BBVA’s widening AI rollout turns staff training into a banking operations question

BBVA expands employee AI access and training. The analysis follows the shift from licences and self-reports to maintaining dependable work.

Giving an employee access to an AI assistant is an event. Helping that employee change a banking task responsibly is a longer process. BBVA's public reporting now makes it possible to follow that transition through several stages: an early group of users, an expanded licence agreement, a bank-wide ambition and a later account of training and use. The interesting story is how an experiment becomes everyday work without losing the people who understand the work.

In a [20 November 2024 account](https://www.bbva.com/en/innovation/bbva-sparks-a-wave-of-innovation-among-its-employees-with-the-deployment-of-chatgpt-enterprise/), BBVA said employees were building specialised GPTs and sharing some through an internal store. It reported that 83% of the initial licensed users used the tool weekly. This is a bank-published adoption account, not an independent productivity study. All sources for this article were checked on 16 October 2026.

The denominator changes as access widens

On [12 May 2025](https://www.bbva.com/en/innovation/bbva-expands-its-agreement-with-openai-to-11000-chatgpt-licences-for-the-banks-employees/), BBVA announced expansion from 3,300 to 11,000 ChatGPT Enterprise licences. It reported that users said they saved an average of 2.8 hours a week and that 83% of licensed users used the tool at least daily. Those are reported experiences among licensed users at that stage. They do not establish that every bank employee achieved the same saving or that the time became a cash saving for the bank.

The early weekly measure and the later daily measure should not be stitched together into a single controlled experiment. The licence pool, reporting moment and use measure differ. A keen early user may also be unlike someone receiving access during a broad rollout. An honest reading keeps each figure attached to its original population instead of multiplying a self-reported average by the entire workforce to manufacture a financial return.

This matters for managers deciding what support to provide. A developer who already knows how to verify generated code faces a different learning task from a relationship manager checking a translated explanation. A risk specialist may need to trace a summary to the original document. The common tool does not make those tasks interchangeable. Training has to connect to the decisions people actually make and the consequences of accepting an incorrect output.

Bank-wide ambition, then a later adoption account

BBVA's [12 December 2025 strategic-alliance announcement](https://www.bbva.com/en/innovation/bbva-and-openai-seal-a-strategic-alliance-to-redefine-banking-with-artificial-intelligence/) said it would extend ChatGPT Enterprise to more than 120,000 employees. The same announcement discussed work on customer experiences and showed a demonstration involving digital banking in Italy and Germany. A rollout intention and a demonstration are different stages from general production availability. This article does not describe the demonstrated customer experience as an already available service for every customer.

A [Spanish-language BBVA page first published on 24 April 2026 and marked updated on 29 September 2026](https://www.bbva.com/es/innovacion/bbva-impulsa-la-adopcion-de-la-inteligencia-artificial-desde-el-talento/), in the version reviewed here, says nearly the whole workforce had access to generative-AI tool licences through global arrangements, and that more than half used such tools weekly. It also reports more than 280,000 hours of AI-related training during 2025. These remain company-reported adoption and training measures. The Spanish source is paraphrased here in English; no interview or independently observed staff experience is implied.

The sequence is more useful than a single dramatic number. The later account suggests attention to access, training and internal adoption roles, rather than treating a licence purchase as the end of the project. It does not show how learning was distributed across teams or what every participant could do afterwards. A training hour is an input. Competence requires evidence that a person can apply the lesson, notice an error and choose a suitable recovery action.

Employees become authors of small tools

The early internal-store account raises another practical issue. When employees create task-specific assistants, useful knowledge can spread from the people closest to the work. A recurring translation, reporting or document-search problem may be spotted by a colleague long before it appears in a central technology roadmap. That is an attractive feature of distributed experimentation, provided responsibility does not vanish when an assistant becomes popular.

The operational questions are concrete. Who maintains the instructions when a procedure changes? Which documents may the assistant read? Can colleagues tell whether a tool is a personal experiment or an approved part of a workflow? What happens when its original author changes jobs? These are analysis prompted by the reported internal sharing, not assertions about deficiencies at BBVA. A shared assistant needs a service owner as well as an enthusiastic creator.

For customers, much of this work is invisible but consequential. A staff member may produce a clearer explanation or prepare more quickly for a conversation. The same tools could also produce a plausible summary that misses an exception. Customer-facing quality therefore depends on what happens between generation and use. An employee should be able to verify the relevant facts without being pressured to treat machine output as authoritative because the bank has invested in the technology.

Workforce implications should be described with the same care. The reporting supports a story about changing skills, wider access and considerable training activity. It does not provide grounds for a job-loss forecast. Time released from a routine task might be spent checking output, supporting customers or doing work previously deferred. Whether it changes staffing depends on choices and circumstances beyond the adoption measures cited here.

A proposed learning example: teach the task, then test the judgement

An original training design for an invented banking team could start with a multilingual customer-information sheet. The team works only with synthetic customers and approved teaching documents. The task is to draft a plain-language explanation in another language while preserving the meaning of fees, conditions and exclusions. This is a learning proposal, not a description of BBVA's curriculum or a legal interpretation of a real product.

Participants first write down what the explanation must preserve. They then use an assistant to prepare a draft and mark each statement against the teaching source. A bilingual reviewer introduces a deliberately awkward expression and a condition spread across two sections. The assessment asks whether the learner notices a missing qualification, rather than rewarding the shortest completion time or the smoothest-sounding translation.

The exercise then moves to a shared assistant. Another learner changes the approved source document, leaving an older version in the teaching archive. The group must decide how the tool identifies its current source, who approves an instruction change and how a colleague knows that an answer needs checking. This turns an abstract lesson about hallucination into a task about maintaining an ordinary banking explanation over time.

Finally, learners practise stopping. They should be able to say that a draft is unsuitable, preserve the useful work and return to a conventional process. A training programme that teaches only how to obtain an answer may leave staff less prepared than one that also teaches how to reject it. The output of this exercise would be a proposed verification routine and a record of skills demonstrated, with no invented productivity percentage.

BBVA's rollout shows why banking AI adoption is becoming an organisational matter rather than a collection of individual prompts. Access creates the opportunity to experiment; staff development determines whether the experiment becomes dependable work. The next useful evidence would connect those learning and maintenance practices to the quality of real tasks. Licence counts, self-reported time and training hours each illuminate part of that story, but none can stand in for all of it.

Explanatory diagram and accessible reading guide


*Original educational infographic based on attributed reports and a fictional proposed exercise; not the named bank’s internal architecture, actual customer decision or measured result.*

[Open the full-size explanatory diagram](/assets/bank-spain-adoption-measures-20261016.svg)

This original diagram explains attributed source context and a fictional learning exercise. The exercise has not been run; the diagram is not an authentic product interface, internal bank architecture or measured operational result.

Access, use and competence are different

Dated BBVA self-reports followed by a proposed teaching pathway. Arrows between publications show chronology, not causation.

November 2024 account

83% of the initial licensed users used the tool weekly.

May 2025 expansion

3,300 to 11,000 licences; 83% of licensed users daily. Users reported 2.8 hours saved weekly.

December 2025 plan

Announced extension to over 120,000 staff. Italy and Germany app experience was a demo.

Reviewed 2026 page

First published 24 April; marked updated 29 September. Keep access, usage and training separate.

Proposed source comparison

Draft a teaching translation, then compare conditions and exclusions with the original.

Use, correct or reject

A reviewer checks meaning. A fluent draft can still need correction or rejection.

Do not combine the two 83% measures, multiply reported time by the entire workforce, or convert training hours into proven competence.

The April 2026 page is marked updated 29 September 2026. Its reviewed version reports nearly all staff with access, more than half using tools weekly and over 280,000 training hours in 2025.

Publication records describe different cohorts and measures. The dashed bridge to the exercise is editorial, not a measured effect.

Solid arrows show the stated or proposed information handoff. Dashed arrows show an editorial bridge or a return for checking. No arrow proves a measured causal effect.

The later publication reports a changed licence pool and daily measure.

The December announcement adds a bank-wide intention and demo.

The reviewed 2026 page is a later account with its own version dates.

Reported adoption prompts this original teaching proposal; it does not prove its outcome.

The proposed exercise assesses source checking and judgement.

A rejected or incomplete draft returns for correction.

Diagram source references

- [S11: BBVA and OpenAI Seal a Strategic Alliance to Redefine Banking with Artificial Intelligence](https://www.bbva.com/en/innovation/bbva-and-openai-seal-a-strategic-alliance-to-redefine-banking-with-artificial-intelligence/) — BBVA.

- [S12: BBVA expands its agreement with OpenAI to 11,000 ChatGPT licences for the bank’s employees - NEWS BBVA](https://www.bbva.com/en/innovation/bbva-expands-its-agreement-with-openai-to-11000-chatgpt-licences-for-the-banks-employees/) — BBVA.

- [S13: BBVA sparks a wave of innovation among its employees with the deployment of ChatGPT Enterprise](https://www.bbva.com/en/innovation/bbva-sparks-a-wave-of-innovation-among-its-employees-with-the-deployment-of-chatgpt-enterprise/) — BBVA.

- [S14: BBVA impulsa la adopción de la inteligencia artificial desde el talento](https://www.bbva.com/es/innovacion/bbva-impulsa-la-adopcion-de-la-inteligencia-artificial-desde-el-talento/) — BBVA.