Measuring AI Exposure in Singapore's Workforce

A task-based, LLM-scored exposure index — with a game/app-studio spotlight

A small, reproducible study that applies an Anthropic Economic Index–style method to Singapore: decompose occupations into tasks, use a frontier LLM (Claude) as the measurement instrument to score each task's automation and augmentation potential, then aggregate to an occupation-level exposure index and weight by Singapore employment.

The point is methodological: using AI itself as a transparent, reproducible instrument to measure AI's economic footprint — the core idea behind the Anthropic Economic Index — and localising it to an APAC market and a sector (games/apps) I know firsthand.

📄 Full write-up: paper/ai-exposure-singapore-working-paper.pdf — a 7-page working paper (method, results, discussion, limitations, references). Source: paper/paper.html.

⚠️ Scope & honesty. This is an illustrative methodology demo, not an official statistic. It models a representative 17 Singapore occupations (+6 studio roles) and 92 tasks. Task scores are LLM judgments (Claude, rubric v1), tasks are O*NET-style and proxy for SG, and employment weights are approximate, based on public SingStat/MOM occupational distributions and normalised across the modeled set. Treat magnitudes as directional.

---

Headline findings

• ~41% of modeled Singapore employment sits in the High AI-exposure band (index ≥ 0.60), ~36% Medium, ~23% Low. Employment-weighted mean exposure ≈ 0.51.
• Exposure tracks knowledge & clerical work, not seniority. The most exposed occupations are Software Engineer, Administrative/Office Clerk, Customer-Service/Data-Entry Clerk, Contact-Centre Agent and Financial Analyst. The least exposed are Cleaner, F&B Server, Construction/Electrical Trades, Retail Salesperson and Driver — embodied and interpersonal work.
• Every modeled occupation is augmentation-leaning, not automation-dominant. On the rubric, augmentation potential exceeds end-to-end automation potential for all 23 roles — consistent with the Anthropic Economic Index's real-world finding that current AI use augments more than it automates. The near-term story for Singapore is task reshaping, not wholesale replacement.
• Studio spotlight (games/apps): the most AI-exposed studio roles are User-Acquisition / Marketing Analyst (0.78) and Game Data Analyst (0.75) — data- and content-heavy roles — while Narrative / Game Designer is least exposed (0.57) but still high, because the writing is highly augmentable even though creative direction stays human.

Figure
	Singapore occupations ranked by exposure
	Every role sits below the diagonal → augmentation-dominant
	Share of modeled employment by exposure band
	Game/app studio roles

---

Method (4 steps)

1. Occupations → tasks. 17 representative SG occupations across all SSOC major groups, plus 6 game/app-studio roles, each broken into 4 O*NET-style tasks (data/).
2. LLM scoring. Claude scores every task on two independent 0–5 scales — automation (AI does it end-to-end) and augmentation (AI assists a human) — per prompts/scoring_rubric.md. Scores are the committed dataset in data/tasks_scored.csv.
3. Aggregate. Exposure index = mean of (automation + augmentation)/10 over a role's tasks (0–1); automation share = mean(A)/(mean(A)+mean(G)).
4. Weight. Combine with approximate Singapore employment shares to estimate the share of work in each exposure band.

Reproduce / extend

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

python src/analyze.py          # rebuild figures + occupation_exposure.csv from committed scores

# Re-score every task from scratch with your own Claude key (transparent instrument):
pip install anthropic
export ANTHROPIC_API_KEY=sk-ant-...
python src/score_tasks.py      # regenerates data/tasks_scored.csv via the same rubric
python src/analyze.py

Extend it by editing the occupation/task lists in src/_build_seed_data.py (e.g. add more SSOC occupations, swap in a different APAC market, or expand the studio taxonomy) and re-running.

Files

data/occupations_sg.csv     occupations, SSOC group, sector, employment weight
data/tasks_scored.csv       92 tasks with Claude automation/augmentation scores + rationale
data/occupation_exposure.csv  aggregated index (generated)
prompts/scoring_rubric.md   the 2-D scoring rubric (the measurement instrument)
src/score_tasks.py          reproducible Claude-API scoring pipeline
src/analyze.py              aggregation + figures
figures/                    charts (generated)

Limitations

LLM-as-scorer carries the model's own biases and a mid-2026 capability snapshot; O*NET tasks are US-derived and proxy imperfectly for Singapore; employment weights are approximate and the occupation set is a representative sample, not the full SSOC. "Exposure" measures technical applicability of AI to tasks, not predicted job loss — adoption, regulation, cost, and labour-market dynamics all sit between exposure and outcomes. This repo is a transparent starting point for that conversation, not the last word.

---

Author: Lei (Lorin) Zhao · Singapore. Method inspired by the Anthropic Economic Index. Built with Claude as the scoring instrument.