BBVA's AI rollout put the GPT builder in front of 100,000 bank employees

The front end is 20,000 employee-built GPTs, not one app

The headline number in OpenAI's write-up is 100,000 employees on ChatGPT Enterprise. The more revealing number is 20,000 — the count of custom GPTs BBVA employees have created, of which roughly 4,000 are used frequently by teams worldwide.

That distinction matters for anyone thinking about how AI actually reaches users. BBVA did not build one bank-wide assistant and route everyone through it. Instead, the interface fragmented into thousands of small, purpose-built tools owned by the people who understand each workflow — legal, risk, customer service, finance, marketing.

A GPT is a thin front end wrapped around a model plus some instructions and knowledge sources. By letting employees assemble those wrappers themselves, BBVA pushed interface design down to the domain expert rather than centralizing it in a product team. Four thousand of those wrappers stuck. The rest presumably did not — which is the expected shape of a build-and-discard experimentation model.

What the surviving GPTs actually do

The source names four concrete examples, and each one is a narrow interface over a specific document or data problem rather than a general chatbot.

Credit Analysis Pro GPT extracts and analyzes unstructured data from annual reports, ESG disclosures, and media coverage — work described as previously manual and time-intensive. A Retail Banking Legal Assistant GPT drafts responses to roughly 40,000 annual client-related legal inquiries by pulling from internal knowledge sources, serving a nine-person legal team. In Mexico, a Client Experience Assistant GPT parses thousands of open-ended survey responses for sentiment and themes.

The most concrete performance figure comes from Peru, where more than 3,000 employees use an internal AI assistant that cut average query handling time from about 7.5 minutes to around 1 minute — the roughly 80% efficiency figure cited in the results summary. Note that this number applies to one deployment, not the whole bank; the org-wide figure OpenAI reports is about three hours saved per employee per week.

Governance was built into the interface layer from the start

BBVA's stated reason for providing enterprise-grade access was to avoid "unauthorized experimentation with consumer AI tools." In other words, the deployment was partly a strategy to give employees an approved front door so they would not use an unapproved one.

We created a secure environment for learning and using AI. —Elena Alfaro, Head of Global AI Adoption at BBVAMontana Labs

The bank pairs that access with an "AI champions" network and advanced users called "wizards" who run hands-on workshops. It also trained 250 leaders, including the CEO and chairman, and reports that executive committee members are now among the most active users. For a self-service GPT-building model to stay safe in a regulated bank, the enablement and governance scaffolding around the interface matters as much as the interface itself.

The implication: distributed GPT-building is a viable interface strategy at bank scale

The reusable lesson here is not that a global bank adopted ChatGPT. It is that BBVA reached six-figure adoption by treating the custom GPT as the unit of deployment — a lightweight, employee-authored interface over a specific task — rather than commissioning a small number of centrally designed applications.

That approach trades polish and control for coverage and speed. Twenty thousand GPTs with a roughly one-in-five frequent-use rate implies a lot of abandoned experiments, and the source offers no detail on how BBVA audits or retires them. But the model surfaced use cases — legal drafting, credit analysis, survey sentiment — that a central roadmap might never have prioritized.

For teams weighing how to put AI in front of a large workforce, BBVA is a datapoint that the interface can be delegated, provided the security, legal, and compliance alignment is set before the building starts rather than after.