The 61-Point Gap: Why AI Can Do Your Job But Isn't Yet

What the gap is

Theoretical exposure answers: if AI could perform all the tasks in a given role, what percentage of those tasks would fall within AI's capability? This is a structural assessment of what the technology can do, based on task composition analysis from Eloundou et al. (2023) at OpenAI, later extended in Anthropic's own research. A 94% theoretical exposure for ICT professionals means that almost every task in a software developer's or systems analyst's job - writing code, reviewing documentation, debugging, writing tests, summarising requirements - is technically within AI's capability as of 2026.

Observed exposure answers something different: what percentage of tasks in this role are actually being performed by AI right now, based on real usage data? Anthropic measured this using its own systems - looking at how Claude and similar tools are actually being used in practice across occupation groups. A 33% observed rate for ICT professionals means that while AI can theoretically do almost everything, it is currently doing about a third. The gap - 61 points - is the space between what is possible and what is actually deployed. That gap is where your runway lives.

The 8-point gap: what disruption happening now looks like

Customer service clerks (ISCO 42) have a theoretical exposure of 78% and an observed exposure of 70%. The gap is just 8 points - and it shows up in the real economy right now, not in projections.

What 70% observed exposure means in practice: automated chat handles first contact for most major consumer brands. AI drafts responses for human agents to review and send. Call routing is largely automated. Tier-1 queries - account balance checks, order status, standard complaints - are handled end-to-end by AI systems without human involvement. The human customer service agent is left with escalations, complex complaints, and emotionally charged situations that AI still handles poorly. Those situations are real - but they are a shrinking share of the total workload as AI capability in emotional and nuanced conversation continues to improve.

The 27-point gap: general clerks still have runway

General and keyboard clerks (ISCO 41) sit at 82% theoretical and 55% observed - a 27-point gap. Larger than customer clerks, which means less immediate disruption, but the direction is the same.

Why is the gap 27 points and not 8? Data entry is largely automated in organisations that have adopted modern systems, but document processing still involves enough variability that human judgment remains in the loop. Scheduling has been substantially automated, but coordination in complex or sensitive contexts - managing a senior executive's calendar, negotiating meeting times across organisations - still involves human judgment that AI handles inconsistently. Correspondence AI is improving rapidly: drafting emails, summarising documents, routing information. This gap will narrow. The 27-point cushion is real, but it is not stable.

The 40-61 point gaps: professionals still have time

Three professional groups show the largest gaps in Anthropic's 2026 data:

• ICT professionals (ISCO 25): 94% theoretical, 33% observed - a 61-point gap
• Business and administration professionals (ISCO 24): 85% theoretical, 45% observed - a 40-point gap
• Business associate professionals (ISCO 33): 72% theoretical, 28% observed - a 44-point gap

Why are these gaps so large? Four factors combine. First, compliance review: AI-generated code, analysis, and financial models require expert validation before they can be deployed or acted on. The cost of a wrong AI output in professional contexts is high enough that removing the human reviewer is not yet economically rational. Second, enterprise procurement timelines: adopting AI tools in large organisations requires security review, legal review, vendor assessment, and integration work that takes 12 to 24 months. The technology is available; the institutional deployment is not. Third, the liability question: when AI-generated code causes a system failure, or AI-generated analysis leads to a wrong business decision, who is responsible? Until liability frameworks are clearer, organisations are cautious about removing the human from the loop. Fourth, the difficulty of replacing senior judgment: a junior developer might be largely replaced by AI-assisted output reviewed by a senior engineer. But that senior engineer's role is not eliminated - it is changed. The judgment work that remains requires exactly the expertise that is hardest to replicate.

Health professionals: regulatory friction as protection

Health professionals (ISCO 22) show a different pattern: 38% theoretical exposure and 12% observed - a 26-point gap, but from a much lower theoretical base. The low theoretical exposure reflects genuine structural limits on AI in clinical work. Patient interaction, physical examination, procedural skill, and the integration of ambiguous signals into a clinical judgment are tasks where AI's theoretical capability is genuinely limited - not just slow to deploy.

Regulatory friction adds another layer. Diagnostic AI exists and is effective in specific domains. Radiology AI is FDA-approved for certain applications and demonstrably accurate. But clinical deployment requires navigating malpractice liability, patient consent frameworks, institutional credentialing, and clinical governance processes that move slowly by design. This is not permanent protection. The FDA's Software as a Medical Device framework is evolving. Radiology AI is already in routine clinical use at major centres. The friction will yield - but it creates a longer runway than most professional groups enjoy.

What this means: timing, not destiny

The gap tells you when, not if. No role in Anthropic's 2026 data shows a stable, non-narrowing gap. The direction for every occupation group is toward observed exposure converging with theoretical exposure over time. What varies is how long that convergence takes - and that depends on the specific failure modes AI needs to overcome before deployment becomes rational.

Source: Massenkoff, M. and McCrory, E. (2026), Labor Market Impacts, Anthropic Research, March 5, 2026. Timing estimates are editorial inference from the gap size and known deployment friction - they are not stated in the original paper.

Roles with gaps below 10 are in active disruption now. That is not speculation - it is reflected in employment data for customer service roles in markets where AI adoption has been fastest. Roles with gaps of 40 to 60 points have years, not decades. The friction is real, the runway is meaningful, and the direction is clear. The most useful question is not "will AI affect my job?" It is "how long do I have to adapt, and what does adaptation look like for my specific role?"