Scaling AI Across Teams: Lessons from BBVA’s Rollout

83% weekly active usage isn’t a “pilot result.” It’s what adoption looks like when AI becomes a default tool instead of a side project.

That number comes from BBVA’s enterprise rollout of ChatGPT, where the bank went from an initial 3,000 employees to 11,000 quickly—and didn’t just hand people a chatbot and hope for the best. They built governance, trained leaders (including the CEO), and created a safe environment for experimentation. The payoff: roughly 3 hours saved per employee per week, 20,000+ Custom GPTs created internally, and workflow tests showing up to 80%+ efficiency gains.

This post is part of our “How AI Is Powering Technology and Digital Services in the United States” series, and here’s why BBVA matters to U.S. SaaS and enterprise teams: the bank solved the same scaling problems American digital service providers are facing right now—security, compliance, employee adoption, and turning productivity wins into real workflow automation.

BBVA’s real breakthrough: AI as a way of working

BBVA’s biggest move wasn’t picking a model or deploying a tool. It was treating AI adoption as organizational change, not “innovation theater.”

Antonio Bravo, BBVA’s Global Head of Data & AI, framed it plainly: the goal was making AI part of the business strategy, not a tech effort off to the side. That’s the difference between a pilot that produces a nice slide deck and a program that changes daily operations.

For U.S. tech and digital services companies, this matters because most AI deployments stall at the same point: a few power users get value, everyone else watches, and leadership calls it “early.” Meanwhile, employees use unsanctioned tools anyway—creating avoidable privacy and compliance risk.

BBVA’s approach is a direct rebuttal to that pattern: bring experimentation into a trusted environment, then scale fast enough to see what actually works.

The results (and why they’re credible)

BBVA reported four metrics that are worth paying attention to because they combine adoption, productivity, and reuse:

• ~3 hours saved per employee per week (time is the universal KPI)
• 83% weekly active usage (adoption is the hardest part)
• Up to 80%+ efficiency improvements in workflow tests (process impact, not just “better writing”)
• 20,000+ Custom GPTs created, with about 4,000 used frequently (reuse and internal productization)

If you’re running a U.S. SaaS platform or a digital operations team, these are the same metrics you should track: time saved, active usage, workflow cycle time, and reusable AI assets.

How BBVA scaled adoption without creating chaos

BBVA scaled quickly because it put structure first. That sounds backwards to teams that equate governance with slowdown. In practice, governance is what makes speed safe.

Elena Alfaro, Head of Global AI Adoption, described building “a safe place to learn and to use AI.” That’s not soft language—it’s an operating principle. When employees feel the official path is safe, they stop going off-road.

1) Governance from day one (security, legal, compliance aligned)

BBVA aligned security, legal, and compliance early. That choice prevents the most common enterprise failure mode:

• A team proves value in a pilot
• Legal/security review begins late
• Rollout gets delayed, watered down, or blocked
• Employees keep using unapproved tools anyway

U.S. companies in regulated or high-trust categories (fintech, health, insurance, HR tech, education, even B2B data platforms) should copy this sequencing. Don’t “win the pilot” and then ask for permission to scale.

Practical step you can take next quarter: create a lightweight AI usage standard that covers:

• Allowed tools and environments
• Data classification rules (what can/can’t be pasted)
• Human review expectations
• Logging, retention, and access controls
• Approval flow for building internal assistants

If it’s too complicated, people won’t follow it. If it’s too vague, it won’t protect you.

2) “Turn shadow AI into safe AI”

BBVA didn’t pretend employees weren’t experimenting. They gave them a secure platform and guardrails so experimentation happened where it could be monitored and improved.

That’s a strong lesson for U.S. digital service providers: shadow AI isn’t a discipline problem—it’s a product problem. If your approved path is slower or less capable than the unapproved path, people will route around you.

Here’s what works in practice:

• Provide a sanctioned AI workspace employees can access in seconds
• Make “approved” the easiest option
• Publish prompt and data handling examples for common roles
• Create a visible channel for sharing internal use cases

Once your internal AI program becomes the easiest way to get work done, governance gets easier—not harder.

3) Train leaders like operators, not spectators

BBVA trained 250 senior leaders, including the CEO and chairman. That’s a clear signal: AI isn’t “for the innovation team.”

In U.S. companies, leadership training is often skipped or treated like a keynote. I’ve found the opposite works: leaders need hands-on reps so they can ask better questions and set better expectations.

A leader who has actually used an internal assistant can:

• Push for process changes (not just “use ChatGPT more”)
• Approve realistic time-to-value plans
• Understand where human review is mandatory
• Fund internal enablement (training, templates, evaluation)

If leaders don’t use the tools, they’ll manage the program as theory.

Custom GPTs: why internal “AI apps” beat generic prompts

The most scalable detail in BBVA’s story is this: employees created 20,000+ Custom GPTs, and around 4,000 are used frequently.

That pattern is familiar to anyone who’s seen internal tooling succeed. Most assets won’t matter; a smaller set becomes the standard.

For U.S.-based SaaS companies, “Custom GPTs” translates to a broader concept: packaged, repeatable AI workflows. Not everyone should be writing prompts from scratch. The highest ROI usually comes from turning common tasks into reusable assistants.

What to standardize first (high ROI, low regret)

Start with tasks that are:

• Frequent (daily/weekly)
• Text-heavy
• Rule-based with clear inputs/outputs
• Easy to QA with spot checks

Examples that map well to U.S. digital services:

• Customer support response drafting with policy grounding
• Sales call summaries with next-step extraction
• RFP and security questionnaire first drafts
• Marketing content variants aligned to brand rules
• Engineering incident postmortem templates
• Internal knowledge base Q&A for frontline teams

The goal isn’t to “automate thinking.” It’s to remove blank-page work and speed up repeatable steps.

A simple operating model for internal AI assistants

If you want your Custom GPT library (or internal agent catalog) to stay useful instead of becoming a junk drawer, adopt three rules:

1. Naming and purpose are mandatory: “Support Reply Helper – Refund Policy” beats “Refund GPT.”
2. One owner per assistant: someone is accountable for updates when policies change.
3. Lightweight evaluation: a small test set of real examples to sanity-check behavior after edits.

This is how AI becomes part of operations rather than a novelty.

Workflow automation: where the biggest gains actually come from

BBVA is moving past individual productivity into workflow automation, operational systems, and customer-facing channels—including a digital assistant called Blue built with OpenAI models.

That’s the correct direction. Productivity gains are real, but they’re often hard to “bank” unless they change the workflow itself.

A concrete example: Peru’s internal assistant

BBVA highlighted an internal assistant used by 3,000+ employees in Peru that reduced query handling time from ~7.5 minutes to ~1 minute—about an 80% reduction.

Notice what’s special about that example:

• It’s specific (a defined workflow: query handling)
• It has clear time metrics
• It’s adopted at meaningful scale

U.S. digital services companies should look for this exact shape: pick a workflow with measurable cycle time, then build an assistant that changes the steps, not just the writing.

The “two-layer” AI stack most companies need

To turn AI into durable operational value, build two layers:

• Layer 1: Copilot layer (personal productivity)
  • Drafting, summarizing, brainstorming, explaining code, writing internal docs
• Layer 2: Workflow layer (system productivity)
  • Intake → classify → route → draft → verify → log → follow-up

Layer 1 creates fast adoption. Layer 2 creates margin.

A practical roadmap for U.S. SaaS teams:

1. Roll out a secure AI workspace and measure weekly active usage.
2. Identify 5–10 workflows where cycle time is painful (support, onboarding, renewals, compliance).
3. Package repeatable assistants tied to those workflows.
4. Add human review gates where risk is real (legal, regulated comms, financial outputs).
5. Instrument outcomes: minutes saved, resolution time, backlog reduction, CSAT impact.

If you can’t measure it, it won’t survive budget season.

People also ask: “How do we scale AI adoption without risking data?”

Answer: You scale AI adoption safely by combining a secure environment, clear data rules, and role-based training—then measuring usage and outcomes continuously.

In practice, that means:

• Use enterprise-grade access controls and identity management
• Define what data can be used, by role and sensitivity
• Provide templates for common tasks so employees don’t improvise risky behavior
• Create a review path for high-impact assistants (customer-facing, compliance-related)

BBVA’s example shows a basic truth: safety is easier when the official path is the path people actually want to use.

What U.S. tech leaders should copy from BBVA in 2026 planning

BBVA’s story lands at the perfect time for U.S. teams planning next year’s budgets. By late December, most roadmaps are being finalized—and AI can easily become a line item without a plan.

Here’s what I’d copy directly:

• Start with trust, not hype: governance and enablement are what make scale possible.
• Scale early enough to see the signal: small pilots hide adoption issues and overfit to power users.
• Empower front lines: the people doing the work create the most useful assistants.
• Measure outcomes weekly: active usage plus cycle-time KPIs beat “number of experiments.”

If you’re building digital services in the United States—SaaS, fintech, marketplaces, IT services—this is the playbook for turning generative AI into operational capability.

What’s your next workflow that should drop from 10 minutes to 2?