Cohort Analysis Report · Instructional Design & eLearning Development · August 2026
Where your tasks, tools and skills sit against AI substitution pressure — and what separates the practitioners whose value compounds from the ones whose value quietly erodes.
There is a version of this report that opens with "adapt or be unemployed." It would be more dramatic and less true, and a cohort that checks the numbers would stop trusting the rest of it. Here is the accurate version, which is more useful and, in one specific way, more alarming.
In a survey of instructional design hiring managers, 89.2% said AI is unlikely to reduce the size of their learning team, even though 92.1% said AI would impact that team within twelve months. The US Bureau of Labor Statistics projects 11% employment growth for training and development specialists from 2024 to 2034 — faster than average — with roughly 43,900 openings a year. PwC found headcount growth at the most AI-exposed companies outpacing the least exposed, 52% against 36% since 2018.
If you tell a cohort they are about to be fired and then they read those numbers, you have lost them. The field is not contracting.
PwC's 2026 Global AI Jobs Barometer, built on more than a billion job ads across 27 territories, found the wage premium for AI skills at 62%, up from 57% the previous year and roughly 25% two years before. Postings requiring AI skills grew 69% against 9% for the market overall. PwC describes a labour market that is no longer adjusting to AI but dividing along it.
The risk is not that your job disappears. It is that the job keeps existing, is done by someone else, and pays that person substantially more than it pays you. That is a slower story and a harder one to notice from inside.
PwC's US analysis found AI-exposed entry-level roles are seven times more likely to require traditionally senior-level skills such as judgement and leadership. Those roles grew 35% since 2019 while other entry-level roles declined 10%. Translated to this field: the "take this SME deck and make it a module" job is consolidating, and the ladder's bottom rung has moved up. If your entire value is manual assembly — converting finished storyboards into screens — you are in the most exposed group in the profession, and that is a direct read of the evidence rather than a scare tactic.
Every chart is interactive: hover or tab to any marker for the reasoning behind its position. Section 02 contains a self-assessment that plots each participant onto the master quadrant with a personalised verdict — run that individually first, then compare placements as a group. Sections 03 to 06 work well as breakout stations, one chart per group, each reporting back on which items surprised them. Section 09 is the exit artifact: each participant leaves with a dated 90-day commitment.
The most productive facilitation move is to ask people to argue with a marker's position rather than accept the chart. The disagreements are where the learning is.
Published figures from PwC, the US Bureau of Labor Statistics, and a survey of instructional design hiring managers. Everything else in this report is interpretation built on top of these.
Dr Philippa Hardman ran a study in which roughly 200 instructional designers blind-scored work produced three ways: an experienced ID working alone, a novice using AI, and an experienced ID using AI. On tasks such as writing learning objectives, the experienced designer plus AI came out on top. This is the single most important finding for a cohort to internalise, because it contradicts both comfortable stories at once. Novice-plus-AI did not beat expertise. And expertise alone did not beat expertise-plus-AI. The two multiply. As the i4cp analyst Tom Stone put it, instructional designers who use AI will replace instructional designers who do not.
Sixteen role archetypes positioned on the two axes that actually determine career outcome. Note that the two axes are independent: being excellent at one does not protect you if you are weak at the other. The two lower quadrants are both losing positions, for opposite reasons.
Not "do you use ChatGPT." Composite of: routine delegation of real work to AI, ability to build repeatable systems rather than one-off prompts, agent and multi-step workflow design, rigorous evaluation of returned output, and integration of AI into tools other people on your team then use.
Composite of: diagnosing whether training is even the right intervention, instructional strategy selection and defence, stakeholder negotiation and pushback, assessment and measurement design, and ownership of a business outcome rather than a deliverable. Every one of these is a decision, not an artifact.
Low AI fluency, low judgement contribution. Order-taker production work, manual module assembly, template filling. This is the group the harsh framing is actually about. Drift here is downward: the tasks get absorbed and nothing replaces them.
Fluent with AI, thin on learning science and business judgement. Fast, cheap, and replaceable by the next person who learns the same tools — which is everyone, quickly. AI fluency alone is a commodity with a short half-life.
Deep craft and real judgement, no AI leverage. Not obsolete — genuinely valuable — but being out-produced by peers with the same judgement and ten times the throughput. Squeezed on volume, not on quality. The slowest and most deceptive decline.
Directs AI across many tasks at once while owning the diagnosis, the strategy and the business result. Production capacity multiplies; judgement becomes the scarce input. This is the only position where AI improvement makes you more valuable rather than less.
Answer honestly rather than aspirationally — the value of this is entirely in the accuracy. Your marker appears on the chart above.
Scores are self-reported and unvalidated. The point is not precision — it is that most people place themselves one quadrant higher than their honest answers put them, and seeing the gap is the exercise.
Thirty-three tasks from the working week of an instructional designer or eLearning developer. Horizontal axis: how much of this task current AI already performs competently. Vertical axis: how much value a skilled human still adds after AI has done its pass. The combination, not either axis alone, tells you what to do.
Voice-over recording, alt text at scale, quiz bank generation, localisation, template module builds, slide assembly from a finished storyboard. AI does these well and there is little residual premium for doing them by hand. These are no longer the basis of a career. They remain in the job; they stop being what you are paid for.
Learning objectives, storyboard drafts, SME transcript synthesis, data analysis, scenario first-passes. AI produces the volume; expert direction and review decide whether it is any good. This is where the Hardman finding lives, and where most of the 62% premium is earned.
Deciding training is not the answer. Performance diagnosis. Stakeholder negotiation and pushback. Defining what competence actually looks like. Owning a measured business outcome. Every one is a decision rather than a deliverable, and decisions are what AI cannot take from you.
Tasks that did not exist in 2022: agent workflow design, AI output review at scale, conversation-simulation authoring, prompt and context system design, AI governance for learning content, human-in-the-loop process design. Small share of the market today, disproportionate share of the differentiation.
| Task | Exposure | What AI does today | What you still own |
|---|---|---|---|
| Topic research | High | Summarises sources, surfaces concepts, compiles findings in minutes. | Verifying accuracy and choosing what actually matters for this audience. |
| Course outlines | High | Generates a solid draft outline from a brief. | Sequencing against the real performance gap and cutting what does not serve it. |
| Storyboards & scripts | High | Usable first drafts of screens, narration and dialogue. | Voice, nuance, and whether the scenario reflects how the job actually works. |
| Quiz & assessment generation | High | Drafts question banks in any format from source content. | Deciding what evidence of competence actually looks like. |
| Alt text & captions | High | Drafts accessibility text at scale. | Accuracy review and meaningful description of complex visuals. |
| Module assembly from storyboard | High | AI-assisted authoring absorbs the build step quickly. | Interaction quality judgement — the most exposed developer task in the field. |
| Image & media generation | Medium | Strong for scene-setting and mood imagery. | Depicting specific processes and procedures accurately; licensing and brand fit. |
| SME interviews | Medium | Drafts question lists, summarises transcripts. | The conversation itself — trust, follow-ups, reading what is not said. |
| Analysis & measurement | Medium | Crunches the numbers and drafts the report. | Deciding what to measure and defending it to leadership. |
| Needs analysis | Low | Organises data, drafts analysis documents. | Diagnosing whether training is even the answer. |
| Instructional strategy | Low | Suggests options from a brief. | Selecting and defending the right approach for the constraint set. |
| Stakeholder management | Low | Drafts emails and summaries. | Negotiation, pushback, alignment and trust. |
| Agent workflow design | New | Nothing — this is the work of directing AI itself. | All of it. The widest current gap between practitioners and organisations. |
| AI output review at scale | New | Cannot reliably judge its own quality. | All of it. Skip this step and the workflow produces confident slop. |
| Conversation simulation design | New | Generates working role-play simulations from a plain-English scenario. | Whether the simulation is realistic, safe, and aimed at the right behaviour. |
Exposure ratings for the first twelve rows follow a published task-by-task analysis of the ID role; the final three are new categories added for this report. The chart above plots all thirty-three tasks.
Forty-five tools plotted on adoption against career leverage. The two are not the same thing, and confusing them is the most common learning-plan mistake in this field. Filter by category to reduce density.
Agentic coding assistants, tool-connection protocols, retrieval stacks, workflow automation platforms, xAPI and learning record stores. Barely adopted in L&D, disproportionately valuable. This quadrant is where a two-year head start is still available — and it will not stay available. Pick one, not five.
Storyline, Rise, Camtasia, Canva, the mainstream chat assistants used casually. Expected on every job description and therefore differentiating on none. The industry-standard authoring suite remains genuinely necessary — it just stopped being an advantage some time ago. Maintain, don't invest.
A frontier assistant used with real depth, project and context systems, an AI video or voice platform, and a data tool. Widely available and still rewarded because most people use them shallowly. The gap between casual and expert use of the same tool is larger than the gap between tools.
The authoring tool market has split in two. Established suites are adding AI to existing manual workflows; AI-native entrants generate a draft course from a brief or an uploaded document in minutes. Both directions are real and both are growing. The strategic point for a practitioner is that the AI-native tools compress the exact production step that used to justify a headcount, while leaving the design judgement entirely with you. As one comparison put it, the AI makes the mechanics easy; the design thinking is still yours. Learn a generator, then be the person who can tell whether its output is worth shipping.
Twenty-three AI disciplines plotted on how quickly a non-technical practitioner can reach competence against how much career differentiation the skill buys. Nobody learns all of these. The chart exists to make the sequencing obvious.
Start upper-right: AI output evaluation, teaching AI literacy to others, context engineering, human-in-the-loop process design. These are reachable in weeks and they compound immediately, because they are judgement skills wearing technical clothing.
Then move up-left deliberately, choosing one: agent orchestration, retrieval design, evaluation frameworks, or conversation-simulation design. These take months and they are where the durable separation happens. Attempting all four produces four shallow competencies and no differentiation.
Basic prompting sits far right and only moderately high — easy to learn, and by 2029 close to worthless as a differentiator because everybody will have it. It is the single most over-invested skill in the field's current training offerings.
Toggle to "with 2029 shift" to see the drift: several disciplines that differentiate today become table stakes within three years. Anything that becomes universally held stops commanding a premium, whatever it once cost you to learn. Plan for the drift, not the snapshot.
Twenty-four instructional design models, frameworks and craft skills, plotted on defensibility against AI substitution and projected demand growth to 2031. Not all of your training holds its value equally — and the parts that hold up best are rarely the parts taught first.
Performance diagnosis, facilitation and stakeholder consulting, evaluation and ROI frameworks, behaviour-change models, scenario and simulation craft, transfer of learning. These are judgement applied to a specific human context. AI can describe every one of them fluently and cannot perform any of them for you.
Naming the phases of ADDIE. Reciting Bloom's levels. Writing a Mager-format objective. Listing Gagné's nine events. A model any AI can reproduce perfectly is no longer evidence of your expertise — it is evidence of a search query. Knowing the taxonomy is table stakes; knowing when the taxonomy is the wrong tool is not.
Your ID education is not devalued. Its centre of gravity has moved from the parts that produce artifacts to the parts that make decisions. The same degree, weighted differently. Practitioners who feel their training is now worthless are usually describing the production half of it and forgetting they were taught the other half too.
Segmentation matters more than a single verdict. A blanket "you're safe" or "you're finished" would be dishonest to a mixed cohort.
| Role today | Exposure | Becomes | What determines which side you land on |
|---|---|---|---|
| Order-taker production ID "Take this deck, make it a module" | Highest | Consolidated. Fewer roles, each covering more scope. | Whether you move up into scoping and diagnosis before the consolidation reaches you. This role has the shortest runway of any in the field. |
| eLearning developer Manual build from finished storyboards | High | Either an AI-assisted build lead, or displaced. | Pure assembly value is the most exposed thing in the profession. Developers who direct AI, integrate simulation and practice layers, and own quality judgement stay valuable. |
| Rapid authoring specialist | High | Absorbed into AI-native authoring workflows. | Template speed was the differentiator and generators are faster. Move toward interaction design and interactive quality review. |
| Multimedia / video producer | High | Creative director over generated assets. | Generation is cheap; accuracy, licensing, brand fit and taste are not. The role survives at a higher level of abstraction or not at all. |
| LMS / LXP administrator | Medium | Learning systems and data owner. | Configuration is automating. Instrumentation, xAPI, data quality and reporting are not — and they sit adjacent to measurement, which is growing. |
| Full-cycle instructional designer | Medium | Transformed, not replaced. | The production middle of your week automates. Whether the judgement layers expand to fill that space, or the role just shrinks, is largely your choice. |
| Learning experience designer | Medium | Experience architect across AI-generated components. | Design taste holds value. Add orchestration and the role compounds; stay analog and you get out-produced by peers with the same taste. |
| Freelance / contract ID | Medium | Sharply bifurcated market. | Rate pressure from AI-native competitors at the low end, expanding opportunity at the diagnostic end. The middle of the freelance market is thinning fastest. |
| Simulation & scenario designer | Growing | In demand and expanding. | AI makes previously impractical experiences — real conversation practice, branching role-play — cheap to build. Someone has to design them well, and AI cannot judge its own realism. |
| Learning analytics / measurement | Growing | Increasingly central. | When production gets cheap, proving impact becomes the scarce skill. This is the safest specialisation in the field on current evidence. |
| Performance consultant | Growing | The most defensible position available. | Diagnosing the real problem and refusing the wrong solution. AI makes your recommendations cheaper to execute, which makes your judgement scarcer by comparison. |
| AI learning architect | Emerging | New role, forming now. | Designs the agent workflows, evaluation standards and governance for a learning function. Barely exists as a job title; the practitioners who define it will set its price. |
| Aspiring ID / career changer | Raised bar | Harder entry, cleaner start. | Entry-level now demands senior-level judgement seven times more often. The compensating advantage is real: you can learn AI-native workflows without unlearning a decade of production habits. |
Three windows. The near horizon is observation, the middle is extrapolation, the far one is genuinely uncertain and labelled as such.
Nobody has a reliable five-year forecast for this and you should distrust anyone selling one. The 2030 column exists to make one point: every scenario in it rewards the same four capabilities — directing AI, designing practice, proving impact, and diagnosing problems. That convergence is the useful signal. If four quite different futures all reward the same skills, those skills are a safe bet regardless of which future arrives.
The exit artifact for a cohort session. Each participant should leave with dates written against the first block.
Log your hours against the six task buckets in Section 03 for two full weeks. Most practitioners discover that 60–70% of their time sits in the absorbed or amplified columns, and they had not noticed. You cannot plan a move without knowing your starting position, and self-estimates on this are reliably wrong.
Take your single most repetitive production task and build a durable workflow for it — a project with proper context and instructions, a documented review rubric, and a defined handoff. The deliverable is that a colleague can run it without you. Casual prompting is not the skill being rewarded; systems are.
Attach a business-outcome measure to a live project, agreed with a stakeholder in advance. Separately, find one training request that will not solve the underlying problem and propose the alternative. These two moves do more for your position on the vertical axis than any course, and they generate portfolio evidence that AI fluency alone cannot.
From the upper-left of the tools chart and the upper-left of the disciplines chart, choose one and go deep: agent orchestration, retrieval and knowledge design, evaluation frameworks, conversation-simulation design, or learning data instrumentation. Depth in one is worth more than familiarity with five, because familiarity is what everybody will have.
In parallel, build the teaching habit. Running an internal session on AI practice for your own team is the highest-return activity available to most practitioners: it forces your understanding into a defensible shape, and it makes you visible as the person who moved the team rather than the person who used a tool.
Own an evaluation standard, an AI governance policy for learning content, or the measurement function for a programme area. The pattern in every growing role in Section 07 is the same: the work moves from producing artifacts to defining what good looks like and holding the line on it.
One check to run annually: if AI got twice as capable tomorrow, would your role become more valuable or less? In the upper-right quadrant the honest answer is "more," because cheaper production makes your judgement scarcer. If your answer is "less," you have located your problem, and you have found it early enough to act on it.
The figures in Section 01 come from PwC's 2026 Global AI Jobs Barometer (more than one billion job advertisements across 27 territories), the US Bureau of Labor Statistics occupational projections for training and development specialists, and a published survey of instructional design hiring managers. The task exposure ratings in the Section 03 table follow a published task-by-task analysis of the ID role. The replacement experiment is Dr Philippa Hardman's blind-scoring study with approximately 200 instructional designers.
One number deliberately absent: there is no verified AI wage premium specific to instructional design. PwC's 62% is a cross-role average spanning every AI-exposed occupation in 27 countries. Any source quoting an ID-specific premium is inventing it, and a cohort should be told so explicitly.
All 140-plus plotted positions, all quadrant boundaries, the 2029 and 2031 drift vectors, the role transformation table and the self-assessment scoring are analyst-assigned. They are constructed to be internally consistent and they are opinions, held with varying confidence. Tool positions in particular will age fastest — this market moves in months.
Scoring rubric, all axes, 0–100: 0–24 absent or negligible · 25–44 present but limited · 45–59 typical for a competent practitioner today · 60–79 clear competitive advantage · 80–100 rare, defensible strength. Treat the charts as a structured argument to disagree with, not a measurement to accept.
Survivorship in the source material. Much of the best writing on AI and instructional design comes from people who sell AI training, AI tools, or ID education. That does not make their data wrong — the PwC and BLS figures are independent — but it does mean the optimistic framing is commercially convenient for its authors. This report is built on the same sources and the same caution applies to it.
Hiring-manager surveys measure intent, not outcome. "We do not expect AI to shrink the team" is a statement of current expectation from people who may not control the budget decision. It is meaningful evidence and it is not a guarantee. The BLS projection is the more independent signal, and it is a projection too.