The challenge
What was deployed
The results
Last Updated: November 16, 2026
BBVA spent two years handing generative AI to employees — 3,300 ChatGPT Enterprise licences in 2024, 11,000 by May 2025 — and then, on 6 October 2026, told analysts and investors that the next phase was a different problem entirely: an internal standard called 'The Frame' for creating, deploying, managing and governing AI agents at scale. The published outcomes are concrete: assistants filtering more than half of relationship-manager conversations in Spain, voice agents resolving close to 90% of interactions in production pilots in Peru and Mexico, complaint registration time down nearly 80%, developer productivity up around 28%. Nearly all of it comes from BBVA's own investor communications, and the bank has not published interaction volumes, costs, error rates or an audited P&L impact. This is third-party analysis of a public record, not an account of work we did.
The problem: 80 million customers and a pile of assistants that do not govern themselves
BBVA is a global bank with more than 80 million customers and operations spanning Spain, Mexico, Turkey, Peru and the US. Its first generative AI phase, beginning in May 2024, was deliberately bottom-up: buy enterprise licences, train people, and let the staff closest to the friction build their own assistants.
That phase worked on its own terms. By May 2025 the bank had expanded from 3,300 to 11,000 ChatGPT Enterprise licences, reported that 83% of licensed users opened the tool at least once a day, counted over 3,000 internally built assistants, and curated roughly 1,000 of them into an internal 'GPT Store' available group-wide. An internal hackathon in Mexico, BBVA Bot Talent, drew over 170 ideas, with winning projects aimed at commercial customer service, SME lending speed and fraud detection in email and SMS.
But thousands of assistants is a literacy result, not an operating model. Nothing in that inventory tells you which bot may read which data, which actions it may take without a human, how its accuracy is monitored over time, or what it costs. In a licensed bank, those are not nice-to-haves. The gap between 'our people use AI' and 'our bank runs on agents' is exactly where most enterprise programmes stall.
What BBVA actually built: 'The Frame'
At a Strategic Talk with analysts and investors, BBVA introduced 'The Frame', a common framework for creating, deploying, managing and governing AI agents at scale. The bank's framing is explicit: after a year of individual use cases, the lesson was that agent work shares many common elements — reusable components, shared functionality, the same management requirements — so the leverage is in standardising the factory rather than hand-building each agent.
According to BBVA's description, the standard covers a common, secure technology environment so teams do not start from scratch; defined roles and ways of working; clearly scoped access to systems and data for each agent; explicit operational boundaries separating what an agent may do autonomously from what requires human oversight; continuous evaluation and monitoring of performance; and systematic measurement of cost and business impact — all inside one security, governance and compliance framework. The bank also says the design preserves technological independence so it can keep working with the most relevant models, tools and platforms rather than being locked in.
everyone has access to AI models. The difference will lie in who is able to integrate them into their systems at scale, manage them safely and responsibly, and use them to create value for their customers
— Antonio Bravo, Global Head of AI Transformation, BBVA, BBVA
How BBVA describes an agent built under 'The Frame'
Teams assemble agents from reusable components inside a common, secure technology environment instead of rebuilding infrastructure per use case.
Each agent gets clearly defined access to the systems and data it needs — and nothing beyond.
What the agent may execute autonomously is separated from what requires human oversight, by design rather than by convention.
Performance is continuously evaluated and monitored after deployment, not signed off once at launch.
Cost and business impact are measured systematically, inside a common security, governance and compliance framework.
The measured outcomes BBVA has published
BBVA's October 2026 update is unusually specific for a bank talking about AI to investors. The figures below are the bank's own.
Two of these deserve a closer look. The relationship-manager filter is a classic deflection metric, but applied to an inbound channel inside the bank's own app rather than to a public chatbot — the assistant triages and answers, and the banker sees what survives triage. The complaint registration figure is a cycle-time metric on intake, which is where regulated firms usually bleed hours: complaints must be logged accurately and promptly, and the logging is tedious structured-data work that language models handle well.
Engineering, not just chat
BBVA is also pointing agents at its own software lifecycle, with the stated goal of compressing release cycles.
Our ambition is very clear: to multiply our capacity to deliver solutions and shorten release cycles from months to weeks, with AI agents helping to orchestrate and execute the development lifecycle while people retain control over context, decision-making and oversight
— Carlos Casas, Global Head of Engineering, BBVA, BBVA
The same update reports that over the past three years BBVA cut significant technology incidents affecting service availability by 53% and average recovery time by 52%, and that close to two-thirds of online transactions were processed in the cloud in 2025, with a target of 90% by 2029. Those resilience numbers are presented as the product of platform engineering work — not as an AI result — and should not be read as agent outcomes.
How the two phases differ
| Dimension | Phase 1: assistants (2024–2025) | Phase 2: agents under 'The Frame' (2026→) |
|---|---|---|
| Who builds | Individual employees and business areas, bottom-up | Teams building on a shared secure environment with defined roles |
| Unit of work | Custom GPTs for drafting, summarising, translation, analysis | Agents that execute steps in a process, with bounded autonomy |
| Control model | Enterprise tenancy, training, usage policy | Per-agent data access, action boundaries, human oversight, continuous evaluation |
| Published metric | Adoption and self-reported hours saved | Process metrics: containment, filtering, cycle time, developer throughput |
| Visible risk | Shadow usage, duplicated effort | Unbounded actions, silent quality drift, unattributed cost |
The customer-facing edge is still deliberately narrow
BBVA has been experimenting at the interface layer too. In December 2025 it demonstrated a conversational app concept connected to ChatGPT for its digital banks in Italy and Germany, letting people ask in natural language about account conditions, card types or savings products. The bank was clear that the initial version handles informational queries, with the ambition of evolving into a digital companion over time; at the time of the announcement, apps in ChatGPT were available to logged-in users outside the EU. Press coverage of the May 2026 launch described it as a discovery and information tool that does not move money, and reported BBVA's intention to scale ChatGPT Enterprise towards 120,000 staff, per PYMNTS.
The bank's own assistant for customers, Blue — described as a multimodal personal adviser available 24/7 — is planned for gradual rollout to individual and business customers in 2027, with AI agents also extending to cross-cutting processes such as claims and financial crime prevention. In other words: internal and semi-internal first, autonomous customer advice later.
BBVA signs with OpenAI and distributes 3,300 ChatGPT Enterprise licences to explore process acceleration and internal innovation.
Licences expand to 11,000. BBVA reports 83% daily usage among licensed users, 3,000+ assistants built, 2.8 hours saved per week on average, and an internal GPT Store of around 1,000 group-wide assistants.
OpenAI names BBVA among early GPT-5 adopters in financial services; OpenAI's Olivier Godement describes strategic tasks that normally take two to three weeks being completed in hours.
BBVA demonstrates a conversational app concept inside ChatGPT for its Italian and German digital banks.
At a Strategic Talk with analysts and investors, BBVA presents 'The Frame' and publishes agent-era metrics: voice agents near 90% resolution in pilots, complaint registration time down ~80%, developer productivity up ~28%.
What the public record does not establish
Being candid about the gaps matters more than the headline percentages.
The voice-agent figure is explicitly described as coming from production pilots, and BBVA has not published interaction volumes, the mix of intents covered, or its definition of 'resolve' — containment inside the agent is not the same as a customer confirming their issue was handled. The complaint metric is registration time, not time to resolution or complaint outcomes, and no baseline is given. The ~28% developer productivity gain has no published methodology and no accompanying defect, rework or review-load data. The time-saving figure is self-reported, and it moved from 2.8 hours per week in May 2025 to 2.4 hours per week in October 2026 across a much larger user base — BBVA has not explained the change, which is a reasonable reminder that survey-based savings are directional at best.
Beyond that: no euro figure for AI spend or returns, no published hallucination or escalation rates, no customer satisfaction delta, no count of how many agents actually run under 'The Frame' today, and no disclosure of which models sit behind which agents. The figures come from BBVA's investor and newsroom communications and have not, as far as the public record shows, been independently audited.
Treating assistant sprawl as progress
Thousands of employee-built bots look like momentum on a slide. They are also thousands of ungoverned artefacts with no shared permissioning, no evaluation, no owner and no cost line. The moment any of them need to take an action rather than draft text, the absence of a platform becomes a blocker — and in a regulated firm, a compliance problem.
What an integration team should take from this
First, sequence matters more than model choice. BBVA's second-phase announcement is almost entirely about control surfaces, not capability: access scoping, action boundaries, monitoring, cost attribution. That is the unglamorous work that decides whether agents graduate from pilot to production in a regulated business.
Second, pick metrics with an owner. Containment on voice, filtering on inbound app messages, intake cycle time on complaints, throughput in engineering — each maps to a team that already has a target. 'AI value' does not.
Third, resist the temptation to front-run autonomy at the customer edge. BBVA's customer-facing surface is still, by its own description, discovery and information; the agents doing real work sit behind employees. That ordering is a choice, and it is the one the published evidence supports.
Finally, read vendor-adjacent numbers the way you would read any unaudited management figure. These are good disclosures by banking standards — and still not enough to build a business case on without running your own baseline. If you are mapping a similar programme, start with the process you can instrument today, not the agent you want to demo. Talk to us or browse other deployment write-ups.
Frequently Asked Questions
What is BBVA's 'The Frame'?
'The Frame' is BBVA's internal standard for creating, deploying, managing and governing AI agents at scale, presented to analysts and investors in October 2026. It provides a shared secure technology environment and reusable components, plus per-agent data access rules, explicit limits on what an agent can do autonomously versus what needs human oversight, continuous performance monitoring, and systematic measurement of cost and business impact.
What results has BBVA actually published?
In its October 2026 update BBVA reported that assistants filter more than half of conversations reaching relationship managers through the app in Spain, that voice agents resolve close to 90% of interactions in production pilots in Peru and Mexico, that complaint registration time fell by nearly 80%, that developer productivity improved by around 28%, and that 75% of employees use AI regularly, saving an average of 2.4 hours per week.
Which AI systems does BBVA use?
Publicly, BBVA's adoption layer is built on OpenAI's ChatGPT Enterprise plus thousands of employee-built custom GPTs, and the bank was named by OpenAI among early GPT-5 testers. Agents are now being built under BBVA's own framework, which the bank says preserves technological independence so it can use different models and platforms. BBVA has not disclosed which model powers each individual agent.
Can BBVA customers talk to an AI agent today?
Only in a limited way. BBVA demonstrated a conversational app inside ChatGPT for its Italian and German digital banks, designed for informational questions about accounts, cards and savings products rather than transactions. Its broader customer assistant, Blue, is planned for gradual rollout to individual and business customers in 2027. Most of the agent work described so far supports employees rather than replacing customer channels.
What should I be sceptical about in these numbers?
They are management disclosures, not audited results. BBVA has not published interaction volumes behind the voice-agent figure, its definition of resolution, the baseline for the complaint-time reduction, the methodology behind the developer productivity number, any cost or ROI figure, or any error and escalation rates. The per-employee time saving is self-reported and declined slightly year over year without explanation.
Twarx analysis
Original interpretationThe interesting part of BBVA's story is not the licence count — it is that the bank publicly reframed its second phase as an engineering and governance problem. Once assistants proliferate, the constraint stops being model quality and becomes the absence of a shared way to permission, bound, evaluate and cost every agent. BBVA built that layer before promising customer-facing autonomy.
Read the numbers honestly
Analysis by
Rushil Shah · AI Systems Builder & Founder, Twarx
