Uneven GenAI Adoption

Why Firms Are Struggling to Support Everyone at Once

The reality of uneven GenAI adoption inside firms

In the same firm, right now, all of the following can be true:

One attorney is struggling to get Copilot to draft a decent email.
Another is building their third custom agent.
A third wants an agent that will scan 5,000 documents on a shared drive and surface the “best NDA.”
A fourth has been using GenAI for six months and is still frustrated that it gives them options instead of the answer.
A fifth is frustrated that access to his favorite tool has been blocked—and now he must complete mandatory GenAI training to regain access, despite never wanting to use GenAI in the first place.

This isn’t an anomaly. It’s the norm.

And it’s not well described by the way we usually talk about GenAI adoption.

Most firms still frame adoption in three buckets: power users, non-users, and everyone else in the middle. It’s a neat and easy-to-understand description. But when you actually look at what’s happening day to day, that “middle” isn’t a group at all.

It’s a wide, uneven spectrum of capabilities, expectations, confidence levels, and constraints—and it’s so wide that traditional training and rollout models are increasingly ineffective.

What looks like a training problem is something else entirely.
It’s a coordination problem across radically different starting points—and most firms don’t yet have the language, structures, or support models to deal with this new reality.

Why the “middle” of uneven GenAI adoption isn’t a group at all

When firms describe their GenAI adoption, they often say something like:

“We have our power users, our non-users, and everyone else in the middle.”

But when you actually look inside that “middle,” you don’t find a shared set of needs. You find dozens of distinct stuck points, for very different reasons:

  • People using Copilot for basic tasks but who are intimidated by anything more complex
  • People excited about agents, but with completely unrealistic expectations of what GenAI can do
  • People getting decent results sometimes, with no idea why it works when it works
  • People who understand the concepts but can’t translate them into their actual workflow
  • People who want to experiment but feel paralyzed by risk concerns
  • People who use GenAI regularly but have hit a ceiling they don’t know how to break through

These aren’t variations of the same problem.
They’re fundamentally different problems that happen to get lumped together.

Treating this group as a single “middle” masks the real work that uneven GenAI adoption creates.

Uneven GenAI adoption: why confidence and clarity don’t correlate

One of the most counterintuitive patterns we see is this:

The people with the most unrealistic expectations are often the most confident about what GenAI can do.

That confidence usually comes from reading about GenAI and inferring what should be possible—not from extensive hands-on experience. As a result, they’re confident precisely because they don’t yet understand where the technology breaks.

They hear that GenAI can “analyze documents” and conclude:

“Great. I’ll have it review all 5,000 NDAs in our system and tell me which one is best.”

When that fails—and it will—the conclusion usually isn’t:

"I misunderstood what this technology is actually capable of."

It’s:

“GenAI doesn’t work.”

Meanwhile, the people with real clarity—the ones who understand probabilistic outputs, limitations, trade-offs, and the need for verification—often sound less confident on the surface. They ask better questions. They hesitate. They know what can go wrong.

In other words, they may not feel confident about their own abilities or about the technology itself—but they are relatively clear about what GenAI can and cannot do, based on first-hand experience.

In uneven GenAI adoption environments, confidence is often mistaken for readiness, while clarity goes unnoticed. Most firms are inadvertently optimizing for the wrong signal.

Mapping uneven GenAI adoption without judging capability

We’ve tried to map GenAI fluency as a journey—not because it’s clean, but because people need some orientation. The intent isn’t to judge capability, but to help people recognize where they are and what to focus on next.

At a high level, the journey looks something like this:

Foundation

  • Understanding what GenAI is and isn’t
  • Using chat tools for basic tasks
  • Getting comfortable with prompting

Integration

  • Embedding GenAI into daily workflows
  • Using it for more complex work products
  • Developing pattern recognition for when it helps—and when it doesn’t

Expansion

  • Building custom agents for yourself
  • Experimenting with new use cases
  • Teaching others

However, even breaking the adoption journey into these stages is an oversimplification.

Two people at the same apparent point on the journey might be stuck for completely different reasons:

  • One lacks time to experiment
  • Another doesn’t understand the probabilistic nature of outputs

Same visible behavior. Completely different support needs.

The value of a journey model isn’t precision or categorization.
It’s decision-making.

It’s giving people language and direction without labeling anyone as “behind” or “advanced.”

And when you know where someone is, you can:

  • Decide what kind of work they should not yet be doing
  • Match support to risk, not enthusiasm
  • Stop over-investing in people who aren’t ready—and under-supporting those who are

Why traditional training fails in uneven GenAI adoption

Firms respond to uneven GenAI adoption by offering more:

  • Intro courses
  • Advanced sessions
  • Office hours
  • Tips and newsletters
  • Agent-building workshops

All well-intentioned. All necessary. And still not enough.

Here’s what tends to happen:

  • People show up to the wrong sessions
  • Beginners attend advanced sessions because they’ve heard “agents are important”
  • Advanced users sit in beginner sessions because they’re not confident enough to self-select
  • Sessions fragment, questions scatter, and no one quite gets what they need

The underlying issue isn’t logistics.

It’s that most people don’t know where they are on the journey and struggle to accurately self-assess. And producing truly differentiated support at scale is exhausting—because “beginner,” “intermediate,” and “advanced” simply aren’t granular enough. Each group still contains multiple, incompatible needs.

And the consequences are pretty obvious—frustration, delayed adoption, underutilization of tools, and a confusing story for clients and talent who listen closely.

The support uneven GenAI adoption actually requires doesn’t scale cleanly

What would help most people is fairly obvious:

Targeted, contextual guidance.
Someone seeing exactly where they’re stuck and giving them the next right thing to try.
Repeated exposure, feedback loops, and safe experimentation.

In other words: something that looks a lot like coaching.

But coaching dozens or even hundreds of attorneys isn’t scalable for most training departments. And behavior change doesn’t happen in a 30-minute session.

That tension—between what works, what can be offered, and what scales—is the core challenge most firms now face with uneven GenAI adoption.

So what now?

You can’t train your way out of this.

Uneven GenAI adoption isn’t primarily a knowledge gap.
It’s a capability development problem across a massive range of starting points.

What firms can do is start by naming the reality:

  • Stop treating adoption as binary
  • Give people language for where they are
  • Make progression visible and achievable
  • Normalize uneven progress instead of fighting it
  • Accept that variation is part of the system, not a failure of it

The real question may not be:

“How do we get everyone to the same level?”

But:

“How do we make uneven capability visible, governable, and productive—without compromising quality, confidence, or trust?”

That’s not something you solve with more training.
It’s something you have to design for.

We don’t have all the answers yet. But acknowledging the real shape of the problem is a better starting point than pretending it’s simpler than it is. We’re also seeing promising ways to provide coaching-like support at scale. The puzzle pieces are starting to fall into place.

If you’re managing GenAI adoption and this feels familiar, you’re not alone. And if you’re experimenting—even imperfectly—with ways to make this variation workable, those are the conversations worth having more of.

We’re always open to comparing notes or offering feedback on what you’re trying.