A Sibos 2026 panel poll in Miami found 61% of respondents have little or no visibility into AI token usage, and only 13% manage costs at the token level. HSBC, BNY and SMBC leaders explain how they judge AI returns.
Most financial firms cannot track their AI token costs, a poll at Sibos 2026 found.
The poll ran during a panel at Sibos in Miami, Global Finance reports. Tokens are the basic unit of text that large language models process.
They are also the main metric model providers use to bill clients. Visibility into AI token costs is where cost control starts.
What Did the Poll Find?
About 61% of respondents reported little or no visibility into AI token usage inside their organization. They said they did understand their overall AI spending.
Only 13% reported detailed visibility and active cost management at the token and model level.
The gap is between the bill and its drivers. Many firms know the total but not which models or agents produce it.
Melissa Tuozzolo, global head of client services at HSBC, described what happens without controls. Some partners “let everyone go crazy with this stuff and build their own agents only to be left with a $100 million bill after a few months.”
Is Token Cost the Right Metric?
Not by itself, said Isabel Schmidt, executive platform owner for payments enablement at BNY. She agreed that clarity on token cost is still missing in many organizations.
But she warned against getting pulled into that detail. “AI is not a goal itself. The ultimate question is: How much does that business process cost me?”
Tuozzolo added other costs to the ledger. They include data governance and cleaning, upgrading existing systems and deploying new technology.
David White, global head of product and data at LSEG, urged product teams to keep the framing simple. He wants them to treat AI as “just another technology to help deliver value to the customer.”
Where Are Banks Seeing Returns?
Schmidt sees AI’s primary value in efficiency, effectiveness and capacity. Efficiency gets most of the attention because processes are easier to automate than they were a year or two ago.
That focus has a side effect, she said. It “also forces us to think more clearly about some of the others.”
HSBC offers a case on effectiveness. About six months ago it deployed an AI-powered “air traffic control” system that routes client queries to the right expert or team.
It cut first-response time for clients in certain markets by 10 hours. Before that, people took in queries and triaged them by hand.
Other deployments freed processing capacity at the bank. Tuozzolo said costs used to rise with transaction volume. “This is the first year in a few years that we’re actually seeing it go the other way.”
She credits HSBC’s cleaner data layer as well as AI.
Sumitomo Mitsui Banking Corp. is deploying an AI-powered call center. It will process calls at 70% of the older model’s cost.
Kazuya Ikeda, a senior executive manager at SMBC, said most of the project’s cost is tokens. “But we are seeing a positive ROI.”
SMBC’s case shows AI token costs can dominate a project budget and still pay back. He also cautioned that value for clients is not always economic value. If peers do the same thing, he said, it is not necessarily creating relative value.
How Does This Land in Miami?
Sibos ran Sept. 28 to Oct. 1 at the Miami Beach Convention Center. MAIN has covered Miami hosting Sibos 2026 for the first time.
MAIN also covered how Sibos banks bet on AI you cannot see, where leaders stressed auditable systems over flashy ones.
Miami’s own fintech companies face the same measurement question. MAIN has profiled five Miami fintech AI firms rewiring how money moves.
Tokens are only one input. Tuozzolo’s list of data, systems and deployment costs sits beside them in any honest ROI calculation.
The poll shows spending is visible and unit costs are not. A firm that cannot price a business process cannot prove its AI return.
Sources: