AI, LLMs, and the next chapter of Spatial Finance

What is Spatial Finance, and why is it accelerating now?

Spatial Finance is the integration of geospatial data into financial theory and practice. It lets organizations quantify physical and transition risk at the asset level by integrating real-time geospatial information into classical modeling and assessment. This reveals how the physical and human forces converging on specific locations (farmland, watersheds, facilities, etc.) affect the value of financial assets and cash flows.

Historically, this form of asset-level environmental intelligence was effectively limited to the domain of the reinsurance and underwriting industries. With the expanding volume of location-specific physical risk data, Spatial Finance provides another source of rigorous information on which to aggregate, process, and base capital allocation decisions.

Advances in satellite constellation data availability and cloud computing and storage have made Earth Observation (EO) insights operationally viable in ways technically impossible even a few years ago. And, AI and agentic workflows now enable the scaling of Spatial Finance by turning disparate data sources and physical risk indices into clear, actionable insights regarding place-based resource allocation.

Spatial Finance’s core function is translation. It translates the language of the physical world – pixels, scores, signatures – and converts it into the business language of exposure, valuation, and credit risk. That translation is visual as much as it is numeric. Rather than relying on spreadsheets, Spatial Finance analysis produces thematic maps and risk models that convey physical realities on the ground with an immediacy that can move decision-makers to act at speed.

Spatial Finance, AI, & LLMs: Current state of play

In a perfect world, an LLM could quickly ingest satellite imagery, geospatial data, and financial statements and produce a rigorous business case for investments targeting identified physical and transition risks (and opportunities!).

However, that is not the reality of how the technology is deployed in practice. LLMs and Spatial Finance can work symbiotically, but with crucial hand-holding by humans at both the start and end of a workflow. Varied data formats and reporting practices, disjointed regulatory information, mismatched timelines of business expenses, and chronic physical and transition risks cloud AI’s ability to identify the most valuable interventions or costly risks.

It should also be noted that LLM capabilities in the realm of processing satellite imagery are still lacking. For an AI model to make sense of that data at scale, it has to rely on a technique called embeddings. These embeddings are a sequence of unique numbers assigned to different attributes, allowing a model to detect patterns and correlations across enormous datasets. As LLMs can read and relate these embeddings, they let a person query that data through a natural language interface, asking questions of the data in plain words.

High-level versus place-based risk narratives

LLMs are strong at gathering, aggregating, and synthesizing data and information. Simple queries can pull spatial datasets and turn asset-level risk into a high-level narrative a capital allocator can read. On the finance side, they reliably handle well-defined, standards-based calculations.

However, they have limits. Out-of-the-box LLMs are, by nature, generalized tools meant to tackle a wide range of applications. They can grasp generalities about a sector by training on a vast ecosystem of existing data and norms. The result is a general risk narrative that is a useful starting point but not a basis for allocating capital, and one still prone to imprecision and errors on high-stakes financial numbers used to make decisions. For instance, they likely could not identify the unpredictable muddle of transition forces – the policy, financial, technological, market, and geopolitical pressures – influencing each individual asset.

AI use cases in Spatial Finance

AI has made it possible to index the Earth at scale. Rather than analyzing each satellite image on its own, large models now compress raw imagery into compact, reusable representations that can be compared across time and geography, turning Earth’s surface into something that can be queried in seconds. Cambridge’s TESSERA maps the globe at 10-meter resolution, and Google’s AlphaEarth offers a general-purpose way to characterize any location on Earth. Together, these models can locate an asset and describe the physical risk it faces with remarkable speed.