Deep Dive — CICC's 100 billion tokens a day, and no ROI number

China International Capital Corporation burns between 70 billion and 100 billion tokens a day, and expects to hit 200 billion next year. The figure comes from the bank's own IT chief, Xie Bisong, at a Tencent Cloud briefing on October 28, and 21st Century Business Herald reported it under a headline that names the real story: the AI race in Chinese finance has moved from counting tokens to counting return.

That return does not yet exist as a number. Xie told reporters he cannot map the spend to output. Asked what the efficiency gain actually is, he said the account probably does not balance if you look at short-term financial return, and that AI should be treated as a strategic investment instead.

Strip the corporate phrasing and you have the most useful datapoint of the year about enterprise agents: a top-tier investment bank has scaled one to more than a thousand daily users, has the token bill to prove it, and still cannot tell you what a finished task costs.

What a reader would type to find this: "AI agent ROI", "enterprise agent token cost", "Tencent WorkBuddy". This is the story of the first large customer cohort hitting that wall.

What Tencent actually sold

WorkBuddy is Tencent Cloud's desktop agent — not a chatbot window, but a program that reads files, calls internal knowledge bases, runs multi-step tasks and hands back finished work. It shipped in March 2026. The enterprise edition arrived June 5, and on September 3 Tencent added four finance-specific builds covering corporate banking, retail banking, individual customer operations, and investment research and advisory.

The customer count moved fast. When the finance edition launched, Tencent said 100-plus financial institutions had signed on. At the digital ecosystem conference in late October that number was 250, spanning underwriting due diligence, research, compliance review, client operations and wealth allocation. Tencent Cloud vice-president Hu Liming said the largest institutions have moved from pilots to expanded procurement, and that more than 500 users inside one company is the point where a deployment counts as real.

The wider product numbers, disclosed to Cailian Press at the same event, are the kind a platform company cites when it wants to look like an operating system: 100 major releases since March, 9 billion task requests served, deployment across more than 50 industries, 50 percent of China's top 500 companies and 60 percent of its top 100, a marketplace of more than 200,000 packaged skills, and over 2,000 partners. Tencent has also built 264 finance-specific skills and 29 connectors with its first ten industry software partners, including Vertex Software, Hundsun's Goldstar, Shenzhou Information and Sinosoft — the integrators that already sit inside every Chinese bank's stack.

That partner layer is the part outsiders skip. WorkBuddy is not being sold as shrink-wrapped software. Tencent's own forward-deployed engineers go into a bank, map the workflow, build a reference deployment, and hand it to partners to replicate. When a vendor has to place engineers on site to make the product work, the customer is buying a project, not a licence — and projects are exactly where ROI arithmetic gets murky.

The bank that voted for it

CICC's path is unusually well documented for a Chinese enterprise deployment. The bank began trialling WorkBuddy in March, ran several desktop agents in parallel for two to three months, then let employees vote. WorkBuddy won, and was folded into the bank's existing in-house assistant, Xiao Jin. Cumulative users of that assistant passed 5,000, with more than 1,000 active WorkBuddy desktop accounts, per Cailian Press.

Then came the number that made the story: 70 billion to 100 billion tokens a day.

Divide one by the other and the average active account is burning 70 million to 100 million tokens a day. That is roughly 50 to 75 million English words, or 500 to 750 novels, per user, per day — arithmetic on the disclosed figures, not a claim CICC made, and obviously not all traffic routes through those desktop seats. But order of magnitude is the point. This is not a person asking questions. It is a fleet of agents reading documents, iterating on drafts and re-reading context windows for hours.

CICC did not stumble into that volume. Xie described a deliberate early strategy the bank internally called "Token Maxing": encourage employees to use the tool as much as possible, and make consumption a performance metric. The bank ran an internal agent-building competition. Staff were pushed to burn.

That worked, which is the problem. When consumption is the KPI, consumption is what you get. Xie's own diagnosis of where the waste sits is almost touching in its ordinariness: employees unfamiliar with the tool paste very large documents into it, and the context window eats the bill. Nobody optimises a number they are being rewarded for raising.

The metric that replaced the metric

The bank is now unwinding its own incentive. Token consumption is still tracked, but no longer as something higher-is-better. CICC is building a shared operations platform and staffing people to run it, tracking who is consuming tokens, in which scenarios, to what result, and where the invalid calls are. In other words, it is doing capacity planning on a workload it invented eighteen months ago.

This is the correct instinct and it is also an admission: the first phase of enterprise agent adoption in China was measured with a metric that had no causal link to value. China's national token statistics show how thoroughly that metric spread. Daily token calls across the country went from about 100 billion in early 2024 to more than 500 trillion in July 2026 — a 5,000-fold rise in two years, per figures from the National Data Administration cited by Cailian Press. Some of that is real work. Some of it is people pasting PDFs into a context window because someone told them to use the tool more.

Tencent group senior vice-president Dowson Tong used the conference to name the correction: enterprises need to stop defaulting to the strongest, most expensive model for every task and start matching model to job. That is a reasonable thing for a cloud vendor to say — it also happens to be the same advice that would shrink the token line Tencent is being paid on.

Cost per token is the wrong denominator

The industry's habit of pricing agents in tokens per million is the reason CICC cannot answer its own board's question. Tokens are an input. Boards buy outputs. Two agents can quote identical per-token prices and produce wildly different bills for the same finished task, because one of them re-reads its own history four times and the other does not.

We worked through a clean example of this in Haiku 5.5's fine print: same price, three times the bill, where two models at identical list prices produced measured per-task costs of $0.21 and $0.07 — a 3x gap created entirely by verbosity. A 40 percent discount on the chattier model is still a loss. At 100 billion tokens a day, a 3x verbosity penalty is not a rounding error; it is the difference between a defensible budget and a scandal.

The fix is unglamorous and mostly organisational: instrument the trace, attach a cost to every completed task, and compare that number against what the task used to cost in human hours. Almost no enterprise deployment does this, because the harness that would emit those traces is the vendor's, and the vendor is billing on volume.

There is a second reason the numbers stay fuzzy. Most of these deployments never leave the first stage of a progression Hu laid out to reporters: point, line, surface. Point is one employee working faster. Line is a whole workflow — gather, analyse, review, output — handed to an agent across several systems. Surface is agents in different departments handing work to each other. Hu said most financial institutions sit between point and line, and that is exactly the boundary where agent value moves from plausible to provable, Leiphone reported.

The blocker there is not model quality. It is that a bank's knowledge is scattered across systems with permission boundaries nobody has ever had to write down for a machine, and that every regulated action needs a trace and a named owner. Getting an agent read-access to a trading system's research archive is a permissions project, not a prompt.

The contrarian read

The generous interpretation of CICC's position is that it is honest. Agent deployment at this scale is genuinely eighteen months old, and the bank is admitting publicly that the accounting lags the rollout. Most Western enterprises in the same position say nothing at all, and their silence gets read as success.

The ungenerous interpretation is that "strategic investment" is the phrase a technology organisation uses when the arithmetic has been run and did not clear. Ninety-nine percent of the token volume at a securities firm is document reading, summarisation and drafting — work that a competently designed workflow with a small model handles at a fraction of the cost. If the bank needs 100 billion tokens a day to reach the same draft, part of that bill is buying the learning curve rather than the output, and the learning curve is not a line item anyone wants to defend in a budget review.

And there is a pricing collision coming that neither side has priced in. Enterprise AI is still sold per seat, and per-seat pricing assumes a headcount that agents are supposed to reduce. We went through how hard that model is already being discounted in ChatGPT Enterprise's $7 seats, where a US bank got seats at just under $7 against a $20 list. Tencent meters WorkBuddy on seats and token consumption, per 21st Century Business Herald, which works while customers are still expanding. It stops working the moment a bank's procurement team does the obvious thing and asks why it is paying per person for software that lets one person do the work of five.

What to watch

Three numbers, in order of usefulness. First, whether CICC's projected 200 billion tokens a day next year arrives with a cost-per-task figure attached — if the bank publishes efficiency in completed work rather than volume, the measurement problem is being solved. Second, whether Tencent puts a per-task price on the four finance Buddy apps, which would force the whole Chinese enterprise agent market onto an output-based meter. Third, whether the 250 institutions behave like reference customers or like pilots that renew once and quietly shrink.

The thing not to watch is the token counter. It went up 5,000-fold nationally in two years and told us almost nothing about whether any of it was worth buying. CICC's most valuable disclosure is not the 100 billion. It is that the bank running the biggest deployment in Chinese finance will not put a number on the return — and that the honest ones are the only ones you can trust to fix it.