OpenAI's Bangkok disaster-management workshop bets on custom GPTs, not new models

What happened in Bangkok on March 29

OpenAI convened 50 disaster-management leaders from 13 countries — Bangladesh, India, Indonesia, Lao PDR, Malaysia, Myanmar, Nepal, Pakistan, Philippines, Sri Lanka, Thailand, Timor Leste, and Vietnam — for a single-day workshop it calls an AI Jam for Disaster Management. The event ran in partnership with the Gates Foundation, the Asian Disaster Preparedness Center (ADPC), and DataKind.

The participants weren't researchers or policymakers at a remove. OpenAI describes them as people directly involved in on-the-ground response: coordinating information, supporting affected communities, and making time-critical decisions. That framing matters, because it shapes what the session actually produced.

The choice to build custom GPTs instead of starting from scratch

The most concrete detail in the announcement is what participants actually did. Rather than commissioning bespoke systems, they worked with OpenAI mentors to build custom GPTs and reusable workflows — targeting situation reporting, needs assessment, and public communication. Three named use cases, each mapping to a recurring task a response team performs under time pressure.

This is a deliberately modest technical footprint. Custom GPTs are configurations of an existing model, not new infrastructure. For teams that OpenAI itself describes as operating with fragmented data, manual processes, and limited infrastructure, the appeal is that a workflow built in a workshop can be reused the next week without an engineering team or a procurement cycle.

The session also flagged responsible use and "building institutional trust" — an acknowledgment that adoption inside government agencies and multilaterals is a governance problem as much as a tooling one.

The usage data that justifies the effort

OpenAI grounds the initiative in two specific numbers from recent storms. During Cyclone Ditwah in Sri Lanka, it reports a 17× increase in cyclone-related messages on ChatGPT. During Cyclone Senyar in November 2025, Thailand saw message volume jump 3.2× compared to prior months.

These figures describe the public already turning to ChatGPT during crises — demand that exists whether or not response agencies are ready for it. The workshop's logic follows directly: if citizens are asking an AI system for guidance during a cyclone, the organizations coordinating the response have reason to bring the same tools into their own decision loops. OpenAI pairs this with regional stakes it cites from external sources: Asia accounts for an estimated 75% of people affected by disasters globally, and the World Bank estimates disasters have cost ASEAN countries more than $11 billion.

A workshop is a funnel, and the pilot phase is where it's tested

OpenAI is explicit that this connects to its OpenAI for Countries Program, expanded at Davos, and that a second phase is planned in the coming months focused on pilot deployments and deeper technical collaboration. That sequencing is the real signal. A one-day event can generate enthusiasm and a folder of custom GPTs; it cannot prove that those tools survive contact with a live emergency.

Sandy Kunvatanagarn, OpenAI's Head of Public Policy, framed the gap plainly:

This session is aimed at closing the gap between what AI can do and how it's actually used in the field. Across Asia, there's strong momentum and interest in AI, but the real opportunity is turning that into practical capability.Montana Labs

The specific implication for anyone deploying AI in constrained, high-stakes environments: the workshop artifacts are worthless unless the pilot phase confirms they hold up when data is fragmented, connectivity is poor, and a decision cannot wait. OpenAI has named that follow-on phase but not yet delivered it — so the honest verdict on this announcement is that it established a promising method and left the hard evidence for later.