Ask a general-purpose AI agent to check whether a plot of land has been deforested, size a buffer zone, or reconcile two coordinate systems, and it will usually give you an answer. That's the problem. Published benchmarks over the past year have measured how often that answer is actually right, and the results are not close.
The Problem Is Measured, Not Rhetorical
Three independent benchmarks, run by different research groups on different model families, converge on the same finding: frontier AI models are unreliable at spatial computation specifically, even when they're fluent at spatial conversation.
GeoAgentBench, a 2026 benchmark built to test AI agents on real spatial-tool tasks, found frontier models scoring below 45% on parameter-execution accuracy — meaning more than half the time, the model calls the right tool with the wrong inputs (arXiv:2604.13888). GeoBenchX found something arguably worse: when a geospatial task is genuinely unsolvable with the available data, agents produce a confident answer anyway more than three times as often as they correctly reject it (arXiv:2503.18129). And a 2026 study of AI-generated Earth Engine code — the kind of code that would compute an actual hectare of forest loss — found that 58% of it fails to execute at all (arXiv:2606.12821).
Put together, these three findings describe the same failure mode from three angles: today's AI agents are good at sounding right about spatial data and bad at being right about it.
What the benchmarks actually found
<45%
frontier-model accuracy calling spatial tools with correct parameters (GeoAgentBench)
>3x
more likely to fabricate an answer than correctly reject an unsolvable task (GeoBenchX)
58%
of AI-generated Earth Engine code that fails to execute (GeoNatureAgent study)
Why This Matters More in Compliance Work Than Anywhere Else
In most applications, a wrong AI answer is an inconvenience. In deforestation-free verification, carbon MRV, or EUDR due diligence, a wrong number is a liability — it's the figure a due-diligence statement is built on, the hectare count a verifier will test, the input to a penalty calculation. A hallucinated hectare isn't a typo; it's the exact kind of error a compliance system exists to prevent.
That's the specific gap TREEO is building toward closing, and it's worth being precise about what "building toward" means here.
What TREEO Already Runs — and What's Still Ahead
TREEO already operates a production AI agent system: multi-model orchestration, autonomous task execution, and structured memory, running today across our own engineering and operations workflows. Separately, TREEO already operates a production GIS and carbon-accounting platform — satellite-verified boundary data, species and allometry databases, project modeling — that our forest-carbon and EUDR work runs on daily.
What doesn't exist yet is the bridge between them: an agent layer that can reason in natural language about a spatial question — is this plot inside the 2020 forest baseline? does this boundary correction need a polygon or a point? — while every actual computation happens in the deterministic GIS engine underneath it, not inside the language model. That bridge is on our near-term roadmap, not in production. We're saying so plainly because the entire point of this piece is that we'd rather under-claim than join the 58% failure rate we just cited.
The Design Principle: Keep the Model Out of the Math
The approach we're building toward is architectural rather than promotional: the language model plans, explains, and writes — it never computes. Coordinate transforms, area calculations, and deforestation-baseline checks run in the same deterministic geospatial engine that already powers our field verification work today; the agent's job is to call that engine correctly and explain the result, not to arrive at a number on its own.
This isn't a novel idea — it's the same lesson every one of the three benchmarks above independently points to: the failure mode is models doing arithmetic they were never designed to do reliably. Removing that step doesn't make the agent less capable; it makes the number it hands you traceable back to an actual computation instead of a plausible-sounding guess.
Where This Is Headed
The near-term target is narrow and specific: EUDR geolocation checks and carbon dMRV plot verification, the two workflows where TREEO's existing GIS platform already does the underlying computation and where a natural-language interface removes real friction for field and compliance teams — not a general-purpose GIS chatbot. We'll publish what we build, including where it falls short, as it ships rather than in advance of it.
Frequently Asked Questions
No. TREEO runs a production AI agent system and a production GIS platform separately today. Connecting them into a natural-language spatial reasoning layer is an active near-term roadmap item, not a shipped product — this article describes the direction, not a current feature.
Building toward compliance-grade spatial AI
We're extending TREEO's production agent infrastructure into geospatial reasoning for EUDR and carbon dMRV work — deliberately, with the computation kept outside the model. If that's a problem you're solving too, we'd like to compare notes.

