88% of companies tried AI. Only 39% saw returns. The bottleneck is not the model.

On the customs-broker case Yuan Xin described on a recent LateTalk podcast, a mid-sized Chinese exporter wired an AI agent into its customs filing. The agent pulled declaration data, matched it against tariff codes, and pre-filled forms for a human reviewer. On the demo, accuracy hit the low 90s. In production, the same agent initially got 60 to 70 percent of forms wrong. The reason was not the model. The reason was that the company's product master data was inconsistent across ERP instances, and the model could not reason across that mess.

By a count circulated in a McKinsey 2025 survey summary republished by QbitAI, 88% of organizations say they have piloted AI in at least one business function. Only 39% say they are seeing meaningful business returns. That gap, between almost-everyone piloting and almost-no-one seeing returns, is the story inside Chinese back offices right now. The next two to three years of enterprise software value will get allocated along that gap, and a lot of it will not get allocated at all.

The vision SAP laid out at its 2026 Sapphire conference, the so-called "Autonomous Enterprise," leans into the same gap. SAP wants users to describe a goal in natural language through Joule Work, and have an agent orchestrate the work across data, process, and downstream tools. It is a coherent pitch. As of mid-2026 it is also mostly a pitch. Yuan Xin made the more interesting argument on the podcast: AI will not make the underlying work of running a company go away. "To use AI well, companies cannot escape data, process, business understanding, and organizational change," he said. He was not selling a vision so much as describing the seam that the vision has to run on.

That seam is why the 88/39 gap looks structural, not transitional. The press around the Sapphire launch, republished by Sina Finance and TMTPost, shows SAP deepening partnerships with Anthropic, Google, Microsoft, AWS, and NVIDIA, and in China working with Alibaba Cloud, Qwen, and DingTalk. The partnerships are real. They do not, on their own, clean up a customs broker's tariff master or reconcile three months of intercompany transactions. Only 6% of companies, by one reading cited by InfoQ, say they have AI governance fully figured out. The plumbing underneath the agent is the hard part.

The new role showing up in job postings is one tell. Forward Deployed Engineers, the FDE pattern popularized by Palantir, are now being hired by Chinese system integrators and SAP partners to sit inside a customer and translate business process into agent behavior. The custom-broker case is a small example. An agent that initially classified only 60 to 70 percent of goods correctly needed a domain expert to keep correcting it; with that loop, the same agent cleared 90% plus. The lift came from data labeling, exception rules, and process redesign, not from a model upgrade. Yuan described this as "the old master still has to be steamed out," an industry phrase for the slow work of turning a senior accountant's or customs broker's gut calls into structured instruction. That work is what FDEs bill for.

The squeeze falls hardest on mid-sized Chinese companies, the group that Yuan Xin said is "most struggling." Large enterprises can pay for a data cleanup that runs into eight figures. Small companies can adopt a thin SaaS stack and move fast. Mid-sized firms sit on years of customizations, multiple ERP instances, and a finance team that knows the system is broken but cannot afford a one-year project to fix it. AI pilots in that layer are the ones that stall. The agent demo works. The agent in production does not. The CFO does not get to mark the project as a return on the P&L, and the project gets canceled before the data team can finish the cleanup.

The longer arc Yuan Xin described is harder to dismiss. China did not grow a global SaaS champion in the previous cycle in part because fast growth left a layer of system debt that Western software-as-a-service models could not absorb. The current AI cycle does not retire that debt on its own. The companies that will capture the value are the ones that can sell the cleanup: large global vendors with a services arm, boutique FDE-style firms, and the system integrators that have lived inside customer ERPs for a decade. The companies that will be left out are the mid-sized firms that cannot pay for the cleanup and do not have a small enough footprint to skip it. The race for AI in the back office is not being won on the model. It is being won on who can sit in the customs office long enough to make the model work.