BNY People uses OpenAI: How 20,000 Employees Built Agents to Scale AI Literacy

🔍 Executive snapshot

I’m reporting on a practical, modern experiment in enterprise learning that has already produced measurable results. At BNY, a learning-first approach to AI has led to widespread adoption: roughly 20,000 employees have created their own agents to build and maintain learning content. The result is faster content creation, a culture of experimentation, and a shift in where value is created inside the organization.

This article explains how BNY approached AI literacy as part of everyday work, the concrete tools they used, the cultural changes that mattered, and a playbook any large organization can use to achieve similar results while keeping governance, quality, and human connection front and center.

🤖 What happened: a concise news-style report

BNY’s people team took a deliberate approach to integrating AI into the employee learning lifecycle. Rather than centralizing AI projects within a small technology team, the organization empowered employees across functions to create and iterate on AI agents. I learned that approximately 20,000 employees have built agents that help generate, rebrand, and correct learning collateral.

One learning manager described a dramatic change in turnaround times: what used to take a month to develop can now be completed in about an hour. That time savings freed the learning team to redirect their work toward deeper employee engagement instead of routine content production.

BNY’s approach included a combination of structured programs—hackathons and promptathons—paired with a philosophy of learning-by-doing. This was not a single playbook; rather the team recognized that learning for AI must be multifaceted and embedded into daily workflows.

🚀 Why this matters: AI literacy at enterprise scale

AI literacy should be a strategic capability, not a special program tucked into L&D. Here’s why the BNY approach is noteworthy:

• Scale without bottlenecks. When 20,000 people can create their own agents, the organization multiplies its capacity to produce and maintain knowledge assets.
• Faster content cycles. Reducing development time from a month to an hour turns learning content from a slow-moving project into a living asset that adapts quickly.
• Democratized innovation. Good ideas no longer need to travel up a rigid chain. People at every level can test and roll out improvements.
• Human-centered outcomes. Time saved on routine production becomes time spent connecting with employees, coaching, and tackling higher-value learning design.

🛠️ The approach: learning by doing, not theory alone

The central thesis of BNY’s method is simple: sustained AI literacy requires practical experience. Training and theory matter, but they must be paired with real tasks that employees actually perform.

I heard this summarize perfectly when a learning lead told me, “We believe in learning by doing, not just theoretical knowledge.” That short sentence guides every practical decision the team made—whether to host a promptathon, build templates for agents, or allow teams to iterate on their own conversational assistants.

How hands-on learning was delivered

BNY used a mix of activities that encourage rapid experimentation and collective learning:

• Hackathons. Short, focused events where cross-functional teams build agent prototypes to solve specific problems.
• Promptathons. Iterative exercises that help participants craft, test, and refine prompts to get better outputs from language models.
• Internal agent marketplaces. Repositories and templates where successful agents are shared, forked, and improved across teams.
• Mentorship and office hours. Quick access to experts who can help troubleshoot model outputs, fine-tune prompts, and design safeguards.

📈 Concrete outcomes and real-world examples

The most compelling evidence of success is in the numbers and the everyday examples.

The most immediate outcome I encountered: the team built an agent that automatically rebrands and corrects learning collateral. This agent handles routine editing, branding consistency, and basic content corrections, producing polished materials in minutes instead of days.

To illustrate the impact:

• Time savings. A task that historically required a month to develop now takes roughly an hour thanks to agent-assisted drafting and rebranding.
• Improved productivity. Learning teams now spend less time on repetitive tasks and more time on strategy, facilitation, and mentoring.
• Scaled reach. Thousands of employees contribute to and benefit from shared learning assets, increasing overall AI fluency across the business.

🤝 Culture and governance: how they balanced openness and control

Scaling AI across an enterprise requires cultural change and practical governance. BNY’s playbook recognizes that there is no one-size-fits-all solution. Instead they layered lightweight governance on top of broad experimentation.

Key governance and culture touchpoints I observed include:

• Guidelines, not gatekeeping. Broad principles for responsible use of AI were shared, but teams were encouraged to experiment within those guardrails.
• Shared responsibility. Teams building agents were encouraged to own their outputs, including accuracy, bias remediation, and privacy considerations.
• Rapid feedback loops. Agents were iterated quickly based on user feedback, reducing the risk of long-lived errors.
• Cross-functional oversight. Product, legal, and learning teams collaborated to ensure that agents met both business and compliance needs.

That combination—empowerment plus accountability—created a virtuous cycle. When people know the rules and still have permission to experiment, they innovate faster and more responsibly.

🧭 Practical playbook: how to scale AI literacy in your enterprise

Based on what I observed at BNY and what I’ve seen work across other large organizations, here is a practical playbook you can use to scale AI literacy and agent building across your company.

1. Start with a clear north star

Define what AI literacy means for your organization and tie it to measurable outcomes. BNY prioritized learning that is built into work, not separate from it, which made adoption feel natural and valuable.

2. Democratize tools and templates

Provide ready-made agent templates for common tasks like content rebranding, FAQ creation, or meeting summarization. Templates shorten the path from idea to working prototype and reduce entry friction.

3. Run focused, short experiments

Use hackathons and promptathons to channel creative energy. Set a clear problem, a short timeline, and a runway for post-event iteration. The event itself is only the start; the real value comes from supported follow-through.

4. Pair autonomy with simple guardrails

Establish a small set of non-negotiable principles—privacy, accuracy checks, and human-in-the-loop for sensitive outputs. Keep governance lightweight so it protects without stifling innovation.

5. Measure what matters

Track both quantitative and qualitative signals:

• Quantitative. Number of agents created, reduction in content production time, user engagement with learning assets.
• Qualitative. Participant confidence with AI, stories of improved employee interactions, quality assessments of agent outputs.

6. Make knowledge sharing routine

Create channels for teams to share successful prompts, agent architectures, and lessons learned. An internal repository where people can fork and adapt agents accelerates learning across the org.

7. Invest in facilitator skills

People who can teach prompt design, model limitations, and evaluation techniques become force multipliers. Investing in these facilitator skills ensures experiments produce reliable, repeatable outcomes.

8. Keep humans central

Agents are amplifiers, not replacements. Keep human judgment, empathy, and contextual understanding in the loop—especially for learning programs that impact careers, performance, and culture.

📚 Example: rebranding and correcting learning collateral

I want to highlight the rebranding agent because it captures the practical power of this work. The learning team at BNY built an agent that performs three core functions:

• Standardize branding. Apply corporate templates, color schemes, and tone guidelines automatically to new materials.
• Correct and refine content. Suggest grammar fixes, clarity edits, and standard phrasing to ensure consistent quality.
• Localize and adapt. Adjust language and examples for different audiences, making materials relevant for global teams.

The agent reduced development time dramatically. Instead of spending weeks coordinating designers, editors, and subject matter experts, a learning manager can produce a polished draft in an hour and then use the extra time to validate nuance, run pilot sessions, and gather feedback.

🔬 What to watch out for: risks and mitigation

Scaling agent creation across thousands of people carries risks if left unaddressed. Here are the most important pitfalls and practical mitigations I recommend.

Risk: quality drift

When many agents are created independently, outputs can diverge and lower the perceived quality of learning content.

Mitigation: implement periodic audits and a lightweight approval workflow for critical assets. Encourage peer reviews and shared quality rubrics.

Risk: data exposure and privacy

Agents trained or prompted with sensitive information can inadvertently expose private data.

Mitigation: clearly define what data can be used in prompts, enforce redaction practices, and route sensitive tasks through secure, governed pipelines.

Risk: overconfidence and misuse

Employees may over-rely on agents for judgment calls, especially in complex or regulated decisions.

Mitigation: require human sign-off on high-stakes outputs and teach employees how to evaluate model confidence and failure modes.

Risk: lack of reuse

Teams might rebuild the same agent patterns independently, wasting effort.

Mitigation: maintain a searchable internal catalog of agents and templates, promote reuse through incentives and recognition.

📊 Measuring success: what to track

Effective measurement shouldn’t rely solely on vanity metrics. Focus on outcomes that reflect learning, adoption, and business impact.

Suggested metrics:

• Adoption metrics: number of employees building or using agents, number of active agents, and template reuse rates.
• Efficiency metrics: average content development time before vs after agent adoption, reduction in manual review cycles.
• Quality metrics: user satisfaction scores with learning content, accuracy audits, and error rates in agent outputs.
• Behavioral metrics: increase in AI-related competency scores, cross-team collaboration rates, and number of improvements originating from non-technical staff.
• Business impact: correlation of learning improvements to business KPIs such as onboarding time, internal mobility, and customer satisfaction where applicable.

🧩 The human element: people first, tools second

One theme kept surfacing for me: technology is an amplifier for culture, not a substitute for it. The most successful programs I’ve seen put people at the center.

At BNY, the team repeatedly emphasized the spirit of learning together. They didn’t wait for perfect governance, and they didn’t restrict experimentation to a small center of excellence. Instead they created opportunities for employees to learn by doing, then invested in coordination and safeguards.

A learning leader put it plainly: “Good ideas come from every level in the organization, and every level in the organization is empowered to bring those ideas to fruition.” That empowerment is the critical ingredient that turns pilot projects into scalable capabilities.

🔁 How to sustain momentum

Initial enthusiasm is easy. Sustaining momentum is harder. Here are practical steps to keep experiments alive and productive.

• Make success visible. Publicize wins, but also surface honest lessons learned so others can iterate faster.
• Formalize lightweight structures. Create small cross-functional squads that act as incubators for promising agents and ensure continuity.
• Invest in upkeep. Agents require updates. Allocate a fraction of operating budgets for maintenance, prompts refresh, and content audits.
• Rotate responsibilities. Let multiple teams take stewardship of shared agents to avoid single points of failure and to promote diversity of input.
• Embed AI into role expectations. Update job descriptions and performance goals to include responsible AI practice and continuous learning.

🔎 A closer look at events that accelerate learning: hackathons and promptathons

Two practical formats proved especially useful: hackathons and promptathons. Each serves a distinct purpose and together they create a strong feedback loop between creativity and craft.

Hackathons: build and ship prototypes

Hackathons accelerate cross-functional collaboration. Set a clear problem, provide basic data and templates, and let teams compete to prototype agents. The focus should be on a working demo and a plan for follow-up, not on perfection.

Promptathons: iterate on prompts and guardrails

Promptathons are shorter, granular workshops that focus on the craft of prompting: writing, testing, evaluating, and improving prompts to elicit reliable outputs. They are low-cost, high-value exercises for teams who rely on language models to generate content.

Both formats emphasize speed, reuse, and documented learning—turning ephemeral experiments into institutional knowledge.

🧾 Real stories: how teams repurposed agents

I gathered several use cases that show the range of agents people built:

• Learning rebrand agent. Standardizes content to company brand and applies editorial style guidelines.
• Onboarding assistant. Curates personalized onboarding playbooks based on role, region, and prior experience.
• FAQ synthesizer. Aggregates information from policies and internal docs to answer common employee questions.
• Meeting summarizer. Condenses meeting notes into action items, owners, and follow-ups with proper context.
• Microlearning creator. Generates short, role-specific learning modules for just-in-time training.

Each agent improved day-to-day work by removing low-value tasks while preserving human oversight for context and judgment. The learning team’s rebrand agent is the clearest single example of the multiplier effect: it reduced time, improved consistency, and allowed staff to spend more time on high-impact activities.

🧠 Lessons I’d pass along

If I had to condense the most important takeaways into a short list for any organization, they would be:

1. Embed learning into work. Programs that sit outside daily workflows rarely scale. Make AI literacy part of how people accomplish their jobs.
2. Empower broadly with guardrails. Permission to experiment unlocks far more value than tight central control, but only when paired with clear, simple rules.
3. Measure outcomes, not activity. Track impact on time, quality, and behavior—then iterate.
4. Share and reuse. Build a culture and infrastructure that makes reuse the default, not the exception.
5. Invest in people. Facilitation and mentoring matter more than the latest model. People who can teach others to use AI responsibly are the best long-term investment.

📣 What leaders should decide now

Leaders don’t need to commit to perfect frameworks before starting. They do need to decide on a few critical items:

• Scope of experimentation. Which functions or domains are early priorities for agent automation?
• Governance baseline. What data is restricted, what approvals are required, and which teams are guardians of compliance?
• Support model. Will there be central funding for templates and mentorship, or will teams self-fund and share learnings?
• Success criteria. How will leaders evaluate progress after three months, six months, and a year?

📬 Closing summary and next steps

BNY’s experiment demonstrates a simple truth: when learning is built into the work, AI literacy scales quickly and responsibly. Empowering a broad set of employees to create agents unlocks creativity and operational efficiency, but it must be balanced with practical governance and a commitment to human oversight.

For organizations starting this journey, the prescription is clear: provide tools and templates, host structured experiments like hackathons and promptathons, measure what matters, and cultivate a culture where good ideas can come from any level. With those elements in place, the technology becomes an amplifier of human potential rather than an isolated technical project.

"Good ideas come from every level in the organization, and every level in the organization is empowered to bring those ideas to fruition."

I expect we will see more organizations follow this path. The difference between piloting and transforming is not an extra technical layer; it is a choice to embed learning into the flow of work and to trust employees to iterate responsibly.

📌 Quick checklist to get started

• Define the north star: what does success look like in 6–12 months?
• Create reusable templates: for content generation, rebranding, and FAQs.
• Run a promptathon: teach teams to write and evaluate prompts.
• Host a hackathon: produce pilot agents and designate owners for follow-up.
• Set lightweight guardrails: privacy, data use, and human-in-the-loop rules.
• Measure outcomes: track time savings, adoption, and quality improvements.
• Share learnings: publish a simple catalog of successful agents and prompts.

I’ll be watching how organizations adopt these patterns. For now, it’s clear that a people-first, hands-on approach transforms AI from a technical curiosity into a practical, organization-wide capability.