Has AI Finally Fixed L&D's SME Problem?
Four practical AI use cases to help manage the SME bottleneck
Hey folks 👋
If you work in L&D, I’d put money on the fact that right now, you’re waiting on a subject matter expert in some shape or form.
Maybe it’s an interview that’s been rescheduled three times. Maybe it’s a 300-page technical manual that arrived with “turn this into training” and no further context. Maybe it’s the sign-off on a design that’s been sitting in someone’s inbox for two weeks.
As a result of AI, our workflows are in some ways faster than anything we could have imagined five years ago. And yet — when I talk with L&D teams in 2026 — the same pain surfaces first, unprompted, every time: the SME bottleneck. Aka, getting the right sort of access to the right expertise in the right moment to extract the information we need to be able to do our jobs.
The numbers back this up. In the State of Instructional Design survey I ran with Synthesia last year with 400+ practitioners globally — SME delays came out as the second-largest barrier to speed, cited by 30% of designers, and a barrier to quality for a further 13%. TL;DR: for the majority of any given project, the designer is working without a live expert in the room — filling gaps, making assumptions, and waiting.
Let’s put that in the context of a typical project. A standard eLearning project runs 10–12 weeks. Each course on average includes an upfront SME interview plus two or three SME review cycles per course — each taking one to two weeks. In practice, this means we’re not losing hours to the SME bottleneck — we’re losing months.
The SME bottleneck isn’t a process inefficiency: it’s a structural tax on the entire function — and it's getting more expensive, not less, as everything else speeds up.
In this week’s post, I explore the answer to the question I get asked a lot: can AI help remove the SME bottleneck? I’ll share the five emerging methods that I’ve seen on the ground for tapping into expert knowledge with AI, the data on where each breaks.
Let’s dive in!
The SME Bottleneck: Access + Articulation
When I ran the State of Instructional Design survey with Synthesia (400+ practitioners globally) it became clear that the SME problem isn’t just an access problem, it’s also an articulation problem. SME delays came out as the second-largest barrier to speed — cited by 30% of instructional designers — and a barrier to quality for a further 13%. The median designer gets SME collaboration during only 38% of project time. Nearly half of us get less.
The access problem — i.e. getting a SME’s time for interview or review - is the problem we feel most powerfully, but it’s the shallower of the two problems we face.
The deeper problem is articulation. Even when we get the expert in the room, most of what makes them expert isn’t available to them for narration. Cognitive task analysis (CTA) literature — developed largely by surgical educators wrestling with exactly this problem — quantifies the size and shape of this challenge: in free-recall descriptions of procedures they’d performed hundreds of times, expert surgeons omitted an average of 71% of clinical knowledge steps, 51% of action steps, and 73% of decision steps (Sullivan et al., 2014).
TL;DR: the majority of what makes someone expert doesn’t get surfaced and captured out in a standard interview. And even when we apply the best available structured cognitive task interview methodologies, a whopping one-third of expert knowledge stays unspoken, in their heads (Sullivan et al., 2014).
This is the so-called Curse of Expertise. When we become expert in something, it becomes compiled and automatic — it exists and runs below conscious access. Ask a great clinician how they knew X and they’ll say something like, “it just looked wrong.” The cues, the rejected alternatives, the reasons they skipped step three: all of these things invisible to the expert performing them.
In practice, this means the standard SME interview mostly extracts only the explicit layer of expertise — the part that’s usually already written in a manual somewhere — and misses the tacit layer, which is the entire point of talking to a human.
When we talk about finding solutions to the SME problem, therefore, we need to measure every solution not just for the access it gives us to the SME’s expertise, but also how well it enables them to surface and share the expertise that we need.
AI + Knowledge Extraction: a short history
This isn’t AI’s first foray into the knowledge extraction problem. In the 1980s, the most hyped idea in artificial intelligence was the so-called expert system. The theory was this: we can use AI to capture a specialist’s knowledge as rules, encode it in software, then scale their judgement to everyone.
Stanford’s Edward Feigenbaum, the field’s leading figure, named the thing that would kill the expert system — the so-called knowledge acquisition bottleneck (Feigenbaum, 1984). The bottleneck he predicted was the one I just described: getting knowledge out of an expert's head and into a machine is far harder than anyone anticipates.
Feigenbaum was right. A cottage industry of "knowledge engineers" who sat with specialists one painstaking interview at a time was built. Months and months was spent on the dream of build “expert brains”, only for the whole approach to collapse. The cause of the collapse was three-fold:
The articulation problem — experts genuinely could not narrate their own reasoning. As Feigenbaum himself noted, an expert system is “expert only because of the knowledge it possesses” — but extracting that knowledge through dialogue proved fundamentally limited. Experts gave knowledge engineers the explicit surface layer; the tacit judgement stayed inaccessible.
Labour cost — the knowledge engineering process was slow and expensive, requiring specialist interviewers working one painstaking session at a time. This is the cost your draft currently names as the sole cause.
Brittleness and staleness — even when knowledge was captured, it was encoded as fixed rules that couldn’t adapt, couldn’t handle edge cases, and rotted as the domain evolved .
Meanwhile in L&D, we have our own history of trying to solve the SME problem which spans over forty years. The headline here is that our field has thrown almost everything at this problem in the hope of solving it. Here’s a brief history:
Hire dual-expert designers — 1960s–1970s onward. The earliest response; predates the modern instructional design field. Military and medical training programmes were building subject-specialist educators long before rapid authoring existed.
Innovate the elicitation — 1970s–1980s onward. Cognitive task analysis and DACUM both emerged in this period (DACUM from the late 1960s; CTA formalised through the 1970s–80s in military and surgical training contexts). This is the contemporaneous “proper” answer to the same problem Feigenbaum was naming.
Capture it in a system — 1990s–2000s. The knowledge management wave is clearly dated to this era — wikis, Lotus Notes, communities of practice, lessons-learned repositories.
Turn the SME into the designer — late 1990s–2000s. Articulate launched in 2002; Captivate in 2004. The “anyone can build a course” era is a product of that tooling wave.
Send the SME a template — 2000s–2010s. Storyboard templates became mainstream as rapid authoring scaled; the asynchronous-fill-in approach is a child of that same era, used when teams couldn’t afford full ID involvement.
Buy the content instead — 2010s–present. Off-the-shelf content libraries (LinkedIn Learning, Coursera for Business, Skillsoft etc.) scaled meaningfully from the 2010s, though the idea is older.
Line them up and you see that every initiative failed due to one of the three knowledge extraction failure modes:
The labour cost of capture — elicitation, KM, expert systems all required expensive human time to extract knowledge
Staleness — whatever got captured dated fast; knowledge encoded once rotted
Missing judgement — pedagogical or domain expertise either walked out the door or never made it in
Using LLMs for SME Knowledge Extraction
So will this wave of AI be different? A question I’m actively exploring with teams in both Higher Ed and corporate L&D is: will LLMs finally allow us to crack the knowledge extraction problem and remove the SME bottleneck?
If we look at the problem honestly and trace the failure modes of everything that came before, the answer is: perhaps.
Here’s my scorecard:
1. The Costs of Information Capture
LLMs — the large language models underpinning tools like ChatGPT, Claude, and every retrieval-grounded assistant you’ve ever built — genuinely collapse the labour cost of capture. They can read a thousand-page document corpus, conduct a structured elicitation interview, and synthesise a transcript overnight. Elicitation now scales.
2. The Staleness Problem
LLMs also plausibly fix staleness. Unlike a frozen rule-base that encoded the expert once and then rotted, a system grounded in a live, living record (i.e. one that monitors how we work) can be refreshed continuously. Knowledge that regenerates doesn’t decay the way a database does.
3. The Articulation Problem
LLMs do not touch the articulation problem — and this is the part that gets undersold. The 66% ceiling on expert knowledge capture isn’t a technology limitation: it’s a human one. Tacit knowledge doesn’t become expressible just because the tool asking the question is more sophisticated. And on top of this, there’s a new failure mode worth naming: where expert systems produced brittle, obviously wrong outputs, LLMs produce confident, fluent, plausible wrong outputs. The failure is harder to spot, not easier.
So, LLMs don't solve the SME problem entirely — but ambient AI might. Ambient AI refers to AI systems that capture expertise by watching experts work, rather than interviewing them about it. Instead of asking a radiologist to explain how they spot an anomaly, you track where their eyes move across thousands of scans — and the AI learns the perceptual pattern directly from the behaviour. Instead of asking a senior analyst to describe their decision process, you log the sequence of actions they take in a system and infer the logic from what they do and don't click. The knowledge is extracted from the work itself, not from a conversation about the work
This is the very leading edge of what’s possible — the vast majority of the ambient AI work is happening in medicine and software engineering right now, not in L&D. But the direction of travel is clear: we are moving, slowly, from a world where knowledge extraction requires a willing expert and a scheduled conversation, toward one where expertise leaves a trace in the work itself and AI learns to read it.
On the Ground: four ways L&D teams are using AI to reduce the SME bottleneck today
On the ground in 2026, there are four common AI use cases emerging which don’t solve but might help to mitigate the SME problem. Here’s the TL;DR:
Approach 1 — AI mines the conversation: record, transcribe & theme SME interviews
Where the knowledge comes from: the SME contact you’re already having.
A simple but effective move that a surprising number of teams still skip: record, transcribe and thematically code every minute of SME contact you get. LLMs now perform comparably to trained human coders on thematic analysis (SAGE, 2026), and ten minutes of processing replaces the selective, lossy notes that used to be a project’s single source of truth. Here’s the prompt, ready to steal:
“Here’s a transcript of an SME interview [paste] and my design brief [paste]. Code the conversation against the brief: (1) which objectives were covered and what the SME said about each, (2) every factual claim that needs verification, with the SME’s exact wording, (3) commitments made — who promised what, by when, (4) questions I asked that got vague or incomplete answers, (5) gaps — what the brief needs that this conversation didn’t cover. Format as a table I can work from.”
What it solves: processing loss on every contact.
Where it breaks: it solves the processing bottleneck, but not access or articulation — perfect capture of an incomplete account is still an incomplete account.
Approach 2 — AI interviews the SME: the async AI interviewer
Where the knowledge comes from: the expert themselves — interviewed by AI.
Picture this: instead of sending a SME a calendar invite, you send them a link to a GPT you’ve trained to run expert interviews on your behalf. At 3pm, between other commitments, the SME opens a conversation with an AI interviewer running a structured elicitation protocol — context, task walkthroughs, critical incidents, common failure points — that probes vague answers (”you said it depends — on what, specifically?”), runs a completeness check per section, and exports a structured document to you overnight.
Software requirements engineering has the same knowledge-extraction problem as we do. Research shows that LLM-based elicitation interviews can capture up to 73.7% of software requirements — specifically 60.94% fully elicited plus 12.76% partially elicited — with error rates comparable to trained human interviewers (LLMREI, 2025). Comparative studies find AI-led interviews are on average more structured and more time-efficient than human-led ones.
So in practice, LLM based interviewing partially solves not just the access but also articulation problem — because an AI interviewer applies the protocol consistently, never gets tired, never gets deferential, and never does the thing every one of us does under time pressure: accept a vague answer and move on. The convenience of this approach gets the most attention, but the consistency and quality of the output may matter more.
One question worth addressing directly, because I get asked it a LOT: will SMEs actually engage with an AI interviewer? The evidence is more encouraging than most practitioners expect. Research comparing AI and human interviewers finds that participants show comparable willingness to disclose, equivalent trust, and no meaningful increase in discomfort — the main difference is less warmth, not more resistance (Curtin University, 2025).
Some SMEs report feeling more candid with AI — less worried about appearing incompetent in front of a colleague or direct report. The bigger adoption risk isn't reluctance; it's the AI-literacy prerequisite. An SME who isn't comfortable with a chat interface won't engage well regardless of how good the protocol is — and that's a design and change management problem, not a technology one.
What it solves: scheduling friction; LD time; interview consistency.
Where it breaks: it’s still interview-based elicitation, so roughly one item in four is still missed. For compliance-adjacent or safety-critical domains, that’s a critical error rate. The access issue persists too: in this approach, the expert is still giving you roughly the same hours — the interview is automated, not removed.
Approach 3 — AI as research assistant: becoming a “junior SME”
Where the knowledge comes from: a published research base, e.g. industry reports, accessed via research platforms like Perplexity, Elicit and Consensus.
Another trend I see on the ground is L&D teams trying to become “partial experts” by working with AI to analyse and interrogate peer-reviewed literature on a topic before they ever meet the real expert. This approach tends not to be a replacement for the human but an augmentation of them which allows the designer to make progress “while they wait”.
In practice: you’re briefed to build training on, say, motivational interviewing, and you have no clinician access for three weeks. So you run a structured “interview” with a research-grounded AI — what do novices get wrong, what does the evidence say about skill decay, where do practitioners plateau — and walk into the eventual expert meeting with a v1 outline.
Research confirms LLMs can generate credible content skeletons grounded in published domain literature (Ullmann, 2024; Tian et al., 2023) — though accuracy for proprietary or emerging domains still requires deep expert review.
What it solves: the cold-start problem and no-access periods. The human SME also shifts from informing to reviewing and correcting: reacting to a draft rather than filling a blank page.
Where it breaks: LLMs trained on published literature reproduce common practice — not your organisation’s practice, and not the frontier. For proprietary or emerging domains, the virtual SME produces confident, plausible, wrong content. And the designers most likely to lean on it are the junior ones least equipped to spot the errors.
Approach 4 — AI analyses the SME’s outputs: building the virtual SME
Where the knowledge comes from: everything the expert has already produced in their day to day work — documents, decks, recorded talks, Slack / teams Messages.
In this approach, the L&D team ingest the expert’s existing output into a GPT or project, with the aim of creating a “virtual SME'“ they can consult on demand: ask it to walk you through the process, challenge it with edge cases, propose a wrong assumption and watch it push back with the relevant passage. A well-built version tells you when you’ve strayed beyond what the corpus covers, rather than improvising.
Again, the aim here typically isn’t to replace the human SME: it’s to do 80% of your thinking and design work before the real conversation, so the real conversation is sharper, shorter and more productive.
The value of this approach may also compound over time: a 2025 simulation study found that LLM agents could reconstruct what an organisation knows by piecing together fragments from across many colleagues — often without ever going back to the original expert. As a simulation, the output of any “AI SME” should be treated as a provocation rather than a finding. But the implication is striking: by the time expertise has spread through a team, it may already be recoverable from the network — not just from the person who first held it.
What it solves: scheduling dependency; live SME hours (roughly halved); a reusable asset.
Where it breaks: the brain only knows what was written down — it compresses and democratises the explicit layer, but the tacit layer was never in the documents, so it inherits the corpus’s blind spots. It also assumes a full and structured footprint: brilliant operators who’ve never written anything down offer nothing to harvest. And a static corpus captures what the expert used to think; currency still requires the human.
Closing Thoughts
For all the change happening in L&D right now — the AI tools, the faster workflows, the shrinking timelines — some of the challenges we face every day are remarkably, stubbornly age-old. The SME bottleneck isn’t new. It isn’t a failure of imagination or a gap in the tooling. It’s a structural problem with a forty-year research history, named by Stanford’s Edward Feigenbaum in 1984 and still unsolved. It’s still the first pain that surfaces, unprompted, every time I work with an L&D team.
Can AI fix it? Maybe — but not automatically, and not without understanding what we’re actually dealing with. The problem has three distinct layers: getting access to an expert’s time, getting them to articulate what they know, and keeping that knowledge current once captured. Most of the solutions being tried right now attack one layer convincingly while leaving the other two untouched. Some solve the visible part while making the invisible part worse — trading scheduling friction for a corpus that confidently reproduces the explicit layer and misses everything that actually makes the expert worth consulting.
Meanwhile, something else is starting to happen. A growing share of L&D practitioners are themselves subject matter experts — designers who’ve absorbed the role, enabled by the same tooling. The bottleneck is beginning to dissolve from the other side, which in turn flips the question: if the expert can now build the training, what does the designer bring? The answer, I’d argue, is the one thing that can’t yet be automated — pedagogical judgement. Knowing when accurate content isn’t the same as effective learning.
Happy experimenting!
Phil 👋
PS: Want to experiment with how AI might solve the SME problem in your context with me and a community of folks like you? Apply for a place on my bootcamp.
