Palm trees sat on the slide two rows above pavement cracks, and a column over, mangroves and oil tanks and swimming pools. Sixty-odd narrow models, each trained to find one kind of thing in one kind of picture, and further down the list, cooling towers, crowd counting, plant leaf disease, insulator defects for a utility inspecting its own power lines.

Nobody in the room reacted, because for the people who work in this software every day, none of it was new. That slide went up during the artificial intelligence track at the Esri User Conference in San Diego, on the second morning, and it is the most honest thing I saw all week.

Rohit Singh presented it. He directs Esri's research and development center in New Delhi, describes his own work as applying deep learning to the science of where, and has public code repositories going back years to single-shot detectors and geospatial notebooks. The collection behind the slide has been growing that whole time. Extraction models in ArcGIS Living Atlas doubled from 30 to nearly 70 in a single year (Esri, 2024), and the library now runs past 100 (Esri, 2026).

Esri put other people's models on its own slide

Under the task-specific list sat a second heading: foundation models. CLIP. Depth Anything. GroundingDINO. Prithvi. Segment Anything. Then a third heading, integration, naming Hugging Face for image inpainting, pixel classification, text translation, and visual question answering. One thumbnail on the right showed canopy height estimation credited to Meta. Esri shipped all of it through Living Atlas as packages a user downloads and runs against their own imagery.

Jay Theodore, Esri's chief technology officer for enterprise and artificial intelligence, had opened the track with the framing that explains the slide. He describes the geographic information system as an operating system that enables enterprise systems, one that leaves enterprise resource planning, customer relationship management, and asset management alone. Companies claiming an AI category rarely credit a rival's research lab on their own product slide.

You pick the model that fits the job. Sometimes Esri trained it, sometimes a research lab did, and Esri packaged it so it runs without you rebuilding the plumbing. Esri's documentation states the tradeoff directly: pretrained models require no fine-tuning and suit an organization that needs to move fast, while foundation models offer more flexibility and demand more expertise (Esri, 2026).

Narrowness is the accuracy argument. The palm tree model wants drone photos sharp enough to resolve a few inches per pixel, in ordinary color, and it will tell you how sure it is about each tree it finds so you can set your own bar for what counts (Esri, 2026). It does nothing else. Each model in the collection ships with its required inputs, its training data, and its expected performance documented (Esri, 2026). Compare that to the pitch enterprise buyers have absorbed for two years, where one large model handles everything and the documentation is a prompt.

The Census Bureau showed what the work looks like up close

Andrea Johnson and Elvis Martinez took the stage next. Johnson is acting chief of the bureau's Geography Division, responsible for the nation's address and spatial data. Martinez is a geographer on her staff. Their job is building the address list the census runs on, and work that consumed more than two years of office and field effort now takes about two weeks (Census Bureau, 2026).

The method deserves following in plain terms, because the phrase artificial intelligence hides what happened. For a block in Florida where something had changed:

1. They started with high-resolution imagery of the block.
2. They ran a pretrained model from Living Atlas against it. The model returned rough outlines of buildings, which Martinez described on stage as squiggly polygons.
3. They fine-tuned that model against a sample they had reviewed by hand, ran it again, and got cleaner shapes.
4. They ran those shapes through a tool Esri has shipped for years, which squares off the wobbly edges into the clean rectangles a building outline should be.
5. Where the new outlines overlapped existing records, they merged and updated based on how confident the model was about each building, plus the information already in their inventory.

Roads followed the same path. Trace the line down the middle of each road, clean up the intersections, then borrow the street names from property records the county already keeps. Addresses and address ranges fell out of that, block by block.

Every step there is a geographic information system professional selecting a model, correcting it against ground truth, and chaining the output through tools the discipline has had for years.

Martinez answered the trust question from the field rather than the lab. Imagery analysis gets hard in the deserts of the southwest and the snow of Alaska, and his team knows this from doing the work, so they fine-tune against what they know. They live in a high-growth area, so they know which roads were paved through which apartment complex last year. Ground truth in his framing means what you already know about the place.

Johnson answered the value question differently. Time and cost savings she expected going in. Accuracy surprised her. Think of the model as an employee, she said, doing the same exact thing every time, where before, different analysts digitizing the same feature produced different results.

Her team kept people in the loop on the hard cases and spent the recovered time on disasters, on rural addresses where nobody uses street numbers, and on Alaska, where the imagery sits under snow half the year and some villages have no road reaching them at all.

A heat wave is a financial event, and your GIS team can already price it

Two weeks before the conference, a July heat wave pushed temperatures across the eastern United States to their summer highs. Helen Turvene pulled up average county temperatures at four in the afternoon on July 3rd, added relative humidity, and applied the heat index equation to symbolize what those counties felt like rather than what the thermometer read. She pasted the equation in and named its source, the National Weather Service, so the science came from the agency rather than the assistant.

County-level felt temperature, computed from two live feeds, rendered across a region in the time it takes to describe the request.

Give that output to a chief financial officer and the questions write themselves:

• Which stores lose a selling day
• Where staffing cost spikes
• What the energy load does to the operating line
• How much to reserve against claims

Turning a weather event into a number on the profit and loss statement means matching the forecast against your stores, your staff, and your assets, one location at a time. The team that can do that matching is already on your payroll.

Alice Benjamin, a solution engineer at Esri, made the same point from a different angle later that morning. She demonstrated a module inside the ArcGIS API for Python, in beta now with general release planned for November, that runs prompts against images and text. Her data was 311 reports from the public, the complaints a city receives about potholes and broken streetlights, each carrying a photo and a comment. One prompt generated alternative text for every image, the description a screen reader speaks aloud, which the reports lacked entirely. Another classified comment sentiment so staff could prioritize.

Accessibility compliance and complaint triage, running inside the mapping platform, against a feature layer.

Inventory positioning across a distribution network has the same shape, since the cost of holding the wrong unit in the wrong warehouse is a distance calculation with a carrying charge attached. Site selection is the older version, scoring a location against demographics and drive time before anyone signs a lease. Neither was demonstrated this week, so treat them as the direction rather than the evidence.

Nobody on stage attached a currency figure to any of it. Census talked about avoiding the cost of a national field operation without saying what that operation would have cost. Your finance lead will ask for the number in the first meeting, and the answer has to come from your own operations.

Two executives put a condition on all of it

Nick Giner, who manages spatial analysis and data science products, had shown the room a technique called embedding earlier that morning. The software takes everything it knows about a place, its population, its income levels, its housing patterns, and boils all of it down to a long string of numbers. Places with similar numbers are similar places. Ask the software to find neighborhoods like this one and it compares the strings, which is faster than any analyst reading through the underlying data by hand.

Sud Menon, Esri's corporate director of software product development, put a condition on that convenience. The string of numbers only holds what went into it. Build it from population and income data and it can tell you which neighborhoods resemble each other on population and income. Ask that same string which neighborhoods face similar flood risk and it will still answer, because the software has no way to know the question moved outside what it was built from. The answer will look identical to a good one.

People using these tools have to understand what went into them, Menon said, and the responsibility to explain that runs through the whole field.

A model that returns a confident wrong answer costs more than a tool that returns nothing.

Brian Cross, who runs Esri's professional services arm and whose teams work with more than 3,000 client organizations (Esri, 2026), described the same market from the delivery side. Everyone is in discovery. Some organizations face pressure to show AI results now, others are testing the water, and everyone has an obligation to experiment when a technology wave arrives. At some point, he said, you pivot to business outcomes, and the organizations that get there stop asking what the technology can do and start asking which problem they are solving.

Ismael Chivite, Esri's senior principal product manager for artificial intelligence assistants, has been at the company since 2002 (Esri, 2024). Between him and Singh, the track was run by a deep learning researcher and a geographer who has spent two decades shipping field data collection software, which tells you how Esri divides the problem.

The rules on the assistants Chivite oversees are clearer than most vendors manage:

• Included, not extra. The assistants come with user types at no additional cost, and three reached general availability in the June 2026 release (Esri, 2026).
• The conversation is free, the work may not be. Assistants consume no credits themselves. Esri's documentation notes that an assistant may recommend a workflow that does (Esri, 2026), so asking for a script costs nothing and running it bills normally.
• Off by default. Access stays off until an administrator turns it on, at the organization level and per user.
• Prompts stay yours. They are not stored and not used for training.
• The model is disclosed. Each assistant reaching general availability ships a card naming the model behind it.

The job description changed while the job title stayed the same

Follow the Census workflow back through and count the skills it took:

• Selecting a model against an imagery specification
• Assembling a review sample by hand
• Fine-tuning the model against it
• Reading the confidence numbers and deciding what bar is defensible
• Chaining the output into the cleanup tools
• Knowing which places the model will get wrong before it gets them wrong

That is a data science job, performed by people whose title says geographic information systems, using ground truth no data scientist in your organization has.

Your GIS team already runs machine learning in production. They have been doing it quietly, on a platform your IT organization probably classifies as mapping software, funded out of a departmental budget, reporting somewhere well below the executives writing your AI strategy.

Meanwhile someone in a business unit is wiring an agent to a spreadsheet export of your building inventory, because the GIS team handed over a file when asked for one. An Esri presenter said he has watched that exchange happen and that it should never happen.

Both problems share a root. The people who hold the authoritative record of where everything is, and who now hold the model selection and fine-tuning skills to make that record machine-readable, sit outside the room where AI decisions get made.

Fixing that takes an org chart and an invitation.