China’s AI Rollout Runs Into an ROI Reality Check

Chinese companies are moving more AI projects into production, but many executives say the financial case still falls short. A new survey reported by Yicai found that 49% of its respondents said AI deployment had not met their return expectations, even as the share of firms using AI in production rose sharply from a year earlier.

The Yicai report is attributed to Deloitte China and the Center for AI, Management and Organization at the University of Hong Kong (HKU). It covered 205 executives in China from various industries, most of them strategic decision-makers. This is a survey of those respondents, not a census of every Chinese company, but it offers a useful view of what happens after an AI pilot leaves the presentation deck and enters daily operations.

The central finding is not that Chinese businesses have stopped adopting AI. It is nearly the opposite. Yicai reports that the share of respondents in an exploration or pilot stage fell to 25% from 37%, while the share in production deployment increased to 64% from 56%. The implementation wave is moving forward. The difficult question is whether companies can measure and capture enough value from it.

The survey examines deployment at company level, a question that complements EastFrontier’s earlier coverage of national AI adoption targets.

More projects have reached production, but returns lag

Yicai says expected returns on investment were mainly in a 10% to 25% range, while actual returns were more often below 10%. A Zhitong Finance report carried by Webull gives a more precise version of the headline figure: 48.5% of surveyed companies said actual AI-project returns were lower than expected. The rounded Yicai number and the more precise secondary report point in the same direction, but they should not be mixed without attribution.

This is a familiar pattern for a technology moving from experiments into routine work. A pilot can show an impressive demonstration because the project team is focused, the data set is prepared, and human experts are standing close by. A production system must handle inconsistent data, integration with old software, security permissions, changing workflows, user training, and the cost of ongoing maintenance.

The survey suggests that many firms are now encountering those harder stages. It does not prove that AI projects are unprofitable overall, or that any particular model is failing. It captures self-reported return expectations and measurement problems among 205 executives. That is still significant because executive expectations are often what determine whether a project receives a second budget, a larger rollout, or a quiet retreat.

Pascal Hua, whom Yicai identifies as Deloitte’s national managing partner of consulting businesses in China, told the outlet that widespread use of AI tools does not mean a company has completed its transformation. Many businesses, he said, have inserted AI into existing processes without changing the operating model around it. The implication is that a new assistant may make an existing task faster without changing the workflow enough to produce a measurable business gain.

The measurement problem may be bigger than the model problem

More than 44% of respondents said they could not clearly calculate the return on an AI project because their value-quantification ability and evaluation frameworks had not kept up with the expanding scope of applications, according to Yicai. The Webull-carried report puts the figure at 44.4%. Yicai also says 60% of respondents lacked a systematic framework for measuring AI value.

Those figures point to an operational gap. A company can count how many employees logged into an AI tool or how many tasks were processed. It is much harder to determine whether the tool improved revenue, quality, speed, risk control, or customer retention after accounting for software fees, compute, data preparation, review time, and process changes.

The most useful metric also varies by job. In customer service, a firm might look at resolution time and escalation rates. In manufacturing, it might track downtime, defects, or energy use. In a research department, the gains may appear in cycles avoided or alternatives considered rather than in a simple labor-saving total. Without an agreed measure, an AI program can look successful to a technical team and disappointing to a finance department at the same time.

This helps explain why the move into production is not the same as a clear return. The new survey keeps the focus on project-level measurement and self-reported expectations.

Workforce expectations reveal the next pressure point

The survey’s employment findings are also revealing, although they are expectations rather than observed labor-market outcomes. Yicai reports that 65% of respondents expected fewer entry-level or junior roles. Mid-level jobs were expected to undergo structural reorganization rather than a simple decline, while top-level roles were expected to remain stable because strategic judgment and accountability are harder to replace.

That does not mean 65% of entry-level jobs have disappeared, nor does it establish that AI will protect senior roles. It shows how executives expect work to change as more AI tools enter daily operations. The survey does not establish how those expectations will translate into observed task reallocation.

The role expectations can be read alongside EastFrontier’s prior coverage of China AI hiring, though the two stories use different sources and should not be treated as one data set. The survey itself does not establish how the expected changes will translate into observed task reallocation.

The survey therefore gives a more grounded picture of China’s AI rollout. Adoption is advancing, production use is rising, and expectations remain high. But the gap between a promising pilot and a measurable return is still wide for many organizations. The firms that close it may not be the ones with the most AI experiments. They may be the ones that define a business metric first, change the underlying process, and keep measuring after the initial excitement fades.