Agentic AI for learning is floundering because the data foundation for enterprise AI in learning is broken. Our customer data consistently suggest that 50-60% of internal content items in a typical catalogue lack descriptions, making any AI retrieval or skill-based recommendations meaningless. Whether you are maximising the utilisation of your LXP or moving beyond it, we need to fix the foundations. This article sets out a simple framework to do that, including the tools you need, and the insights you’ll gain at each stage.

The days of disillusionment in genAI and agentic AI have arrived. Recently Salesforce, whose share price took a beating over AI this year, responded with justified commentary on the very real risks of rolling out AI without the proper data foundations or guardrails in place:

Rolling out artificial intelligence tools without the right systems in place could have “catastrophic” consequences for businesses, the UK head of the cloud-computing company Salesforce has warned.

“AI is making such profound decisions on behalf of people,” Zahra Bahrololoumi, chief executive of Salesforce’s UK and Ireland business, told The Times. “You can’t just release it in the wild. Unfortunately, some companies do. They think they can build it all themselves, but they learn very painful lessons that you actually do need guardrails and safety precautions.

“If AI doesn’t have access to the right data, or doesn’t have the right guardrails, it could be catastrophic for businesses.”

Salesforce has good reasons to call out the risks in ungoverned AI adoption, and we see the same challenges. Filtered implements agentic AI across its product, and we apply AI in close concert with other vendors which use their own AI: Degreed, SAP SuccessFactors, Workday, and many content providers.

Everywhere, the same pattern appears: AI agents without tight guardrails, organisational context or strong data foundations go haywire, creating risk and wasting a lot of money on genAI tokens.

Why is this? Well, one problem is that particularly in HR, we treat agents as if they were people, with skills we can impart to them, and the ability to learn from experience (and we thus assume that we can ‘train’ them to behave with the same reflexes as people do). In other words, we expect them to behave in a smart, non-stupid way.

But genAI is a form of complex machine based on a set of numbers or ‘weightings’ and no more than that: it can reason, but it does not think in the holistic way we do and certainly does not learn the way a human learns. The mechanism that drives all agentic output is simple: a prompt and a response. The complexity comes from giving agents access to tools and generating new prompts from the agent’s own responses in cycles that enable agents to reason their way through tasks.

 But that is all agents are: if the prompts go haywire at any point, if the right tools aren’t available at any point, it breaks. Don’t expect the resourcefulness, tenacity or common sense of a human operator. Don’t expect an LLM to learn any skill that is not recalled and supplied as a prompt when needed, part of its explicit training or its fine-tuning (and those two are weak influences).

What does it mean for talent development?

The right foundations for AI that works are:

• Clean data about people and content to draw on for its responses
• The right tools to get that data: particularly, vector search
• Governance: skills, rules, shared memory

These foundations are almost never in place for enterprise talent development where, our customer data consistently shows, 50% of content assets do not have even have a human-readable description. As a judge of the AI in learning technologies awards for the past few years, I can tell you that the successful projects are fundamentally grounded in clean databases, appropriate search and guardrails – not in using the fanciest frontier models or interfaces.

Because of its weak foundations, we can expect learning and talent to be deluged in AI slop: content pathways, development plans and recommendations that seem feasible, only to miss crucial details. This is indeed what is happening. That’s because the agent you switch on inherits every gap in your content and skills data, and then acts on those gaps with total confidence.

An agent building a development plan does not know that half your catalogue has no usable description (which we see in nearly every enterprise customer). It cannot see that a course is tagged to a skill it barely touches. It retrieves what it can reach, ranks it by whatever signal happens to be available, and presents the result as considered advice. Unlike a human advisor, the AI does not hesitate: it will recommend the wrong thing to the wrong person, at scale.

The plumbing is being laid over a mess

is Learning data is increasingly being opened up to agents through connectors and MCP servers, so that a plan built in Word, or a query put to a copilot, can pull from the whole library rather than from a static course list. That direction of travel is the right one, and it is where enterprise learning should be going. The difficulty lies in what is actually being connected.

A data layer is only as trustworthy as the data flowing through it, and the quality and accuracy benchmarks that layer needs have not been embedded. The trust stack that should sit underneath, scoring, checking and endorsing every asset before an agent is allowed to touch it, is largely absent, which means new pipes carrying the same contaminated supply.

As the larger SaaS players start to reckon with the risks of the AI turn, HR and learning vendors are racing in the opposite direction, switching on MCP servers before fixing the underlying content and skills data. The instinct is understandable, because an MCP endpoint is quick to stand up and it demos well, but an MCP over bad data serves bad data faster. You have not solved the retrieval problem so much as automated it.

Dan Tesnjak, Lori Niles-Hofmann and Sandra Loughlin have all written on the topic of laying the right foundations for AI. The whole connection between skills data and AI implementation depends on the right infrastructure, and currently these infrastructure layers, to the extent they exist, are full of holes (such as 50% of content not having the right metadata). At the best, all you will have organised and mapped is the encoded, official knowledge about how the business should run. 

But, as Loughlin writes, capturing what has historically been seen as tacit knowledge – the things that don’t get written down that constitute a deeper layer of organisational intelligence – is also part of the challenge. However, it’s hard to do this in practice when so much of that knowledge is wrapped up in day-to-day work, private conversations and confidential documents. You need more than a skills infrastructure layer. You need, as our CEO Mal has written, a trust stack that protects any information used by the system with content-specific rights and containment.

What does quality scoring look like? Our agent examines the material in any learning content package or pathway and evaluates it on the basis of five criteria. Here is what good looks like for those criteria (these are excerpted from our actual agent prompt – you’re welcome!):

There are other areas to include. We are also planning to incorporate usage data to understand the engagement and retention characteristics of content based on its start vs completion rates. And we are planning to allow customers to customise the rubric by adding an additional prompt on top of ours. This will capture important organisational nuance like including stale terminology which applies to specific internal initiatives and platforms in the freshness evaluation.

This data is made available at a summary level and as a detailed breakdown through our API and MCP server, so that AI retrieval is able to apply quality as a parameter to what it achieves. For a long time Filtered has been able to determine what is most relevant. Now, thanks to developments in AI, we can also determine what is best – just at the point when the same AI developments make knowing that essential.

The route to AI that works runs through the data

The vendors switching on MCP servers are right about the destination, because agentic retrieval is where enterprise learning is heading. The route to it runs through the data, though, and the data is where the work has to be done.

In turn, you need to be able to trust the vendors and people who work directly with your most sensitive content and people data: HR agentic readiness without a trust stack will never reach full scale production across your workforce. Whether you are maximising the utilisation of your current LXP or moving beyond it entirely, the foundations have to be fixed first.  The alternative is a strategy that results in the mass production of AI slop and expensive agents that are far less useful than human learning advisors. That kind of failure is the risk of staking everything on AI without fixing the foundations.

“But my stack has its own AI”

One possible objection to the need for a separate data layer is from teams who have committed to a single vendor is: We have chosen our platform. It has its own AI, its own copilot, its own MCP server. The data lives in one place. Why do we need a separate layer to fix anything?

Because the AI does not fix the data, it only reads it. A single vendor’s AI assistant retrieving from a single vendor’s catalogue is still retrieving from a catalogue where half the internal learning items have no description and the skill tags were never checked against a benchmark.

There are plenty of new AI learning platforms that use genAI both to generate quality content and to enrich its metadata. We can expect this as a standard going forward (as well as a good deal of poorly-designed AI slop!). But new content isn’t the gap we need to fix. Instead, the need is to audit the vast mass of legacy content data (usually at least 10,000 PDFs, videos, SCORM packages) and extract what is valuable from that content.

This content is largely unreadable to AI in its present form. So an MCP over an unenriched catalogue gives you confident answers drawn from the same broken foundation and is likely to favour generic content over your own stuff because generic content has far better metadata (often actually over-promising on what the content can deliver for a user). And a single stack or a new AI vendor doesn’t exempt you from fixing the data layer any more than building a new house exempts you from laying its foundations.

A framework to fix the foundations

You can get started with fixing the foundations in an Excel spreadsheet tomorrow. A fully traceable and retrievable data foundation for learning requires more embedded and specialised infrastructure. But none of it needs to cost anywhere near the sum of an enterprise-level skills product, with per-seat licenses. And the whole point of fixing the foundations is to keep a tight lid on runaway costs from agentic AI. Here are the stages involved:

A framework to fix the foundations

Fixing the foundations is necessary to get agentic AI for learning delivering a return on investment to your organisation. Filtered customers, from leading high street banks to big four consultancies and the world’s biggest FMCG firms, are realising multimillion dollar savings from clean data foundations for learning that enable:

• Consolidation of learning content libraries into single provider through tagging that enables a direct comparison (= £400k – £1.2m saving on licenses)
• Realising the value from their current investments in LMS and learning experience platforms by generating measurable engagement and shifting from third party to internal content (= £1m+ saving on content licenses)
• Self-build upskilling hubs via agents and AI-assisted coding on top of clean, indexed learning content metadata (= £1.5 – £3m saving vs investing in skill platform licenses)

And this is simply at the level of fixing the foundational layer that supports the discovery of relevant learning content. Beyond the foundations, there is huge potential in agentic AI to support better business performance. That requires something more than clean foundations: it requires a shared insights layer across the different data platforms and AI agents involved in supporting performance so you can measurably improve the efficiency of onboarding programmes, go-to-market initiatives or mergers & acquisitions.

 And we will be announcing an entirely new approach to that problem soon.