介入程度后，Payal 对“AI 代理权（Agency）”给出了精准定义：**AI 在清晰划定的边界内，代表人类做出决策并端到端完成工作的能力，期间不需要每一步都停下来请求人工批准**。例如，AI 可以自主理解员工问题、引用权威知识库、完成请求闭环，或是自动校验员工的带薪休假（PTO）余额、提交请假申请并抄送直属主管。基于此，Mission Control 将企业中的 AI 划分为三个清晰的层次： ```mermaid flowchart TD accTitle: Mission Control 3-Layer Architecture subgraph L3["第三层：战略智能层 (Strategic Intelligence Layer)"] A3["宏观态势感知 / 告警聚类 / 根因分析 (RCA)"] B3["人类管理者掌舵：基于洞察做战略决策"] A3 --> B3 end subgraph L2["第二层：代理协作层 (Agency Assistance Layer)"] A2["Freddy Copilot 实时辅助人工坐席"] B2["提炼工单上下文 / 推荐回复 / 智能分类与路由"] A2 --> B2 end subgraph L1["第一层：AI 基础层 (AI Foundation Layer)"] A1["自主智能体 (Autonomous AI Agents)"] B1["低风险高频任务：密码重置 / 制度问答 / 全天候自助闭环"] A1 --> B1 end L3 -.->|"指导与策略输入"| L2 L2 -.->|"数据与工单升级"| L1 ``` - **第一层：AI 基础层（AI Foundation）**：处理高重复性、低风险的基础请求，实现无需人工干预的自主闭环。 - **第二层：代理协作层（Agency Assistance）**：在面对较复杂工单时，AI 作为 Copilot 为人工坐席收集上下文、提炼摘要并推荐下一步最佳操作。 - **第三层：战略智能层（Strategic Intelligence）**：人类完全掌控大局，AI 负责自下而上提炼业务全局趋势与潜在风险，为管理者提供行动建议。 ### [13:13](https://www.youtube.com/watch?v=kU20tH50Yb8&t=793s) 现场成熟度摸底与 Freshworks 统一平台基石在深入讲解架构落地前，Payal 再次发起现场调查： - 大多数与会者表示自身企业仍停留在“AI 仅提供初步辅助、绝大部分工作依然由人工完成”的阶段； - 只有极少数的一两个人表示已经实现了“AI 自主解决日常工作”； - 现场所有人均已涉足 AI，没有人处于完全未起步状态。 Payal 安慰大家，企业所处的成熟度阶段各不相同是完全正常的，切忌妄想“一夜之间直接跳跃到完全自主运作”。正确的路径是将合适的业务流逐步匹配到对应等级的代理权上。随后，她介绍了 Freshworks 的平台理念与体量数据： - **公司创立与使命**：成立于 2010 年，始终坚信“企业级软件无需复杂繁琐也能极其强大”。 - **企业规模**：现为上市公司，服务全球超过 75,000 家客户，拥有 4,500 多名员工，年度经常性收入（ARR）指引已突破 8.1 亿美元。 - **平台架构核心**：Freshworks 并非简单外挂 AI，而是构建了跨全套件的统一平台。其核心依托配置管理数据库（CMDB），全面打通企业中的人员、资产、IT 服务及它们之间的拓扑关联。正是这种深厚的上下文关联，使得其搭载的 **Freddy AI** 能够跨越“单纯回答问题”的聊天机器人阶段，进化为能够安全做出决策并执行操作的智能体。 ### [16:14](https://www.youtube.com/watch?v=kU20tH50Yb8&t=974s) 深入第一层与第二层：自主闭环与 Copilot 坐席赋能 Payal 结合具体业务场景详细拆解了前两个层级的落地方式： #### 第一层：AI 基础层（自主服务闭环） - **多渠道即时响应**：员工可通过 Microsoft Teams、Slack、电子邮件或自助服务门户随时发起请求。 - **端到端自主执行**：若员工在凌晨 2 点遗忘密码，Freddy AI 能自动完成身份鉴权、校验安全策略并在后台直接完成重置，全程零人工等待。 - **深度整合与可信溯源**：打通 Google Drive、Confluence、SharePoint 等分散知识源。在回答育儿假等政策咨询时，Freddy 不仅能主动追问补充细节，还会明确标注引用的文档出处，保障内容可信度。 - **多语言个性化**：原生支持跨国企业多语言环境，根据员工偏好自动切换交互语言。 #### 第二层：代理协作层（Freddy Copilot 赋能坐席） - **消除传统工单准备时间**：在传统模式下，运维坐席在打开工单后的前几分钟都在费力阅读背景和搜索解决方案；而在 Copilot 辅助下，坐席打开界面的瞬间，Freddy 就已经理解了全貌。 - **智能推荐与一键采纳**：自动生成回复草稿、推荐工单分类与路由标签，并关联展示历史相似工单的处理方案，坐席只需一键确认即可发送。 - **自愈式反哺知识库**：持续分析反复出现的高频痛点，主动向 IT 管理人员推荐应当新建或优化的知识库条目。 ### [19:45](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1185s) 深入第三层：战略智能层与异常归因分析在传统运营中，CIO 与 IT 领袖往往饱受“数据庞杂却缺乏有效洞察”的困扰，不得不花费大量时间手动翻阅报表。Mission Control 的第三层——战略智能层，彻底改变了这一现状： Freddy AI 能够主动监控全链路运行状态。当系统侦测到特定基础设施的大量关联告警信号时，它会自动进行告警聚类并直接输出根因分析（RCA），在潜在重大事故爆发前向管理者发出预警。 Payal 展示了一个实际控制台案例：系统监测到平均解决时间（MTTR）异常激增了 200%。Freddy AI 不仅敏锐地捕捉到了该指标异动，更直接给出了精准的归因解释——指出**本次时长激增主要是由 IT Ops 团队负责的网络与硬件相关工单暴增所引起的**。这种具备业务上下文的战略级智能，让管理者无需深陷数据泥潭即可迅速做出精准决策。 ### [20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) IT 运维深度根因排查与核心业务成效在 IT 运维（IT Ops）场景中，AI 赋能的控制中心架构赋予了团队深度下钻与调查异常的能力。当系统监控到工单量或报警突增时，运维人员不再盲目救火，而是能够清晰判断“背后的核心驱动因素是什么”、“引发当前连锁反应的源头何在”。系统能迅速关联分析近期是否发生过底层基础设施的变更（Infrastructure Change），帮助团队真正追根溯源，秒级定位技术故障的根本诱因。这一整套任务控制体系在 Freshworks 客户的实际落地中展现出了颠覆性的业务回报。多项关键运营指标实现了质的飞跃： - **解决时效大幅缩短**：部分客户的平均问题解决速度提升了 76%（76% faster resolution time）。 - **常规工单自主拦截**：实现了高达 65% 的平均工单拦截率（65% average ticket deflection），大量日常重复诉求由 AI 在无需人工干预的情况下自主闭环。 - **释放海量人力工时**：为一线服务坐席节省了难以计数的工作时间，让员工从机械化的流程中解脱出来，专注于更具业务价值与创造性的工作。 ### [21:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1297s) 曼彻斯特城市大学实践案例与三层架构闭环作为真实落地的典型标杆，曼彻斯特城市大学（Manchester Metropolitan University）为多达 22,000 名在校师生和教职员工提供 IT 服务支持。在面临工单总量呈现爆炸式增长的严峻挑战下，该校完整部署了前述的三层 AI 代理架构，实现了规模化服务与精细化管控的平衡： ```mermaid flowchart TD accTitle: 曼彻斯特城市大学的三层 AI 架构实践 subgraph L1["第一层：自主行动层 (Autonomous Action)"] A["Freddy AI Agents"] --> B["22,000 名师生员工自主服务 / 常见请求秒级闭环"] end subgraph L2["第二层：人机协同层 (Assisted Decisions)"] C["Freddy Copilot"] --> D["为服务坐席提供实时建议 / 极速响应复杂支持"] end subgraph L3["第三层：战略洞察层 (Human-Led Strategy)"] E["AI Insights 洞察引擎"] --> F["管理层全面掌控全局态势 / 决策未来优化行动"] end ``` - **第一层（Freddy AI Agent）**：部署自主服务智能体，确保学生与教职员工的日常需求在自助服务渠道得到即时、精准的解决。 - **第二层（Freddy Copilot）**：为所有一线 IT 服务人员配备 AI 副驾驶，通过智能建议与一键采纳，助力坐席以极高效率向学生交付个性化支持。 - **第三层（AI Insights）**：校方管理领导层利用全景 AI 洞察分析系统趋势与运营瓶颈，以此作为依据制定下一步的管理策略与改进决策。演讲最后，讲者总结了这套“既赋予 AI 充分自主权、又牢牢掌握战略掌控力”的控制中心架构精髓，并邀请参会嘉宾前往展厅展位展开进一步技术交流与互动问答。 ## Notable quotes > The spacecraft didn't ask permission to adjust trajectory for every solar wind fluctuation—that would be absurd. But when an oxygen tank exploded on Apollo 13, human expertise took over immediately. — Payal Patel > [07:56](https://www.youtube.com/watch?v=kU20tH50Yb8&t=476s) Complexity is a choice to grow slower than you should. — Dennis Woodside > The agency paradox: why giving AI more autonomy (in the right places) actually increases human control over outcomes. — Payal Patel > [00:08](https://www.youtube.com/watch?v=kU20tH50Yb8&t=8s) 我们究竟应该让 AI 承担多少工作？如何从中榨取最大价值？以及人类应该在哪些环节牢牢保持控制权？ — Payal Patel > [01:41](https://www.youtube.com/watch?v=kU20tH50Yb8&t=101s) 企业应该赋予 AI 多大的责任？ — Payal Patel > [10:37](https://www.youtube.com/watch?v=kU20tH50Yb8&t=637s) AI 在清晰划定的边界内，代表人类做出决策并端到端完成工作的能力，期间不需要每一步都停下来请求人工批准 — Payal Patel > [13:13](https://www.youtube.com/watch?v=kU20tH50Yb8&t=793s) 企业级软件无需复杂繁琐也能极其强大 — Payal Patel > [20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) 76% faster resolution time — Payal Patel > [20:54](https://www.youtube.com/watch?v=kU20tH50Yb8&t=1254s) 65% average ticket deflection — Payal Patel ## Mentioned Payal Patel, Dennis Woodside, Freshworks, Freddy AI, Freddy Copilot, Freddy AI Agents, AI Insights, NASA, Apollo 13, Manchester Metropolitan University, Microsoft Teams, Slack, Google Drive, Confluence, SharePoint, CMDB, Ai4 ## Video description Payal Patel - Senior Director, Solution Engineering, North America, Freshworks The AI conversation has become uncomfortably binary: either hand everything to AI and hope for the best, or resist automation entirely and fall behind. But the organizations achieving real results are doing something different—they're building Mission Control architectures where AI has genuine agency to act autonomously, while humans maintain strategic oversight. This is the lesson from NASA's Mission Control: autonomous systems handled thousands of micro-decisions during Apollo missions, but flight directors always knew what was happening and could intervene when judgment mattered. The spacecraft didn't ask permission to adjust trajectory for every solar wind fluctuation—that would be absurd. But when an oxygen tank exploded on Apollo 13, human expertise took over immediately. Service operations need the same architecture. AI agents that deflect 65% of routine tickets without human review. Copilots that suggest responses agents can accept with one click—or override when context demands it. Insights that surface trends proactively but let leaders decide what action to take. This is AI with agency: sophisticated enough to act independently, architected carefully enough that you never lose control. Through real examples and the Mission Control framework, we'll explore how to calibrate agency correctly—knowing when AI should act autonomously (password resets, routine requests, policy questions), when it should assist human decision-making (complex troubleshooting, escalations, edge cases), and when humans must stay in full control (strategic decisions, sensitive situations, novel problems). The result is service operations where AI handles exponentially more work while leaders feel more in control, not less. Key takeaways are, the agency paradox: why giving AI more autonomy (in the right places) actually increases human control over outcomes. Mission Control architecture: the three layers of agency (autonomous action, assisted decisions, human-led strategy). Calibrating agency correctly: frameworks for deciding what AI handles independently vs. where humans stay in the loop. Real benchmarks from organizations getting agency right: 65% autonomous deflection, 76% faster resolution, zero loss of quality or control. Governance without bureaucracy: how to build oversight that enables speed rather than blocking it. The agentic AI maturity curve: progressing from AI-assisted to AI-autonomous operations while maintaining strategic control [Source](https://www.youtube.com/watch?v=kU20tH50Yb8)

Source quotes

Payal Patel: The spacecraft didn't ask permission to adjust trajectory for every solar wind fluctuation—that would be absurd. But when an oxygen tank exploded on Apollo 13, human expertise took over immediately. [7:56] Dennis Woodside: Complexity is a choice to grow slower than you should. Payal Patel: The agency paradox: why giving AI more autonomy (in the right places) actually increases human control over outcomes. [0:08] Payal Patel: 我们究竟应该让 AI 承担多少工作？如何从中榨取最大价值？以及人类应该在哪些环节牢牢保持控制权？ [1:41] Payal Patel: 企业应该赋予 AI 多大的责任？ [10:37] Payal Patel: AI 在清晰划定的边界内，代表人类做出决策并端到端完成工作的能力，期间不需要每一步都停下来请求人工批准 [13:13] Payal Patel: 企业级软件无需复杂繁琐也能极其强大 [20:54] Payal Patel: 76% faster resolution time [20:54] Payal Patel: 65% average ticket deflection

Source quotes

[7:56] Dennis Woodside: Complexity is a choice to grow slower than you should.

Original transcript

[0:08] All right. Well, I get to be the lucky one. That's the last one of the day. So, [0:13] I'm holding you up from your drinks probably in the gambling. By the way, see me at the [0:18] craps table cuz I won two grand last night. So, that's a fun exciting night. Um, all right. Well, [0:24] let's kick it off. So over the last year, the AI conversation has changed dramatically and the [0:31] question is no longer about can AI help us. It's more about organizations have seen the potential [0:39] of AI already and they're trying to think about how much should AI actually do? How much value [0:46] can I get from AI and where should people remain in control outside of AI? And that's what we're [0:52] going to be um talking about. Now before I get started um tell uh show me by hands how many of [1:00] you already have AI in production. Awesome. Uh now keep your hand up if you're completely comfortable [1:09] with the level of autonomy you've given to AI. Okay. Um that's usually where all of this kind of [1:17] starts right to come down. It's exactly why we're here. Um today I'll intro introduce this concept [1:23] of mission control a framework for giving AI enough autonomy to create the value while keeping [1:30] people in control of a lot of that decision making that you would need that matters most. [1:41] All right so uh just a little bit about me. Um as it was just stated I lead the solution [1:46] engineering team at Freshworks. Um my team works with customers uh especially CIOS and CTO's [1:54] uh to really understand how we can help from an AI uh operations and service operations [1:58] perspective across across their portfolio. Um and a lot of the conversations I get to have is [2:05] uh with customers leads us to some of the biggest challenges we're going to be [2:08] talking about today. And it's not whether AI is capable. It's more about where AI [2:14] can act independently but then also still help you stay in control when you need to. [2:23] So, one of the things that I think we've done wrong over the years is that we're maybe asking [2:28] the wrong question. The conversation has focused a lot on what AI can do. [2:34] Should we use it? But most organizations have already moved way past that at this point. [2:39] They're deploying co-pilots. Um they're embedding AI into everyday actions. Um a [2:46] lot of you already mentioned you're using AI and leveraging it. The real question is how [2:51] much responsibility should we give AI within an organization. Um and that's where agency comes in. [2:59] um agencies is AI's ability to act, make decisions and complete work without waiting for any human [3:08] intervention in any way. The organizations that are seeing the greatest impact are really the [3:14] ones that are have the ability to leverage AI the most. They're being very intentional [3:20] about how they leverage AI, where it can act independently, and where humans should stay [3:26] involved to ensure that it's controlled and structured. And this isn't about replacing [3:32] people. It's more about how AI can handle some of those repetitive inquiries, how it can create [3:37] well- definfined work so people can focus on more of the complex situations within an organization. [3:47] So one of the biggest obstacles that we face today isn't the technology, it's the mindset around AI. [3:55] The conversation has become very binary. On one side, there's a pressure that everybody probably [4:02] has in this room, which is to automate and that um there needs to be clear governance around [4:10] how AI is going to be helping. Now there's a lot of inconsistent experiences at times and [4:16] questions around trust and accountability. And then on the other side, some organizations move [4:21] too cautiously. And what that's delaying is it's they're keeping AI too limited. The pilots aren't [4:27] getting to the scale that they want, but on they feel more safer with how they're implementing it [4:32] and how they're getting the organization to trust AI. Um, one moves fast without control, the other [4:38] one pretty much moves slow and falls behind. So what's the right method? Um it's not about how a [4:46] we should use AI. It's more about where should AI act autonomously within your organization? [4:53] Um and this is the decision every organization is making today. Okay. With that being said, [5:00] how many are actively looking to expand AI into more of their business processes? Okay. Um, and [5:10] how many of you have already done that and gotten broad approval within the organization of those [5:16] business processes and adoption? Saw two hands go up, not the same hands. That was interesting. Um, [5:23] okay. It's not about how fast you move. It's really about how are you getting intentional with [5:29] it and how are you ensuring that you do have the adoption because you can buy all the AI you want [5:35] but if within the organization there isn't enough adoption you're not going to be able to get the [5:40] most value from that AI that you're leveraging. So let's look at a a totally different extreme. Many [5:49] organizations are moving quickly with AI. They're launching pilots. They're buying tools. They're, [5:54] you know, automating wherever they can. The problem isn't speed. It's speed without strategy. [5:59] I'm sure you've been hearing this throughout this conference, but strategy is really important in [6:03] terms of how you're thinking about your AI pilots. Um, the first one, it's all about the governance. [6:09] Um, many AI initiatives never reach their full potential um because of ownership, accountability, [6:17] making sure you have the right operating models and stakeholders aligned on it. Second is making [6:22] sure you can drive business value. How many of you are actually measuring if the AI is doing what you [6:30] intended it to do from an outcome perspective? Okay. Is it working for you? You can give me a [6:37] raise of hand if it's a yes. Ah, it's a no. It's a maybe. Yeah. It's a maybe. Okay. Yeah. I think [6:45] it's really critical that that the organization continues to measure if it's actually working for [6:52] them. And when I talk about measuring it, it's making sure you're getting the right gains from [6:56] it. Did you even know the metric that you were going to measure pre and post implementation? [7:02] It's super critical because AI is not a one-time turn it on. You have to continue to refine, [7:08] make it better, get more value from it throughout. And then finally, complex. Complexity is a big [7:14] factor as well. Um, long implementation cycles and extensive customization can turn AI into [7:22] another project. How many of you have had a legacy system that took you one year to implement? Sure, [7:27] many of you have had it. Now, AI shouldn't be taking a year to implement. So if you invested [7:32] in a software that's taking you way too long to get AI to come to life, then that's something [7:37] that you should be looking into because you want to be able to get speed to be able to leverage [7:42] those AI capabilities. The lesson isn't that AI doesn't work. It's more um that governance and [7:50] clear architecture and those business outcomes need to be set up before you get into it. [7:56] So now let's quickly look at um the other end of the spect spectrum and I'll call about think about [8:03] caution and how AI from a security compliance and so on. Organizations that delay AI are not [8:10] standing still. They're often failing behind in speed. Um in service operations AI can resolve [8:18] uh routine issues. So think about simple inquiries like reset passwords and um tell [8:23] me about my benefits things that are internal specific. Um and they can also surface up the [8:28] right knowledge articles. It also reduces the time that employees spend searching across different [8:34] systems for quick information. Now choosing to adopt some of these capabilities can really help [8:41] reduce the amount of time that people are spending to get the information that they need. This is an [8:48] interesting stat. Um, some numbers behind it will give you even a better perspective. For [8:53] a 500 person organization, the cost of complexity can exceed $2.5 million each year. That's a lot of [9:02] money for an organization. First, productivity. Employees can lose nearly seven hours a week [9:08] searching for information, switching between tools, and navigating lots of manual processes. [9:14] Um, if you think about how many tools you've invested in, there's so many different tools [9:19] that people have to go to to get what they need. And without simplifying that, you're creating [9:25] another complexity layer. And finally, there's the opportunity cost. Organizations that simply um [9:32] simplify operations can actually move faster. So, consolidating tools, bringing tools into one place [9:37] instead of having multiple places for those tools itself. Our CEO Dennis Woodside said it well. [9:43] Complexity is a choice um to grow grow slower than you should. That is the point that I reflects to [9:51] me is because if you really want to grow you need to remove that complexity and friction [9:58] okay so we've now looked at both extremes. Um the question now is should we use AI and how helpful [10:06] it is. Um the answer depends on the task. If you can find repeatable, low-risk work like resetting [10:13] passwords like I mentioned earlier or answering common questions or routine tickets in the [10:18] um in the service operations model, then you're in a good place. For more complex work, AI [10:23] should assist people. Um make your agents faster because they can get the information they need to [10:29] get their work done a lot faster. And that's where co-pilots come in in a majority of the situations. [10:37] Okay. Now I want to define the most important word in today in today's presentation. It's this [10:43] concept of agency. AI agency is the ability to make decisions and complete work on your [10:50] behalf within clearly defined boundaries without requiring approval at every step. So we've been [10:57] looking into a lot of autonomous workflows. How can it make the decision without having [11:03] uh to get any approval? And AI agency specifically is very different. It means AI can complete that [11:10] without any influence at all. So let's take an example. AI can um answer an employees [11:18] question. It can cite the correct knowledge. It can resolve the request really quickly or [11:23] it can validate PTO balance or help you figure out what you need to do to uh take time off and [11:30] notify your manager. So it can do a lot of things that would simplify processes itself. [11:38] Now that we've defined agency, what does it kind of look like in practice? AI is not [11:43] simply autonomous. Um on the left side, what you're seeing is AI suggests draft responses, [11:50] surfaces up information. It gives you recommended actions. In the middle, it assists you. And how [11:56] it can assist you is that it gathers contextual information. and it gives you what you need to [12:00] make better decisions faster. And on the right, it starts to act, taking all those actions that [12:06] it can take, whether it's doing that PTO for you or setting up um your service catalog for you, [12:12] whatever it might be. Now, that doesn't mean that people lose control. It actually helps govern [12:19] there's obviously govern governance behind it, but also helps you operate and be a lot efficient as [12:25] you're going through your business. So, how much should we give access to AI? Um, this is a little [12:34] bit of a foundational slide here. Layer one is the AI foundation. It handles all of those repeatable [12:40] things that we talked about. Layer two is going to be that agency assistance layer that we talked [12:46] about where assists someone summarizing context, giving them information, recommending the next [12:52] best course of action. And then layer three is the strategic intelligence layer. People [12:58] remain in control while they understand what's going on with AI. It's telling you what's going [13:04] on within your business so that you can then add layer 1 and two into the mix as needed. [13:13] All right. So before we continue on, I'd love to get a quick poll from the room. By a show [13:18] of hands, how many of you um are mostly in option A where AI is helping but people are still doing [13:26] most of the work? Good amount of the group. Um and how many of you have a have AI autonomously [13:36] resolving work today? Wow, very few hands. Okay, one. Um, anyone just getting started, just started [13:46] to leverage AI in here? Wow, everybody's in. Okay. Well, that's good. There's no wrong answer. [13:52] Every organization, I think, is at a different journey. And I think that's a good thing. Um, [13:56] that's why you're all here today this week, right? Learning about that. Um, the mistake [14:01] is thinking you need to jump all the way to C overnight, right? I think the goal is to move [14:06] the right workflows to the right level of agency so that you're getting the most value from AI. [14:13] Okay, so everything we discussed this is kind of brings us to um what whether you're evaluating [14:19] AI or thinking about AI um at Freshworks what we've done is we've built our platform around [14:26] the same principles that I just talked about. We think about those layers very consistently. So [14:32] um we just to give a quick intro to everyone we were founded in 2010 with a simple belief. We [14:37] think that enterprise software doesn't have to be complicated to be powerful. Today, we're a public [14:43] company. We have more than 75,000 customers with over 4,500 employees, and we're continuing to grow [14:51] more than 810 million in annual revenue guidance. It's um how we're thinking about our unified [14:58] platform as well. Um so the way that we've been thinking about this is AI is not just um embedded [15:07] into our platform. It's across our entire suite. So the biggest thing is that we wanted to create [15:13] a unified platform that can be accessed across various um stakeholders within an organization. [15:20] At the center is our CMDB. The system is understanding your people, your assets, um all the [15:27] services and how they're interconnected. Um, the context allows AI to move beyond just answering [15:33] questions. It's actually making informed decisions and taking actions very safely. That's from an [15:39] agentic perspective. And then across the platform, Freddy AI can resolve routine work, assist agents. [15:47] It can surface up those insights to leaders and it can also um provide you to scale your teams as [15:54] needed. And everybody's talking about trust these days. So one of the things that we take pride in [16:01] is that whether you're using autonomous agentic services from us or assisted co-pilots, we want to [16:07] make sure that you can trust it and we have layers of security embedded across our entire platform. [16:14] Okay, so let's bring this all together. We talked a lot about the mission control layer one which [16:19] was the AI foundation where it handles the routine tasks. Layer two is that agency assistance and [16:25] then layer three is that strategic layer that we talked about. So how does that kind of fit into [16:30] our portfolio? Um layer one from an AI foundation perspective, we provide um first anon um service [16:41] from our AI agents. Employees can ask for help through teams, through Slack, through email, uh [16:47] through a portal and immediately get a resolution. So, imagine your employee forgets their password [16:54] at 2 a.m. Freddy's able to help them identify who they are. Are they are they somebody who should [16:59] get access to resetting the password? They apply all of the security policies in there and they [17:04] reset the password automatically for that employee without having to wait or no human intervention. [17:11] Second, Freddy is contextaware. It's deeply integrated. It can search across uh knowledgebased [17:17] articles, Google Drive, Confluence, uh SharePoint, other connected services. And think about if [17:24] you're asking for parental leave, it Freddy can bring in all of that information together and [17:29] follow up and ask other questions, but it can also site a reference of where that content is coming [17:34] from so that you can trust it. And then third, it's all very personalized. So if you are a global [17:41] organization, it's multilingual. It can speak in various languages. We support many of those within [17:47] our portfolio. And it can help give the right answers based on the language of your choice. [17:56] All right. Layer two is that agency layer. Um this agency assistance is more about helping the [18:03] agent be a lot faster. So let's talk about some of those capabilities. Um this is layer two. Um [18:09] these are our co-pilot capabilities. This helps agents become faster at what they're doing. So, [18:16] um, think about as they're responding to an employee. Um, it can accelerate the resolution by [18:22] suggesting responses. It can draft communications. It can recommend categorization and routing. Um, [18:29] it can also provide context. So, if the inquiry came in before, it can suggest [18:34] similar tickets so that your agents know how to respond to that issue. Um, so you're not having [18:40] to search for that content. And then finally, Freddy improves self-service over time. It can [18:46] identify reoccurring issues. It recommends knowledge articles. It helps you continue to make [18:51] your IT service management uh better by giving you more information and actions that you can take. [19:00] And this is a a quick view of what that can look like within a traditional desktop or uh UI [19:07] of an agent. In a traditional workflow, an agent spends the first few minutes reading the ticket, [19:13] searching for information. But with Copilot, that changes. As soon as the ticket opens, [19:18] Freddy understands the context. It can surface up relevant knowledge. It can [19:23] um give responses. It can categorize it. it can um give the agent the control it needs but guide [19:30] it every step of the way. So they can accept a recommendation, they can also um be able to [19:37] send content and responses to the customer all within one place. All right. And then the final [19:45] layer is the strategic intelligence layer. In this one, we've seen AI resolve that routine work, [19:53] make agents more productive. And how does that happen? A lot of CIOS face um a lack of data. How [20:00] many of you don't have the right data to make the decisions that you need to make? And what Freddy [20:05] AI does is it can surface up the the capabilities within the solution without you having to sift [20:11] through a report. It's going to tell you here's a knowledge article for you to create based on [20:16] historical context or here's a major incident that could happen within your infrastructure because [20:22] we're seeing signals of a lot of alerts coming in and share the root cause analysis of it. So more [20:28] strategic in behavior because it helps you make those decisions of what to do next a lot better. [20:35] All right. And this is a quick visual of how that looks. If you're noticing in this example, average [20:41] res resolution time has increased by 200%. Freddy not only identifies the change, it explains that [20:48] the increase is driven by a network and hardware ticket within the IT ops team. So, it's giving [20:54] you the ability to investigate what's going on from the IT ops perspective. It's you can decide [21:00] what's driving the increase um what's creating the effect of it. Is this related to a recent [21:05] struct infrastructure change? you can really dig in and understand what's causing the issue itself. [21:14] All right, we've covered the framework, we've shared a lot. These are some results that [21:17] we've seen from some of our customers. Um, some customers have seen 76% faster resolution time, [21:24] 65% average t ticket deflection, and even um these capabilities have saved many, [21:31] many agent hours for individuals as well. And the final story I'll leave you on is [21:37] one from Manchester Metropolitan University. They support about 22,000 students and staff [21:43] at T and ticket volumes continue to grow significantly. They applied the same three [21:48] layers that we talked about. Um at layer 1, they have Freddy AI agents making sure [21:53] that employees and students are getting what they need from a self-service perspective. [21:58] We have Freddy co-pilot helping all the all the agents so that they can make sure the students [22:03] are getting the support that they need really quickly. And from an AI insights perspective, [22:08] the leaders within Manchester are using this to make decisions on what they should do next. [22:15] Okay. Um and that was it. I'm happy to take any questions. Um and thank you all. Feel [22:22] free to visit our booth um which is going to be in the expo hall. Thank