How Do Enterprises Scale AI Workforce Upskilling? Bank of America's 3-Level Playbook for 213,000 Employees

TLDR: AI workforce upskilling at enterprise scale is not a training problem, it is an operating model problem. Bank of America's experience deploying AI tools and learning programs across 213,000 employees offers the clearest available evidence of what actually works: a layered capability model, dedicated institutional infrastructure, and a deliberate decision about where humans remain in the loop. This post breaks down their approach and what it means for operations leaders planning their own programs.

Best For: COOs, Chief People Officers, and transformation directors at mid-to-large enterprises currently stalling on AI adoption because employees are not using the tools, or using them inconsistently, and leadership cannot figure out whether the problem is the technology, the training, or the culture.

AI workforce upskilling is a structured capability-building program that prepares employees at every level to work alongside AI tools in their day-to-day roles. It differs from traditional technology training because the goal is not to teach people how software works. The goal is to change how work itself gets done, which requires a combination of tool exposure, skill development, and explicit guidance on where human judgment still matters. For enterprises with thousands of employees across dozens of functions, getting this right is the difference between an AI program that lives in the pilot phase indefinitely and one that actually changes operating margins.

Why Most Enterprise AI Upskilling Programs Stall Before They Scale

Most enterprise AI upskilling programs stall because they are designed as events rather than systems. A two-hour training session, a lunch-and-learn, or a vendor-led demo gets employees familiar with the tool but does not change what they do on Monday morning. Adoption requires habit, and habits require reinforcement at the workflow level.

The data on this is consistent across industries. McKinsey's research on AI adoption finds that companies with strong AI adoption practices invest not just in training but in workflow redesign and manager enablement alongside it. Without all three components working together, training completion rates look good on a dashboard while actual tool usage stays low.

The failure mode is predictable: organizations measure activity (courses completed, licenses activated) rather than behavior change (prompts generated, errors caught, decisions supported by AI-generated analysis). These are different things. Courses completed is a procurement outcome. Behavior change is a transformation outcome. The companies that confuse them never get past the first one.

The Training-Without-Workflow Problem

Even well-intentioned programs fail when they train people on AI tools in isolation from the workflows where those tools apply. An employee who learns how to use a generative AI assistant in a classroom setting still needs to understand, specifically, which tasks in their actual job should now involve that assistant and which should not. That specificity requires function-level design work, not just enterprise-level rollout.

BCG's research on generative AI at work found that employees given AI tools without task-level guidance used them for the wrong things at approximately the same rate as the right things, with roughly equal probability of producing better or worse outcomes. The tools themselves are not enough. The application map matters just as much as the capability.

Manager Resistance Kills More Programs Than Employee Resistance

The second underrated failure point is middle management. Individual employees are generally not resistant to AI tools, at least not initially. They are cautious and want to understand the implications for their role. Managers, by contrast, often have more at stake because AI programs frequently reorganize workflows in ways that shift accountability and authority.

Programs that brief the C-suite and train frontline employees but skip manager enablement routinely fail in the middle. Managers who do not understand the new workflow expectations either ignore the program or actively discourage tool use in their teams. Neither outcome shows up in training completion dashboards until it is too late.

How Bank of America Built a 3-Level AI Upskilling System

In the Season 13 finale of MIT Sloan Management Review and BCG's podcast "Me, Myself, and AI", Bernard Hampton, the leader of Bank of America's Academy, described the upskilling architecture the bank built to prepare its 213,000-person global workforce for AI. The episode is worth listening to in full. What follows is an analysis of the approach through the lens of what enterprises in traditional industries can actually replicate.

The Academy is Bank of America's internal learning and development organization, staffed by more than 1,000 people. Its mandate covers onboarding, professional development, compliance education, and increasingly, AI capability building. It is not an ad-hoc function standing up training modules in response to a new vendor contract. It is a permanent institutional infrastructure for workforce capability that has AI upskilling built into it as a core ongoing function.

Level 1: AI Literacy for Everyone

The first level covers the entire workforce and asks a single question: does every employee understand what AI is, what it can do in their specific context, and what it cannot do? This is not a technical education. It is a conceptual grounding that removes the fear and misunderstanding that cause passive resistance.

According to Bank of America's April 2025 press release, over 90% of the bank's 213,000 employees now actively use Erica for Employees, the bank's internal AI-driven virtual assistant. That number did not happen by accident. It reflects years of deliberate literacy-first investment that made AI feel like a normal part of the work environment rather than an experimental technology that some employees had to opt into.

The bank launched Erica for Employees in 2020, initially for technology support functions. By 2023 the platform had expanded to cover health benefits queries, payroll questions, and more. The gradual scope expansion is a deliberate literacy strategy: it gave employees repeated, low-stakes contact with AI assistance before more complex AI tools arrived in their workflows.

Level 2: Prompt Engineering and Functional Application

Once baseline literacy exists, the second level focuses on specific skill development: how to work with AI tools effectively in particular job functions. At Bank of America, this is where the training becomes role-specific. Software developers receive coding assistance tools and training on how to use them effectively; they have achieved efficiency gains of over 20% on coding tasks. Relationship managers in Merrill and Private Bank have access to the ask MERRILL and ask PRIVATE BANK platforms, which generated more than 23 million interactions in 2024.

The Academy uses AI-powered conversation simulators to provide interactive coaching for client-facing roles. Employees practice client interactions in a simulated environment and receive real-time feedback before those skills get deployed in live client situations. In 2024, employees completed over 1 million of these simulations, and the feedback from participants consistently highlights the value of practicing before performing.

This is a materially different approach from most enterprise training programs, which ask employees to learn new capabilities while simultaneously using them in live client-facing or business-critical situations. The simulation layer reduces the cost of mistakes and accelerates the time it takes for new capabilities to feel natural.

Level 3: AI Design and Development Capability

The third level is the narrowest in terms of the employee population it covers, but arguably the most strategically important. This is where a subset of employees develop the ability not just to use AI tools but to shape how they work, identify new applications, and contribute to the organization's AI product roadmap.

Bank of America holds more than 7,400 granted patents and pending patent applications, more US granted patents than any other financial services company. Over 1,200 of those patents are AI and machine learning focused, representing 17% of the total portfolio. A meaningful share of that innovation comes from employees at the design and development capability level who are close enough to front-line operations to identify what AI can solve and technically equipped enough to contribute to building the solution.

The talent pathway that supports this level is not separate from the broader upskilling program. Bank of America has filled 44% of jobs in recent years through internal mobility, which is partly a consequence of the upskilling investment. Employees who develop AI skills at Level 2 become candidates for Level 3 roles without the bank having to go to a tight external talent market for every AI-capable position.

What the Numbers Actually Tell Operations Leaders

The output of Bank of America's program is not just a training completion percentage. The outcomes are measurable at the operational level in ways that most enterprises have not yet achieved.

IT service desk call volume dropped by more than 50% as a result of Erica for Employees adoption. That is not a training metric. That is an operational efficiency outcome that shows up in headcount requirements, call handling times, and employee satisfaction data. Approximately 150,000 active users generate more than 1.5 million AI prompts per week, which means the tools are embedded in daily work rhythms rather than being used occasionally when someone remembers they exist.

The bank spends $13 billion annually on technology, with approximately $4 billion directed to new technology initiatives. The size of that investment reflects both the bank's scale and its strategic commitment. But the ratio matters more than the absolute number: when learning and development is structurally integrated into the technology investment rather than funded separately as an HR program, the two things reinforce each other. Technology investment without capability investment delivers tools that do not get used. Capability investment without technology investment trains people on systems that never arrive. The integration is the strategy.

Before launching a scaled AI upskilling program, most enterprises benefit from an honest AI readiness assessment to understand where the real gaps sit across data, process, talent, and governance. Attempting to upskill a workforce before the supporting infrastructure exists produces trained employees with nowhere to apply what they have learned.

The Internal Mobility Signal

The 44% internal mobility statistic deserves more attention than it usually gets. Internal mobility at that rate is not primarily a recruitment efficiency story. It is an indicator of capability fluidity, which means the organization can redirect talent toward higher-value work as AI changes what certain roles require, rather than being forced to hire externally every time a new capability is needed.

For operations leaders in traditional industries, this is the most replicable and strategically valuable outcome of an AI upskilling investment. Manufacturers, logistics companies, and financial services firms with strong internal mobility can absorb the workflow changes that AI forces without the disruption of constant external hiring or, at the other end, unmanaged headcount reductions that damage institutional knowledge and employee trust.

The AI transformation roadmap that most enterprises need to build should treat internal mobility rates as a lagging indicator of workforce AI readiness, not as a separate HR initiative. The two are the same program measured at different time horizons.

What Skeptics Get Wrong About Enterprise AI Upskilling

"Our employees are not ready for this"

The most common objection from operations leaders is that their workforce is not ready for AI tools, either because of age demographics, technical literacy, or cultural factors specific to their industry. Bank of America's experience suggests this concern is largely misplaced when the rollout design is right.

The bank's workforce includes tens of thousands of employees across retail banking, call centers, back-office operations, and field roles, none of which are naturally high-tech populations. The 90%+ adoption rate for Erica for Employees did not happen because Bank of America hired AI-native employees. It happened because the tool was introduced gradually, the use cases were immediately practical, and employees experienced a direct personal benefit from using it. Readiness is mostly a design problem, not a demographic one.

"We cannot afford the infrastructure"

The objection that an enterprise cannot build Academy-level infrastructure is understandable when the comparison is Bank of America's 1,000-person learning organization. But the key structural insight is not the headcount. It is the integration of learning into operations rather than treating it as a separate function.

Most enterprises already have some form of learning and development infrastructure. The change required is not building something new from scratch but redirecting existing L&D investment toward AI-specific capability building and, critically, connecting that investment to the technology deployment schedule so training and tool access arrive at the same time. Effective AI change management is as much about sequencing as it is about content.

"We will just wait until the tools mature"

A third common objection is timing-based: the tools are changing so fast that training people now means retraining them in twelve months. This is accurate but leads to the wrong conclusion. The foundational skills in Bank of America's Level 1 and Level 2 curriculum are not tool-specific. They are transferable: understanding how to identify tasks suited to AI assistance, how to evaluate AI-generated outputs critically, and how to integrate AI into a workflow without introducing new errors. These skills apply regardless of which specific platform is in use.

Companies that wait for the tools to mature will find themselves in the same position they are in now, except their competitors will be twelve months further into the capability-building curve. The most common reason AI pilots fail to scale is not technical immaturity but organizational unpreparedness that accumulates while leadership waits for the right moment to start.

The 3 Structural Decisions Every Enterprise Must Make Before Scaling AI Upskilling

Bank of America's approach resolves three structural decisions that most enterprises leave unresolved when they launch upskilling programs. Getting these right early determines whether the program scales or stalls.

1. Where humans stay in the loop

The clearest signal of a mature AI upskilling program is a documented set of decisions about where human judgment remains non-negotiable and why. Bank of America is explicit about this: not every decision in a 213,000-person organization is suited for AI assistance, and the Academy's training materials address this directly rather than leaving it to individual employee discretion.

For a mid-market manufacturer or logistics company, this translates to a practical exercise: before deploying AI tools in any workflow, map the decision points in that workflow and explicitly assign each one to either AI-assisted or human-only status. This is not a limitation on the technology. It is a governance step that prevents errors and builds employee trust, because people are more willing to use AI tools when they understand the boundaries clearly.

2. How capability levels connect to career pathways

The internal mobility data from Bank of America is evidence of a deliberate design choice: AI capability development is tied to career advancement, not just to job retention. Employees have a reason to develop Level 2 and Level 3 skills because those skills open internal opportunities that would otherwise require external hiring.

Most enterprise upskilling programs do not make this connection explicit, which reduces motivation and limits adoption. When an employee cannot see how developing AI skills connects to anything beyond keeping their current role, the incentive to invest serious effort is low. When the connection to internal advancement is clear and credible, the incentive flips.

3. The measurement framework

Training completion is the wrong primary metric. The right metrics are behavioral: prompts generated per user per week, tasks shifted from manual to AI-assisted, decision cycle times in workflows where AI is embedded, error rates in AI-assisted versus unassisted work. Bank of America tracks active usage at the level of weekly prompt generation by 150,000 users. That number is meaningful because it reflects whether AI is embedded in daily work, not whether employees attended a session.

Building this measurement framework before the rollout starts ensures the program is accountable to outcomes rather than activities. It also gives leadership a credible basis for the AI investment conversation with the board, since the metrics map directly to operational efficiency rather than to HR program participation.

Frequently Asked Questions

What is AI workforce upskilling?

AI workforce upskilling is a structured program that prepares employees across all levels to work effectively with AI tools in their specific roles. Unlike general technology training, it targets behavior change in actual workflows. According to Bank of America's 2025 data, over 90% of its 213,000-person workforce now actively uses AI-powered tools daily.

How do you scale AI upskilling across a large enterprise?

Scaling AI upskilling requires a layered capability model, not a single training program. Bank of America uses three levels: universal AI literacy for the full workforce, role-specific skill development for functional teams, and design and development capability for a smaller population. Infrastructure matters as much as content: the Academy's 1,000-person team makes continuous learning the default, not an event.

What is the 3-level AI upskilling approach?

The 3-level model covers literacy, functional application, and design capability. Level 1 ensures every employee understands what AI does in their context. Level 2 builds role-specific prompt and workflow skills. Level 3 develops the capacity to contribute to AI tool design and application discovery. Each level builds on the previous, and each targets a different population size.

How long does enterprise AI upskilling take?

Meaningful behavioral change at enterprise scale typically takes 18 to 36 months. Bank of America launched Erica for Employees in 2020 and did not reach 90% active usage until 2025. The timeline reflects gradual scope expansion and repeated low-stakes contact with tools before high-stakes application. Enterprises that try to compress this into 90 days typically achieve training completion without behavioral adoption.

What metrics should you track for AI upskilling success?

Track behavioral metrics, not activity metrics. Active users per week, prompts generated per user per week, tasks shifted from manual to AI-assisted, and decision cycle time reductions in targeted workflows are the right measures. Bank of America reports 150,000 active users generating 1.5 million prompts per week, which is a behavioral signal. Training course completion rates are an activity signal that does not predict adoption.

Why do AI upskilling programs fail?

Most AI upskilling programs fail because they are designed as training events rather than systems. BCG research shows that employees given AI tools without task-level guidance use them incorrectly at roughly the same rate as correctly. Additional failure points include skipping manager enablement, launching before tools are embedded in actual workflows, and measuring completion instead of behavior change.

How did Bank of America achieve 90% AI adoption across its workforce?

Bank of America achieved 90% adoption through gradual tool expansion and practical use cases, not mass training campaigns. Erica for Employees launched in 2020 for IT support, expanded to HR queries in 2023, and grew scope steadily. Each expansion gave employees repeated, low-stakes contact with AI assistance that built familiarity before more complex applications arrived. The IT service desk call reduction of over 50% reinforced adoption with a visible benefit.

What is the role of an internal AI academy in enterprise upskilling?

An internal AI academy provides the institutional infrastructure that makes upskilling continuous rather than episodic. Bank of America's Academy is a 1,000-person organization responsible for onboarding, professional development, compliance, and AI capability building. The key design principle is integration: learning and development is built into the technology deployment schedule so training and tool access arrive simultaneously, not months apart.

How do enterprises connect AI upskilling to internal mobility?

Enterprises connect AI upskilling to internal mobility by making AI skills a qualifying criterion for advancement, not just job retention. Bank of America filled 44% of roles through internal mobility partly as a result of upskilling investment. When employees can see that developing AI capability opens internal opportunities they would otherwise miss, motivation shifts from passive compliance to active investment. The career pathway must be visible and credible before training begins.

What is the right governance structure for AI in the workforce?

The right governance structure explicitly assigns every AI-assisted decision to one of two categories: AI-assisted or human-only. This is not a limitation on capability. It is the governance step that prevents errors and builds employee trust. Bank of America's training programs address this explicitly rather than leaving boundaries to individual discretion. For most enterprises, the right starting point is a workflow-by-workflow decision map completed before tools are deployed.

How do smaller enterprises replicate Bank of America's model?

Smaller enterprises replicate the structural logic, not the headcount. The key decisions are: connect learning to technology deployment timelines, define the three capability levels and who targets each, measure behavioral outcomes rather than activity, and make AI skills visible in career pathways. A mid-market manufacturer does not need a 1,000-person Academy to get these right. It needs clarity on which workflows AI will touch and a plan to prepare the people in those workflows before the tools arrive.

What role does AI play in training itself at Bank of America?

AI-powered conversation simulators are a core training tool at The Academy. Employees practice client interactions in a simulated environment and receive real-time feedback before applying those skills in live situations. More than 1 million simulations were completed in 2024. The effect is a reduction in the cost of learning through errors in high-stakes environments, and an acceleration in the time it takes new capabilities to feel natural and consistent.

How do you address employee resistance to AI tools?

Address resistance through practical benefit, not communication campaigns. Employees who experience a direct personal benefit from using an AI tool, like saving time on an IT ticket or getting a faster answer on a benefits question, adopt the tool and continue using it. Resistance is typically driven by uncertainty about what the tool does and fear of being evaluated on outputs they do not yet control. Gradual tool introduction with clear scope and immediate practical wins resolves both concerns more effectively than internal marketing.

What does AI upskilling mean for job security?

AI upskilling, when designed correctly, improves job security by making employees more valuable, not more replaceable. Bank of America's internal mobility rate of 44% is evidence that employees who develop AI capability move into roles with greater scope and impact. The risk of displacement concentrates in roles where employees do not develop AI-relevant skills, not where they do. The practical implication for enterprise programs is that upskilling communication should center on opportunity rather than job protection.

What should enterprises avoid when launching AI upskilling programs?

Avoid launching upskilling before the tools, the governance structure, and the measurement framework are in place. Training employees on AI capabilities six months before they have access to the tools produces a gap during which enthusiasm fades and organizational memory of the training evaporates. Similarly, avoid measuring success by training completion instead of behavioral adoption. And avoid skipping the manager enablement layer: managers who do not understand the new workflow expectations will passively undermine adoption in their teams without intending to.

How does AI upskilling connect to AI transformation strategy?

AI upskilling is the human implementation layer of AI transformation strategy, and it is the layer most frequently underfunded relative to technology investment. A transformation roadmap that specifies the tools to deploy and the workflows to target but does not specify who will be trained, at what level, and by when, will stall at the adoption phase even when the technology deployment goes smoothly. Effective change management treats workforce capability as a workstream with the same level of rigor as the technology deployment itself.