Alibaba Cloud Executive Says Most Enterprise AI Projects Fail on Bad Use Cases, Not Weak Models

Alibaba Cloud Executive Says Most Enterprise AI Projects Fail on Bad Use Cases, Not Weak Models

TREE NEWS reports: Alibaba Cloud Intelligence Group Vice President Huo Jia used a media briefing during the company’s annual Yunqi conference to deliver a blunt assessment of enterprise AI: most failures are not caused by models that are insufficiently smart, but by companies choosing the wrong scenarios, overestimating their own data, and managing AI projects like traditional IT rollouts.

Huo, who took over the team in August last year, said he has received zero customer complaints and has never had to apologize in person — a record he attributes to strict project selection. “Having money and a project doesn’t mean we’ll take it,” he said.

Three Common Failure Patterns

He described three recurring failure modes. First, enterprises attempt to train vertical-domain models from scratch, only to find their data is unprepared, their compute clusters are unsuited to training, and they cannot keep pace with frontier labs that now ship new model versions roughly once a month.

Second, companies believe their existing IT data is ready for AI. Huo’s formulation has evolved from “IT data is not equal to AI data” for training to “IT data is not equal to agentic data” for applications — much of what sits in enterprise systems cannot be used directly in agent environments.

Third, firms build agents for a handful of users. He cited a large enterprise with a B2B e-commerce platform handling just over 100 transactions a day that wanted a full AI transformation plan. “Then why do you need AI?” Huo asked. Manual processing takes only minutes to hours.

Say No, Shorten Cycles, Deploy Engineers

Huo’s prescriptions are twofold: say no to customers, and push engineers to the customer site. He has imposed two additional conditions on projects — short delivery cycles, with monthly or even weekly releases instead of quarterly launches, and a “One Team” model that abandons the traditional client-vendor split.

Alibaba Cloud began experimenting with a forward-deployed engineer (FDE) model before the current hype cycle. In a late-2024 energy-sector project, solution architects were dispatched to the field, validated scenarios on-site, and ultimately delivered a production-grade system. The approach was formally approved for piloting at an internal strategy meeting in December.

Huo is cautious about the FDE label itself, noting it risks being dismissed in China as old wine in new bottles. “I’ve told my team I don’t really want to talk about this term anymore,” he said.

Business Models: Only Two

On monetization, Huo is direct. He sees only two viable models: long-term service contracts and product-based businesses. He is betting on outcome-based pricing — Result as a Service — arguing that only genuine delivery of results justifies a premium. His conviction strengthened after May as model capabilities improved.

For traditional software firms, his message is harsher: if core product competitiveness and margins cannot rise, per-person-day billing is the only option left. “What exactly is your moat?”

Alibaba Cloud itself does not expect FDE work to be a major profit center. The team’s core KPI is whether it drives product adoption. “Domestically, expecting to make big money from this is probably unrealistic,” Huo said.

Market Implications

Huo’s comments carry several signals for investors. First, token consumption is surging and prices are rising, driven by AI-native startups and high-tech customers — a demand signal for cloud and compute infrastructure. Second, enterprise adoption remains slow and scenario-by-scenario, meaning the gap between AI-native demand and mainstream enterprise revenue may persist longer than bulls expect.

Third, the shift toward outcome-based pricing and service contracts could reshape margins across the enterprise software and IT services sector, pressuring legacy per-seat SaaS models. Fourth, the FDE hiring surge — with related roles up more than 800% between January and September — suggests a structural increase in delivery costs that vendors must absorb or pass on.

For now, Huo’s framing is that token price curves and enterprise deployment curves have yet to converge. “Don’t lose your imagination just because you don’t know,” he said. “After four years of training, the starting gun has only just fired.”

Key Takeaways

• Enterprise AI failures are driven by scenario selection and data readiness, not model capability.
• Rising token prices reflect genuine demand from AI-native and high-tech customers, but mainstream enterprise adoption remains early.
• Outcome-based pricing and long-term service contracts may become the dominant monetization models, challenging traditional SaaS economics.
• The rapid growth of forward-deployed engineering roles signals higher delivery costs and a closer link between product development and customer reality.