Shahar Merom

My work

Australia's AI Opportunities

2025Funded by OpenAISupported by ACS, AIIA, BCA, COSBOA and Women in Digital

What I did

Estimated the economic value of AI for Australia from the task level up. I mapped O*NET tasks to ANZSCO occupations and classified each task's exposure to AI following Eloundou et al. (2023), validating the classification against a sample of Claude conversations. I estimated the token requirements of each task with lightweight LLMs, then used a Cobb-Douglas production framework to translate task-level effects into output, aggregated with ABS data.

What I found

AI could add about $142 billion a year by 2030 (annual gross value added, constant 2025 dollars). One sentence on where the gains concentrate: which industries or occupations

Limitation

The model is partial equilibrium, so it captures no general price effects or capital accumulation. What you would do differently next time

The Pounder Review: NSW Innovation System Assessment

March 2025With Kate PounderIndependent review commissioned by the NSW Government

What I did

Co-authored an independent review of the NSW innovation ecosystem, commissioned by the NSW Government as the evidence base for its Innovation Blueprint. We assessed the ecosystem against five elements, estimated innovative firms' contribution to the state economy using ABS firm-level microdata, benchmarked NSW against peer jurisdictions using venture capital and startup data, and reviewed the state's existing innovation programs. Your part: which analysis you led

What I found

NSW leads Australia on innovation: it attracted 57% of Australian venture capital over the past 24 years and has produced 17 tech unicorns, more than any other state. But its programs have been fragmented and hard to navigate, and it has been slower than peer jurisdictions to adopt an integrated innovation strategy. The review set out three major opportunities and 25 potential actions for the Blueprint. Optional: a headline number from the economic analysis

Limitation

One limitation of the review's approach, and what you would do differently

Task requirements in entry-level and experienced job ads

ForthcomingLLM-based classification

What I did

Used LLMs to classify tens of thousands of job ads by their task requirements, comparing entry-level and experienced roles. As a reliability check I measured agreement across three models (Claude Haiku, Claude Sonnet and DeepSeek) and used confidence labels to flag uncertain classifications. Data source, period and number of ads

What I found

Headline finding, once you're ready to share it

Limitation

Cross-model agreement is a reliability check, not validation against hand-labelled data: models can agree and still be wrong. Any further limitation