GIS's Scarce Skill Is Now Checking Code

An AI assistant, asked to write an ArcPy script, guessed that a parcel layer held a field called "Improvement_Value". The geodatabase held one called "Improvemen", truncated to 10 characters, probably a leftover from an old shapefile. The script failed. The machine had made what looked like a reasonable guess from common naming conventions, and the data disagreed.

Eric Pimpler, founder of Geospatial Training Services, used the episode as the worked example in a guide published Feb 5, 2026. It is a small failure, but it shows where geospatial work is heading. The scarce skill is moving from writing ArcPy and ModelBuilder logic to checking what a machine wrote. Esri's own design choices point the same way.

Code on every desk

The ArcGIS Pro assistant is a natural-language helper built into the application. Esri says it can run supported in-application actions, answer documentation questions and generate code, including ArcPy scripts, SQL query layers and OpenCypher queries for knowledge graphs. ArcGIS Pro 3.7 was released on May 14, 2026, according to Esri's roadmap. Starting with 3.7, the beta is open to all eligible ArcGIS Online named users, no invitation required.

So syntax is no longer the barrier. Pimpler says as much: assistants don't remove the need to understand GIS concepts, but they do remove the need to memorize syntax. What stays is knowing what the data looks like, what the tool should do to it, and whether the output makes sense.

Constrained by design

Esri has not tried to hide the limits. Its post on the beta, published May 27, 2026 and written by product manager Vedant Bajaj, says support for arcpy.mp and arcpy.sharing is limited. It says responses can vary as you rephrase a request, which it calls expected behavior with generative AI. And it says complex, multi-step jobs such as building a 3D city scene can't yet be handled as a single request. In ArcGIS Pro 3.7, the prediction of data validation checks covers only water and electric network datasets.

Esri's Ismael Chivite, speaking to xyHt about the agents aimed at non-technical users, was blunt. "The types of things that these agents will do will be limited, and they will be limited by design, on purpose." he said. He also said that without GIS fundamentals, the assistants are not designed for the non-technical user. This is a product built for people who can read its output.

Esri's materials say the same, according to xyHt: a human should always be in the loop to judge whether generative output is appropriate, because the technology can produce inconsistent results. ArcNews (Winter 2026) put it plainly: "While AI assistants are powerful tools, Esri encourages users to validate the results before making decisions based on them."

Where the work actually sits

Pimpler's guide describes what the new job looks like. AI assistants don't know your schema, he writes, so he recommends running an inventory script first and pasting the results into the conversation. The analyst supplies the ground truth. "Your GIS expertise is the validation layer that AI cannot replace." he writes.

The older tools carry the same lesson. In a separate guide dated June 25, 2026, Geospatial Training Services says a ModelBuilder diagram doubles as documentation that stakeholders and reviewers can follow. Export to Python, it says, drops iterators and the model-only logic tools, so you add loops, branching and error handling yourself. It calls the export "a scaffold, not a finished product." Generated code gets the same treatment. Treat it as a draft and review it before it runs. ArcGIS Pro 3.7 even lets geoprocessing history entries carry descriptions, a place to note what a step was meant to do.

The clearest commercial example comes from xyHt's report on Esri's 2026 User Conference. At NextEra Energy, an in-house tool called Boundary Hunter uses Claude to pull site-boundary information from documents and generate map polygons. The company says tasks that took hours or days now take minutes. A person still owns the resulting boundary. xyHt's analysis found the pattern repeated across the agentic mapping app, the Copilot Studio integration, Boundary Hunter and Ask NWS. Each is scoped to a specific authoritative dataset, and each keeps a human or a defined data boundary between the model's output and the decision that follows.

Consider the stakes in a Harris County, Texas demonstration, also reported by xyHt. ArcGIS Pro's file knowledge graph traced land-ownership networks among 292 parcels within 500 feet of a proposed transmission easement. In one case it followed a single owner through a general partnership to 172 related parcels. Whoever writes the query that finds those links is no longer the bottleneck. Whoever can say whether the links are real is.

The main objection

The counter-argument is that this is a transitional stage. Esri's roadmap points to assistants that build and configure custom agents, plus a mid-term conversational analysis experience meant to help users explore results and create reproducible workflows. xyHt reports roughly a dozen more assistants planned before the end of 2027, after three shipped in June 2026, with native MCP arriving in ArcGIS Enterprise in November 2026. As the machines chain more steps, the schema guesses should get rarer.

They should. But better chaining does not shrink the reviewer's job; it makes errors harder to see. A wrong field name stops a script and announces itself. A wrong assumption buried in a ten-step workflow produces a plausible map. Esri's own answer to that risk is limits and human sign-off, not confidence. Chivite said the licensing for agentic AI is not settled either: "The licensing is actually not finalized yet."

What follows

If this read is right, the premium goes to the analyst who knows the geodatabase well enough to catch a truncated field and understands the question well enough to doubt a tidy result. Training programs built around syntax will age badly. Those built around schema, data quality and validation will not.

Pimpler's script failed on a field called "Improvemen". That is the kind of detail a machine will keep missing and an experienced analyst will keep catching.