BBVA turned 20,000 Custom GPTs into its AI front end

The front end is a GPT, not an app

The most concrete detail in OpenAI's write-up isn't the productivity figure. It's that BBVA employees created more than 20,000 Custom GPTs, with roughly 4,000 used frequently. Each of those is a small purpose-built interface — a wrapper with instructions, context, and a defined task — sitting on top of ChatGPT Enterprise.

That reframes what 'the product' is inside BBVA. The bank didn't commission a catalog of bespoke internal apps with their own screens and login flows. It gave people one surface and let them shape thousands of narrow front ends on top of it. The interface layer became a thing employees author, not something a central team ships and maintains.

For teams building applied AI, that's a meaningful design decision. A Custom GPT is the lightest possible front end: no deployment pipeline, no UI framework, no release cycle. The cost of creating a new interface drops close to zero, which is why the count runs into the tens of thousands.

What the Peru assistant tells you about the interaction surface

The single clearest workflow result comes from Peru, where an internally built assistant serves more than 3,000 employees and cut query handling time from about 7.5 minutes to roughly 1 minute — an ~80% reduction, per the source.

Read as a front-end story, that's a lookup-and-respond task that used to require navigating systems or documents and now happens in a single conversational exchange. The gain isn't a smarter model in the abstract; it's collapsing several manual interface steps into one query box that already holds the right context.

That's also why the ~3 hours saved per employee per week figure is plausible rather than aspirational: it's the aggregate of many small interaction shortcuts like Peru's, spread across the 83% who use it weekly.

Stickiness comes from the interface being trusted, not just useful

BBVA's framing repeatedly ties adoption to safety. Elena Alfaro describes giving employees 'a platform that was safe so they could start experimenting,' with security, legal, and compliance involved from day one. The stated goal was converting shadow AI into sanctioned AI by making the official surface the path of least resistance.

Once you start using it, it's very sticky and… you feel it helps you a lot. Montana Labs

The frontend lesson underneath the stickiness quote: an interface only becomes the default if using it is easier and safer than the alternatives employees would reach for anyway. BBVA competed with shadow tools by making its own surface the more convenient one, then paired it with hands-on training for 250 senior leaders, including the CEO and chairman.

The implication: front-end authorship is now a distribution strategy

BBVA's numbers describe a specific pattern worth naming. When you push interface creation to the front lines — the people closest to each task — you get thousands of narrow tools quickly, and the useful minority (the ~4,000 frequently used GPTs) surfaces through actual usage rather than a roadmap committee.

The tradeoff is fragmentation: 20,000 interfaces means 16,000 that mostly don't get used, and no guarantee of consistency across them. BBVA's answer, per the source, was governance as the foundation and sharing frequently used GPTs across the bank — curation after creation, not before.

For applied teams, the takeaway from this specific rollout is that letting users build their own front ends can outpace centrally designed apps for adoption, provided the shared platform underneath handles the security and the winners get promoted. BBVA's next phase, the roadmap it calls 'The Eight,' will test whether that bottom-up interface sprawl can be consolidated into customer-facing products without losing the momentum that got it there.