Agent: Oranges and Tangerines: 171% Return, 86% Failure Range?

The same kind of Agent project, born in Huainan or Huaibei, meets a completely different fate. — Shui Ge

Research by IDC and Microsoft shows that AI Agents that successfully reach production scale achieve a global average return of 171%, and the U.S. market reaches as high as 192%.

Yet data from Forrester and Anaconda reflects the other side of the same market: 86% to 88% of Agent pilots never make it into production. The high commercial returns of Agents stand in sharp contrast to the widespread failure of mass production.

The question is no longer whether AI can create value, but through what vehicle and under what governance conditions an organization can stably absorb the output of Agents.

Tang Daosheng, Senior Executive Vice President of Tencent Group, echoed this at the 2026 Global Digital Ecosystem Summit: "The biggest problem for enterprises deploying Agents today is not that the models aren't strong enough, but that there are too many breakpoints."

I. Broken Transmission: Individuals Got Faster, Enterprises Did Not

McKinsey's 2026 global survey provides the most precise statistical skeleton for this fault line: 80% of respondents report that AI has improved their personal productivity, but only 37% of organizations can attribute at least part of their EBIT impact to AI, a proportion exactly flat with the previous year. Only about 6% are true high performers that attribute at least 5% of EBIT to AI with a "significant" impact. The most notable difference between high performers and other organizations is not models or budgets, but workflow redesign. Nearly three-quarters of high performers fundamentally redesigned their workflows, versus only about one-quarter of other organizations.

Personal efficiency measures "whether a person does things faster"; corporate profit measures "whether a process produces attributable, repeatable, auditable results." There is no automatic transmission pipeline between the two; it must be built through workflows, permission structures, and accountability relationships.

When Agents move from "answering" to "executing," the failure mode undergoes a qualitative change. Errors by personal Agents are locally reversible; errors by service Agents propagate along the process. But the control layer for service Agents is being built far faster than auditability. A survey by Nylas in the second half of 2026 shows: 54% of enterprises have experienced Agent errors in the past 90 days, the most common consequence being triggering workflows that require teams to execute in reverse; only 53% of enterprises can fully produce audit logs; and the incident rate accelerates with deployment scale, at 45% for organizations with 1 to 5 Agents and 70% for organizations with more than 20 Agents. A Deloitte survey shows that only 21% of organizations have established mature governance models for autonomous Agents. Gartner predicts that by the end of 2027, more than 40% of Agentic AI projects will be canceled due to rising costs, unclear business value, or insufficient risk control.

The speed at which Agents are deployed will ultimately not exceed the speed at which they are accepted, audited, and attributed. Wu Yongjian, Vice President of Tencent Cloud and Head of R&D for Tencent Cloud Intelligence, put this boundary more concretely: Personal Agents are centered on the individual, with users commanding them and judging the results themselves. Service Agents stand from the enterprise's perspective and must consider different permissions, different security, and who they can be shared with. The former is a 2C business; the latter is a 2B business.

Microsoft CEO Satya Nadella's argument likewise emphasizes the importance of separating permissions: "The most trusted agentic systems will not be the ones whose models we trust the most, but the ones designed so that we can trust their models the least." The core principle is "separating the supply of intelligence from the permissions over intelligence"; the controls that determine what an Agent can access and what actions it can take must lie outside the model. Gartner's analysis likewise lists governance and identity infrastructure as prerequisite investments for scaling Agentic AI.

The direction of runtime governance is shifting from "post-hoc auditing" to "runtime prevention." ServiceNow's AI Passport uses cryptographically signed credentials to verify an Agent's authorization scope at runtime. Tencent Cloud's productization path is moving toward the same goal: controllable behavior, traceable processes, and manageable data. Governance is not a policy document attached after deployment, but an attribute that an Agent carries from birth.

The return on Agents depends heavily on "whether they can cross the mass-production threshold," not on the model itself.

II. Moving Service Agents from Difficult Pilot-to-Production to Large-Scale Mass Production

In 2026, the center of gravity of the Agent market is shifting from "building" to "operating." Futurum Research's judgment is: Agent capabilities have become abundant, while structural governability remains scarce; without the latter, enterprises only have half a solution.

The standardization of the connection layer between Agents and business is maturing rapidly, but this has not translated in sync into maturity at the governance layer. A survey by Stacklok in July 2026 shows that although 41% of software organizations report that MCP has entered production, governed enterprise-grade production environments account for only 5%. The stability gap, the governance vacuum, and the security paradox are not three isolated technical problems, but the same structural gap: the dynamic behavior of Agents is not being dynamically governed, and enterprises are still using control systems designed for deterministic software to manage a non-deterministic executing entity.

The real gap only becomes exposed after entering the production environment, embedding into enterprise production processes and high-frequency interaction entry points. Competition among enterprise-grade Agents is shifting from model capability to delivery entry points, orchestration governance, and commercialization systems.

For an Agent to cross the chasm from pilot to mass production, it must answer a completely different set of questions: When no one is watching over its shoulder, can it still run stably? Can every action it has taken be traced afterward? When a failure occurs, who bears the responsibility? Wu Yongjian summarizes the difficulties of landing Agents in enterprises into four categories: data (contracts, invoices, and update records are scattered, with inconsistent standards), identity and systems (networks, identities, and permissions need to be connected repeatedly), effectiveness and cost (after launch, effectiveness, cost, and failures must all be visible, trackable, and fixable), and usage and collaboration (with so many Agents built, how can employees find them, use them, and provide continuous feedback). These four categories of difficulty correspond to the same thing: for service Agents to move from pilot to production, five capabilities—access, build, execution, constraint, and distribution—need to interlock with one another.

Tencent Cloud's ADP platform is laid out precisely around this value chain, and the accompanying ADP FDE delivery methodology summarizes project implementation into five stages: access, build, validation, promotion, and governance, and has accumulated 25 standardized scenario packages covering five major industries.

Access defines the identity under which an Agent enters enterprise systems, emphasizing admission and authorization rather than mere connectivity.

Build addresses "how implicit process rules become deliverable business applications." The most critical change in ADP is moving toward AI-native, letting natural language run through the entire process of generation, debugging, evaluation, release, and tuning. People who understand the business only need to describe requirements, and the platform can automatically understand goals, break down tasks, and generate agents.

Another important change is the expansion of build targets, combining agents with site services. Agents are responsible for analysis and execution, while site services carry operation pages, business logic, and data state. Enterprises can directly build end-to-end complete business applications, extending from "building an Agent" to "building a business system that can actually run."

Operation and governance make Agent behavior observable, correctable, and rollback-capable. Open integration lets ADP embed into enterprises' existing systems rather than starting a separate set. Distribution and operations solve the problem of "multiple Agents have been built but employees don't know which one to use." The value of Agent Portal's unified management, intelligent routing, and observability dashboards lies not in the dashboards themselves, but in the management actions they trigger.

H3C's data from 1.3 million conversations per day is used to adjust Token allocation and incentive strategies. Suning.com reduced Token consumption by 10% by locating abnormal time consumption and adjusting prompts and model configurations. What these cases demonstrate is how "operations language" is actually used—not "whether the Agent is running," but "which stage the cost is spent on and which step the latency is stuck at."

The two-way connection between ADP and WorkBuddy establishes an institutionalized conversion channel between the individual entry point and the enterprise middle platform. Enterprise-grade Skills created in ADP can be synchronized to WorkBuddy in real time, and employees can call them directly at their daily work entry point. High-quality workflows validated by employees in WorkBuddy can be converted into ADP agents with one click through connectors, bringing them into enterprise-grade governance and distribution. The individual entry point produces validated workflows, and the enterprise middle platform consolidates them into governable, reusable assets.

The route for Agents to cross from pilot to mass production has been opened up, and the platform is in place. But to truly achieve large-scale landing in enterprises, the platform's strategy and resolve will still be tested—whether it can mobilize the enthusiasm of all parties in the market.

Handing over half of one's life to partners to maximize the extension of one's own business boundaries is Tencent's consistent strategy. Today, in order to accelerate the large-scale landing of Agents in enterprises while expanding its own leading advantage, Tencent has also chosen to hand over half of its life.

Standardized scenario packages only solve half of "replicability"; the other half is "who replicates." Wu Yongjian's judgment on this is: models and foundational capabilities solve only about 30% to 40% of the problem, and the last mile of accessing enterprise internals is the 60% to 70% difficulty. "A chef making a dish" and "running ten thousand stores" are two different things. This means the last mile cannot be completed remotely by the platform; it must be filled in on site by partners close to customers, completing the prerequisites of data, processes, and permissions. "Tencent has experienced this model many times—whether internet, AI, or TO B—our team is trying to lay it out systematically, and we also hope Agent landing can proceed at scale and efficiently like this." This is how he analyzes the leverage effect of partners on large-scale business landing.

This determines that ADP's partner strategy is not channel agency but a restructuring of division of labor. Tencent consolidates the 70% that can be standardized into scenario packages, the FDE methodology, and evaluation use cases, while partners complete the 30% of on-site customization and organizational access on top of the standard. Wu Yongjian's judgment on the trend of share is direct: "Right now the share between Tencent and partners may be 9:1, and later it may become 7:3 or 6:4. I think this is a good thing." Making the cake bigger rather than zero-sum is premised on the standard being accumulated thickly enough. The more share partners take, the more solid the replicable foundation is, and the more it shows that large-scale landing is being rapidly realized.

In an interview with Qiushui Bitan, Wu Yongjian further pointed out that Tencent Cloud is currently focused on TO B landing opportunities, and that enterprise landing of Agents "has only just begun," with the real penetration rate not high. For Tencent, this is an opportunity to build a platform, with the goal of becoming a leading domestic enterprise AI middle-platform provider. For our downstream partners, this is also an opportunity to seize: combine ADP to build applications on top of it, plus the FDE delivery methodology, to undertake enterprise landing.

Over the past decade or so, every change in the rules of Tencent's products has often been accompanied by changes in the business landscape, traffic, and capital flows, and now such change is appearing once again and stirring the market. Wu Yongjian regards the deep integration and collaboration recently promoted by several Tencent products as the key for partners to "seize this wave of Tencent's ecosystem dividend." He told Shui Ge, "I also told our partners yesterday that the biggest dividend in 2027 lies in Hunyuan, WeCom, WorkBuddy, and ADP driving simultaneously in different tracks, and the synergy will amplify scenarios that a single product cannot cover."

III. The Shift in Procurement Logic: The Watershed Is in the Organization, Not the Model

Putting ADP back into the competitive landscape makes the route it has chosen more directional. Compared with the general-purpose builder route of most platforms fighting over "building Agents faster," ADP takes another, more differentiated path: it does not compete over who builds faster, but over whose output can be truly taken on by the organization.

To this end, Wu Yongjian also strictly defined ADP's positioning: "It is not a general-purpose Agent development platform, but one that helps enterprises build Agents serving data and business scenarios, capable of producing clear business results." This means Tencent Cloud has proactively given up the market where anything can be built, drawing the platform boundary on the side of "being responsible for business results."

The industry is also reaching the same consensus. Gartner's analysis of 107 agentic AI deployments offers a directional judgment: by 2028, 80% of tangible ROI will come from specialized, domain-specific Agents rather than general-purpose Agents. Robert Hetu's exact words were: "Organizations must scale successful domain-specific Agents to enterprise-level deployment." Gartner also gave a case: an industrial service provider achieved an annual ROI of $3 million and saved 90,000 hours through a digital employee that automated parts ordering. All of this proves the importance of process governability.

Agents are managed as "execution entities," but the auditability deficit in their workflows is still widening. Gartner data points out that although token costs are falling, the inference cost of agentic workflows will grow more than fivefold by 2028, so competition will shift from "unit token price" to "the full-load cost and reliability of each task delivery." Gartner's review of nearly a thousand cases further shows that advanced reasoning models consume 5 to 30 times the tokens per task of basic chatbots, yet currently fewer than 10% of deployments can deliver robust positive output results.

The pricing power of service Agents depends on which layer of tasks they undertake and how much that layer of tasks is worth. Bain's survey of 951 global enterprises provides an organizational footnote: only 7% of enterprises run fully autonomous Agents in production, 38% require human approval, and nearly 40% of enterprises see actual cost savings of 0% to 10%, far below the target of 11% to 20%.

Gartner's ranking of payback periods—SDR 3.4 months, customer service 4.1 months, finance 8.9 months, engineering 9.3 months—is essentially a ranking of process governance maturity. The fastest payback is not in the most intelligent scenarios, but in those with the clearest boundaries and the shortest accountability chains.

Just as the businesses with the most mature engineering are the first to get results, the enterprises that have already achieved the strongest results are also those that have already treated governance and process design as CEO-level issues. The procurement logic for enterprise Agents will shift from IT tool budgets to operating budgets. The decision-maker for tool budgets is the CIO, measuring deployment numbers and adoption rates. The decision-maker for operating budgets is the CFO and business line leaders, measuring process cycles, unit costs, and business results.

The essence of Agent failure is governance failure. The diffusion of Agents itself reflects the uneven levels of organizational governance across enterprises, requiring deploying enterprises to have process complexity and governance infrastructure to assign responsibility to individuals. Yang Niannong, Senior Vice President of iSoftStone, put it more bluntly: "AI will not be responsible for business results; it must be people who are responsible."

English text is being prepared.

The accountability structure of service Agents should not make Agents bear legal responsibility, but should ensure that every Agent action can be traced back to a responsible human authorizer. A Gravitee survey shows that 85% of organizations have no formal accountability arrangements for AI Agent behavior, and only 7.2% have designated responsible persons.

Between the 171% ROI and the 86% of pilots that never entered production, the distance that has not been crossed is not insufficient model capability, but that organizations are not yet ready to take on the runtime identity, governance boundaries, and operational engineering of Agent output. ADP turns service Agents from "projects" into "operable assets," which is essentially answering this question. And the value of operating assets can only be realized through the business results they continuously operate.

The watershed is not in the model, but in the organization. And the scaling of organizational capability cannot be covered by a single platform alone. It requires both data governance and process mapping on the customer side, and division-of-labor undertaking on the ecosystem side: the platform accumulates standards, FDE completes the translation, and partners walk the last mile.

This article comes from the WeChat public account "Qiushui Bitan," author: Shui Ge, published with authorization from 36Kr.