400,000 prompts per day: What the big banks' interim AI results mean for regular operations

For a long time, the rule of thumb was: banks experiment with artificial intelligence, but little of it makes it into regular operations. The quarterly reports and announcements of this summer paint a different picture. On the second-quarter 2026 analyst call on July 14, Bank of America CEO Brian Moynihan cited figures that make people sit up and take notice: More than 200,000 of the institution's employees work with AI tools, generating over 400,000 prompts per day. More than 300 AI use cases are approved internally, including over 100 based on generative AI; several dozen of these are already fully rolled out. Shortly before that, in mid-June, HSBC and Google Cloud had announced a multi-year partnership designed to enable more than 200 new AI use cases over the next two years, delivering more than $100 million in value. And BBVA, together with Visa, reported the first payment triggered by an AI agent in Europe. The transition from pilot to production operation is no longer a declaration of intent for the majors — it is happening.

What the numbers actually say

What is remarkable about Bank of America's statements is less the sheer scale than the structure behind it. The institution does not speak of a general "AI rollout," but of a maintained inventory of approved use cases — each one reviewed, prioritized, and backed by an economic value proposition. Moynihan emphasized that all approved use cases had "good economics": advisors prepare client meetings faster because research and presentation materials are generated automatically; developer teams use coding assistants; service units work with agentic workflows.

HSBC's approach seems structured in a similar way: The partnership with Google Cloud and Google DeepMind's engineering teams deliberately starts with a few key areas — highly personalized wealth advisory and earlier detection of financial crime — and builds on an existing base of several hundred applications already running in the cloud. Here, too, the rule is: prioritization and infrastructure first, then scaling.

The message behind both reports: The difference between experiment and regular operation does not lie in model access. High-performance models can be purchased by any institution today. It lies in the governance of use cases, in the database, and in the ability to make benefits measurable.

Regular operation has three requirements

Three recurring patterns can be read from the reports of the pioneers:

• An actively managed use case portfolio. Approval processes, a central inventory, and clear utility calculations per use case. This sounds bureaucratic, but it is the foundation for getting the right dozen out of hundreds of ideas into production — and for enabling regulators to understand at any time what is in use and where.

• A resilient database. Personalized advice, fraud detection, and agentic workflows all access the same resource: up-to-date, linked, quality-assured customer data. Institutions whose customer data is scattered across silos can build individual pilots — but not regular operations.

• Broad enablement instead of isolated solutions. 400,000 prompts per day do not originate in an innovation lab, but when the entire workforce receives tools, guardrails, and training. The productivity impact comes from the broad base.

Why this is not just a major bank issue

One might object: With the budgets of Bank of America or HSBC, many things can be industrialized. But the logic applies regardless of size — it just shifts. Mid-sized institutions do not need to operate 300 use cases. They need the few that have the greatest leverage in their own business model — in the retail business, typically along the customer lifecycle: reaching out at the right time, relevant product recommendations, early detection of churn, automated follow-up processes in service.

It is precisely here that the gap between pilot and production is often smaller than expected once the data issue is resolved. Segmentation based on daily updated transaction and interaction data supports several use cases simultaneously — from campaign management to churn prevention. Conversely, even the best model remains ineffective if it operates on outdated or incomplete data.

Where scaling fails in practice

Experience from projects in recent years shows: It is rarely the models that fail the step into regular operation. Three stumbling blocks keep appearing:

• The pilot proves the wrong thing. Many pilot projects show that a model works — but not that it fits into existing processes, systems, and responsibilities. A prototype on a data export is quite different from a service integrated daily into campaign management or advisory services.

• No one owns the use case. If the business department, IT, and data team share responsibility without any one entity being responsible for benefits, operation, and further development, the case is left abandoned after the pilot. The pioneers solve this with clear ownership per use case.

• Success is not measured. Without defined key performance indicators — conversion, processing time saved, churn prevented — it is impossible to prioritize or justify expansion. What the big banks are demonstrating is, at its core, consistent benefits controlling.

Keep the regulatory flank in mind

The transition to regular operation falls into a phase in which the supervisory authority is clarifying its expectations. The transparency requirements of the AI Act under Article 50 apply as of August 2, 2026 — anyone using AI in customer contact must disclose this. The ECB is requiring significant institutions to submit action plans against AI-powered cyberattacks by the end of October. And the postponed high-risk obligations for applications such as credit scoring will return to the agenda from December 2027. Anyone setting up production structures now — inventory, approvals, documentation, data quality — is simultaneously fulfilling a large part of what regulation will require anyway. Governance here is not a stumbling block, but the ticket to scaling.

Conclusion

The summer reports from Bank of America, HSBC, and BBVA mark a leap in maturity: AI in banking in 2026 is no longer a collection of pilots, but is becoming an operating model. The pioneers show what matters — managed use cases, a viable database, and enablement across the board. For all other institutions, this is less a question of budget than of sequence: first consolidate customer data and prioritize the most valuable use cases along the customer lifecycle, then scale. The tools to do so are available. The competitive advantage is being created right now — among those who get started.