At Volcano Engine’s FORCE conference this week, Zhuoyu Technology highlighted a 2025 MIT study: 95% of enterprises investing in generative AI have not achieved measurable returns, with most projects remaining stuck at the pilot stage. What matters to us is less a particular company’s scorecard than the diagnosis it put forward—enterprise AI is not failing to make money because of the models; the models are already strong enough. The bottleneck has shifted to governance, or more bluntly, organizations are not yet ready to bring AI into the workplace.
What this is
Zhuoyu Technology—formerly DJI Automotive—currently has more than 1,000 R&D staff, 34 core customers, and over 130 vehicle models in partnership, making it a typical enterprise handling highly sensitive data. Source code, customer materials, and R&D documents are all stored in its internal Feishu environment. During its presentation, the company offered the following formula:
Organizational capability = technical readiness × governance maturity × people readiness.
The first factor is not the problem: models are stronger, platforms are more numerous, and tools are more mature. The other two have not kept pace. Based on this, Zhuoyu identified four boundaries for governing Agents—AI programs that can autonomously call tools to execute tasks: the data boundary (what data can be used and what cannot leave the internal network), the permission boundary (who can invoke what), the accountability boundary (logging behavior and making it traceable), and the supply chain boundary (whether the operating environments of plugins and MCP—the standard protocol models use to connect to external tools—are trustworthy). Its implementation architecture has three layers: egress control (ArkClaw), runtime governance (ClawSentry), and confidential inference (AICC). All three come from Volcano Engine’s product portfolio.
Industry view
Zhuoyu cited three sets of external data to support its diagnosis of a governance gap: MIT’s 95% no-return finding, Gartner’s 2026 projection that only 41% of Agent projects will be ROI-positive after one year, and McKinsey’s finding that employees save 6.4 hours per week but AI contributes more than 5% of profits at only 6% of enterprises. Deloitte’s 2026 figures are even more direct: 74% of enterprises expect to use Agents within two years, yet only 21% have mature governance frameworks.
However, we note that this was a case presentation at a supplier conference, and several points should be discounted. First, Zhuoyu’s solution is tightly coupled to Volcano Engine’s products, so its notion of “governance maturity” effectively amounts to promoting a specific architecture. Second, MIT’s 95% figure is disputed within the industry because different studies use different definitions of “measurable return.” Third, “organizational capability = the product of all three factors” sounds appealing, but a zero-product diagnosis often masks the question of whether a solution fits the specific use case.
Impact on regular people
For enterprise IT: procurement logic is shifting from “buying technology” to “buying certainty.” Zhuoyu’s selection checklist for CIOs has three items—governance discipline, deterministic engineering, and depth of security architecture. Model benchmark scores alone are no longer enough.
For individual professionals: until Agent governance frameworks mature, employees who use AI privately to process sensitive data such as contracts, customer materials, or source code face compliance risks. This is not a technical problem; it is an organizational-boundary problem.
For the consumer market: Zhuoyu works on more than 130 vehicle models. Once Agents are deployed at scale, autonomous-driving R&D and in-vehicle interaction will be the first areas to change, while ordinary vehicle owners are unlikely to notice any direct impact in the short term.