AI in L&D: What Comes After the Pilot Phase? How to Scale AI

Enterprise AI adoption is accelerating rapidly, but scaling remains far more difficult than starting. McKinsey’s 2025 research found that 88% of organizations now use AI in at least one business function, yet only 7% report having fully scaled AI across the enterprise. Deloitte’s 2026 State of AI research echoes this pattern, showing that while access and experimentation continue to expand, most organizations are still struggling to convert isolated pilots into operational capability at scale.

In learning and development, this gap is especially visible. AI tools often generate strong early enthusiasm because they can accelerate drafting, simulation design, assessment creation, and knowledge transformation. But once teams try to move from promising experimentation to sustained enterprise use, momentum often slows.

In a previous article, we explored why AI adoption in L&D often slows after the pilot phase, highlighting how governance gaps, security concerns, and unclear ownership create structural friction at the point where scale should begin.

This article builds on that analysis by examining what large organizations need to put in place to move beyond that friction. It argues that the issue is not the pilot itself, but whether the organization has designed the conditions required to support trust, capability, governance, and repeatable use at scale.

This article draws on emerging findings from a joint research initiative by CommLab India and researchers at Lancaster University exploring how artificial intelligence is shaping workplace learning across large organizations. The patterns discussed here are informed by anonymized interviews with enterprise learning leaders across multiple industries and AI adoption contexts.

One of the easiest mistakes organizations make with AI is assuming that a successful pilot naturally leads to scale. In practice, it rarely does.

A pilot can demonstrate that a tool is useful. It can show that content can be drafted faster, simulations can be generated more quickly, or repetitive learning design tasks can be accelerated. It can even create real enthusiasm inside a learning team.

But scale asks a different question. Not “Can this work?” But “Can this work reliably, safely, repeatedly, and legitimately across the organization?”

That is a much harder threshold to cross.

Across the organizations studied, many of the early AI experiments in L&D did, in fact, work. Teams found practical value quickly. They reduced development time, improved responsiveness, expanded experimentation, and opened up new possibilities for simulation, practice, and content transformation.

Yet in many cases, that momentum slowed once the organization had to decide whether AI should become part of the actual operating model of learning.

This is the point at which many organizations discover that the real challenge is not adopting AI. It is building the conditions under which AI can become trusted and usable at scale.

That is not a tooling challenge. It is a design challenge.

One of the most common reasons AI adoption in L&D stalls is that organizations approach scale as if it were primarily a software deployment problem.

They focus on tool selection, license access, procurement, platform approval and rollout communication. Those things matter. But they are not enough.

Scaling AI in L&D is not simply about getting more people access to a tool. It is about redesigning how work gets done once the tool becomes part of the workflow.

That means asking more structural questions:

These are operating model questions, not just technology questions.

That is why many organizations remain stuck in pilot mode even after approving tools.

Access alone does not create transformation. Scale begins only when AI is treated not as an add-on, but as part of how the function operates.

This is where many organizations first begin to move beyond superficial adoption.

One of the strongest patterns across the cases was that organizations often treated too many AI use cases as if they carried the same level of risk.

That creates unnecessary friction.

Not every AI use case in L&D should trigger the same degree of review, governance, or hesitation. A low-risk use case such as generating draft headlines for an internal learning module is not the same as using AI to create regulated learning content, customer-facing simulations, or decision-support materials in sensitive environments.

When organizations fail to distinguish between these levels of risk, experimentation slows unnecessarily.

What many large organizations need is not unrestricted AI use. They need a safe lane for low-risk experimentation.

This means creating clearly bounded spaces where learning teams can test AI in ways that are useful, visible, governed, low-risk and easy to learn from. This is important because scale does not begin with enterprise-wide transformation. It begins with well-structured learning.

And learning requires experimentation.

Low-risk L&D use cases that are often good starting points

The point is not that these uses are risk-free. It is that they are easier to govern, easier to learn from, and easier to scale responsibly than high-stakes use cases introduced too early.

What Helps AI in L&D Move Beyond the Pilot Phase

One of the most consistent patterns across enterprise AI adoption is that governance often enters the conversation too late.

In many organizations, experimentation moves ahead quickly inside small teams. Then, once momentum builds, the work is handed over to legal, compliance, privacy, or security for review. At that point, what could have been a constructive design conversation often becomes a bottleneck.

This is not because governance functions are inherently restrictive. It is because they are often being introduced only after assumptions have already solidified.

When governance appears only as a gate at the end, it is far more likely to slow progress. But when governance is brought in earlier, in proportion to the level of risk, it becomes much easier to shape experimentation in ways that are both useful and safe.

This is one of the most important shifts organizations need to make.

Governance should not operate only as a final checkpoint. It should become part of the design of experimentation itself.

Gartner’s 2025 research suggests that organizations that conduct regular assessments, implement AI-specific usage policies, and provide tailored guidance are significantly more likely to achieve stronger AI value outcomes. In other words, governance is not just a risk function. It is increasingly a value-enabling capability.

This is one of the clearest differences between organizations that remain stuck in pilot mode and those that begin to scale more deliberately.

In several of the organizations studied, meaningful progress only began once AI could be used inside secure, enterprise-approved environments.

This mattered enormously.

In the early stages, many teams were blocked from using public AI tools because of concerns around IP protection, prompt retention, data leakage, external model training, and unclear terms of use. That hesitation was understandable. But blanket restriction often created a second problem: teams were left without a safe and sanctioned way to experiment productively.

This is where trusted AI environments become essential.

Large organizations do not necessarily need to open every tool to every employee. But they do need to create environments in which useful AI-supported work can happen with enough security, visibility, and control to make adoption legitimate.

That might include enterprise copilots, sandboxed AI workspaces, private model instances, approved prompt environments, and function-specific AI workspaces for L&D.

Once these environments exist, the conversation shifts. AI is no longer framed simply as a risky external tool. It becomes something the organization can govern, observe, and improve.

That is often the turning point between blocked experimentation and trusted operational use.

This is often one of the most practical ways to unlock movement after a stalled pilot.

One of the quieter reasons AI adoption slows is that many organizations tell teams to “use AI responsibly” without ever defining what responsible, useful, or high-quality AI use actually looks like in learning work.

That creates ambiguity.

If teams are given access to AI but no shared guidance around:

then experimentation remains scattered and difficult to scale.
Scale becomes possible only when experimentation becomes legible.
This means organizations need to move from vague encouragement toward clearer, role-relevant guidance.

For example, the expectations for an instructional designer using AI to accelerate early-stage design work should not be identical to those for a facilitator, a manager, or a subject matter expert. The learning function needs usage models that reflect the actual shape of work.

This is one of the most overlooked but most important steps in moving beyond pilots.

One of the clearest lessons from the research is that access does not automatically create adoption, and adoption does not automatically create value.

People need to know how to use AI well inside the realities of enterprise learning work.

That means organizations need to build not only tool familiarity, but also the human capabilities that allow AI to be used effectively, responsibly, and strategically.

Across the cases, the teams making the most meaningful progress were not necessarily the ones with the most advanced tools. They were the ones with stronger capability in areas such as framing the right problem, judging output quality, contextualizing AI-generated material, orchestrating AI within workflow, and maintaining oversight and accountability . These are not secondary capabilities. They are increasingly the core of effective AI use in L&D.

This is why capability-building cannot stop at “how to prompt.” Prompting matters. But it is not the same as professional readiness.

The capabilities that matter most after the pilot stage

This is where AI adoption begins to mature from experimentation into professional practice.

Eventually, every organization that wants to scale AI in L&D faces the same shift:

It must move from a collection of interesting experiments to a coherent way of working.

That is what an operating model provides.

Without one, AI remains fragmented:

This creates uneven adoption, uneven quality, and very little cumulative learning.

What large organizations need instead is a clearer AI operating model for the learning function. Not necessarily a rigid framework, but a repeatable logic for how AI is used, reviewed, integrated, and governed across the work.

That includes:

This is the point at which AI stops being a side experiment and starts becoming part of the actual infrastructure of learning work.

And that is usually the point at which scale becomes possible.

What L&D Needs to Scale AI Responsibly

For many organizations, the challenge is not knowing whether AI matters. It is knowing what to do next in a way that is practical and credible.

The good news is that moving beyond the pilot phase does not require a full enterprise transformation plan on day one. But it does require deliberate movement.

Five practical actions L&D leaders can take now

These are modest moves, but they matter because they begin to turn AI from a scattered set of experiments into a managed capability.

And that is how scale usually begins.

In L&D, AI pilots often create exactly the kind of momentum organizations hope for. They generate curiosity, speed, possibility, and visible wins. But pilots alone do not create transformation.

The organizations that move forward are not necessarily the ones that run the most experiments. They are the ones that build the conditions that allow useful experimentation to become trusted, repeatable, and scalable.

Across the organizations studied, one pattern became increasingly clear: The pilot proves possibility. Scale requires design.

That design includes governance, trusted environments, clearer usage models, stronger human capability, and a more coherent operating model for how AI fits into the work of learning itself.

This is why the real challenge is not whether AI can help L&D. In many cases, it already can. The deeper question is whether L&D is building the structures that allow AI to become more than a promising pilot. Because in large organizations, scale is rarely a consequence of enthusiasm alone. It is the result of deliberate organizational design.

—RK Prasad (@RKPrasad)