With the conversations complete and the stakeholder map built, this week was about drafting the brief that would pull everything together. The central challenge was deciding what to say and, just as importantly, what not to claim. AI discussions tend toward optimism, and I wanted the brief to be as clear about limits as it was about opportunities, a commitment I made in my original proposal.

I structured the brief around the stages of air quality governance and asked what AI could realistically contribute at each one. I looked at five areas: mapping airsheds, calibrating low-cost sensors to improve monitoring, forecasting air quality further ahead, supporting compliance and enforcement, and responding to pollution episodes, including insurance schemes that compensate informal workers when pollution crosses a threshold. Real examples grounded each one, such as machine learning approaches to identifying airshed boundaries and Delhi's AI-enabled platform for monitoring construction sites.

The more I wrote, the more a pattern emerged. Every application depended on foundations that are not yet reliably in place. One in five cities in the national clean air programme has no continuous monitoring station. Many industrial emissions monitors are missing large amounts of data. Data is not shared consistently between agencies, monitoring is often disconnected from any mechanism for acting on it, and the institutions responsible for enforcement are short on staff and technical expertise. This led me to the brief's central argument: AI for air quality should be built on shared digital public infrastructure, which is the kind of work eGov already does, and it should be introduced where the supporting data and institutions are ready.

My background in American government policy contributed something unexpected this week. Drawing on that background, I identified a software package developed by the US Environmental Protection Agency that could be replicated or adapted for an Indian context. I was careful about how I framed this. A tool that works in one governance system is not automatically suitable for another, and the brief says what would need to be true for it to work.

The tangible outcome of this week was a full draft of the brief, which I shared with my manager and the head of the Policy team for review. I welcomed their feedback, because I knew it would make the work better.

The Laidlaw attribute most visible this week was rigor and evidence-based thinking. Judgment and integrity were equally important, because the temptation in a document like this is to make AI sound more useful than the evidence supports.

My role this week was that of an author, taking responsibility for claims that other people might act on.

Next week, I will incorporate the feedback I receive and complete the final version of the brief.