What L&D Leaders Told Us at CLO Exchange: AI Is Ready. Most Programs Aren’t.

This week we spent three days in Park City at CLO Exchange, sitting down with dozens of senior L&D leaders and Chief Learning Officers from some of the largest organizations in the country. We also had the chance to co-present on AI trends and real-world results with Cara Halter of GP Strategies. This meant we got to hear these themes echoed back from the audience in real time, and dig into the strategy behind them on stage.

Here’s what we heard, what we presented, and what we think it means for anyone building in L&D right now.

Where most leaders really are

Three themes came up again and again in conversations. 

Interest in AI roleplays tools is skyrocketing, and it’s easy to see why. Leaders are watching the need get more urgent, watching the capabilities expand, and starting to picture the value this could unlock for their teams. A lot of programs started with AI on basic workflows: drafting a first pass of a course outline, generating content faster than before. That’s real value, and it’s just the on-ramp. The bigger story is how far the platforms themselves have moved. AI can now train and coach in the flow of work, join live calls, score conversations against a rubric, and apply deep context about a specific role, industry, or compliance environment. We heard that curiosity turn into real plans, in specific, human terms: a learning leader mapping out what it would take to rebuild her org’s entire learning stack from the ground up with AI as the foundation, and a team asking whether an AI tutor could teach content directly and then verify whether someone mastered it. The questions in that room were about how fast to move.

Scale is the bottleneck, and the opportunity. Some of the most respected programs we heard about are also the most fragile, because they depend entirely on a small number of skilled humans. One leader described coaching programs that get consistently strong results but can’t grow because they’re built around a handful of trainers who can only be in one place at a time. Another manages a learning function with a couple dozen people supporting a workforce of thousands, and is actively rethinking how performance management and development can work when an old, top-down evaluation model doesn’t scale to the org anymore. The leaders getting the most out of AI right now are rethinking the structure of the program itself, so their best human coaches and trainers can focus on the highest-value moments while AI handles the repetition and reach. Snowflake ran into this same wall with manager coaching and got roughly 1,200 hours a quarter back once AI took on the repetition.

The data to prove ROI exists now, but most programs aren’t designed to capture it. For years, L&D has leaned on surveys and self-reported confidence as the main proof that a program worked. That’s starting to change because the tools now make better data possible, and most program designs haven’t caught up yet. A leader in a highly regulated, compliance-heavy field talked about wanting to catch and coach on issues in the moment, during real interactions, without running into the data and privacy restrictions tied to sensitive records. The fix is designing practice scenarios around the hardest conversations, ones that recreate the same pressure and complexity, so people can build the skill before it matters for real. Another leader running training across a set of highly distinct, siloed business units was stuck on a related problem: how to build one strong training foundation that can be reskinned quickly for each group without starting from scratch every time. In both cases, the underlying opportunity is the same. Set a baseline, apply consistent scoring across every learner, and show growth that ties directly back to business results instead of a satisfaction score.

The three phases of AI adoption in L&D

On stage, we laid out a framework for where organizations really sit on the AI adoption curve, and it’s rarely where they think.

Phase 1: productivity and efficiency. This is where most teams are today. AI is used to build the same deliverables faster: scenarios, assessments, images, video. It’s useful, but it’s also the safest possible use of the technology, and a lot of teams default to the familiar formats because AI still feels new and a little uncertain.

Phase 2: learners using AI directly. This is the next step, and where a growing number of leaders are starting to experiment: roleplays, AI coaches, learning bots that learners interact with themselves. The catch is that most organizations in this phase are still treating learning as an event or an assigned activity, something a learner does and finishes, rather than something ongoing.

Phase 3: learning as the work. This is the emerging frontier, and almost nobody is fully there yet. It’s AI embedded directly in the flow of work, surfacing relevant context and coaching in real time as part of doing the job itself. Very few products or organizations have made it this far. Naming it now still matters, because it changes what you choose to build first. It’s the thinking behind why we built Yoodli around Learn to Practice to Do instead of a standalone course.

The paradigm shifts underneath

A few bigger shifts are reshaping how L&D gets designed, and they explain why the phases above matter.

Courses used to be discrete units with a start and an end. Now they’re giving way to continuous learning agents and ecosystems that don’t work that way. LMS and LXP vendors are starting to respond, and language like “headless SaaS” is showing up in how the space describes itself.

Content used to sit at the center of program design. AI is now good enough at answering the “what”: here’s the content, here’s the information. That frees L&D to focus on the “how”: how does this specific person build the skill. Personalization that used to be cost-prohibitive at scale is affordable now.

The harder shift is control. It’s moving from designers to learners, with AI driving more of the moment-to-moment experience. That asks instructional designers to give up some control they’ve held for a long time, and it only works if they trust the data layer underneath it.

None of this is binary, and it shouldn’t be treated that way. Think of it as a spectrum, with tight control on one end and learner-driven exploration on the other. A compliance topic needs consistency and precision above all, so it sits closer to the controlled end. Something exploratory can tolerate more flexibility, so it drifts toward the other side. What’s missing is a shared vocabulary for the space in between, which is part of why so many of these conversations start from scratch every time.

What it looks like when it works

Two examples from our own work show what these shifts look like in practice.

At Ring Central, managers grading AI roleplays by hand for a new rep used to spend a full day per rep on certification. With Yoodli, RingCentral cut that certification time by 90%. Managers walk into a debrief with objective analytics already in hand instead of grading calls by ear, and reps can practice as many times as they need before that debrief happens. The result is a faster time to field, without adding headcount to get there.

At Ochsner Health, the challenge was interpersonal skills at scale across 10,000 people, where coaches didn’t have a consistent way to give feedback across that many individuals. The team built 12 custom goals aligned to Ochsner’s values and what already worked for their strongest coaches, established a baseline score before training, and measured the lift afterward. The areas that moved: trust between practitioners, ownership, and how clearly people set expectations around their own careers.

Both are the same underlying problem: a program that works but can’t scale without cloning your best coaches ten times over. That’s what Yoodli is built to solve for L&D teams broadly, standardizing practice and evaluation across roles, regions, and cohorts so every rep or learner gets the same rigor, whether there are 50 of them or 50,000.

Building an ecosystem beats buying a tool

A few practical threads tie this all together for anyone starting to build an AI strategy for L&D.

Look for partners over vendors. The ones worth keeping push your program design forward. The ones to skip just sell you a feature.

Compose the stack on purpose. Go deep with specialized tools where the stakes are high, like regulated or safety-critical topics, and stay broad with general tools where one platform can reasonably span many use cases.

Test for hallucination risk anywhere precision matters most, particularly compliance and safety. Not every use case needs that scrutiny, but the ones that do need it applied rigorously.

Start with what you already have. Prove value with the tools already licensed across the org, like Copilot, before making the case for a dedicated platform.

And remember that the core questions haven’t changed. What skill are we building, who needs it, how will we know it worked. AI just gives you a new lens to ask them through.

What this means for L&D right now

The organizations that pull ahead over the next year will be the ones that used this moment to rethink the program itself: where humans add the most value, where AI can extend reach without diluting quality, and how to build measurement in from the start instead of bolting it on at the end.

Grateful to Cara Halter and the team at GP Strategies for co-presenting, and to everyone who stopped by to talk shop, share a pain point, or push back on an idea. Those conversations are exactly why events like this matter.

Interested to hear how Yoodli can help grow your training and L&D programs? Get Connected.