--- title: "78% of ICBC's remote service now runs without a human — the bank calls that the easy 78%" date: 2026-11-09 category: AI in Action site: NeuroAI canonical: https://neuroai.site/a/na-app-icbc-gongxiaozhi-ai-service language: en --- # 78% of ICBC's remote service now runs without a human — the bank calls that the easy 78% > ICBC says AI now handles 78 percent of remote customer service and absorbs work equal to 55,000 staff-years behind a filed customer-facing model. A call reaches a bank in China at eight in the evening. The caller is seventy, her card has been declined at a pharmacy, and she is not sure whether the problem is her password, her balance or the machine. The first voice she hears is synthetic. It asks her to state her problem, understands the sentence, checks the card status, and offers to unblock it — then, because her voice is shaking, hands her to a person in Hefei. That handoff is the part of the story the numbers do not capture, and it is where the next two years of bank AI will be decided. ## Key takeaways - Industrial and Commercial Bank of China (工商银行) says AI now carries 78 percent of its remote service workload, and that its "digital employees" absorb work equal to 55,000 full-time staff-years a year. - Its customer-facing assistant 工小智 covers nine business areas and 50 functions, and the bank says it was among the first batch of customer-facing large model (大模型) applications filed with China's cyberspace regulator. - In the first half of 2026, AI across remote channels served 133 million customer interactions, while the bank reported 600-plus AI scenarios in production. ## What "78 percent" is counting The figure comes from the bank's 2025 annual report: 78 percent of remote service is handled by AI. "Remote service" is a specific thing in Chinese banking — phone banking, in-app chat, text and video assistance — and it excludes branch counters and relationship managers. Most of that traffic is routine: balance checks, limit changes, card replacement, statement requests, password resets. Those are exactly the tasks a model with account access can now finish end to end. The bank describes its architecture as a hierarchy: a general model for language, financial micro-models for regulated actions, and an orchestration layer that picks between them. A balance query takes the fast path; a dispute is meant to escalate rather than improvise. The 55,000 staff-year figure is the bank's conversion of that volume into labour, not a headcount cut. It sits in the annual report beside the phrase 数字员工 — digital employees, a framing Chinese banks use for software that occupies a seat in the workflow rather than a chatbot on a website. ## The gate nobody outside the industry notices Before any of this could face a customer, it had to be filed. China requires generative AI services offered to the public to be registered with the Cyberspace Administration, and a bank putting a model in front of retail customers is squarely in scope. ICBC says its customer-facing application was among the first in the industry to complete that filing, and that the model runs behind a set of controls: retrieval restricted by a least-privilege rule so the model can only see the slice of a customer's data relevant to the query, semantic screening of user input, and real-time screening of generated output before it is spoken. This is the least glamorous part of bank AI and the most binding. A hallucinated interest rate is a mis-selling complaint; a hallucinated instruction about a deceased relative's account is a lawsuit. The filing does not guarantee accuracy, but it forces the bank to document how it would catch itself. ## Where the model still fails The remaining 22 percent is not a random sample. It is complaints, fraud alerts, bereavement, disputed charges, elderly customers who cannot navigate an app, and the long tail of products that were sold with an oral promise a decade ago. These are the calls where the customer is already angry and the correct answer depends on context that is not in any database. The bank's answer is a split: 工小慧 assists human agents with knowledge retrieval, suggested scripts and sentiment detection, while 工小智 faces the customer. The bank says the assistant cut call duration in priority scenarios by about 10 percent — a modest figure, and the most believable one disclosed, because it describes a human still doing the work with better tools. Handoff design, not model quality, is what customers actually feel. A system that routes a frightened caller to a human in twenty seconds is better than a system that answers 95 percent of questions and traps the other five. ## The arithmetic behind the ambition ICBC spent 28.588 billion yuan (≈ US$4.03B / HK$31.4B) on technology in 2025, equal to 3.41 percent of revenue, and the bank has said it built a domestic-chip compute (算力) pool of over a thousand accelerators with minute-level switching between training and inference. In the first half of 2026, revenue was 446.163 billion yuan (≈ US$62.8B / HK$491B), up 9.1 percent, and net profit was 176.475 billion yuan (≈ US$24.9B / HK$194B), up 4.5 percent. The operational numbers are more interesting than the financial ones: voucher data entry 25 times faster than manual work, replacing 100 million manual entries in six months; remote authorisation reviewed five times faster, now covering about half the bank's volume; private-banking proposal generation cut from five days to three hours. None of those are customer-facing. They are back-office tasks where the cost of an error is a correction, not a complaint — which is exactly where a large bank should be deploying models first. ## The next test is the branch The offline pilot is small: 21 high-frequency self-service tasks upgraded with a model, and more than 300,000 transactions completed in pilot regions in the first half. That is a rounding error against 630 million mobile banking customers and 290 million monthly active users. It is also where the hard cases live. A branch customer who needs help is, almost by definition, someone who could not complete the task alone. Automating that interaction means automating patience — and patience is the one thing a bank cannot file with a regulator. ## What readers can do now - **When a bank bot offers to "handle it for you", get the confirmation in writing** — an in-app message or SMS — before relying on it. - **Know the escalation phrase**: asking for a human agent is a right, not a favour, and most Chinese bank apps now expose it in the menu rather than the chat. - **For elderly relatives**, set up the branch or family-assisted channel in advance; the 22 percent of calls that defeat the bot are disproportionately theirs. ## Honest limitations The 78 percent and 55,000 staff-years figures come from ICBC's 2025 annual report as reported by Chinese financial media; we did not review the workload methodology, and "remote service" is a bank-defined category that excludes branches. The 133 million interactions, 600-plus scenarios and 100 million replaced entries come from the 2026 interim report. The claim of being among the first customer-facing model filings is the bank's; the regulator publishes no ranked list, so "first batch" cannot be independently confirmed. The 10 percent call-duration figure is a company number. Financial totals are as reported. Nothing here evaluates whether AI service is better or worse for customers — only what the bank says it deployed. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-app-icbc-gongxiaozhi-ai-service Free to quote with attribution and a link to the original.