China's biggest banks are pouring money into in-house foundation models. The smaller ones have found a cheaper route. Anhui Xin'an Bank recently disclosed a procurement plan with a budget capped at 118,800 yuan, according to a November 13 report by Beijing Business Today, buying 50 seats of an AI office agent from Tencent Cloud, meant to sit inside daily collaboration as an "office companion." Across the small and mid-sized banking sector, seat subscriptions, token purchases and OA-embedded agents have quietly become the standard playbook.

The seats in question are Tencent's WorkBuddy enterprise edition, on a one-year service term that can stretch to two more. Before committing, Anhui Xin'an Bank ran a two-month internal trial and judged the results up to expectations. WorkBuddy comes from Tencent Cloud's CodeBuddy team, launched in March, and handles document summarization, data analysis and slide generation. One detail mattered more than the feature list. The bank has been running on WeChat Work as its unified collaboration platform since 2022, so native integration carried the decision.
Anhui Xin'an is hardly alone. Bank of Beijing's 2026 intelligent office collaboration tool project went to Tencent Cloud for 1.5 million yuan. Xiamen International Bank bought public-cloud LLM token services and AI tools from Alibaba Cloud. Shanghai Rural Commercial Bank is soliciting POC vendors for an OA subsystem with office and document assistants. Jinhua Bank's OA upgrade landed with Weaver Network in August for 820,000 yuan.
Why this pattern repeats is not hard to grasp. Wu Zewei, a researcher at Sunshines Bank, points out that pay-as-you-go seat and token models keep upfront costs near zero, contain sunk costs, and plug into existing office systems and account structures without heavy customization. Banks get to validate results first, then decide whether to scale.
But inside banks' strict control regimes, lightweight tools remain the co-pilot, not the driver. Frontline staff say intranet-deployed in-house assistants are still the primary tool, with public models reserved for first drafts. Marketing scripts, client communications, weekly reports and meeting minutes are the most common use cases. Anything involving number verification still gets checked by hand in business systems. Data leakage, model hallucination, permission overreach and accountability draw the line on how far these tools can travel.
How that tension between cost saving and compliance resolves will determine where the "office companions" end up. One thing already seems settled. When AI procurement is counted in hundreds of thousands of yuan instead of billions, smaller banks have voted with their wallets.