Ant Digital Will Eat the Accuracy Risk on Agentar 2.0

Ant Digital, the enterprise-technology arm of Ant Group, used the opening day of the World Artificial Intelligence Conference in Shanghai on July 17 to put a new bet on the table for Chinese enterprise AI: customers will pay only when its industry-tuned agents demonstrably deliver, with the vendor absorbing the accuracy risk if they fall short.

The product behind the promise is Agentar 2.0, Ant Digital's agent-building platform. It now ships with nearly 200 role-level "digital expert" templates and a library of several hundred Skills-level agent tools that customers can assemble into vertical workflows. Ant Digital vice president Sun Lei framed the launch in three vectors on 36kr's "Kryptalk Future" stage at WAIC: industry-specific foundation models built on Ant Group's two decades of work in payments, finance, wealth, and insurance; scenario-specific translation layers that turn business questions into technical tasks; and an ecosystem layer that opens the platform to outside partners through APIs (36kr interview with Sun Lei at WAIC 2026).

The commercial move breaks from the standard Chinese enterprise-AI playbook. Sun Lei told 36kr that Ant Digital is shifting to a "value delivery, pay-for-results" model, where customers are billed on whether the agent's outputs actually hold up. The benchmarks Ant Digital uses to define that outcome are its own: a 90% to 95% accuracy rate on professional-domain questions in finance, and an 85% or higher "win-tie rate" when its agents' answers are compared head-to-head with human specialists. Both figures are vendor-reported, drawn from internal evaluation suites rather than independent audit. That gap is the load-bearing caveat of the launch.

Finance is the lead use case. Banks, insurers, and asset managers are the verticals where a hallucinated answer is most expensive, and where procurement teams have spent two years pushing back on general-purpose LLMs. Sun Lei, in the same interview, named three recurring objections from financial-services buyers: security and compliance, professional accuracy versus the model's tendency to produce plausible but wrong answers, and the return-on-investment math when a pilot does not convert to production. Ant Digital's answer to all three is a vertical foundation model trained on Ant Group's payment and finance data, layered with engineering to make outputs auditable, and a contract that prices on whether the customer judges the result usable (36kr interview with Sun Lei).

Two third-party handles exist for the launch, both thin. The cleanest is the selection of an Agentar-powered finance case as an "international standard financial application excellence case" in 2025 (Xinhua, October 2025). The other is a partnership with Linyang, a Chinese energy-metering and grid-services company, to build AI agents for electricity trading. Linyang has published its own announcement of the collaboration (Linyang news, 2025), and Chinese financial press has reported on the resulting "virtual trader" agents (STCN, Gitcode/CSDN). Neither handle constitutes an independent benchmark of the 90-to-95% or 85% figures. They show that real customers are wiring Agentar into production workflows, not that the agents hit the numbers Ant Digital claims.

The "China's industry-version Harness standard" framing in the launch is the vendor's own positioning, not an external comparison. Harness is a US AI-infrastructure company that sells engineering and governance tooling for software teams; the analogy Sun Lei draws is that Ant Digital wants to occupy a similar "how enterprises actually build and ship AI" layer in China. 36kr's interview does not lay out why Harness is the right reference rather than, say, a Chinese cloud stack or a vertical SaaS incumbent.

Other Chinese vertical-AI vendors now face a choice: match the outcome-pricing bet, or stay on seat licensing and be priced as legacy SaaS. The vendors who push hardest on outcome pricing in the next two quarters will signal which way the market runs.