Enterprise Agents are facing challenges in large-scale deployment.
McKinsey disclosed in 2026 that 62% of enterprises globally are experimenting with AI Agents, yet only 23% have truly achieved large-scale deployment. The domestic market is at a similar turning point. According to research by iFenxi, 56% of enterprises in China are in the pilot quick-win stage, while only 13% have entered the full-scale rollout phase. Management's focus is gradually shifting toward scenario selection, implementation paths, scalable replication, and value realization.
These shifts indicate that enterprise Agents are transitioning from "capability readiness" to "business readiness". Bridging this gap requires navigating three critical checkpoints: business context, collaborative execution, and value feedback, ultimately forming a sustainably operational production system across data, processes, organization, and governance.
01 、From Assisted Decision-Making to Business Closed-Loop: Four Stages of Enterprise Agent Deployment
The integration of large models into enterprises merely marks the starting point for Agent deployment. As execution authority, process scope, and business responsibilities expand, enterprise Agents generally evolve into four distinct forms.
Copilots provide information, suggestions, and content, leaving the final judgment and execution to employees. They are well-suited for integration into high-frequency tasks such as knowledge retrieval, call summarization, tag extraction, and content generation, with their value primarily reflected in enhanced individual productivity.
Task-oriented Agents begin to invoke tools, capable of completing tasks with well-defined boundaries, such as analyzing customer conversations over a specific period, identifying potential risks, or generating work orders. At this stage, the Agent possesses a certain degree of autonomous action, yet task objectives, inputs/outputs, and operational pathways remain relatively fixed.
Collaborative Agents engage in longer business processes, working alongside employees or other Agents to accomplish tasks. They must determine when to execute independently, when to request confirmation, and when to hand over to other roles. Consequently, employees shift from hands-on operations to review, judgment, and exception handling. The primary challenge at this stage often stems from legacy processes: enterprise SOPs, system interfaces, and approval mechanisms are predominantly designed for human workflows. With the introduction of Agents, task allocation and handover protocols require fundamental restructuring.
Business-oriented Agents further drive processes centered on operational objectives, using business outcomes as the basis for evaluation. In scenarios such as meal ordering, appointment scheduling, customer callbacks, and sales lead nurturing, the Agent can start by understanding requirements and seamlessly proceed through decision-making, tool invocation, task execution, and result feedback. The metrics for evaluating these Agents expand beyond response accuracy to include business KPIs such as task completion rate, conversion rate, work order resolution time, and customer coverage scale.
These four forms will coexist within enterprises for the long term. The key differentiator among them is the extent of action and responsibility the Agent can assume. The closer the Agent operates to a business closed-loop, the more employees must concentrate their efforts on objective management, strategy formulation, complex decision-making, and risk mitigation.
02 、Three Checkpoints for Large-Scale Deployment: Context, Collaborative Execution, and Value Feedback
Once Agents enter core business operations, they must complete a continuous chain from comprehension to action and finally to result feedback. A break in any segment of this chain prevents model capabilities from being reliably translated into business value.
The disconnect in business context is the first to emerge. Customer profiles, orders, contracts, work orders, conversation logs, and operational metrics are scattered across CRM, ERP, contact centers, and knowledge bases, each with distinct data structures and definitions. Even if an Agent can access these systems, it may fail to grasp the underlying business relationships among the data. Without comprehensive context, it is prone to generating judgments that are locally accurate but globally distorted.
The disconnect in collaborative execution is more insidious. Legacy systems and SOPs dictate how employees input, query, approve, and hand over tasks, yet rarely define the permission boundaries and takeover mechanisms for Agents. If Agents can only provide recommendations outside the business systems, employees must still manually transfer information; conversely, if systems directly grant execution permissions, it may introduce compliance and operational risks. Clear task delineation is required between Agents and employees, ensuring that customer information, historical conversations, rationale for decisions, and task statuses are synchronized during handovers.
The disconnect in value feedback determines the sustainability of the project. Metrics such as intent recognition rate, response speed, and model invocation costs can reflect technical performance but fail to directly demonstrate business ROI. If the outcomes of Agent execution do not feed back into knowledge bases, strategies, and models, the system struggles to achieve continuous optimization. Enterprises must translate operational objectives into executable and trackable tasks for Agents, and subsequently feed the results back into the next cycle of decision-making.
At this stage, scenarios characterized by clear objectives, measurable outcomes, high labor input, and the ability for timely human takeover are more likely to form a value closed-loop. Such boundaries help enterprises control deployment risks while accumulating verifiable experience for the subsequent expansion of Agent authority.
03 、Building an Agent Production System: Integrating Intelligence into Business and Human Collaboration
To address these three checkpoints, enterprises need to build an Agent production system that connects business comprehension, task execution, and result feedback, enabling Agents to reliably participate in daily operations.
Cloopen's full-stack Agent product architecture illustrates a viable construction approach. The foundational layer integrates communication, model, and data capabilities; the middle layer consolidates business processes, business knowledge, industry skills, and scenario-specific experience; and the top layer hosts specific applications such as Collaborative Agents, Voice Agents, Quality Assurance Agents, and Private Domain Operations Agents, further serving diverse industry scenarios.
Building Agents in isolation around single scenarios often leads to redundant data integration, poor capability reuse, and inconsistent governance standards. A unified foundation with shared business capabilities helps reduce subsequent replication costs.
The key to resolving the context disconnect is transforming fragmented data into business objects and their relationships that Agents can comprehend and invoke. Clarifying how customers relate to orders, the status of contracts, the completion of historical work orders, and the implications of operational metrics are all essential to forming a clear business context.
The business semantic layer undertakes this task. It extracts core objects such as customers, orders, contracts, work orders, and metrics from disparate systems, uniformly describing their attributes, relationships, and states, before providing comprehensible and callable data to the Agent. Cloopen's business semantic layer architecture connects large model capabilities, data semantics, and legacy business systems: the large model handles comprehension, reasoning, planning, and generation; the semantic layer describes customers, orders, contracts, profiles, and metrics; and the foundational layer maintains connections to CRM, work order, financial systems, and knowledge bases.
Following business comprehension, Agents must translate objectives into concrete actions. Different tasks impose varying requirements on autonomy and determinism. Open-ended communication can rely more heavily on prompts, preserving the model's space for comprehension and generation; tasks with relatively explicit rules can be consolidated into skills; and processes with strict compliance requirements and fixed execution paths are best controlled via workflows. Enterprises can establish action boundaries of varying strictness for Agents based on task complexity, risk levels, and the degree of openness.
These actions must also integrate into the enterprise's existing business systems. Agents can perform information extraction, knowledge retrieval, content generation, automated form filling, and process monitoring within CRM, contact centers, and work order platforms. Employees continue to use familiar workbenches, with their operational focus gradually shifting from data entry to review and confirmation, complex judgment, and exception handling. Consequently, business systems become a shared workspace for Agents and employees.
Human-Agent collaboration naturally emerges during this execution process. The system must clearly define which tasks can be completed independently by the Agent, which nodes require employee confirmation, and under what conditions tasks should be escalated to human handlers. Agents are well-suited for highly repetitive, rule-based, and data-intensive tasks, while employees focus on objective setting, key customer communication, complex decision-making, and risk mitigation. During handovers, customer information, historical conversations, decision rationale, and task statuses must be synchronously transferred to prevent employees from having to reconstruct the context.
Governance mechanisms permeate the entire Agent execution lifecycle. Prior to task initiation, access management restricts the data the Agent can access and the tools it can invoke; during execution, quality assurance and auditing continuously log the decision-making process, promptly triggering alerts or circuit breakers in the event of anomalies; upon task completion, employee confirmations, customer responses, and business outcomes enter the feedback loop to refine knowledge, rules, and strategies.
Feedback outcomes are critical to forming a closed-loop in the production system. Enterprises must translate business metrics such as conversion rates, task completion rates, work order resolution times, and customer coverage scales into Agent objectives, subsequently adjusting the next cycle of actions based on execution results. Cloopen's publicly available business closed-loop engine architecture integrates professional knowledge, business processes, KPI assessments, industry practices, and outcome evaluations into a unified system, ultimately driving result delivery.
When comprehension, decision-making, execution, and feedback operate in a continuous loop, enterprises acquire a sustainably optimizable Agent production system, paving the way for Agents to enter more business scenarios and manage longer processes.
04 、Cloopen Practice: Exploring the Value Closed-Loop of Two Types of Business Agents
Cloopen has long served contact centers, customer service, and marketing operations, accumulating extensive customer interaction data, business process insights, and agent operational experience. Its recently launched Voice Agent and Private Domain Operations Agent target real-time interaction and long-term customer nurturing, respectively, providing two concrete examples for the aforementioned production system.
The Voice Agent targets voice scenarios such as marketing, customer service, appointment scheduling, and callbacks. It must continuously monitor customer status during calls, generate communication strategies, dynamically adjust its expressions based on feedback, and invoke business systems to complete tasks. Voice interaction is highly real-time, with customer intents and emotions subject to sudden changes, making human-Agent handover particularly critical. The Agent can handle mass outreach, intent recognition, and standardized tasks; when encountering complex issues, high-value customers, or critical conversion nodes, human agents intervene promptly, leveraging the existing conversation context to continue the service.
According to research by iFenxi, Cloopen's Voice Agent currently processes an average of approximately 20,000 conversations online daily, achieving an overall conversation completion rate of about 94%. In voice marketing scenario calculations, a human agent working an 8-hour shift typically handles 150 to 300 calls at a monthly cost of roughly 5,000 to 8,000 RMB; in contrast, the Agent can operate continuously at a monthly cost of approximately 2,000 to 3,000 RMB. Exact costs and completion rates vary depending on scenario complexity, task definitions, and the human takeover ratio, but these figures effectively illustrate the marginal benefits of large-scale execution.
The Private Domain Operations Agent manages longer-cycle customer relationships. It integrates membership, product, transaction, and interaction data to build dynamic customer profiles, subsequently generating operational strategies focused on objectives such as activation, repurchase, and sales volume. It maintains continuous engagement via Enterprise WeChat, official accounts, mini-programs, apps, SMS, and emails. Every piece of customer feedback alters their status, requiring the Agent to adjust communication timing, content, and recommendation strategies accordingly.
Taking the health and wellness scenario as an example, after a user purchases a three-month supply of fish oil, the Agent can schedule a callback based on the purchase date, inquire about usage, answer product-related questions, and drive repurchases at the optimal time. While it is challenging for operational staff to maintain relationships with a massive customer base simultaneously, the Agent can maintain continuous records and deploy differentiated strategies for each individual. In a case disclosed by Cloopen, the Agent helped manage approximately 80,000 private domain customers, generating a monthly GMV of around 1.2 million RMB, representing a GMV increase of approximately 87%.
These two products exemplify two types of business closed-loops. The Voice Agent centers on a single real-time interaction, completing identification, communication, execution, and handover within a shorter cycle; the Private Domain Operations Agent focuses on continuous insight and operations throughout the customer lifecycle, influencing repurchases and business outcomes over a longer period. They share underlying capabilities such as data integration, industry knowledge, strategic decision-making, tool invocation, human-Agent handover, and performance evaluation.
According to research by iFenxi, business Agents still exhibit clear capability boundaries. In certain business scenarios, Agents can achieve approximately 70% to 80% of the performance level of top-tier human agents. Complex emotional handling, non-standard issues, and high-risk decision-making still require human involvement. This division of labor aligns better with current enterprise deployment conditions. Agents scale up service volume while human agents safeguard experience, judgment, and risk boundaries, with both jointly delivering business outcomes.
05 、The Future of Enterprise Agents: Human-Agent Collaboration, System Restructuring, and Value Pricing
As Agents penetrate core business operations, enterprise software, business processes, organizational structures, and business models will evolve in tandem.
Enterprise software will gradually transition from primarily serving manual operations to concurrently serving both employees and Agents. Beyond data entry and process routing, systems must provide Agent-recognizable business objects, callable tool interfaces, explicit permission rules, and traceable operational logs. Whether an enterprise's digital foundation can support Agent actions will directly dictate their scope of applicability.
Industry knowledge will become an increasingly critical differentiating asset. While foundational models provide general comprehension and reasoning capabilities, it is enterprise knowledge, industry rules, scenario-specific data, and best practices that determine an Agent's performance in specific business contexts. As these assets accumulate, the gap between vendors will increasingly manifest as differences in industry acumen and business engineering capabilities.
Human-Agent collaboration will further drive the restructuring of organizational mechanisms. Enterprises must redesign job responsibilities, approval nodes, handover standards, and performance evaluations to ensure employees and Agents collaborate toward unified business objectives. As tasks continue to be decomposed, multi-Agent systems and human employees may form more complex collaborative networks, separately handling insights, communication, execution, quality assurance, and exception handling.
The scope of governance will expand from individual Agents to multi-Agent collaborative networks. Cross-Agent permission inheritance, accountability tracking, anomaly isolation, and auditing mechanisms will become new governance priorities. Enterprises must also define clear information transmission boundaries between different Agents to prevent localized errors from amplifying across the collaborative chain.
Once business outcomes become quantifiable, the commercial models for Agents may evolve accordingly. Scenarios such as marketing and customer operations may pioneer pay-for-performance and revenue-sharing models, forging a more direct link between vendor value and metrics like conversion rates, repurchase rates, and operational growth. This will propel Agent vendors from software capability delivery toward continuous operations, while also requiring enterprises and vendors to jointly define data definitions, responsibility boundaries, and risk-sharing mechanisms.
The focal point of competition in the next phase of enterprise Agents will be business depth. While models provide general capabilities, data semantics, process restructuring, human-Agent collaboration, and result feedback determine the extent of the business chain an Agent can penetrate. Only enterprises capable of integrating these elements into a cohesive production system will have the opportunity to translate localized efficiency gains into sustained business value.