Over the past three newsletters, we have built an argument step by step.

We started with purpose. Analytics exists to support decisions, not to report activity. Without that clarity, even the most sophisticated data infrastructure produces dashboards that get opened on Monday and forgotten by Thursday.

We then looked at what most organizations already have. More decision-relevant data than they realize, waiting to be connected to the decisions that actually matter. The gap between available data and useful data is real, but it is smaller and more manageable than it first appears.

And we looked at how to close that gap deliberately, by designing technology, processes, learning programs and roles around the data we need rather than hoping it emerges as a by-product.

Each of these steps has value on its own terms. A learning analytics practice built around decision intelligence, grounded in a realistic data inventory, and sustained by design for data habits will outperform the conventional approach regardless of what else is happening in the world.

But something else is happening in the world. And it changes the stakes for everything we have discussed in ways that I think are still not fully appreciated inside the L&D profession.

That something is AI.

The quiet shift that is already underway

Most conversations in L&D about AI focus on what AI can do for learning. Personalization at scale. Content generation. Intelligent performance support embedded in the flow of work. Automated administration. These are real and significant changes, and I have written about them at length in the vision series that preceded this one.

But there is a dimension of AI that gets far less attention, and it is arguably the more consequential one for the argument I want to make here.

Organizations are already outsourcing significant portions of their decision-making to AI. Not in a dramatic, visible way. Not through a formal announcement or a strategic initiative. But quietly, incrementally, and continuously, in ways that are easy to miss precisely because they feel unremarkable.

An employee asks an internal AI assistant which procedure applies to their situation. A manager uses an AI-generated summary to prepare for a performance conversation. A recruiter relies on an AI ranking to decide which candidates to advance. A leadership team reviews an AI-produced analysis of skills gaps before deciding where to invest in development. In each of these moments, a decision is being shaped, at least in part, by an AI output.

This is not a future scenario. It is the present reality in a growing number of organizations. And the pace of adoption is accelerating, not stabilizing.

The question that this raises, and that almost nobody in L&D is yet asking seriously, is this: what is the quality of those AI-supported decisions? And who is responsible for ensuring that quality over time?

The problem with outsourcing decisions to AI

AI is genuinely impressive. I have said this before and I mean it. The capability of current AI systems continues to surprise me in positive ways. But capability and reliability are not the same thing. And in organizational decision-making, reliability is what matters.

When an AI system supports a decision, the quality of that support depends entirely on what the AI has learned from. Not just the sophistication of the algorithm, but the data it was trained on, the examples it was exposed to, the definitions it internalized, and the context it was given.

If an internal AI assistant is trained on company documents that are outdated, inconsistently written, or poorly structured, it will give outdated, inconsistent, and poorly structured guidance. Confidently. At scale. To everyone who asks.

If a skills intelligence system infers capability gaps from data that was never designed to capture the right signals, it will recommend development interventions based on flawed assumptions. Those interventions will be funded. People's time will be spent on them. And the gap between what was recommended and what was actually needed will remain invisible, because nobody connected the training outputs back to the performance outcomes they were meant to improve.

If a performance support tool recommends a course of action based on patterns in historical data, but that historical data reflects how work was done two years ago rather than how it is done today, the recommendation will be subtly but systematically wrong. And because it comes from a system rather than a person, it will be trusted more than it deserves.

This is not a theoretical risk. It is the predictable consequence of deploying AI on top of data infrastructure that was not designed with decision quality in mind. And it is happening right now in organizations that believe they are making progress on AI adoption.

Garbage in, garbage out is not a new principle. But it has never applied at this scale, this speed, or with this degree of invisible confidence before.

Why this is a learning problem, not a technology problem

Here is the connection that I think changes everything for L&D.

At its core, AI learns. We do not program AI step by step the way we program traditional software. We expose it to data and allow it to learn patterns, meanings, relationships and behaviors from that data. The quality of what AI produces is therefore a direct function of the quality of what it learned from.

This is exactly the same logic that applies to human learning. The quality of what a person can do is a direct function of the quality of the learning experiences they have had, the examples they were exposed to, the feedback they received, and the context in which their learning was applied.

L&D understands this logic deeply. We have spent decades thinking about how to design learning experiences that produce reliable, transferable capability. We know that poorly designed training produces fragile knowledge that does not transfer. We know that learning from bad examples produces bad habits. We know that without feedback loops and contextual application, even good learning degrades over time.

All of that expertise applies directly to AI. The challenge of ensuring that AI learns from high-quality, current, well-structured, contextually appropriate data is a learning design challenge. It requires exactly the kind of thinking that L&D is supposed to be good at. And yet it is currently being treated almost exclusively as a technology and data engineering challenge, handled by IT and data science teams who are highly skilled at building systems but who have no particular expertise in learning quality, knowledge transfer, or behavioral change.

This is a gap that L&D is uniquely positioned to fill. And it is a gap that will only grow larger as AI becomes more deeply embedded in organizational decision-making.

The organization that built decision intelligence first

Let me describe two organizations and ask you which one is better positioned to benefit from AI.

The first organization has invested heavily in AI tools. It has deployed an internal large language model trained on its document library. It has implemented an AI-powered skills platform. It has piloted a generative AI assistant for its L&D team. The data underpinning all of these systems is the same data that has always existed: a mix of LMS completion records, HR system exports, manually maintained skills frameworks, and a document library that is comprehensive in volume but inconsistent in quality and currency. Analytics is still largely built around completion rates and satisfaction scores, because that is the data that is clean and accessible. Nobody has systematically asked which decisions the analytics should be supporting, because the conversation about analytics has always been about reporting rather than decision quality.

The second organization started differently. Two years ago, before significant AI investment, it worked through the kind of process described in this newsletter series. It got clear about which decisions mattered most. It took an honest inventory of the data it had. It identified the gaps between available data and needed data. It started designing for data across its technology choices, its program designs, and its core processes. It built a Learning Data Pond that connected data from across its learning and HR systems into a clean, consistent, analytics-ready layer. And it started using that layer to produce analytics that actually informed decisions, gradually earning credibility with business and HR leadership in the process.

Now both organizations are deploying AI. But the second organization is deploying it on top of a data infrastructure that was designed to produce high-quality, consistent, decision-relevant signals. Its AI assistant gives more reliable guidance because it is trained on better-structured, more current content. Its skills intelligence is more accurate because it is inferred from performance data rather than self-assessments. Its analytics AI produces more trustworthy outputs because the data it draws on was designed to support decisions rather than to satisfy reporting requirements.

The AI tools available to both organizations are essentially the same. The gap in outcomes between them is not a function of the technology. It is a function of the decision intelligence infrastructure that one of them built and the other did not.

L&D as the guardian of decision quality

This brings me to the argument I want to make directly, because I think it is the most important strategic reframe available to L&D right now.

In an AI-driven organization, the quality of decisions depends on the quality of what AI learns from. And ensuring that quality over time, as the organization evolves, as work changes, as strategies shift and roles are redefined, is not a one-time setup task. It is an ongoing responsibility that requires deep understanding of how learning works, how knowledge becomes behavior, how context shapes performance, and how data should be designed to capture what actually matters.

That responsibility has no obvious owner in most organizations. IT builds the infrastructure. Data science teams build the models. Business leaders consume the outputs. Compliance and risk functions monitor for specific categories of error. But nobody owns the question of whether the organization's AI is continuously learning from the right things, in the right context, with the right quality.

L&D should own that question. Not because it is the only function with a stake in the answer, but because it is the function with the deepest expertise in exactly the capabilities the question requires.

This is what I described in the vision series as L&D becoming the guardian of decision intelligence. It is not a role that involves controlling AI or owning technology infrastructure. It is a role that involves owning the learning responsibility behind AI-supported work. What AI is allowed to learn from. What must be corrected or updated as reality changes. How AI guidance evolves as work evolves. Whether the outputs AI produces are actually supporting good decisions or subtly undermining them.

That is a learning problem. And Learning and Development is the right function to lead on it.

What this means in practice

I want to be concrete about what this role looks like, because it is easy to describe it in abstract terms that sound compelling but give no indication of where to start.

The first practical implication is that L&D needs to get involved in AI governance conversations that it is currently not part of. When the organization is deciding how to train its internal AI systems, what data to include, how to update that data over time, and how to monitor output quality, L&D should be at the table. Not as a passive observer, but as the function with the deepest expertise in learning quality and knowledge transfer.

The second implication is that the design for data habits described in the previous newsletter apply directly to AI readiness. Every decision about technology, process, program design, and role that is made with data quality in mind is also a decision that improves the quality of what AI can learn from. The two agendas are not separate. They are the same agenda pursued at different levels of ambition.

The third implication is that L&D needs to develop the capability to monitor AI output quality as a learning problem. When an AI system gives guidance that is outdated, contextually wrong, or subtly misleading, that is a signal that something in what it learned from needs to be corrected. Identifying those signals, diagnosing their cause, and correcting the upstream data or framing is exactly the kind of work that L&D should be equipped to do, because it is structurally identical to identifying that a training program is producing the wrong behaviors and redesigning it.

None of this requires L&D to become a data engineering function or an AI development team. It requires L&D to extend its existing expertise in learning quality, knowledge transfer and performance support into a new domain where that expertise is urgently needed and currently absent.

The window is open, but not indefinitely

I want to close with something I feel strongly about, because I think the timing matters more than it might appear.

The organizational functions that will own the decision intelligence agenda in AI-driven organizations are not yet determined. In most organizations, that question is genuinely open. IT has a claim based on infrastructure ownership. Data science has a claim based on technical capability. Strategy has a claim based on its proximity to executive decision-making. HR has a claim based on its ownership of people data and workforce planning.

L&D has a claim based on something none of those functions has: deep expertise in how learning works, both for humans and increasingly for AI systems. That expertise is genuinely distinctive. It is directly relevant to the most important challenges that AI adoption is creating. And it is currently underutilized in almost every organization I am aware of.

But windows do not stay open. As AI adoption accelerates and the decision intelligence agenda becomes more visible, other functions will move to claim it. Some already are. The organizations that position L&D as the guardian of decision quality now, before the territory is fully contested, will find it much easier to sustain that position than those who try to claim it after the fact.

This is not a call to move fast and break things. It is a call to move deliberately, starting with the foundations described in this series, and to be clear-eyed about where those foundations are leading.

Decision intelligence, done well, makes L&D indispensable. Not because it makes a compelling argument for L&D's relevance, but because it makes L&D the function that the organization genuinely cannot do without when it comes to the quality of the decisions that shape how work gets done.

That is the destination this series has been building toward. And the path to it starts with the same question we asked in newsletter one.

What is analytics actually for?

Peter Meerman

SLT Consulting — Learning Analytics Made Easy