Two AI adoption challenges

There are two distinct challenges that any organisation serious about AI adoption has to face.

1. The first is a data transformation challenge: getting your content, skills data and knowledge assets into a state where AI agents can work with them. This week Leon Butler, IBM UK & Ireland general manager, talked to LinkedIn about the importance of the right data: investing in a powerful AI model will still produce poor results when that model is fed with ‘poor data, inaccurate data, old data’. Fixing L&D data This means extraction, atomisation, tagging, scoring, governance and integration, applied continuously rather than as a one-off migration. And right now, across every enterprise we work with, the current landscape is similar: thousands of SCORM packages no agent can read, very incomplete content metadata and governance workflows that are either not present or not followed. AI strategy conversations orbit a different altitude: which models, which agents, which skills will survive the next five years. But we forget the basic data transformation challenge at our peril; without AI-ready data, all AI strategies will fail

    

2. The second is an organisational design challenge: restructuring decision-making, team workflows and approval hierarchies so that the organisation can absorb the throughput that AI-augmented individuals now produce. Individual productivity gains from AI are visible and sometimes dramatic: in some of our customer organisations, individuals report working 30% faster with AI. But those gains are not translating into organisational performance because the organisations themselves are still shaped around human processing speeds: monthly steering committees, not daily ones; approval chains that assume someone reads every page; reporting cycles that cannot absorb what AI-augmented individuals now produce in a morning. Azeem Azhar has made this case persuasively: AI creates individual productivity, but the organisations around those individuals are legacy structures that were never designed for this throughput.

This article does not attempt to address the second challenge. Filtered provides learning infrastructure, not management consultancy, and the work of redesigning organisations around AI pace is a different discipline. Where we help is in the critical first step for any strategy: the data preparation work that has to happen before AI agents become useful, and the specific gap in existing tooling that makes that work harder than it needs to be.

The surface for learning has shifted

The most visible change in enterprise learning over the past eighteen months is where people go. They no longer open the LMS or LXP as a first instinct (if they ever did at all!). They open Copilot, ChatGPT, Claude, or an internal equivalent.

In an LMS, content is browsed: titles, thumbnails, pathways designed for human scanning. In a conversation with an AI agent, content is retrieved, reassembled and served in context. A catalogue designed for human browsing is of limited use to an agent that needs structured fragments, usable descriptions, quality scores and skill tags to select the right material.

The AI tool itself is a commodity. Any organisation can procure a Copilot licence. The differentiator is the proprietary knowledge behind it. An employee using Copilot connected to the open internet might learn something useful, but it will not be what the organisation needs them to know. A sales leader does not want their team learning about retail execution from generic sources. They want the team drawing on the organisation’s own methodology, case studies and standards. There is also serious business risk in the use of generic model knowledge to address internal work. Actions undertaken by humans or agents which aren’t based on the approved internal policy and training documentation because the AI can’t see it will inevitably expose the organisation to legal and regulatory challenges. 

The model is interchangeable, the knowledge base is the asset (and necessary defense against risk). Most enterprise knowledge bases are nowhere near ready.

What AI surfaces require

Bob Mosher and Conrad Gottfredson did the world a service when they coined the term ‘five moments of learning need’: 1) learning something new; 2) going deeper; 3) applying; 4) solving problems; and 5) dealing with change. But decades later, most enterprise catalogues are still built around ‘new’ and ‘more’. The apply, solve and change moments are barely served, because the content was never designed for them and the delivery surface (a portal with a search bar) was never going to get there.

The shift from portal to AI chat interface is, among other things, an opportunity to finally serve those apply and solve and moments, because AI agents can retrieve a targeted fragment and deliver it in context rather than recommend a one-hour course. That is what context-aware retrieval means in practice:

• Answer a question using only approved, quality-scored content
• Cite the precise source, module or extracted segment
• Deliver a short learning fragment rather than recommend a full course
• Ask follow-up questions to test comprehension
• Offer a scenario, role-play or short assessment where appropriate
• Record what was used, understood and demonstrated
• Recommend the next asset based on the learner’s response
• Return anonymised insight to L&D about gaps, weak content and unmet demand
• Operate in Copilot, Teams, Claude or another endpoint through MCP and APIs

None of this is exotic. Every line on that list is feasible with current technology. What makes it hard is that the content and the infrastructure underneath it are not ready. Now, this infrastructure is being built. In fact, everywhere start-ups are working on organisational memory for AI: retrieval-augmented generation, vector databases, graph-based knowledge systems. However, learning has specific challenges that the general-purpose solutions will not address: opaque data formats (SCORM, AICC, xAPI), painful integration gaps (try getting a bulk SCORM export out of most LMS vendors), and the need to answer questions that go well beyond ‘did the employee access it?’ to ‘did they understand it, can they apply it, and what should they receive next?’

Existing learning platforms are fundamentally unprepared for this shift. Which brings us to the question of what has to happen next.

Two speeds of transformation

There is a tendency in conversations about AI to collapse the future and the present into a single problem. The future is speculative: which roles will AI reshape, what new skills will matter, how should we think about career development when the half-life of technical knowledge is shrinking by the quarter? These are real questions, but they are hard to act on now because the answers keep moving.

The present is different. The here-and-now challenge is specific, and we would summarise it as data preparation. Without it, nothing else works. Unfortunately, AI itself will not ride to the rescue here. AI capabilities in content extraction, analysis, tagging and scoring have advanced substantially (more on that below), but increasing AI capability does not eliminate the organisational, system-level challenges that sit around the AI:

• Getting content out of the LMS in a usable form
• A data structure for atomising it 
• A rubric for quality scoring it
• Establishing governance workflows
• Feeding content effectively into AI agents (and governing those agents)

These are architectural problems. They require tooling, process and sustained attention. A better model does not make them go away.

Built on the Filtered trust stack

These lifecycle governance capabilities extend the core Filtered Intelligence infrastructure our customers already use: skills tagging via Skills Framework Manager, content atomisation and quality scoring via SCORM Intelligence, and agent delivery via Connect AI, which includes our forthcoming Microsoft Copilot Marketplace Connector. The result is a step change: from a layer that enriches content to one that governs the full content lifecycle from source to agent.

None of this works if the infrastructure itself is not trusted, and trust is where many AI-enabled vendors fall down. Filtered runs all AI inference on self-hosted models via AWS Bedrock. No customer content is sent to a third-party inference provider. No customer data is used to train any model. Every scored asset is traceable: the reasoning behind each quality score and skill tag is visible, so the layer can show its working.

We have held ISO 27001 certification continuously since 2018 (currently ISO/IEC 27001:2022, UKAS-accredited, with zero nonconformities at our most recent audit) alongside Cyber Essentials certification. Our August 2026 penetration test returned zero critical and zero high findings, and the MCP server specifically had zero findings of any severity. Filtered is approved on the UK Government’s G-Cloud framework, central government is among our customers, and we are a certified Future Plus Impact organisation. For organisations handling sensitive content and people data, the trust stack underneath the AI is as important as the AI itself, and we have built accordingly.

What we are not claiming to solve

We are not pretending to solve the problem of designing organisations around AI. That second challenge (restructuring decision-making cadences, approval hierarchies and team workflows for AI-speed throughput) is real and difficult, and the people best placed to address it are the senior leaders inside those organisations, many of whom are our customers. The best ideas we hear about organisational redesign come from them, not from us.

What we can do is close the lifecycle content governance gap so that the shift from portal to AI agents does not leave your organisational knowledge behind. The surface of learning has changed. The competitive advantage belongs to the organisations whose proprietary knowledge is structured, governed and accessible to agents, because the knowledge base is the asset and the model is interchangeable.

The alternative is AI tools that serve confident answers from broken foundations: the mass production of AI slop, at scale, with your brand on it.

If you are ready to close the gap, starting with a demo and test on a batch of 1,000 SCORM packages for a meaningful test with budget you can easily find before the end of the year, talk to us.

 1. My problem with the five moments is that they a) assume learners are motivated to work through the first four, and b) tack on ‘dealing with change’, as if change can be delivered by learning programmes alone. We tend to be highly motivated, efficient learners when something matters to us, and completely useless learners when it does not (see Tough: a better-researched framework than Mosher and Gottfredson’s). A framework based around the five moments for a new system, without the deep change management work to tie that system to employee needs, pains and desires, will not land. Change is an organisational design problem, and attitude is not a capability problem. Work like AssembleYou’s AI Confidence Method addresses precisely this: attitudes first, then capability. Will-do before can-do. Unless your machinery is wired along the desire lines that make people tick (ambition, pride, curiosity, desire), it will not work, no matter how neat it is.