AI 智能體進入組織的速度快於營運模式建立問責安排

Organizations can deploy agent capability faster than they can make delegated work accountable.
AI agents are entering organizations as technology, but they are changing organizations as participants in work.

Most executive conversations still begin with capability. What can the agent do? How quickly can it be deployed? Which platform should the organization choose? How many workflows can be automated?

Those are reasonable questions.

They are no longer sufficient.

Once an agent can receive an assignment, use tools and information, create or change a work product, hand work to another participant, encounter exceptions, and influence a consequential outcome, the organization is no longer managing only a software feature.

It is managing a new participant inside a work system.

That transition is happening faster than many operating models can absorb.

A fictional but familiar scenario
Consider a corporate-lending team preparing for a client renewal.

An AI agent has been connected to relationship notes, financial information, risk indicators, prior credit documents, and market research. It produces a polished briefing in minutes.

The briefing presents an old expansion plan as a current client priority. It also turns an analytical signal about working-capital pressure into a statement of fact.

An experienced relationship manager catches both problems before the material reaches the client and rebuilds the briefing manually.

Nothing harmful leaves the organization.

It is tempting to call this a successful human review or a simple model-quality issue.

But when leaders ask what happened, the operating gap becomes visible.

The business thought technology had authorized the data sources. Technology believed the business owned the use. Data governance could identify the records, but not why an inference had been presented as fact. The reviewer had no explicit acceptance criteria. Nobody could point to the complete evidence chain that should have accompanied the briefing.

The agent produced the error.

The operating model determined whether the organization could detect, challenge, contain, and learn from it.

That distinction matters.

Governance has to reach the decision
Across regulated data, analytics, architecture, and transformation work, I learned that organizations can govern individual components well while still losing accountability between them.

A dataset can have ownership, lineage, quality rules, access controls, and an approved platform while the decision built from it still lacks clear purpose, accountable acceptance, or sufficient evidence.

Governed data does not automatically create a governed decision.

The missing element is often not another control on the data or model. It is the operating connection between trusted inputs and consequential use.

Someone must define what the work is trying to accomplish. Participants must operate within explicit boundaries. Evidence must travel with the work. The level of challenge must match the consequence. And someone must remain accountable for how an outcome becomes acceptable for use.

Agentic AI makes this old problem harder to ignore because it can compress the distance between information, inference, action, and decision.

Governance therefore has to travel through the work, not stop at the data or model.

The organization acquired capability, not a work system
Most organizations already know how to govern components.

Security governs access. Data teams govern information. Model-risk functions evaluate certain models. Technology governs platforms and integrations. Business teams own processes and customer outcomes. Procurement governs vendors. Audit tests selected controls.

Each function can perform its own job responsibly while the end-to-end work remains unowned.

The dangerous space is often between the controls:

The outcome is described as “produce a briefing” rather than defining what makes the briefing acceptable for its intended use.
The agent can retrieve information, but source eligibility and purpose are not bound to the assignment.
A human is expected to review the result, but the decision, evidence, authority, independence, and time required for that review are not designed.
A polished output exists, but its handoffs, intermediate states, uncertainty, and exceptions cannot be reconstructed.
A pilot is considered successful because the agent completed a task, even though the complete service required hidden human repair.
These are not isolated technical defects.

They are operating-model defects.

A feature is not the right management metaphor
A conventional feature waits for a person to use it.

An agent can participate across time. It can receive a goal, select steps, call tools, use context, produce artifacts, hand work to another participant, encounter exceptions, and change what a person or system will do next.

Calling an agent a participant in work does not make it human, an employee, or a legal person.

It establishes a management boundary.

Any participant that can materially affect an outcome must be visible, assigned, bounded, observed, and governed in proportion to the consequences of its work.

The same principle applies to human participants, deterministic systems, specialist teams, external providers, and assurance functions.

They are not identical.

Become a Medium member
But they all participate in the complete work system that produces — or fails to produce — the outcome.

Press enter or click to view image in full size

From agent capability to accepted outcome: dependable enterprise work emerges from the work system between them.
The unit of management has to move
If leaders manage the agent as the primary unit, they will naturally measure model performance, task completion, latency, cost per call, adoption, and the number of deployed use cases.

Those measures are useful.

None of them proves that the organization can rely on the resulting work.

The better unit of management is the accepted outcome: a defined work product or service result that satisfies explicit acceptance criteria for its intended use through an accountable acceptance mechanism.

That distinction is important.

Human accountability does not require human approval of every transaction.

For some work, acceptance may be automated within clearly delegated authority and explicit tests. Other work may require exception-based review, independent challenge, or direct human acceptance because the consequence is greater.

What matters is that the organization has deliberately designed how acceptance occurs, who is accountable for that design, what evidence supports the decision, and when human intervention becomes mandatory.

The accepted outcome is produced by a work system: the end-to-end arrangement of participants, information, tools, rules, handoffs, evidence, controls, and decisions required to create, challenge, accept, recover, and change the work.

This shifts the executive conversation.

Instead of asking only whether an agent can draft a credit memo, leaders ask whether the complete mixed workforce can produce an acceptable memo under explicit authority, using eligible sources, with reconstructable evidence, an appropriate level of challenge, an accountable acceptance mechanism, and a safe response when conditions change.

That is a much harder question.

It is also the question that determines whether agent capability can become dependable enterprise work.

This is an enablement argument
Operating models are sometimes treated as structures that slow innovation down.

For agentic work, the opposite should be true.

Organizations cannot delegate consequential work confidently when every exception depends on an improvised meeting, every reviewer must reconstruct the evidence, and every team interprets authority differently.

That design creates approval queues, hidden supervision, rework, brittle pilots, and autonomy that disappears as soon as something unusual happens.

Explicit outcomes, bounded authority, evidence-bearing handoffs, designed acceptance mechanisms, and safe states allow more work to proceed without constant executive intervention.

Good operating design gives leadership the conditions under which it can responsibly say yes to more autonomy.

The purpose of the operating model is not to eliminate autonomy.

It is to make greater autonomy supportable.

Five questions for the next executive review
A leader does not need a complete enterprise framework to expose the operating gap.

Start with one consequential agent-enabled workflow and ask:

1. What exact outcome is the combined workforce authorized to produce, and what makes that outcome acceptable for its intended use?

A task description is not enough. The organization needs explicit criteria for when the result can actually be relied upon.

1. Who or what participates in producing the outcome, and under whose responsibility?

That includes people, agents, deterministic systems, specialist teams, external providers, and assurance functions.

1. Which information, tools, actions, purposes, and decisions are permitted for this assignment?

Access alone does not establish authority. The organization must define what a participant may do with what it can reach.

1. What evidence must travel with the work, and what level of challenge does the consequence require?

The organization should be able to reconstruct why the outcome exists, what information supported it, what assumptions were made, and what form of challenge was appropriate.

1. Who owns acceptance of the outcome, when may acceptance be automated, and what conditions require human intervention, rejection, pause, or stop?

Human accountability and human approval are not the same thing. The operating model should determine when automation is supportable and when consequence requires a human decision.

If the organization cannot answer those questions, the next priority may not be another agent, platform feature, or policy.

It may be the operating design that connects capability to accountable work.

The leadership question has changed
AI strategy often asks where agents can be deployed.

The more consequential question is where the organization is prepared to let them participate in work.

That requires leaders to see beyond the agent.

The real object of management is the complete system that turns intent into an accepted outcome — including the people and systems that direct, make, challenge, accept, intervene, recover, and learn.

Organizations will continue acquiring agent capability quickly.

Their advantage will depend on whether their operating models can turn that capability into work the enterprise can responsibly trust.

Take one consequential agent-enabled workflow in your organization. Which of these five questions would be hardest to answer today?

This article is part of Operating the Mixed Workforce, a leadership series on accountable work across humans and AI agents. It is inspired by my forthcoming book of the same name. The series explores how organizations can redesign authority, accountability, governance, and operating models as AI agents become participants in enterprise work.