AI and Jobs — 80% of Workers Have Tasks Exposed── Viewed task by task rather than occupation by occupation, AI replaces some tasks and augments the rest rather than eliminating whole jobs

01The "vanishing occupation" framing is too coarse

Discussion of AI and employment tends toward binary predictions: this occupation will disappear, that one will survive. A 2023 study by Eloundou, Manning, Mishkin and Rock moved the unit of analysis from occupations to tasks. They assessed every occupation in the U.S. labor market by measuring how much each constituent task could be accelerated by a large language model. The result: about 80% of U.S. workers are exposed on at least 10% of their tasks, and roughly 19% are exposed on 50% or more.

The implication is that few occupations become entirely redundant, yet most occupations see a portion of their tasks shift to AI. The effect shows up not as extinction or survival but as a rearrangement of job content.

02Task exposure spans all wage levels

A striking feature of the Eloundou study is that task exposure does not track neatly with wages. Earlier waves of automation tended to displace mid-wage, routine tasks — assembly-line work, clerical data entry. LLM exposure is different. It reaches into high-wage, cognitive tasks such as legal drafting, financial analysis, translation and code generation.

The IMF's January 2024 staff discussion note extended this analysis internationally. It estimated that about 40% of global employment is exposed to AI, with rates of 60% in advanced economies, 40% in emerging markets and 26% in low-income countries. The higher rate in advanced economies reflects the larger share of cognitive tasks in those labor markets.

03Substitution and complementarity split within the same job

Not every exposed task is one that AI performs instead of a person. The IMF estimated that roughly half of exposed jobs face a complementarity opportunity — AI augments the worker, raising productivity and potentially wages — while the other half face a substitution risk. Acemoglu, in a 2024 paper, took a more conservative view of the automatable share, projecting a GDP increase of 1.1% to 1.6% over ten years. That figure rests on the premise that only a fraction of tasks can actually be automated at acceptable quality.

The substitution-complementarity boundary runs through individual jobs, not between them. Consider a regulatory affairs specialist in pharma. Drafting submission documents is a task that AI can accelerate substantially through first-draft generation. Interpreting an agency's review posture and constructing a filing strategy requires contextual judgment and interpersonal trust — tasks where AI remains an aid, not a replacement.

04Japan's labor market carries two structural conditions others do not

Applying the task-exposure framework to Japan requires acknowledging two features that distinguish it from most advanced economies. First, demographics. In 2023 the share of the population aged 65 and over stood at 29%, the highest in the world. The working-age population fell 16% from a peak of 87.3 million in 1995 to 73.7 million in 2024. In a market where labor supply is contracting, task substitution by AI is less a threat to employment and more a means of maintaining output with fewer workers.

Second, adoption speed. The OECD's 2024 report on AI and the Japanese labor market noted that Japanese workers show lower AI exposure relative to peers in other advanced economies. Among Japanese AI users, 34.7% reported engaging in reskilling or upskilling to work with AI, and middle-aged and older workers reported lower rates of performance improvement from AI use than their counterparts elsewhere. Population decline creates demand for AI; adoption and reskilling pace constrain supply.

05Three pharmaceutical roles are reshaped task by task

Task-level analysis applies directly to three core pharmaceutical roles.

Medical representatives (MRs) perform a bundle of tasks: information delivery to physicians, call preparation, CRM entry, internal reporting. Information retrieval, summarization and report drafting are tasks AI can accelerate. Building physician trust and tailoring proposals to individual prescribing patterns remain interpersonal, judgment-intensive tasks that AI augments rather than replaces. Japan's MR headcount peaked at 65,752 in fiscal 2013 and has declined for ten consecutive years, reaching 46,179 at the end of fiscal 2023. The decline predates generative AI — channel diversification is the primary driver — but AI will accelerate the reorganization.

Regulatory affairs specialists draft submissions, monitor regulatory intelligence and prepare responses to agency queries. First-draft generation and cross-country regulatory scanning are substitutable tasks. Reading an agency's review posture and shaping a filing strategy demand institutional knowledge and remain in the complementarity zone.

Medical affairs professionals conduct literature reviews, draft medical-inquiry responses and identify key opinion leaders. AI speeds the extraction of findings from large bodies of literature. Clinical-trial design advice and long-term academic relationships with KOLs depend on expert judgment and interpersonal engagement, keeping them in the augmentation category.

06Three conditions determine whether task substitution becomes productivity

Substituting a task with AI does not automatically raise organizational productivity. A 2023 study by Brynjolfsson, Li and Raymond measured the effect of deploying a generative AI tool in customer support: issues resolved per hour rose 14% on average, and 34% among less experienced agents. The study points to three conditions that must hold.

First, time freed by substitution must be redirected to higher-value activities. If a report that once took two hours now takes thirty minutes, the remaining ninety minutes need a defined destination — more physician calls, deeper case preparation — or the gain evaporates. Second, the organization must be able to verify AI output. In pharma, regulatory submissions and promotional materials carry legal liability; unchecked AI-generated text is a compliance risk, not a productivity gain. Third, workflows must be redesigned around the new task allocation. Inserting AI into an unchanged process captures only a fraction of the potential; rethinking task sequences and role boundaries captures more.

07Three steps a pharmaceutical company can take now

Applying the task framework inside an organization follows a practical sequence.

First, decompose each role into its constituent tasks. For every role — MR, regulatory, medical affairs, pharmacovigilance — list the tasks: drafting, searching, summarizing, negotiating, advising, recording. For each task, assess whether AI can cut the time required by at least 50% without reducing output quality. This is the criterion Eloundou and colleagues used, and it provides a concrete, testable threshold.

Second, design the reallocation of freed time. Decide in advance where the hours saved will go. For MRs, that might mean more face-to-face physician meetings or more personalized information packages. For regulatory staff, it might mean earlier engagement with agencies or deeper analysis of comparable precedents. Without an explicit reallocation plan, freed time dissipates into low-value administration.

Third, embed AI-output verification into existing review processes. In pharma, materials undergo regulatory and medical review before external use. AI-generated content must enter the same pipeline with explicit flagging. Build a step that distinguishes AI-drafted text from human-drafted text so that reviewers apply appropriate scrutiny, and ensure that no AI output reaches a regulator or a physician without documented human verification.