Baixing pays its AI agents by results, and call volumes surge

As foundation models keep getting more capable, one question increasingly decides enterprise buying decisions: what, exactly, are companies willing to pay for? Baixing, a Hong Kong-listed Chinese AI company, gave a concrete answer in its 2026 first-half report: not model capability by itself, but a measurable, billable outcome. Total revenue fell as regulation bit into its lending-related business, but its AICC (AI Contact Center) unit grew 52%, with revenue from new, non-financial sectors up 195%.

Two live deployments illustrate the shift. At a logistics client, the daily volume of calls handled by AI agents rose from under 1,000 at launch to more than 15,000 by the end of August. At a major securities brokerage, the number of deployed agent seats grew from about 30 early in the year to about 210. Baixing calls these agents, which take on real query, complaint and sales work, 'silicon employees,' and charges for them by results rather than by license or seat — a rare, financially documented showcase of enterprise agent commercialization.

Two opposite curves in the books

In the first half of 2026, Baixing posted revenue of 914 million yuan, down 43% year on year; gross margin fell from 73% to 53%, and the company swung from a 201-million-yuan profit a year earlier to a 322-million-yuan loss. The report attributes this to regulation on internet-assisted lending that took effect in October 2025, tightened further by rules rolled out through the first half of 2026, which pushed several financial institutions to suspend certain products — hitting Baixing's core AI-decisioning and credit-marketing businesses directly. AI-decisioning revenue fell 26% to 372 million yuan; other business revenue fell 62% to 377 million yuan.

AICC moved the opposite way: revenue of 166 million yuan, up 52%, taking its share of total revenue from roughly 6.8% to about 18.1%. At the same time, Baixing chose to raise R&D spending 31% to 394 million yuan (43% of revenue) even as overall revenue shrank; of 1,477 employees at the end of June, 901 — 61% — work in R&D. CEO Zhang Shaofeng called the shift a 'strategic upgrade' on the earnings call.

From scripted outbound calls to conversations that actually resolve

The previous generation of contact-center bots mostly handled scripted outbound work — reminders, standardized marketing campaigns. Complex inbound calls are a different problem: the caller shows up with an open-ended question, an unpredictable tone and situation, and the agent has to understand it, hold a multi-turn conversation, then query a knowledge base and internal systems to actually finish the job. Zhang said on the call that traditional bots hand off 70-80% of these complex cases to a human — precisely the gap the new generation of AICC is built to close.

Baixing's technical bet is to train its own specialized foundation models rather than build an application layer on top of a third-party model: it runs BR-Voice (multimodal voice), BR-LLM-Proactive (proactive dialogue) and BR-Vision-Doc (document understanding), with parameter counts deliberately kept between 0.4B and 30B to favor accuracy, stability, latency and inference cost over raw scale. Zhang explicitly contrasted this with Western AICC players like Sierra or Decagon, which build largely on top of third-party models from OpenAI or Anthropic: Baixing controls pretraining, post-training and inference optimization in-house, so when something breaks in production, it can tell faster whether the problem sits in the application layer or the model itself, and move straight into the next training cycle. On the multilingual dialogue speech-understanding task of the INTERSPEECH 2026 MLC-SLM Challenge, Baixing claims 94.84% accuracy — third worldwide, first among Chinese industry players. On the ground, a team of Forward Deployed Engineers maps a client's existing workflows and reorganizes them into agent-run ones.

Charging for the outcome, not the subscription

Baixing's RaaS (Results as a Service) model bills, depending on the case, for calls or tickets successfully resolved without a human handoff, for the actual scale of agents deployed, or as a share of the business revenue an agent generates. On the logistics case, management said on the earnings call that per-resolved-call cost is already more than 50% lower than a traditional human-staffed call center. That differs sharply from software licenses, SaaS subscriptions, or token-based billing: once software ships, the operating risk stays with the buyer, while outcome-based pricing pulls the technology vendor directly into the client's business process, making the model's performance, stability and cost the vendor's problem too.

This is not a new idea for Baixing. On December 18, 2025, its parent Bairong Yunchuang formally launched an enterprise AI-agent strategy and a platform called 'Results Cloud,' built around a concept it calls 'silicon-carbon co-governance' — AI ('silicon') taking on standardized, process-heavy, compute-intensive work, while humans ('carbon') focus on strategy, creativity and relationships. The platform has three layers — proprietary infrastructure and models, an operating system for enterprise agents, and a store of ready-made agents — and launched with four flagship roles: a sales-and-service agent credited with cutting one team's annual attrition from over 70% to 0% and lifting conversion 217%; a recruiting agent that cut hiring cycles from 28 to 2 days; a professional-services agent for cross-border expansion that compressed turnaround from 90 to 14 days; and a knowledge-production agent that cut deep-report delivery from 20 to 4 days.

The next stage of enterprise AI isn't about talking better, it's about doing more; it's not about delivering a feature, it's about delivering a result.

From finance to logistics, airlines and brokerages

Baixing's AICC capability, refined over seven or eight years in finance, is now being replicated into logistics, aviation, brokerage, banking, telecoms and insurance — service-heavy sectors with high, recurring volumes of customer interaction and enough industry concentration that a proven deployment at one large client can spread quickly, first internally, then to competitors. By July, 13 brokerage-sector institutions had signed or were in contract talks with Baixing. Management acknowledged, though, that the time-consuming part of cross-industry expansion now sits in large enterprises' procurement and validation cycles and in systems integration — volume only ramps once security, compliance and service boundaries are signed off.

For a company evaluating an AI-agent vendor, the practical lesson is that the purchasing criteria are shifting from feature checklists to verifiable outcome metrics — resolution rate without human handoff, cost per interaction handled, and evidence the vendor can replicate a deployment at scale. For workers in customer service, recruiting or back-office legal and financial support — roles that are already highly standardized — this case is an early-warning signal: once one piece of technology can be replicated from one industry to the next, those are the jobs automated first, and HR teams have reason to plan the shift of staff toward judgment-heavy, relationship-driven work now, rather than wait for enterprise AI agents to already be in production.