GeoAI Assessment: Separating Ready from Promising

During the GeoAI hype cycle, every GovTech roadmap conversation acquired the same pressure: add AI. Esri was releasing AI-powered geospatial capabilities — object detection from aerial imagery, change detection, automated feature extraction — with obvious surface appeal for government workflows like plan review and code enforcement.

The question wasn't whether these were impressive. It was which of them a government platform could actually put in front of agencies: deployable on the infrastructure customers really run, reliable enough for consequential decisions, and valuable enough to justify the integration cost. Nobody had cut through the marketing material to answer that.

I assessed all fifteen capabilities hands-on against three axes: customer value (which government workflows does this materially improve), technical readiness (what does it take to run this against real customer deployments), and operational constraints (what breaks, what needs human review, what the failure modes cost in a government context).

The pattern that emerged: most capabilities failed on operational readiness, not model quality. The demos were real; the gap was everything around them — deployment models, infrastructure prerequisites, and the reliability bar that consequential government decisions demand.

I prioritized a small set worth pursuing and recommended explicitly against the rest for now — with the evidence attached, capability by capability. The phrase that did the most work in the leadership brief was "promising but not operationally ready": it acknowledged the potential while making the gap concrete.

Recommending against the most ambitious, most demo-friendly options was the hard part. The discipline that made it stick was separating three claims that hype tends to blur: what a capability can do, what it can do reliably in a customer's environment, and what it can do reliably enough that an agency should act on its output.

The core AI product skill in this market is saying no with evidence. Anyone can say yes to a demo; the value a PM adds is a defensible map of what's ready now, what's ready next, and what the real blockers are.

This assessment is also why I build with AI tooling daily and prototype in public: readiness judgment doesn't come from reading release notes. It comes from putting your hands on the capability and watching where it breaks.