From Tool to Collaborator: Experience, Platform Multi-Homing, and the Emerging Evaluation Layer among AI Builders in Southeast Asia

Most claims about artificial intelligence (AI) capability in Southeast Asia rest on proxies. They say little about how practitioners actually combine human judgment with AI systems. This working paper analyses registration data for the first 2,719 approved registrants of Agentic AI Build Week 2026, a buildathon in Ho Chi Minh City, Vietnam, with participants from 55 countries. Descriptively, tool use is portfolio-based. Of builders with tool data (n = 2,615), 80.6% use more than one AI platform, OpenAI (77.5%) and Anthropic Claude (76.3%) are at near-parity, and 27.6% use at least one Chinese model. The pipeline is young: 74.8% of builders have under two years of AI/ML experience. Inferentially, regressions control for role, organisation type and country (n = 2,614). AI/ML experience is associated with larger toolkits (incidence-rate ratio 1.42 at 3–5 years versus under one year) and with higher odds of using Chinese models (odds ratio 2.75) and six or more platforms (odds ratio 4.88). Paired McNemar tests show that OpenAI leads Claude only among first-year builders, by 8.3 points; from the second year the two are statistically indistinguishable. I interpret these associations through the human–AI collaboration literature and suggest that model evaluation is emerging as a distinct human capability. The sample is self-selected and cross-sectional, and the author's organisation organised the event. Version 2 note: Version 2 (27 October 2026) is a corrections release following a full reference check; the data, analysis, results and findings are unchanged. It corrects one misattributed citation: the finding that intensity of technology use after arrival, rather than adoption speed, drives long-run income divergence is now credited to Comin and Mestieri (2018, American Economic Journal: Macroeconomics), with Comin and Hobijn (2010) retained only for its own finding on adoption lags (Sections 2.3 and 6.3). It also updates bibliographic details: Dell'Acqua et al. now cites the published Organization Science (2026) article instead of the 2023 working paper; the Holm (1979) entry uses the JSTOR stable URL in place of an unregistered DOI; and the GenAI Fund (2026) report entry adds its title and retrieval date. Two literature statements were reworded to match their sources more closely (NASEM 2022 on oversight and accountability in Section 6.3; Amershi et al. 2019 on human-AI interaction guidance in Section 6.2), and a short revision note was added.