The BNY Playbook: Why 20,000 Employee-Built AI Agents Beat Your Centralized AI Team
BNY handed agent-building tools to 20,000 employees instead of building agents for them. The result: thousands of custom AI workflows deployed in months, not years. While most Fortune 500 companies are still running pilot programs with their 30-person AI Centers of Excellence, BNY's employees are shipping production
BNY handed agent-building tools to 20,000 employees instead of building agents for them. The result: thousands of custom AI workflows deployed in months, not years. While most Fortune 500 companies are still running pilot programs with their 30-person AI Centers of Excellence, BNY's employees are shipping production agents—without waiting for IT approval.
This inverts everything we thought we knew about enterprise AI deployment.
The thesis is simple: Enterprise AI's deployment bottleneck isn't technology—it's the centralized team model. BNY's Eliza platform proves that democratizing agent development to end users achieves faster adoption, better ROI, and more sustainable transformation than top-down AI initiatives. The winners in enterprise AI won't be those selling finished agents to IT departments, but those selling agent development platforms directly to employees.
The Centralized AI Team is the Bottleneck
Walk into any enterprise with a serious AI initiative and you'll find the same organizational structure: a centralized AI team—maybe 30 to 50 people—building solutions for business units across the organization. Marketing needs a content generation agent. Compliance needs document review automation. Sales needs lead scoring. Legal needs contract analysis. Finance needs fraud detection.
The queue is 100 items deep. The team can ship maybe 10 solutions per year if they're fast.
This is a fundamental queue theory problem disguised as a technology adoption problem. Even well-funded AI teams can't scale linearly with demand. Each use case requires requirements gathering, custom development, testing, security review, and deployment. The math doesn't work: thousands of potential use cases divided by dozens of developers equals years of backlog.
BNY recognized this constraint as a solvable architecture problem rather than accepting it as inevitable. Their hypothesis: employees closest to the work understand their workflows better than any central team ever will. A compliance officer knows which document patterns trigger regulatory flags. A client service manager knows which questions consume 80% of their team's time. A treasury analyst knows which data reconciliation tasks are high-volume and low-complexity.
What if you gave those employees the tools to build agents themselves?
How BNY Built Eliza
BNY's answer was Eliza, an agent development platform built on OpenAI's models. Not ChatGPT access—an actual development environment where employees can create custom agents for their specific workflows.
The scale of adoption is what matters: 20,000+ employees now building and deploying agents. This isn't a pilot program. It's not "select power users in three departments." It's enterprise-wide democratization of agent development.
The use cases span the operational spectrum. Compliance teams built agents that flag regulatory issues in client communications. Client service managers created agents that draft personalized responses based on account history and transaction patterns. Treasury analysts deployed agents that reconcile data across systems and surface discrepancies for human review.
The pattern repeats: employees identify high-frequency, well-defined tasks in their workflows, build agents to handle them, iterate based on results, and deploy to production—often in days rather than the months required for centralized IT development cycles.
The governance model likely balances enablement with control. Broad access to the platform, but guardrails around data permissions, model capabilities, and deployment approval workflows. The technical architecture suggests a managed environment where employees can configure agents without writing code, but can't accidentally expose sensitive data or bypass security protocols.
What makes this possible in 2026 is the maturity of the underlying AI capabilities. OpenAI's models handle multi-step reasoning, tool use, and context management well enough that non-technical users can build reliable agents by defining objectives and providing examples rather than writing traditional software.
The Numbers: What Success Looks Like
The headline metric is adoption: 20,000+ employees actively building agents. For context, most enterprise AI initiatives consider themselves successful with 200 users. BNY is operating at 100x that scale.
This adoption rate creates a compounding effect. As more employees build agents, organizational AI fluency increases. Teams share effective patterns. Use cases become templates. The platform improves based on aggregate usage data. The learning curve flattens for new users because they're entering an environment where agent development is normalized rather than experimental.
The time-to-value comparison is stark. Centralized AI teams measure project timelines in quarters. The democratized model measures in weeks or days. An employee identifies a workflow inefficiency on Monday, builds and tests an agent by Wednesday, and deploys to their team by Friday. No backlog. No requirements document. No sprint planning.
The productivity metrics matter less than the velocity metric. The actual efficiency gain from any single agent might be modest—maybe it saves 30 minutes per day. But when thousands of employees are deploying agents simultaneously, the aggregate impact is massive. And more importantly, the organization develops muscle memory for identifying AI-solvable problems and deploying solutions quickly.
This is what enterprise AI adoption looks like when the deployment model matches the scale of opportunity.
Why This Model Works
The economics are simple: the marginal cost of one more use case approaches zero in the democratized model. In the centralized model, each new use case requires developer time, which is the scarce resource. In the platform model, each new use case requires employee time, which scales with the size of the organization.
The information asymmetry advantage is even more significant. Central AI teams can't possibly understand the nuanced context of every workflow across a 50,000-person organization. They rely on requirements documents and stakeholder interviews, which compress complex operational knowledge into simplified specifications. The resulting agents work, but they're generic.
Employees building agents for their own workflows don't have this problem. They understand the edge cases. They know which data sources are reliable and which require validation. They recognize patterns that wouldn't make it into a requirements document because they're tacit knowledge rather than explicit process.
This mirrors the platform transformations we've seen in other categories. Airtable didn't win by building the perfect database for every use case—it won by giving non-technical users the tools to build databases themselves. Figma didn't win by designing every company's interface—it won by making design tools accessible to product managers and engineers, not just designers.
The agent development platform category will follow the same pattern. The winners won't be those selling pre-built vertical agents to IT departments. They'll be those selling agent development platforms directly to end users.
What Dies
If democratized agent development is the future, several current enterprise AI strategies face obsolescence.
Pre-built vertical agents sold to IT departments hit an adoption ceiling. The value proposition is speed and expertise—"we've built the perfect compliance agent, just deploy it." But standardization is the enemy of adoption. Compliance workflows at JPMorgan differ from compliance workflows at BNY, which differ from compliance workflows at Wells Fargo. The pre-built agent handles 70% of the use case, and the remaining 30% requires customization that the vendor can't economically provide. The democratized platform handles 90% because the end user customizes it themselves.
The consulting model for AI transformation gets compressed from 18-month engagements to platform enablement sprints. Traditional AI transformation programs involve assessment, strategy, pilot development, change management, and scaled deployment—each phase measured in quarters. When employees can build and deploy agents themselves, the consulting engagement shifts from "we'll build your AI capabilities" to "we'll set up your platform and train your teams." That's a three-month project, not an 18-month program.
Centralized AI Centers of Excellence don't disappear, but their mandate shifts fundamentally. Instead of building agents, they govern platforms. They establish security frameworks. They maintain the infrastructure. They provide advanced support for complex use cases. They analyze usage patterns to optimize the platform. But they stop being the bottleneck for every deployment.
For AI startups, the implication is clear: product-led growth beats enterprise sales in this category. The founders building Eliza competitors should optimize for end-user adoption, not IT procurement processes. The champions are employees who want to solve their own problems, not CIOs evaluating vendor proposals.
The Playbook: How to Replicate This
Democratized agent development isn't universally applicable, but the decision framework is straightforward.
Prerequisites matter. You need baseline AI fluency across the organization—not expert-level, but comfortable enough that employees understand what agents can and can't do. You need governance frameworks that can operate at scale without creating approval bottlenecks. You need technical infrastructure that supports rapid iteration without compromising security.
BNY likely spent 12-18 months building those prerequisites before deploying Eliza at scale. That groundwork isn't optional.
The make-versus-buy decision depends on your constraints. Building your own platform gives you control and customization but requires significant engineering investment. Leveraging existing platforms—whether purpose-built agent development tools or extensible AI infrastructure—accelerates deployment but constrains customization. For most enterprises, the answer is probably "buy the platform, customize the governance layer."
Success metrics must extend beyond adoption. Twenty thousand employees building agents is impressive, but the business cares about outcomes. Track agent quality: how often do agents produce results that humans trust without modification? Track business impact: which workflows show measurable efficiency gains? Track organizational learning: is the average agent quality improving over time as employees develop better intuition for effective agent design?
Commonwealth Bank, another early adopter of enterprise-wide AI deployment, reports similar patterns: adoption accelerates when employees control the tools, and quality improves faster than centrally-developed solutions because feedback loops are tighter.
The playbook isn't complicated. Build the foundation. Deploy the platform. Govern outcomes, not processes. Measure impact, not activity.
What This Means
By 2027, I expect the enterprises that adopt democratized agent development platforms will have 5-10x more AI use cases in production than those maintaining centralized AI teams. The gap won't be incremental—it'll be structural.
The biggest enterprise AI opportunity over the next 24 months isn't building vertical agents for specific functions. It's building agent development platforms that knowledge workers can operate independently. That's a different product, a different go-to-market strategy, and a different competitive dynamic than the current enterprise AI landscape.
BNY's 20,000+ employee agent builders represent the largest documented case of democratized AI deployment. The model is no longer theoretical. The benchmarks are established. The question for every enterprise is whether they'll adopt this architecture before their competitors do.
The centralized AI team model made sense when AI capabilities required specialized expertise to deploy. That constraint no longer binds. The organizations that recognize this shift earliest will build sustainable advantages in operational efficiency, innovation velocity, and employee capability development.
The ones that don't will spend 2027 explaining why their 50-person AI team hasn't delivered the transformation their board expected.
Key Takeaway: Enterprise AI's bottleneck is organizational structure, not technology—BNY's deployment of 20,000+ employee agent builders proves that democratized platforms beat centralized teams for adoption velocity, customization depth, and sustainable transformation at scale.
📚 Related Articles
Enterprise AI Deployment Model Scorecard
Assess whether your organization is built for centralized AI teams or democratized agent development.
0/6Evaluate Your AI Architecture
Recommended Actions
Share ResultRetake AssessmentCopied to clipboard!
Enterprise AI Deployment Model Scorecard
Assess whether your organization is built for centralized AI teams or democratized agent development.