Foundation models and copilots: how AI is redrawing GIS in 2026

Over the last eighteen months, artificial intelligence stopped being just another technique inside GIS and started changing the very structure of geospatial work. We're not talking about promises: we're talking about models published with dates, licenses and benchmarks, and assistants that already operate inside ArcGIS Pro and QGIS. This note summarizes what is happening on two fronts — foundation models and natural-language copilots — with the verifiable numbers and, just as importantly, the documented limits.

If you're looking for applied use cases in Colombia, we covered those in AI applied to GIS: real cases. Here the focus is the industry's underlying transformation.

First front: geospatial foundation models

A foundation model is pre-trained once on massive volumes of satellite imagery and then fine-tuned for specific tasks with relatively little labeled data. Three releases define the state of the art:

Prithvi-EO-2.0 (NASA + IBM, December 2024)

Trained on 4.2 million global time-series samples from NASA's Harmonized Landsat-Sentinel (HLS) archive at 30 m resolution, Prithvi-EO-2.0 outperformed its predecessor by 8% on the GEO-Bench benchmark (75.6% average at release, ahead of six other geospatial foundation models at the time). It was designed with domain experts for disaster response, land-cover and crop mapping, and ecosystem monitoring. What matters most for an agency or consultancy: it is Apache 2.0, free, available on Hugging Face with IBM's TerraTorch fine-tuning toolkit.

AlphaEarth Foundations and the Satellite Embeddings (Google DeepMind)

Google DeepMind and Earth Engine published the Satellite Embedding dataset: every 10 m pixel of the planet's land surface, represented as a 64-dimensional vector condensing a year of observations from Landsat, Sentinel-1 (radar), Sentinel-2 (optical) and LiDAR. The Earth Engine catalog covers 2017–2024; the official Google Cloud Storage archive also includes 2025. The data is CC-BY 4.0, but the GCS bucket has operated under a provider pays model since July 2026.

The conceptual shift is enormous: the official documentation states the embeddings are designed for classification, clustering and change detection, and that comparing two years reduces to computing the dot product between vectors. In other words, mapping land cover or detecting change at national scale no longer requires training a deep network per project: it requires linear algebra on precomputed features, plus judgment to validate.

Major TOM (ESA Φ-lab + CloudFerro, January 2025)

The European Space Agency took the data route: Major TOM includes the largest ML-ready Sentinel-2 collection published as of early 2025, plus an expansion of more than 169 million precomputed embeddings from processing over 62 TB of Copernicus data, built with four different models, all open on Hugging Face.

The common pattern across all three: publicly accessible models or data with open licenses. That does not make the full operation free: storage, download, compute and commercial use of platforms such as Earth Engine can still cost money. Even so, the entry barrier fell and competitive advantage shifted toward local data quality and ground validation.

Second front: conversational GIS

ArcGIS: a dozen assistants in beta

By October 2025 Esri had roughly thirteen AI assistants spread across the ArcGIS platform, all in beta or preview. In June 2026, the Survey123, Business Analyst and translation assistants reached general availability. The ArcGIS Pro 3.7 assistant entered open beta for eligible users on May 27, 2026: it generates ArcPy code, SQL and OpenCypher queries, and executes supported in-application actions from natural language.

One detail that matters especially to us: among the 3.7 assistant's actions are managing parcel fabric records — Esri's cadastral data model — and creating and realigning LRS routes. Conversational AI is already touching cadastral and infrastructure workflows, not just thematic maps.

QGIS: GIS Copilot, peer-reviewed

In the open-source world, the GIS Copilot / SpatialAnalysisAgent plugin (Penn State, published in the International Journal of Digital Earth, 2025) embeds an LLM agent in QGIS: it decomposes the task, selects among ~390 compiled QGIS/GDAL tools (600+ in the December 2025 v1.3 release), generates and executes PyQGIS code, and self-debugs. In the paper's evaluation — 110 spatial analysis tasks, up to three attempts — it achieved roughly 92–95% success on basic tasks, 80–83% on intermediate, and 75% on advanced ones.

The limits, documented

The same peer-reviewed evaluation that validates GIS Copilot documents its failures, and they are exactly the ones a GIS professional would recognize:

It omits reprojections unless explicitly asked: the agent overlays layers in different coordinate systems without blinking.
It confuses data models: assigning vector inputs to raster-only tools (the cited case: gdal:proximity).
It trips over format quirks, like shapefile's 10-character field-name truncation, which silently breaks joins.

The conclusion isn't "AI doesn't work"; it's that human geospatial judgment shifted from producing to auditing. Someone who doesn't know what a CRS is won't catch that the copilot skipped the reprojection. In Colombia, where we work in MAGNA-SIRGAS / national origin (EPSG:9377), that kind of error isn't theoretical: it's the difference between a defensible appraisal and a challenged one.

What it means for Colombia

Here the public evidence is more qualitative, and we say so honestly: there are no verifiable adoption figures for Latin America. But the institutional signals exist: the national planning department (DNP) has explored AI to support the multipurpose cadastre update, and Colombia's spatial data infrastructure (ICDE) already discusses geospatial intelligence as a trend for territorial development.

For a public agency or an agribusiness, the practical reading is threefold:

The input became far more accessible. Global annual 10 m embeddings under CC-BY cover Colombia from 2017 onward; storage, download and compute may still cost money. Projects like deforestation monitoring or Sentinel-2 environmental monitoring now start from a baseline that in 2023 would have cost months of processing.
Local validation is the differentiator. A global model doesn't know Chocó's cloud cover or the land covers of the Llanos foothills. Ground truth, well-designed sampling and knowledge of the territory are worth more than ever.
The cadastre is on AI's agenda, from Esri's parcel fabric to institutional pilots. For processes like the multipurpose cadastre, AI will prioritize and pre-classify; the legal and technical decision will remain human and auditable.

Where we stand

At GeoSAT we've spent 30 years watching the tools change: from analog restitution to drones, from local servers to Earth Engine. The perennial lesson applies now too: technology gets adopted, judgment gets built. We combine these open models with real Colombian data and ground validation, in our agricultural platform Geobristol and in our clients' projects.

If you want to evaluate where foundation models or GIS copilots can accelerate your operation — no smoke, with metrics — check out our geospatial AI service or write to us.