The Next Phase of Enterprise AI: From Deploying Models to Managing a Workforce of Digital Employees

According to IDC's forecast, China's AI agent market size will be approximately 212 billion yuan in 2025, reach an estimated 449 billion yuan in 2026, and grow to 3320 billion yuan by 2029, with a compound annual growth rate exceeding 100%. On the same basis, the global number of daily active agents is projected to rise from 2860 million in 2025 to 22.16 billion in 2030.

At first glance, these numbers suggest another AI boom. But the real story isn't about more models being deployed—it's that companies are now managing digital workers like employees: no badges, but they have accounts, access rights, and KPIs.

This article clarifies three key points: what large-scale digital workforce operations entail, the core pillars that support them, and how enterprises can assess their readiness today.

01 The real change isn't in the model—it's in the team.

Over the past two years, enterprises have approached AI through deployment: purchasing compute power, deploying models, and integrating a few use cases.

Since 2026, the landscape has shifted. Gartner predicts that by the end of 2026, 40% of enterprise applications will include task-specific AI agents. When an organization runs dozens, hundreds, or even thousands of agents, the challenge is no longer about model strength—it's a completely different set of questions:

Who grants them permissions? Who is accountable for their actions? Can their activities be audited? How is their effectiveness evaluated? How are errors remediated or rolled back?

In short, the next phase of enterprise AI isn't about model parameters—it's about operations.

02: From Tool to Productivity — The Three-Stage Leap

Integrating AI into enterprises typically involves three stages.

Phase 1: Tooling for point efficiency. Write copy, research info, generate code—humans use tools; tools do not own process accountability.

Phase 2: Agent — Task Automation. Performs complete workflows independently, such as auto-replies, automated reconciliation, and report generation.

Phase 3: Digital Employees — They become integral team members with defined roles, clear access boundaries, embedded workflows, performance evaluations, and fully traceable actions.

The first two stages address efficiency; the third addresses organizational capability. In manufacturing, a simple truth holds: AI's value lies not in how powerful the tools are, but in whether they can become productive force. The leap from tool to productivity is bridged by operations.

03 Scale operations with four pillars

Once digital employees scale up, effective management becomes a core capability. Breaking it down, four pillars support scalable operations.

First, embed into workflows. Digital employees are not tools left on the sidelines; they are integral parts of business processes—appearing in approval flows, ticket queues, and reconciliation streams. Upstream and downstream teams can see them and seamlessly integrate with them.

Second, permission governance. Clear rules must define who provisions digital employee accounts, what access levels they receive, and which data they can touch. Alibaba Cloud's AgentRun platform implements a three-tier multi-tenant structure for permissions, coupled with single sign-on and token quotas, to limit the operational scope of digital employees at the source.

Third: Data and Audit. Every action taken by digital employees must be logged. Qianxin's Intelligent Agent Security Platform creates a traceable chain linking who issued the command, what the agent did, and its impact, using five-dimensional tracing and immutable audit logs.

Fourth, operational closed loop. Digital employees require version control, performance evaluation, decommissioning and recycling, plus SLA agreements—Microsoft has realized this as an Agent 365 product, essentially a human resources system for digital workforce.

The 04 industry is voting with action

This isn't just a concept on paper. Over the past year, leading players have turned operations into a core product capability.

Alibaba Cloud launched AgentRun, offering governance capabilities such as multi-tenant permissions, token quotas, and SSO. Tencent categorizes enterprise agent maturity into four levels (L1–L4), enabling organizations to benchmark their current stage. Microsoft introduced Agent 365, unifying agent governance, deployment, and operations under a single management framework. Security vendors are also rapidly adopting this trend by establishing dedicated product lines for agent security.

These actions point to the same conclusion: as agent counts scale from single digits to hundreds or thousands, management concerns—security, permissions, and auditability—take precedence over model performance.

05 Manufacturing is the ultimate testing ground.

Among all industries, manufacturing may have the most direct perception of scalability.

One production line, one workshop, one order—each represents a tightly linked chain of processes. Digital employees here are not just optional assistants; they are fully integrated into R&D, process engineering, scheduling, quality control, and supply chain operations.

More critically, the two wheels are integrated: one drives production execution, the other manages business operations. Data flows bidirectionally between them—production data informs business decisions, while business decisions orchestrate production. Only when AI is embedded in both wheels does it truly evolve from a tool into productive power.

06 Determine if you're ready by answering four questions

For most enterprises, the priority now is not to stockpile more models, but to answer four key questions:

1. Is there a clear business owner? Whether digital employees are managed by IT or the business unit determines whether they can be truly integrated into workflows.

2. Is there an access control and audit baseline? Even for a pilot, you must first define what actions are permitted and ensure all activities are logged.

3. Are there quantifiable outcome metrics? There must be a measurable baseline for time savings, cycle reduction, or error rate decrease.

4. Is there an exit mechanism? The ability to reclaim and roll back underperforming agents determines the upper limit of trial-and-error costs.

Boundaries must also be upheld: data security, accountability, and industry evaluation standards are still evolving. Any claim of fully automated scaling warrants skepticism. Digital employees in demo environments differ significantly from those in real-world operations, separated by an entire operational framework.

Final thoughts

What truly matters is not how many models we deploy, but whether an enterprise has the capability to manage a workforce of digital employees that continuously evolves.

Models can be bought, compute can be rented—but an operational system cannot. The first to build management capabilities for this team secures the entry ticket for the second half of enterprise AI.