Fast answers were never the hard part. Defensible ones were.
Organizations are pointing AI at the map services they already run, and finding that nobody can stand behind the answers.
The model is not what fails. A GIS estate that grew over twenty years cannot tell anyone, person or machine, which of its duplicate layers is the authoritative one, what its codes mean, or which shape a place name refers to. So the AI guesses at every step and hands back a number with no record of the guesses.
One ordinary request shows all of it.
Find fatal crashes in 2024 within 1 mile of Fairfax.
It is one short sentence and a compound analysis: an ambiguous place, a buffer, two filters, and the right dataset to begin with. Requests like it take up staff time every week. Seven decisions sit inside it.
Which layer?
Several webmaps hold a layer called “Crashes”, backed by different services with different fields and update schedules. It picks one.
Data ExplorerOne dataset is designated for the concept, and an administrator has tested that the question lands on it.
Which Fairfax?
Virginia has a Fairfax City and a Fairfax County, and the city sits inside the county. It assumes one.
Data ExplorerIt offers both as a one-tap choice and records the one you pick.
What counts as near Fairfax?
It looks for a jurisdiction field, because that makes an easy filter. That finds crashes coded to Fairfax and misses every one in the mile around it. The count is quietly wrong.
Data ExplorerThe boundary comes from the one boundary dataset your organization approved. The mile is a drawn ring around it, shown on the map and kept with the answer.
What do “fatal” and “in 2024” mean?
It guesses a severity field, a code for fatal and one of several date fields. It may guess differently next time.
Data ExplorerThey resolve to the fields and the code your crash expert approved, and the answer says which.
Who checks the query?
Nobody. The model writes it.
Data ExplorerThe model never writes one. It picks from fixed operations, and every column and value is checked against the real data first.
Where does it run?
On your production service, competing with the map viewers that service exists for.
Data ExplorerOn a local copy, in seconds, with no load on the source.
What is left afterwards?
A number.
Data ExplorerThe number and its record: dataset, boundary, filters and the time the data was read. Rerun it next month, or make it 2 miles and ask again.
The same sentence goes in both times. In the demo, for Fairfax County, it comes back as 62, with the ring on the map and the filter in plain view. Only that number can be defended. If the other one is wrong, nobody may ever find out.
The AI proposes a plan. It never just answers.
For every question, the AI's job is to say how approved data would answer it: which dataset, which boundary, which filters, which operations. Ordinary software then checks that plan against what your people have approved, and runs it. The plan is shown with the answer and kept.
Protocols such as MCP solved connection: how an AI reaches your systems. Three things are still missing once it is connected, and better prompting cannot supply them, because the information does not exist anywhere a model can read it. Which source is true. What the data means. How to prove what was done. Data Explorer exists to create that information, let your people govern it, and enforce it every time a question is asked.
Seven problems, and where each one goes
No mechanism for truth
GIS estates collect duplicate layers across webmaps, services and portals. Nothing in the stack can say “this is the authoritative source for this concept”. That fact has nowhere to live.
The library gives it a home: one dataset per concept. Each description is drafted from the services, the webmaps and the data itself, then reviewed by an administrator, who can run real questions and confirm where they land before any user sees an answer.
“Fairfax” is not a place until someone decides which one
Duplicate names, city limits against metro areas, boundaries that changed, features that sit on the line. Each tool settles these silently today, and two competent analysts get different, equally plausible counts. Matching records to places by a city field is the easy shortcut, and those fields are chronically blank or stale.
Place names resolve against a catalog with one approved source for each kind of boundary. An ambiguous name comes back to the user as a choice. Spatial questions are answered with geometry, and when a follow-up implies a different area, the system asks.
Meaning lives in people's heads
Operational data is full of undocumented columns and codes. Either someone who knows the data translates, or a model works the meaning out again on every question, where nobody can see it, correct it or keep it.
That work moves from question time to setup time. The system proposes what each field and code means, shows its evidence, and marks the proposal as a machine's guess. An administrator approves or edits it once. Fields nobody can make sense of are left out of search, with the reason recorded.
An answer is not a method
Any released number can be challenged by a records requester, a regulator, a court, an auditor or a board. A paragraph from an AI, even a correct one, is a dead end. You cannot reproduce it, hand over how it was made, or change one part without starting again.
Every answer carries its plan, readable by a person and by a machine. The plan reproduces the answer and can be changed and rerun. A refusal carries its reasons in the same form. What you hand the requester is the record.
AI with raw query access
The obvious design lets the model write a query and treats safety as a list of instructions in a prompt. Instructions are not boundaries.
Here the model selects from fixed operations, and each argument is checked against real columns, approved values and known boundaries before anything runs. Expensive work needs an explicit yes. The limits are built in.
Source services can't carry the load
The questions worth asking, such as buffers, statewide rankings and large joins, are the ones map services cannot serve. They hit record limits, paging and timeouts, on servers that may handle one request at a time, and they compete with the viewers those services exist for.
Administrators choose which datasets to copy into an analytical store. Heavy questions run there in seconds. The sources remain the system of record, the copy is read-only, and every answer says when its copy was taken.
Expertise walks out the door
Which layer is right, which Fairfax they mean, which code means fatal, which questions deserve a careful no. That lives with a few people. When they are out, answers wait. When they retire, it leaves with them.
It becomes structure in the library, with a name and a date on each decision. The same question gets the same answer whoever asks it.
The person who knows the data is the one in charge
None of this is configured by a vendor or tuned in a prompt. A subject-matter expert drives it from an admin screen built for them, and no code is involved.
- The system does the first draftIt reads the services and the data and proposes which dataset answers which kind of question and what each field and code means. Every proposal is marked as a machine's guess until a person approves it.
- The expert approves, edits or switches offOne screen for each dataset: its plain-language description, the words people use for it, what each code means, which fields can be searched.
- The effect can be shownAsk a real question before and after a change. The answer's record names the dataset, the field and the code that were used, so the change is visible and so is its cause.
- Every decision has an authorEach one carries a name and a date, and a saved decision is never overwritten by the system's next guess.
That is the difference between a system your experts hope is right and one they can demonstrate is right, to a manager, an auditor or a records requester.
Where your data goes, and what it costs to run
Every AI procurement asks both questions, and here they have one answer, because both follow from one decision: the intelligence lives in the library, not in the model. The hard work is done once, at setup, and written down. What is left at question time is a small job: turn a sentence into a plan, using vocabulary the system hands the model. In our development measurements, small fast models produced the same plans as the largest ones once that vocabulary was in place.
So the model is a setting. An administrator chooses a commercial service, a model in your own cloud, or a private one on your own hardware, and can change the choice later. The system reads your services read-only and runs as a container where you approve. Administrators decide field by field what can be searched, and what it builds up is kept in open formats you hold.
- Most of an answer spends no tokensBoundaries, buffers, checks, execution and drawing are ordinary computation. The AI is consulted briefly, for the plan.
- The stable context is cachedThe vocabulary the model needs is not paid for again on every question.
- Small models are enoughThe most expensive tiers are an option and never a requirement.
- Plans are reusedRepeating or adjusting an earlier question reruns its plan. Nothing is worked out again.
- Nothing expensive happens by accidentA large scan is an explicit, confirmed act.
Why not just use what we have?
Esri's own tools and assistants
They work on the webmaps and services as they are, so they inherit the duplicates, the stale descriptions and the missing authority. They answer “what is on this map”. Data Explorer works alongside them: it reads the services your GIS team already publishes, read-only, and nothing about how those are run has to change.
Our services wrapped with MCP
That gives an agent a line to your services. The agent still makes all seven silent decisions, now automatically. Data Explorer speaks MCP too. What differs is the other end of the line: a governed library, where the alternative is forty raw endpoints.
A text-to-SQL or BI copilot
It has no spatial operations, and the boundary, buffer and distance questions are the expensive ones. It also has the model write the query.
A metadata cleanup project
Documentation nobody enforces drifts. Here the descriptions are what the system checks against, administrators can test them, and the first draft is written for you from the data.
See it answer a question an analyst would be asked
The demo runs on publicly available map services. A draft statement of requirements that a public body can adapt for its own procurement is available on request.