ANALYSIS · LABOR ECONOMICS · OCCUPATIONAL RISK
Which Jobs Survive
AI displacement is not random. It follows structural patterns the literature now documents with enough precision to make useful predictions about which occupational categories are likely to persist and which are not.
AI LABS
FEB 27, 2026
ESTIMATED READ: 15 MIN
The question is almost always asked wrong. "Will AI take my job?" is the wrong unit of analysis. The right question is: "Does my job require the kinds of things AI systematically cannot do — and will that remain true as the capability curve moves forward?" The literature gives us enough to answer this with more precision than most people expect.
After reviewing 26 sources across AI research, macroeconomics, diffusion theory, and historical labor transitions, a nine-variable framework emerges. Jobs that score well across these variables are likely to survive the near-to-medium term transition. Jobs that score poorly are not.
The Nine-Variable Framework
These variables were introduced across our two-pass literature review. Each one independently predicts displacement risk. Together, they form a composite profile that is more predictive than any single dimension.
V1
AGE / EXPERIENCE STRATIFICATION
Displacement is not uniform across career stages. Entry-level roles in high-exposure occupations face 13–20% employment decline (Brynjolfsson ADP data). Experienced workers with tacit knowledge and client relationships are stable. Jobs that require years of accumulated judgment resist substitution better than jobs that can be learned from documentation.
SURVIVAL SIGNAL: Roles where seniority creates genuine capability differentiation
V2
AUTOMATING VS. AUGMENTING SPLIT
AI that replaces workflows drives displacement. AI that augments worker capability drives productivity with stable or growing employment. The key question: does this role require judgment that AI enhances, or execution that AI replaces? Customer service scripts → automating. Medical diagnosis support → augmenting. The same occupational category can split on this dimension.
SURVIVAL SIGNAL: Roles where AI tools make workers more effective, not redundant
V3
CAPABILITY-TO-IMPACT LAG
Humlum's Danish data shows a 2–4 year lag between AI adoption and measurable labor effects. This is not protection — it is a delay. Jobs in sectors currently experiencing adoption but not yet measurable displacement may be in the lag phase rather than the safe zone. The JOLTS professional services compression suggests the lag is ending for high-exposure knowledge work.
SURVIVAL SIGNAL: Sectors with no documented adoption, not just no documented effects yet
V4
PRODUCTIVITY INCENTIVE CEILING
Acemoglu's TFP ceiling (0.53–0.71% over 10 years) limits the economic incentive for broad AI adoption in augmenting applications. Employers pursuing AI primarily for cost reduction have stronger incentives for automating deployment. Jobs in sectors where the economic incentive is primarily augmentation (complex outputs, bespoke work) have structural protection that routine-output jobs do not.
SURVIVAL SIGNAL: Roles where output quality, not output volume, drives value
V5
WAGE-DISPLACEMENT PARADOX
Falling labor share in the economy may raise average wages for remaining workers even as employment contracts — particularly if current wages are below the wage-maximizing labor share. This means survivor jobs may pay significantly more. Jobs in sectors where labor share reduction and wage growth are compatible represent a viable long-term outcome.
SURVIVAL SIGNAL: High-skill, low-supply roles that benefit from scarcity effects
V6
S-CURVE NATURAL CEILING
Technology adoption saturates at 75–85% of addressable market. Jobs in the 15–25% of occupational territory that AI is structurally unlikely to reach — due to regulatory constraints, physical requirements, relational complexity, or extreme contextual variability — have durable protection regardless of capability improvements.
SURVIVAL SIGNAL: Roles that exist in the structural AI ceiling zone
V7
ABSORPTION SINK AVAILABILITY
Historically, displaced workers survived when a new sector absorbed them (agriculture → manufacturing, manufacturing → services). AI targets knowledge workers — the historical absorption sink. Surviving jobs must exist in sectors that are themselves absorbing displaced workers from AI-affected industries. Healthcare, skilled trades, and human-contact services are the primary candidates.
SURVIVAL SIGNAL: Roles in sectors actively absorbing displaced workers
V8
CREDENTIAL REVERSAL
Prior automation waves displaced low-credential workers; high-credential workers adapted. AI reverses this: it disproportionately targets high-credential knowledge work (legal research, financial analysis, coding, writing). Historical recovery models — "get more education" — are structurally inapplicable when it is the educated workers who face displacement. Jobs in physically-embodied, legally-regulated, or socially-complex spaces have credential protection AI cannot replicate.
SURVIVAL SIGNAL: Roles where credentials gate physical, legal, or social access AI cannot obtain
V9
ENTERPRISE EXECUTION GAP
75% of AI initiatives fail to deliver expected ROI. Organizations struggle to execute the human change management required for large-scale AI deployment. Jobs in organizations with low AI maturity, weak change management capacity, or strong regulatory friction have practical protection from the execution gap — not from AI capability itself, but from organizational inability to deploy it effectively.
SURVIVAL SIGNAL: Roles in high-friction deployment environments (healthcare, government, regulated industries)
Occupational Risk Assessment
Applying the nine variables to occupational categories produces a risk profile that diverges significantly from the simple "routine vs. non-routine" framing that dominated early AI displacement analysis. The results are counterintuitive in several places.
SOFTWARE DEVELOPMENT (ENTRY)
Junior developers, code review, documentation, testing
VERY HIGH
2025–2028
V1 (no seniority moat), V8 (credential reversal)
SOFTWARE DEVELOPMENT (SENIOR)
Architecture, complex problem-solving, cross-functional integration
MEDIUM
2029–2034
V1 (seniority moat), V3 (lag protection)
LEGAL SUPPORT
Paralegals, legal research, document review, contract drafting
HIGH
2026–2030
V2 (automating), V8 (credential reversal)
LEGAL PRACTICE (SENIOR)
Trial attorneys, complex negotiation, client counsel
LOW-MEDIUM
2034+
V6 (relational ceiling), V9 (regulatory friction)
FINANCIAL ANALYSIS
Research analysts, report writing, data processing, modeling
HIGH
2026–2030
V2 (automating), V3 (short lag)
ADMIN & OFFICE SUPPORT
Secretaries, data entry, scheduling, customer service scripts
VERY HIGH
2025–2028
V2 (fully automating), BLS confirmed
HEALTHCARE (CLINICAL)
Physicians, nurses, physical therapists, hands-on care
LOW
2035+
V6 (physical ceiling), V7 (absorption sink), V9 (regulatory friction)
HEALTHCARE (ADMIN)
Medical billing, scheduling, insurance verification, records
HIGH
2026–2030
V2 (automating), contradicts sector narrative
SKILLED TRADES
Electricians, plumbers, HVAC, construction, welding
LOW
2035+
V6 (physical ceiling), V7 (absorption sink)
MANAGEMENT (MID-LEVEL)
Project managers, team leads, coordination roles
MEDIUM-HIGH
2028–2033
V2 (coordination automating), V9 (org friction)
EDUCATION (K-12)
Teachers, classroom instruction, student relationship
LOW-MEDIUM
2033+
V6 (relational ceiling), V9 (public sector friction)
CONTENT CREATION (ROUTINE)
Copy, reports, summaries, templated journalism
VERY HIGH
NOW
V2 (fully automating), already occurring
The Jobs That Survive: Common Factors
The literature converges on a set of properties that predict occupational survival. They are not about education level, which is the most important departure from historical analysis.
HIGH SURVIVAL PROBABILITY
Physical Embodiment + Contextual Variability
Skilled trades represent the most robust survival profile in the model. Every call-out is different. The environment is not digitally modeled. The work requires physical presence in variable conditions, integrating sensory information AI cannot receive. The S-curve ceiling (V6) is structural, not temporary. These roles also occupy the absorption sink position (V7): they are growing as knowledge workers are displaced, not declining alongside them. BLS projects electricians (+10.8%) and construction laborers (+8.2%) to grow faster than average through 2034.
PHYSICAL CEILING
ABSORPTION SINK
HIGH CONTEXTUAL VARIABILITY
LOW AUTOMATION INCENTIVE
HIGH SURVIVAL PROBABILITY
Clinical Healthcare — The Structural Exception
Clinical healthcare survives not because AI cannot assist in diagnosis — it can, and does — but because the delivery of care remains physically, legally, and ethically bound to human presence. The regulatory friction (V9) is extraordinary: FDA approval pathways for AI clinical tools are measured in years, not months. The relational dimension of care has documented clinical outcomes. And demographic demand (aging population) is expanding the sector faster than AI can contract it. The critical distinction: healthcare admin is not protected by any of these factors and faces very high displacement risk.
REGULATORY MOAT
DEMOGRAPHIC DEMAND
PHYSICAL PRESENCE REQUIRED
RELATIONAL CLINICAL VALUE
MEDIUM SURVIVAL PROBABILITY
Senior Knowledge Work — The Seniority Moat
The V1 finding is decisive here: displacement is disproportionately concentrated at entry level. Senior practitioners in law, medicine, finance, and engineering maintain a seniority moat built from client relationships, institutional knowledge, and judgment accumulated over careers. This moat is real but temporary — it does not protect against the next generation of AI that trains on the outputs of those senior practitioners. The question is not whether the moat exists (it does) but how long it holds.
SENIORITY MOAT (V1)
TACIT KNOWLEDGE
CLIENT RELATIONSHIPS
SHRINKING OVER TIME
HIGH DISPLACEMENT RISK
Entry-Level Knowledge Work — The Credential Trap
This is the most important finding of the Pass II literature. Prior automation waves displaced low-education workers; high-education workers were the absorption sink. AI inverts this: it disproportionately targets the tasks performed by entry-level knowledge workers — research, drafting, analysis, review, documentation — because these tasks are well-defined enough for AI to execute but require enough intelligence that only educated workers could previously perform them. The credential reversal means historical recovery mechanisms ("retrain for a better job") are structurally inapplicable. There is no better job to retrain for when the educated knowledge economy is the sector under displacement pressure.
CREDENTIAL REVERSAL (V8)
AUTOMATING DEPLOYMENT (V2)
NO ABSORPTION SINK
BLS CONFIRMED
The Question the Framework Cannot Answer
A framework that tells you which jobs survive today cannot tell you whether the surviving jobs will remain intact tomorrow.
The recursive improvement scenario — covered in detail in The Recursive Displacement Loop — introduces a caveat that no occupational survival framework can adequately address. If AI capability growth is self-reinforcing (AI systems that improve AI research, that improve AI systems), then today's survival factors may not be tomorrow's.
Physical embodiment is currently structural protection. But robotics and sensor capability are improving. The S-curve ceiling of 75–85% addressable market may itself shift as the addressable market expands into physical domains.
The framework above describes a relatively near-term horizon (2025–2034). For longer horizons, the honest answer is that the recursive scenario makes occupational survival prediction structurally unreliable — not because we cannot analyze the factors, but because the capability assumptions underlying the analysis may themselves change.
A note on individual career planning: This analysis describes aggregate occupational risk at the sector and role-category level. Individual outcomes within any category vary enormously based on employer, geography, specific skills, relationships, and adaptability. The goal of this framework is not to predict any individual's outcome — it is to identify where the structural pressures are concentrated so that individuals, organizations, and policymakers can make better-informed decisions.