“
AI Agent 是一次从"附加项"到"AI 优先"的跨越——这是工作方式的根本改变，需要思维模式与企业文化的深刻转型。 AI agents are the leap from being an 'add-on' approach to being an 'AI-first' process — a fundamental change in workflow that requires a profound shift in mindset and corporate culture.

Oliver Parker · Google Cloud 生成式 AI 全球 GTM 副总裁VP, Global GTM for Generative AI, Google Cloud

00全局判断The Big Call

报告的价值判断有三层：其一，价值不在"把旧流程做快"，而在重新设计流程本身；其二，AI 驱动一场"代际重构"——人从执行步骤中退出，转向高价值的跨业务战略编排；其三，临界点已过——52% 的受访组织已将 Agent 投入生产，88% 的早期采用者已看到正 ROI，讨论"要不要用"已经过时，问题只剩"从哪里开始"。 The report's judgment has three layers: value lies not in making old processes faster but in redesigning them; AI drives a "generational refactoring" moving humans from execution to strategic orchestration; and the threshold is crossed — 52% of organizations run agents in production and 88% of early adopters see positive ROI, so the question is no longer "whether" but "where to start".

01趋势一 · 人人都有 AgentTrend 1 · Agents for every employee

意图式计算：说出目标，剩下交给 AgentIntent-based computing: state the goal, let agents deliver

2026 最显著的业务转变不是效率，而是"以员工为中心的根本转型"：每个个体被放大为创新与增长的主引擎。其底层，是人机界面的行为迁移。 The most significant business shift of 2026 isn't efficiency — it's an employee-centric transformation: every individual becomes a primary engine of innovation and growth, powered by a behavioral shift in the interface.

指令式计算 · 过去Instruction-based · Before

  • 人分析表格、写代码、拉数据Humans analyze spreadsheets, write code, pull data
  • 一步步告诉电脑"怎么做"Step-by-step, telling the computer "how"
  • 产出上限 = 个人的时间与精力Output capped by personal time and energy

意图式计算 · 2026Intent-based · 2026

  • 员工陈述期望结果Employees state the desired outcome
  • LLM + Agent 决定"怎么完成"LLMs + agents determine "how to deliver"
  • 新职责 = 制定策略、监督 Agent 系统（开票、签约）New role = set strategy, oversee agent systems (invoicing, contracting)
数据 AgentData

筛选数百万结构化与非结构化数据点，发现市场趋势中的可行动模式。Sifts millions of data points for actionable market patterns.

分析 AgentAnalyst

7×24 监控趋势、竞品与社交舆情，每天早晨投递一页纸洞察。Monitors trends and sentiment 24/7; a one-page brief every morning.

内容 AgentContent

按品牌语气起草社媒帖文与博客文章，交经理审阅。Drafts posts and articles in the brand voice for review.

创意 AgentCreative

依据策略与创意规范生成配图与视频。Generates images and videos from strategy and guidelines.

报告 AgentReporting

连接分析平台，每周五交付一页纸战役数据摘要。Pulls weekly campaign data; a one-page summary every Friday.

"10x 营销经理"示范了个人如何编排这套系统——五个 Agent 各司其职，经理专注品牌叙事与策略。三个落地参照： The "10x marketing manager" shows how an individual orchestrates this system — five agents with clear divisions of labor while the manager focuses on brand storytelling and strategy. Three reference points:

Suzano

全球最大纸浆制造商，用 Gemini Pro 将自然语言转为 SQL 查询 BigQuery 上的 SAP 数据——5 万名员工的查询耗时下降 95%。The world's largest pulp maker turns natural language into SQL over SAP data on BigQuery — query time down 95% for 50,000 employees.

TELUS

5.7 万名员工常态化使用 AI，每次交互节省 40 分钟；Gemini Enterprise 让知识工作者自建 Agent，SI 合作聚焦复杂长期项目。57,000+ employees save 40 minutes per interaction; Gemini Enterprise lets knowledge workers build their own agents.

边界 · 人是最终决策者Boundary · Humans decide

"Agent 不受控"是误解——尤其在媒体等内容行业，讲什么故事、如何变现，人的创意与策略判断不可替代。"Agents act without control" is a misconception — in media especially, what stories to tell and how to monetize them remain human calls.

02趋势二 · 流程即数字装配线Trend 2 · Workflows as digital assembly lines

多 Agent 端到端跑通业务流程Agents run processes end to end

Agentic 系统是一条"数字装配线"——由人引导、多步骤、多 Agent 编排的端到端流程，让整个业务 7×24 更聪明地运转。真实价值在于打通孤岛：电信业把网络运营、现场服务与呼叫中心连成一条链——Agent 自主修复网络异常、自动开现场服务工单、同步通知客服告知客户，一气呵成。 An agentic system is a "digital assembly line" — a human-guided, multi-step workflow orchestrated across agents, running the business smarter 24/7. Real value comes from dissolving silos: in telecom, an agent remediates a network anomaly, opens a field-service ticket and alerts the contact center — in one integrated sequence.

A2AAgent2Agent

装配线的"关节"The joints of the line

开放标准，让 Agent 之间无缝集成与编排——即使来自不同开发者、构建于不同框架、属于不同组织。未来图景：媒体公司的 Agent 直连零售商的 Agent 展示剧情内商品；医院 Agent 在患者授权下与实验室、保险 Agent 直接协作。 An open standard enabling seamless integration and orchestration between agents — even across developers, frameworks and organizations. The vision: a media company's agent connects to a retailer's agent to surface products shown in content; hospital agents work with labs and insurers under patient consent.

MCPModel Context Protocol

装配线的"神经"The nerves of the line

解决 LLM 两大天生局限——知识冻结于训练时点、无法触达外部世界。MCP 提供标准化双向连接，让 AI 应用接入 Cloud SQL、Spanner、BigQuery 等实时数据源与工具，这是 Agent"接地"于企业数据的前提。 Solves two inherent LLM limits — frozen knowledge and inability to act. MCP provides a standardized two-way connection to live data and tools (Cloud SQL, Spanner, BigQuery) — the prerequisite for grounding agents in enterprise data.

AP2Agent Payments Protocol

代理式商务的"信任层"The trust layer of agentic commerce

现有支付体系默认"人在直接下单"。AP2 回答三个关键问题：如何证明用户授权？商家如何确认 Agent 请求不是幻觉？欺诈发生时谁负责？——为代理式商务提供安全、开放、可扩展的框架。 Today's payment systems assume a human initiates the purchase. AP2 answers three critical questions: how to prove user-given authority, how merchants verify an agent's request isn't a hallucination, and who is accountable for fraud.

Salesforce × Google Cloud

基于 A2A 开放协议构建跨双平台的协作 Agent，迈向开放、可互操作的 Agentic 企业基础。Cross-platform agents on the A2A open protocol — toward an interoperable agentic enterprise.

Elanco

动物健康巨头：Elanco.ai 每站点自动整理、提炼、比对 2,500+ 份非结构化制度文档，规避最高 130 万美元的生产力损失。Auto-processes 2,500+ unstructured policy documents per site, avoiding up to $1.3M in productivity impact.

PayPal

采用 AP2 构建代理式购物："黑色款补货就买，超过 100 美元不买"——Agent 监控价格与库存，条件满足且有人类预授权时安全成交，捕获本会流失的高意向订单。"Buy when the black one is back, not above $100" — with human pre-approval, agents securely close high-intent sales that would otherwise be lost.

金融业的下一站：多步 Agentic 合规系统自动监控法规变化、定位受影响政策、更新内部流程并生成完整审计链——"2026 年，Agentic 生态将在金融业从试点走向生产。"（Toby Brown，Google Cloud） Finance's next stop: multi-step agentic compliance that monitors regulatory change, identifies impacted policies, updates workflows and creates full audit chains — "In 2026, agentic ecosystems scale from pilots to production in financial services." (Toby Brown, Google Cloud)

03趋势三 · 代理式客户礼宾Trend 3 · The agentic concierge

对话的起点，不再是"请输入订单号"Conversations no longer start with "enter your order number"

差别不在 AI 本身，而在数据：礼宾型 Agent 接入企业愿意为该客户开放的上下文——CRM 购买记录、物流轨迹、账单系统——因此记得你是谁、买过什么、正在经历什么。49% 的受访高管以客户服务为 Agent 首要用例。 The difference isn't just the AI — it's the data. The concierge is grounded in the enterprise context you choose to share per customer — CRM history, logistics tracking, billing — so it remembers who you are and what's happening now. 49% of executives adopt agents for customer service first.

传统 ChatbotLegacy chatbot

  • "请输入您的 12 位订单号""Please enter your 12-digit order number"
  • 脚本选项，答不了复杂问题Scripted options; fails on nuance
  • 客户要反复喊"转人工"Customers repeat "operator!" to reach a human
  • 被动等待投诉Waits for complaints

代理式礼宾Agentic concierge

  • "您好 Elizaveta，您是问上周买的蓝色毛衣吗？已签收——要退货还是换货？""Hi Elizaveta — calling about the blue sweater from last week? A return or an exchange?"
  • 记住偏好与历史对话，真正一对一Remembers preferences and history; truly one-to-one
  • 自然语音交互回归（1–3 年内）Natural voice interaction returns (within 1–3 years)
  • 监控触发器，在投诉前主动解决Monitors triggers and resolves before complaints
① 核实1 · Verify

物流代理下午 3 点标记配送"失败"——查后台确认：配送车抛锚。A delivery is flagged "failed" — the backend confirms: the van broke down.

② 改约2 · Reschedule

登录物流系统，预约明早最早时段重新配送。Rebooks the first available slot tomorrow morning.

③ 补偿3 · Compensate

登录账单系统，自动发放 10 美元服务积分。Applies a $10 service credit automatically.

④ 告知4 · Notify

短信致歉并确认新时段；复杂或情绪化问题则"智能转接"人工并附完整摘要。Texts an apology and confirmation; complex cases get a smart handoff with a full summary.

Home Depot

"Magic Apron" 全天候 AI 专家：详细安装指南、产品推荐、评价摘要，让家装更简单。"Magic Apron" offers 24/7 expert guidance — how-tos, product recommendations, review summaries.

Danfoss

覆盖 100+ 国家的制造商：邮件订单处理自动化，80% 交易决策自动完成，响应从 42 小时到近实时，5 套系统合一。80% of transactional decisions automated; response from 42 hours to near real-time; five systems into one.

医疗 · 预测性健康系统Healthcare · Predictive system

Agent 整合影像、EHR 与理赔数据，把主动式洞察送入临床工作流，实现人群层面的前瞻性风险管理——从"被动响应"走向"预测性学习型健康系统"。Agents integrate imaging, EHR and claims to push proactive insights into clinician workflows — from reactive care to a predictive learning health system.

04趋势四 · Agentic SOCTrend 4 · The agentic SOC

安全运营：从"看警报"到"先出手"From watching alerts to acting first

攻击者只需对一次，防御者必须全对——而 82% 的分析师正被警报与数据过载淹没。"警报疲劳"正是攻击者最大的优势。SOAR 只能带来增量改善；能推理、行动、观察、调整的 Agent，才可能扭转攻防不对称。 Attackers need to be right once; defenders must be right every time — yet 82% of analysts are drowning in alerts. Alert fatigue is the attacker's greatest advantage. SOAR delivers only incremental gains; agents that reason, act, observe and adjust can flip the asymmetry.

警报Alert→ 检测Detection→ 分诊与调查Triage & investigation→ 威胁狩猎Threat hunt→ 恶意软件分析Malware analysis→ 检测工程Detection engineering→ 响应Response→ 建议 + 升级 · 人管Recommendation + escalation · human

评估→行动→再评估的动态循环让系统实时适应变化的环境。多个 SOC Agent 共享企业上下文与安全遥测数据，通过 A2A 与 MCP 互通，并持续吸收安全专家的实战经验。Google 的 Secure AI Framework 2.0 用于应对自主 Agent 带来的新风险。Agent 成为"力量倍增器"后，人的价值上移到三件事： The evaluate→act→re-evaluate loop adapts in real time. SOC agents share enterprise context and telemetry, communicate via A2A and MCP, and continuously absorb real-world expert insight. Google's Secure AI Framework 2.0 addresses the new risks of autonomous agents. With agents as force multipliers, human value moves up to three things:

威胁狩猎Threat hunting

用直觉与经验引导 Agent："我预感这台服务器有问题，找一下有没有异常的对外数据传输。"Guide agents with intuition: "I have a hunch about this server — hunt for unusual outbound transfers."

监督 AgentSupervising agents

调优 Agent 的"交战规则"，对其自动响应做绩效评审。Fine-tune agents' "rules of engagement" and performance-review their responses.

构筑防御Defending

聚焦长期安全态势：设计更好的防御架构、预判下一波攻击，而不是追着当前警报跑。Focus on long-term posture: architect better defenses, anticipate the next wave — not chase current alerts.

Torq · Socrates

AI SOC 分析师运行于 Google Cloud 基础设施：90% 一线任务自动处置、手动任务减少 95%、响应快 10 倍。90% of tier-1 tasks auto-remediated, manual tasks down 95%, response 10x faster.

Specular

进攻性安全平台：基于 Gemini 2.5 Pro 自动化攻击面管理与渗透测试，帮企业快速排序并响应威胁。Built on Gemini 2.5 Pro to automate attack-surface management and penetration testing.

CodeMender · DeepMind

AI 代码安全 Agent 已能在"测试充分"的软件中发现新零日漏洞——"随着攻击者把 AI 纳入作战，这将是我们迎接新挑战的最佳工具。"（Sandra Joyce，Google Cloud）An AI code-security agent that already finds new zero-days in well-tested software — "our best tool to meet this new challenge."

关键提醒（Francis deSouza，Google Cloud）：模型、数据与 Agent 正在急剧扩大企业攻击面——安全人才必须同时精通 AI 与安全两个领域，才能防住使用 AI 的攻击者。 A key reminder (Francis deSouza): models, data and agents dramatically expand the attack surface — security professionals must be deeply bilingual in AI and security to defend against AI-enabled attackers.

05趋势五 · 规模化靠人才Trend 5 · Scale via talent

最容易被忽视的趋势：瓶颈不是模型，是人The most overlooked trend: the bottleneck is people

盯着模型、平台与提示词是诱惑，但错过了最关键的要素——人。专业技能半衰期已缩至 4 年，科技领域仅 2 年；而"Agent 编排者""AI 参谋长"这类新角色，市场上根本没有现成人才（Shweta Maniar，Google Cloud）。从业者与决策者共识：技能决定能不能被录用、被晋升、把职业做长——同时直接拉动生产率、创新与营收。 It's tempting to fixate on models, platforms and prompts — that misses the most critical element: people. Skills expire in 4 years (2 in tech), and the market has no ready-made "agent orchestrators" (Shweta Maniar). Practitioners and decision-makers agree: skills drive hiring, promotion and career growth — plus productivity, innovation and revenue.

确立目标Establish goals

找到最重要且可测量的目标（如"100% AI 工具采用率"），并与组织大局对齐。Find what matters most and what you can measure (e.g., "100% AI tool adoption"), aligned to the bigger picture.

争取赞助Secure sponsorship

三角角色：高管赞助人（资金与背书）、氛围负责人（"AI 大喇叭"）、AI 加速器（把想法变成可用方案）。The trio: an executive sponsor (funding & backing), a groundswell lead ("AI megaphone"), an AI accelerator (ideas into solutions).

持续激励Sustain momentum

"数字枢纽"+游戏化排行榜持续收集与奖励用例；高管周报、季度奖项保持动能，同侪示范跨角色传播。A "digital hub" with gamified leaderboards; weekly exec emails and quarterly awards keep the story going.

融入工作流Integrate workflows

内部黑客松让创意赛出来，优胜概念可进入正式项目；Field Days 在实战场景中练会用 AI 工具（线上+线下）。Hackathons let ideas compete into the official program; Field Days practice new tools in realistic settings.

可信框架Trusted frameworks

安全是所有人的责任：培训哪些数据可用（与不可用）、识别 AI 社会工程等新型威胁。Security is everyone's job: train on what data can/cannot be used and how to spot AI-powered social engineering.

TELUS

96% 员工表示使用 AI 工具的信心提升并承诺在工作中应用；培训项目影响力从 2 月到 9 月翻倍。96% report higher AI confidence and commit to applying it; program impact doubled from February to September.

金融服务 · 内部先行Finance · inside-out

银行可先把客服 Agent 部署给内部员工试用——员工用真实问题测试，真正理解客户体验后再对外发布（Georgina Bulkeley，Google Cloud）。Roll out the support agent internally first — employees test with real questions to genuinely understand the customer experience (Georgina Bulkeley).

理念Philosophy

"AI 采用应当是一个持续的故事，而不是书里的一章。"（Gina Narvaez，TELUS 数字人才团队总监）"AI adoption should be an ongoing story, instead of just one chapter in a book." (Gina Narvaez, TELUS)

06数据附录Data Appendix

关键数据一览Key data at a glance

52%受访组织已将 AI Agent 投入生产环境organizations with agents in productionROI of AI 2025
88%早期采用者至少一个用例获得正 ROIearly adopters: positive ROI on ≥1 use caseROI of AI 2025
49% / 46%用例第一名客服 / 营销与安全运营#1 use case customer service / marketing & securityROI of AI 2025
82%分析师担心警报过载漏掉真实威胁analysts fear missing real threatsForrester 2025
95%Suzano 查询耗时下降（5 万员工）Suzano query-time reduction (50k staff)报告案例Report case
40 minTELUS 每次 AI 交互节省（5.7 万人）TELUS saved per interaction (57k staff)报告案例Report case
$1.3MElanco 规避的潜在生产力损失（每站点 2,500+ 文档）Elanco potential impact avoided (2,500+ docs/site)报告案例Report case
80%Danfoss 交易决策自动化，响应 42h→近实时Danfoss decisions automated; response 42h→real-time报告案例Report case
90% / 95%Torq 一线任务自动处置 / 手动任务减少，响应快 10 倍Torq tier-1 auto-remediation / manual tasks down, 10x faster报告案例Report case
96%TELUS 员工 AI 使用信心提升，培训影响力半年翻倍TELUS: higher AI confidence; program impact doubled报告案例Report case
4 / 2 年专业技能半衰期 / 科技领域skill half-life / in techForbes 2024
核心观点Core Insights

十大核心观点Ten core insights

意图式计算取代指令式计算Intent-based computing replaces instruction-based computing

员工陈述目标，Agent 决定执行路径——人机界面的范式转变，而非效率工具的迭代。Employees state the outcome; agents determine the path — a paradigm shift, not an efficiency patch.

员工角色重定义：从执行者到编排者From doers to orchestrators

核心职责变为设定策略、监督 Agent 系统；人始终是编排者与最终决策者。Set strategy, oversee agent systems; humans remain the orchestrators and final decision-makers.

Agent 已跨过生产临界点Agents have crossed the production threshold

52% 组织已投产；客服（49%）、营销/安全（46%）、技术支持（45%）领跑。52% in production; customer service (49%), marketing/security (46%), tech support (45%) lead.

企业转向"数字装配线"Businesses shift to digital assembly lines

多 Agent 端到端编排、7×24 运转；88% 早期采用者已见正 ROI。Multi-agent orchestration end-to-end, 24/7; 88% of early adopters see positive ROI.

互操作性是规模化的前提Interoperability is the prerequisite for scale

A2A 跨组织协作、MCP 接入实时数据与工具、AP2 支撑代理式商务——没有开放标准，装配线无法跨组织拼接。A2A for cross-organization collaboration, MCP for live data and tools, AP2 for agentic commerce — assembly lines cannot span organizations without open standards.

客服从脚本机器人进化为代理式礼宾From scripted bots to agentic concierges

基于企业上下文的记忆与个性化，加上主动式服务——在客户投诉之前解决问题。Enterprise context, memory, personalization — plus proactive service that resolves issues before customers complain.

安全从"警报响应"走向"主动行动"Security advances from alerts to action

82% 分析师受警报疲劳困扰；Agentic SOC 把分析师升级为战略防御者。82% suffer alert fatigue; the agentic SOC elevates analysts to strategic defenders.

AI 攻防两端同时升级AI escalates both offense and defense

Agent 扩大攻击面同时成为防御者最强武器（CodeMender 发现零日）；安全人才须"AI+安全"双语。Agents expand the attack surface while becoming the defender's strongest weapon; security talent must be bilingual in AI and security.

技能半衰期缩短，人才是最大价值杠杆Shrinking skill half-life makes upskilling the ultimate lever

技能半衰期 4 年（科技 2 年）；"Agent 编排者"无现成人才，须建五大学习支柱。4-year half-life (2 in tech); no ready-made "agent orchestrators" — build the five pillars of AI learning.

2026 的机遇本质是"人的解放"The 2026 opportunity is fundamentally human

把员工从重复低价值工作中释放，聚焦创造性、战略性与同理心工作；今天开始实验的企业正在积累管理、治理与规模化 AI 的内部能力。Free teams from repetitive work for creative, strategic, empathetic work; early experimenters build the expertise to manage, govern and scale AI.

行动路线Action Path

给决策层的自查清单A self-check for decision-makers

流程Workflow

选一条高价值流程，做装配线试点Pilot an assembly line on one high-value process

多系统、多步骤、规则清晰（工单 / 采购合规 / 安全分诊）；用 A2A + MCP 打通数据与工具，90 天内拿到可量化 ROI 证据。Multi-system, multi-step, rules-rich; wire via A2A + MCP, produce quantifiable ROI evidence in 90 days.

客户Customer

把"上下文数据"当作礼宾体验的地基Treat context data as the concierge's foundation

盘点 CRM、物流、账单数据能否按客户授权开放给 Agent；设计主动服务闭环与"智能转接"机制。Audit whether CRM, logistics and billing data can be shared per-customer with consent; design proactive loops and smart handoffs.

安全Security

评估 Agentic SOC 的三件事Assess three things for an agentic SOC

警报分诊自动化率、人机交接点设计、以及"AI + 安全"双语人才的招聘与培养计划。Auto-triage rate, human hand-off design, and your pipeline for bilingual "AI + security" talent.

人才Talent

把技能投资写进预算科目Put skills investment on the budget line

五大支柱逐项对照：可测量目标？三角角色到位？激励机制？黑客松与 Field Days？可信框架？Walk the five pillars: measurable goals? trio in place? incentives? hackathons & Field Days? trusted frameworks?

结语Closing

更快的公司，最终是更"人"的公司。A faster company is, ultimately, a more human one.

Agent 化的终点不是无人化，而是把人放到只有人能站的位置：创造、策略与同理心。员工的重新定义、流程的重新装配、客户体验的重新想象、安全防线的重新部署、人才体系的重新投资——五条线并进，才是 2026 的完整答卷。 The endpoint of agentic transformation is not headcount zero — it is placing humans where only humans stand. Redefining employees, rewiring workflows, reimagining customer experience, redeploying security, reinvesting in talent — five lines, advanced together, make the full 2026 answer.

数据来源：Google Cloud《AI Agent Trends 2026》，及其引用的 The ROI of AI 2025（n=3,466）、Forrester 威胁情报基准 2025、Forbes 2024。本页为独立解读汇编，与 Google 官方无隶属关系。 Sources: Google Cloud "AI Agent Trends 2026", incl. The ROI of AI 2025 (n=3,466), the Forrester Threat Intelligence Benchmark 2025, and Forbes 2024. This page is an independent editorial digest, unaffiliated with Google.