• China Merchants Bank says average daily token throughput rose more than 78% while AI contributed 13.88 million equivalent staff hours in the first half
  • The figures show banks beginning to measure AI consumption against work performed rather than treating token growth alone as progress

The fact

China Merchants Bank said its average daily AI token throughput increased by more than 78% in the first half of 2026, while its AI applications contributed an estimated 13.88 million equivalent employee working hours.

The bank had deployed 256 specialised models and 1,386 intelligent scenarios by the end of June. Its AI systems are being used across retail banking, corporate banking, risk control, operations and office work. The equivalent-hours figure is the bank's estimate of work performed or time saved by those applications; it is not a count of jobs removed or an audited payroll saving.

Other Chinese banks are also paying closer attention to AI operating costs as model use expands. Industry reporting shows institutions comparing token consumption, computing costs and business output when deciding where to deploy new AI applications.

The assessment

China Merchants Bank's figures put two useful numbers next to each other. Daily token throughput rose by more than 78%, showing that employees are using AI much more heavily. The 13.88 million equivalent staff hours try to answer the harder question: what did all that computing actually do?

That figure still needs care. An hour saved on drafting, searching or processing does not automatically become an hour of payroll savings. Staff may spend the time on customers, difficult cases or other work, and some AI output still has to be checked before it can be used. The value will therefore differ from one task to another.

For BTW readers, the better comparison is the cost of getting a piece of work finished to an acceptable standard. Token prices are only part of that bill. A model that looks cheap can become expensive if it needs repeated attempts or leaves employees with more checking afterwards.

What to watch

Watch whether banks begin reporting task-level measures alongside token use, including completed cases, review time and operating cost. Procurement terms will be particularly revealing if banks start linking model or computing payments to accepted output rather than consumption alone.