Skip to main content

Automation risk assessment

Secretaries and Administrative Assistants: AI exposure and automation risk

Secretaries and Administrative Assistants score 57% on the Eloundou et al. LLM exposure measure (high exposure). 25 of 32 rated tasks are exposed. BLS projects −6.0% employment, 2025–35.

Data compiled 2026-10-09 · BLS 2025–35 · O*NET 31.0 · AEI May 2026

LLM exposure (β)

57%

High exposure

BLS projects −6.0% employment for secretaries and administrative assistants from 2025 to 2035.

Eloundou et al., Science 2024

Exposure scores

How much of this occupation's task list large language models could speed up

57%

High exposure (β, human-rated)

Share of this occupation's tasks that annotators judged an LLM could do at least 50% faster at the same quality, counting tasks that need extra software at half weight.

Eloundou et al., Science 2024

64%

Same measure, rated by GPT-4

The study's model-rated β for comparison. Directly exposed tasks alone (α) score 37%; all exposed tasks (ζ) score 77%.

Eloundou et al., Science 2024

45%

Observed exposure in real AI usage

How much of the theoretically exposed work actually shows up as automated, work-related Claude usage.

Anthropic observed exposure (Mar 2026)

Exposed and unexposed tasks

Human ratings of 32 O*NET task statements, Eloundou et al., Science 2024

Of 32 rated tasks, 13 are directly exposed (E1), 12 are exposed once complementary software is built (E2), and 7 are not exposed (E0).

Exposed tasks

  • Answer telephones and give information to callers, take messages, or transfer calls to appropriate individuals.

    E1 · Directly exposed

  • Greet visitors or callers and handle their inquiries or direct them to the appropriate persons according to their needs.

    E1 · Directly exposed

  • Create, maintain, and enter information into databases.

    E2 · Exposed with added software

  • Operate office equipment, such as fax machines, copiers, or phone systems and arrange for repairs when equipment malfunctions.

    E2 · Exposed with added software

  • Set up and manage paper or electronic filing systems, recording information, updating paperwork, or maintaining documents, such as attendance records, correspondence, or other material.

    E2 · Exposed with added software

Tasks not exposed

  • Use computers for various applications, such as database management or word processing.

    E0 · Not exposed

  • Compose, type, and distribute meeting notes, routine correspondence, or reports, such as presentations or expense, statistical, or monthly reports.

    E0 · Not exposed

  • Make copies of correspondence or other printed material.

    E0 · Not exposed

  • Train and assist staff with computer usage.

    E0 · Not exposed

  • Order and dispense supplies.

    E0 · Not exposed

Job outlook

U.S. projections, BLS Employment Projections 2025–35

−6.0%

Projected employment change, 2025–35

From 1.9 million jobs in 2025 to 1.8 million in 2035. All occupations: +3.5%.

BLS Employment Projections 2025–35

163,900

Openings per year, 2025–35 average

Includes openings from growth and from workers who retire or change occupations.

BLS Employment Projections 2025–35

$47,540

Median annual wage, 2025

BLS Employment Projections 2025–35

High school diploma or equivalent

Typical education for entry

BLS Employment Projections 2025–35

What BLS expects to change

Productivity change - share decreases as technology, including artificial intelligence (AI) systems, assists workers with routine back office functions, such as preparing documents, scheduling, and notetaking.

BLS note on factors affecting secretaries and administrative assistants employment, Table 1.12, BLS Employment Projections 2025–35.

Primary profiles: BLS Occupational Outlook Handbook · O*NET OnLine

How AI is used on this work

Claude.ai conversations on this occupation's tasks, Anthropic Economic Index (May 2026)

Augmentation (AI works with the person)29.4%
Automation (AI does the task)70.6%

Augmentation and automation are shares of conversations with a classified collaboration pattern, excluding unclassified conversations. The individual pattern shares include them. 70.4% of all conversations on these tasks were for work. See the full collaboration breakdown

Sources & methodology

Every figure on this page comes from the public datasets below. We do not edit the source values; derived figures are explained here.

  1. Exposure is the human-annotator β score from Eloundou et al.: the share of an occupation's O*NET tasks that a large language model could do at least 50% faster at equal quality, counting tasks that need extra software (E2) at half weight. We label β under 0.25 low, 0.25–0.5 moderate, and 0.5 or more high; these bands are ours, not the authors'.
  2. Task labels (E0 not exposed, E1 directly exposed, E2 exposed with complementary software) are the study's aggregated human ratings of each O*NET task statement for the occupation. We show up to five tasks per group.
  3. Observed exposure (Anthropic, March 2026) combines that theoretical feasibility with real Claude usage on the occupation's tasks, weighting automated and work-related use more heavily than augmentative use, then averaging by time spent on each task. Values run from 0 to 1.
  4. Employment, growth, openings, wage, and education are the BLS 2025–35 projections as published; BLS utilization notes are quoted verbatim.
  5. Collaboration shares come from Anthropic Economic Index data for Claude.ai conversations in May 2026, classified to the O*NET tasks they match. "Automation" groups directive and feedback-loop conversations; "augmentation" groups task iteration, learning, and validation. Both buckets are shares of conversations with a classified collaboration pattern, excluding the unclassified "none" pattern; the six individual pattern shares include unclassified conversations. They describe how people use Claude on this occupation's tasks — not what share of workers in the occupation use AI.
  6. BLS occupations are matched to O*NET-SOC codes with the BLS O*NET-to-NEM crosswalk; where a BLS occupation spans several O*NET occupations, we average their values.
  7. Exposure measures what AI could help do, not a forecast of job loss.

  • Eloundou et al., Science 2024

    Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). "GPTs are GPTs: Labor market impact potential of LLMs." Science, 384(6702), 1306–1308. doi:10.1126/science.adj0998. Occupation- and task-level exposure data from the authors' public repository (openai/GPTs-are-GPTs).

    Vintage: Published June 21, 2024; human annotations of O*NET tasks collected 2023. License: Data repository: MIT License. Retrieved 2026-10-09.

  • Anthropic observed exposure (Mar 2026)

    Massenkoff, M., & McCrory, P. (2026). "Labor market impacts of AI: A new measure and early evidence." Anthropic, March 5, 2026. Occupation-level observed exposure (Anthropic Economic Index, labor_market_impacts/job_exposure.csv).

    Vintage: Published March 5, 2026. License: CC BY (data). Retrieved 2026-10-09.

  • BLS Employment Projections 2025–35

    U.S. Bureau of Labor Statistics, Employment Projections program. Occupational projections, 2025–35, and worker characteristics, 2025 (Table 1.2); fastest growing occupations (Table 1.3); factors affecting occupational utilization (Table 1.12).

    Vintage: 2025–35 projections, released August 27, 2026. License: Public domain (U.S. federal government work). Retrieved 2026-10-09.

  • Anthropic Economic Index (May 2026)

    Massenkoff, M., Lyubich, E., Sacher, S., Hitzig, Z., Zhang, S., Heller, R., & McCrory, P. (2026). "Anthropic Economic Index report: Cadences." Anthropic, June 26, 2026. Claude.ai usage metrics by occupation, global, May 2026.

    Vintage: Release of June 26, 2026; Claude.ai conversations from May 2026. License: CC BY (data). Retrieved 2026-10-09.

  • O*NET 31.0 Database

    This page includes information from the O*NET 31.0 Database by the U.S. Department of Labor, Employment and Training Administration (USDOL/ETA). Used under the CC BY 4.0 license. O*NET® is a trademark of USDOL/ETA. Metaintro has modified all or some of this information. USDOL/ETA has not approved, endorsed, or tested these modifications.

    Vintage: O*NET 31.0, August 2026 release. License: CC BY 4.0. Retrieved 2026-10-09.

Return to navigation