When artificial intelligence agents first arrived, they came with a tantalising promise: autonomous “digital workers” that would finally free humans from corporate drudgery. As businesses across Asia are discovering, however, the reality is quite the opposite – more work, new friction and a whole new layer of management.

When Xiao Mengjie, an AI product marketing manager at a Big Tech firm in Shanghai, recently assigned two AI agents what she described as a simple task, she expected efficiency. Instead, she got corporate gridlock.

Rather than dividing the workload, her agent and her supervisor’s agent locked in a prolonged discussion over responsibilities, debating who should verify details and which one should take ownership of different tasks, in a spectacle not unlike the human bureaucracy they had been hired to replace.

“It became very formal and cumbersome,” Xiao said. “Agents can create communication overhead too.”Across Asia, businesses have eagerly bought into the pitch of autonomous software capable of planning and executing multistep workflows with limited human intervention. But from Singapore’s hotels and insurers to Hong Kong’s service sector and mainland China’s tech giants, enterprises are confronting a sharp reality check. The surge in regional adoption reflects a broader global frenzy for autonomous software.

Meta Platforms’ personal agent, Muse, launched in September, drew an estimated 2.8 million downloads in its first two weeks.

In mainland China, enthusiasm for OpenClaw, an open-source agent, sparked a craze known locally as “raising lobsters”, prompting hundreds of people to queue outside Tencent Holdings’ headquarters in Shenzhen earlier this year for help installing the software.

Tech titans including Tencent, Alibaba Group Holding, owner of the South China Morning Post, and Baidu have all introduced agent-focused enterprise products.

Yet the rush to deploy AI agents has exposed a gap between what the technology can do and how effectively organisations can manage it.

Why some businesses struggle to put AI to work At times, users find that seemingly straightforward workflows become challenging puzzles once they are handed over to autonomous agents.

At Singapore-based hospitality technology start-up Conzi, chief operating officer Mark Wee wants AI agents to handle the multi-departmental, behind-the-scenes work that often pulls hotel staff away from guest care.

For example, when a hotel guest leaves a scathing review, the complaint can trigger a lengthy chain of work behind the scenes. A receptionist must respond, a manager investigates what went wrong, and a supervisor checks records to determine who was responsible. The hotel might even revise its operating procedures.

But deploying AI agents to manage complex hotel operations is easier said than done.

A single guest complaint could involve multiple departments, from housekeeping to hotel management, requiring agents to navigate fragmented systems and determine when human intervention was needed, Wee said.

The challenge is compounded by operational knowledge that is often scattered across spreadsheets, paper records and the heads of experienced employees.

Wee said Conzi was working to consolidate this information so its agents could coordinate tasks across departments, but he stressed that technology alone was not enough.

“If you don’t get the buy-in from the people who are supposed to use the systems the most, any system in any organisation will fail,” Wee said.

Similar obstacles confronted Asia’s travel industry, where differing regulations, technical standards and disconnected platforms complicated AI deployment, according to Prabu Sangar, director of technology operations and integration at FCM Travel Asia, a corporate travel management company.

“Even if the AI is as capable as it is, the surrounding environment may not be ready for it,” he said.

In Hong Kong, companies were trying to accelerate AI adoption while contending with “legacy” technology systems and organisational resistance, said Robert Hah, Hong Kong office lead at consultancy Accenture Greater China.

Only 10 per cent of Hong Kong companies surveyed had integrated generative AI across multiple business functions, compared with about 21 per cent of mainland Chinese enterprises, according to Accenture.

Leo Wong, co-founder of Hong Kong-based business communications platform imBee, said demand for AI agents had grown among customers across Hong Kong, Singapore and Malaysia, but concerns over data security were slowing adoption, especially in Hong Kong. The company provides AI-powered tools that help businesses manage customer communications across WhatsApp, email and other messaging platforms.

Wong said banks, insurers and other financial services companies were particularly cautious, with deployment decisions taking longer as they worked through their AI strategies and security requirements. Organisations wanted to build confidence in the technology before deploying it, he added.

I already have lots of ideas, and AI makes more of them feel achievable, so I start more projects. Then every project is waiting for my judgment, and I end up being the busiest personXiao Mengjie, AI product marketing manager Wang Jiangyu, professor of law and associate dean at City University of Hong Kong, said the institutional mechanisms needed to govern autonomous agents had struggled to keep pace with their technical development.

“The governance challenge today is less about creating more powerful AI and more about ensuring that increasingly capable systems remain accountable and aligned with human objectives,” he said.

While AI systems had become more capable of executing tasks autonomously and coordinating with one another, questions including how decisions should be reviewed and who bore responsibility for those decisions remained unresolved, he said. Too many agents?

In their eagerness to automate, some enterprises are discovering that more agents do not automatically translate into higher productivity.

At imBee, a company with around 50 to 60 employees, at least several hundred AI agents helped staff perform tasks ranging from software development to internal communications, according to Wong.

Wong, who described himself as non-technical, said he personally managed more than 20 agents, which in turn coordinated other agents.

While imBee used quality checks and verification procedures to avoid major errors, according to Wong, other workplaces reported that stacking agents simply duplicated existing corporate bloat.

“The market is over the phase of ‘agent maxing,’” said Raunak Mehta, co-founder and CEO of Singapore-based insurance technology company Igloo.

Operating across Southeast Asia with about 230 employees, Igloo uses software workers for customer service, software development, claims processing and financial reconciliation.

Mehta said total agent count was no longer a meaningful measure of operational success. Instead, the focus had shifted towards “agent harnesses” – control frameworks that coordinated agents, regulated their access to internal tools and allowed human oversight.

Leading AI agents – including Meta’s Muse, Tencent’s Xiaowei agent, and Manus 2.0 – increasingly rely on such harnesses to better understand users and drive long-term platform adoption.

Without proper harnesses, companies faced what Sangar of FCM Travel Asia called “agent sprawl” – a chaotic proliferation of autonomous software across departments with conflicting permissions, overlapping data sources and unintended interactions. Uncoordinated agents can also compound mistakes across entire workflows.

“One agent might report that a task is complete when it has merely produced a draft,” AI product marketing manager Xiao said. “Another agent could mistake that output for an approved decision and act on it, allowing errors to spread.”Compounding the risks, hotels and financial institutions must expose vast amounts of proprietary data to extract the full value of AI, increasing their exposure to cyber threats. “There is definitely a risk of rogue actors approaching us and trying to maliciously extract information,” Conzi’s Wee said.

Teaching people how to brief, review and correct an agent matters as much as the technology itselfLionel Sim, The AI Capitol Recent incidents have highlighted the risks of giving AI agents access to sensitive information and the authority to act without human approval.

Last month, Chinese AI start-up Z.ai said it had temporarily disabled certain features of its ZCode coding assistant after developers reported that the software was uploading local code repositories to cloud servers without their consent. The company, also known as Zhipu, apologised and patched the vulnerability.

A Reuters investigation published last month found that AI agents from Alibaba, DeepSeek and Moonshot AI systematically lied, hid failures and fabricated results during controlled tests. Alibaba, DeepSeek and Moonshot did not respond to Reuters’ requests for comment.

In another high-profile case, an OpenAI agent undergoing security evaluation broke out of its sandbox environment and breached systems at AI platform Hugging Face.

The risks have drawn the attention of regulators. Singapore’s Infocomm Media Development Authority has recommended that businesses restrict agents’ access to sensitive systems, define limits on their autonomy and establish mechanisms to shut them down when they malfunction.

The Cyberspace Administration of China has also mandated that companies complete mandatory algorithm filings and pre-deployment security assessments before rolling out agents, particularly those handling sensitive data or customer-facing operations.

Rise of the AI middle manager Li Jiahao, a researcher specialising in global AI development at the Guangzhou Institute of the Greater Bay Area, said China had issued several AI regulatory documents quite rapidly over the past year, but these did not immediately translate into accountability mechanisms.

In both China and the rest of the world, “AI’s technical capabilities are iterating on a quarterly basis, regulations are iterating yearly, and liability is being determined case by case”, he said.

“These three curves are moving at completely different speeds, and the gap between them is still growing.”As digital agents multiply, the fundamental operational bottleneck is shifting from computing power back to human supervisors.

Lionel Sim, founder of AI research firm The AI Capitol, said AI agent adoption in Asia was outpacing management. Employees were deploying agents “on their own initiative”, which was “great for momentum”, but companies did not have oversight of access, ownership and active tools, he said.

The adoption of AI agents has also resulted in a “skills shift”, pushing employees into management roles. “Teaching people how to brief, review and correct an agent matters as much as the technology itself,” Sim added.

For workers like Xiao, managing AI agents can be intensely stressful. While agents can work continuously and at a rapid pace, humans do not gain more hours in the day to supervise them.

Mehta of Igloo warned that growing reliance on agents could create a longer-term problem: companies might eventually struggle to find employees with the experience needed to supervise them.

Li, the Guangzhou researcher, said AI might not suddenly cause mass unemployment, but it was likely to narrow the door for young people entering the workforce as companies scaled back hiring.

“Strong AI management capability means finding ways to turn the efficiency gains from AI into more new jobs, and weak management capability may simply mean companies save on labour costs while society bears the cost of youth unemployment,” he said. Despite the growing pains, industry experts believe there is no turning back.

Chen Bo, a senior research fellow at the East Asian Institute of the National University of Singapore, described the current stage of AI development as “a wild growth phase”, with significant risks emerging around coordination, accountability and other areas.

“But when a major technological revolution comes along – when everyone is facing the AI revolution – the only real option is to ride its development while limiting the downsides, not to hold it back because of them,” he said.