Executive summary: agentic operating model

企业组织的核心变化不是“少一些人”，而是管理跨度、决策权和能力边界同步重构。

01 | Adoption has crossed the threshold

75% 知识工作者已在工作中使用生成式 AI；78% AI 用户自带工具进入工作场景，说明组织正式治理落后于实际使用。

75%

global knowledge workers

02 | Agents become a new labor layer

Gartner 预测 2028 年 33% 企业软件将内置 agentic AI；IBM 调研中 99% 企业 AI 开发者正在探索或开发 AI agents。

33%

enterprise apps by 2028

03 | Org chart shifts to decision flows

从岗位职责转向任务包 ownership
从层级审批转向 guardrail-based delegation
从 IT 项目转向业务域 agent portfolio

04 | Case evidence is already visible

Microsoft	top 5% Teams users saved 8h meetings/month via Copilot summaries
Klarna	AI assistant reported doing work equivalent to 700 FTE in support interactions

Implication map: where organization design changes first

Source: Microsoft Work Trend Index 2024; Gartner press release Jan 2025; IBM/Morning Consult 2025; Klarna company update 2024 | Estimates and reported metrics

Adoption pressure: bottom-up AI is outrunning formal operating models

When 78% of AI users bring their own tools, the enterprise has already reorganized informally.

Adoption funnel in knowledge work

Leadership bottleneck

AI needed for competitiveness	79%
Worry productivity gains hard to quantify	59%
Worry leadership lacks plan / vision	60%

Gap: adoption is individual; value capture requires redesigned teams, data access, governance and process ownership.

Risk moves from IT to organization design

Shadow AI creates unmanaged data pathways
Managers cannot see true productivity baselines
Career signals change: AI aptitude begins to rival tenure
New control points sit in business processes, not only model platforms

Case: Microsoft 365 Copilot changes meeting economics

Microsoft observed the heaviest Teams users (top 5%) summarized eight hours of meetings in March using Copilot — roughly one workday/month. The structural implication is fewer “information relay” meetings and more asynchronous decision artifacts owned by agents.

Management question

The first design problem is not choosing a tool; it is deciding which decisions may be delegated, which work products become agent-generated, and which human roles become reviewers, trainers, exception handlers or orchestrators.

Inventory high-volume decision loops
Define authority boundaries by risk tier
Reprice roles around judgment and accountability

Source: Microsoft & LinkedIn Work Trend Index | 2024 global survey of 31,000 people in 31 countries; Microsoft 365 productivity signals

Operating model: from hierarchy to human-agent mesh

AI agents insert a new execution layer between strategy, processes, systems and frontline work.

Target-state architecture

Design primitives

Agent as digital worker, not chatbot
Task API replaces job description
Human owner per decision domain
Audit trail as default work product

Governance split

Low risk	Auto-execute
Medium	Human approve
High	Human decide

Span of control expands selectively

Managers can supervise more work because agents make status, exceptions and drafts visible. But spans expand only where processes are instrumented and decision criteria are explicit.

+20–40%

illustrative capacity uplift in codified workflows

Middle layer changes shape

Coordination-heavy roles shrink; “agent product owner”, “workflow architect”, “AI risk controller” and “data steward” become repeatable roles embedded in business units.

What does not change

Accountability remains human and managerial
Strategy ambiguity still requires executive judgment
Culture and incentives determine adoption quality

Source: IBM Think 2025; Gartner 2025 agentic AI forecast; McKinsey/QuantumBlack State of AI 2024–2025 | Synthesis by author

Workforce impact: jobs decompose into automatable tasks and judgment nodes

Agents alter task mix faster than they eliminate whole roles; redesign must start at activity level.

Task decomposition by role archetype

Blue=automatable task share; cyan=judgment / exception share (illustrative).

WEF labor-market context

The Future of Jobs 2025 reflects >1,000 employers representing 14M workers across 22 sectors and 55 economies, with technology change one of the main drivers through 2030.

2025–2030

re-skill window

Role redesign rules

Map tasks, not headcount first
Move humans to exceptions, judgment, relationships
Retain “process memory” during transition
Budget reskilling as operating expense, not training event

Case: customer service

Klarna reported its AI assistant handled two-thirds of customer service chats and did work equivalent to 700 full-time agents. The redesign pattern: agents absorb routine resolution; humans handle emotional, regulatory and edge cases.

Case: software engineering

GitHub Copilot research found developers completed a coding task 55.8% faster with Copilot in a controlled experiment. Structure shifts toward smaller teams with higher review, integration and architecture responsibility.

New capability stack

AI fluency	prompt, critique, evaluate
Process literacy	codify steps and exceptions
Risk literacy	know escalation triggers

Source: World Economic Forum Future of Jobs Report 2025; Klarna company communications 2024; GitHub/Microsoft Copilot research 2023–2024

Governance: authority design becomes the critical control system

The operating model must specify what an agent may know, decide, do and learn from.

Delegation matrix

RACI becomes RAACI

Responsible	Human / agent
Accountable	Named human
Agent	Tool + scope + logs
Consulted/Informed	Workflow subscribers

Control objectives

Prevent unauthorized tool use and data exfiltration
Make decisions explainable after the fact
Separate test, approval and production execution
Measure drift in quality, cost and risk

EU AI Act benchmark

EU AI Act entered into force in 2024 with staged obligations. Even outside Europe, it creates a reference model: classify use cases by risk, document controls, and define human oversight for high-risk systems.

Metrics to install

Automation rate	% tasks agent-completed
Exception rate	% escalated to humans
Loss rate	errors × business impact

Board-level question

Which business decisions can be executed by software under management-approved policy, and how will the board know when that authority is creating unacceptable concentration, bias or control risk?

Source: European Commission AI Act timeline 2024; IBM Think 2025; Gartner agentic AI governance commentary 2025 | Governance synthesis

Functional archetypes: impact varies by repeatability and data access

Agents scale first in functions with repeatable workflows, digital records and measurable outcomes.

Functional impact heatmap

Where value pools form

Customer service: resolution + routing
Sales: lead research + next-best action
Finance: reconciliation + commentary drafts
HR: employee queries + skills matching

Org structure response

Each function needs a small “agent cell”: process owner, AI engineer / builder, data steward, risk partner and frontline champion. Cells should report to business leaders, with shared platform standards from IT.

Case: Salesforce Agentforce

Salesforce positions Agentforce as autonomous agents embedded into CRM workflows across service, sales, marketing and commerce — evidence that enterprise software vendors are packaging agents by business function, not by technical component.

Case: ServiceNow / enterprise workflows

Workflow platforms are moving from ticket routing to agentic resolution. Organization implication: shared-service teams become exception-management centers with agent operations dashboards.

Decision criterion

Prioritize functions where outcome can be measured weekly, access rights are explicit, and escalation paths are clear. Avoid starting where accountability is diffuse or data lineage is weak.

Source: Salesforce Agentforce product materials 2024–2025; ServiceNow AI agents announcements 2024–2025; Microsoft Work Trend Index 2024 | Market synthesis

Economics: value capture requires moving from pilots to portfolio management

The unit of investment shifts from “model project” to “agent portfolio with accountable P&L outcomes”.

GenAI value architecture

BCG framing: majority of value depends on people/process transformation.

Portfolio dashboard

Service deflection	cost + CX
Sales research	conversion
Finance close	cycle time
Procurement sourcing	savings

Funding model

Move from annual AI budget to rolling venture-style portfolio: kill weak pilots within 8–12 weeks; scale agents that demonstrate quality, cycle-time or cost improvements with measurable baselines.

Cost side

Model inference and orchestration
Human review / exception handling
Data integration and permissions
Monitoring, audit and risk testing

Benefit side

Cycle-time compression
Labor reallocation to high-value tasks
Higher consistency of routine decisions
Revenue lift from faster response

CFO test

Do not approve an agent unless it has a business owner, a baseline metric, an exception budget, a kill criterion and a named accountable human for adverse outcomes.

Source: BCG From Potential to Profit with GenAI 2024; Microsoft Work Trend Index 2024; McKinsey State of AI 2024–2025 | Economics synthesis

Roadmap: redesign the organization in waves, not as a big-bang restructure

The practical path is controlled scaling: codify work, delegate low-risk tasks, then rewire roles and incentives.

12-month transition roadmap

Wave 1: work inventory

Map top 50 workflows by volume and cost
Identify decisions, data, tools and exceptions
Define AI risk tiers and owners

Wave 2: agent cells

Stand up 3–5 domain agent cells
Run 8-week pilots with kill metrics
Publish playbooks and audit logs

Wave 3: role migration

Move people from repetitive drafting, routing and reconciliation into exception handling, customer judgment, data stewardship and agent quality management.

Wave 4: incentives

Performance metrics must reward safe delegation and value capture, not just individual output. Managers should be evaluated on portfolio ROI, quality and control incidents.

Operating cadence

Weekly	quality + exception review
Monthly	portfolio funding decisions
Quarterly	org design / role updates

Source: Microsoft Work Trend Index 2024; WEF Future of Jobs 2025; IBM/Morning Consult 2025; Author implementation synthesis

Summary: six decisions executives should make now

Agentic AI is an organization design problem disguised as a technology wave.

1 | Define delegation boundaries

Which decisions can agents execute autonomously, with approval, or never?

2 | Create domain agent cells

Place small cross-functional teams inside business functions, not only in IT.

3 | Rebuild role architecture

Move people toward judgment, exceptions, customer trust and agent supervision.

4 | Install portfolio governance

Fund, scale and kill agents based on explicit P&L, quality and risk metrics.

5 | Upgrade management systems

Measure process outcomes, exception rates and control incidents weekly.

6 | Reskill around AI aptitude

Make AI fluency and process literacy part of career progression.

Recommended next step: launch a 2-week organization diagnostic covering 50 priority workflows, then select 3 agent cells for 8-week controlled pilots with board-visible metrics.

Source: Synthesis based on Microsoft 2024, Gartner 2025, IBM 2025, WEF 2025, BCG 2024, EU AI Act 2024, company case disclosures

AI Agent 对企业组织架构的影响