Talking to a map: what an AI GIS assistant can actually do

The short answer: it is not the model looking at your data and guessing. It is the model choosing which GIS operation to run, running it on the real geometry, and telling you the result. That distinction is everything.

"Talk to your map" is a phrase doing a lot of work in marketing at the moment, so it is worth being precise about what the useful version is.

The two designs, and why only one works

Design one: send the data to the model and ask. Upload your features, ask "how many parcels are within 500 m of a well", let the model work it out.

This does not work. Models do not do geometry. It will produce a confident number that is not right, and you will have no way of knowing. It also means shipping your entire dataset to somebody else's server.

Design two: give the model tools. The model never sees the geometry. It sees a description of your layers, and a list of operations it can invoke. You ask a question; it decides "this is a buffer, then a select-by-location, then a count"; the software runs those, on the real data; the model reports what came back.

The arithmetic is done by the same geometry library that a desktop GIS would use. The model is a translator between your English and the operations, not a calculator. When it gets something wrong, it gets the choice of operation wrong, which is visible and checkable, rather than the number wrong, which is not.

Geo Hound's assistant is the second design: 26 tools, all executed in your browser.

What it can carry out

Questions of the data. "Which suburb has the most building consents?" "How many parcels are over two hectares?" "What is the average depth of the wells?"

Styling, which is the surprising one. "Make the flood zones orange and label them by name." "Colour the parcels by zoning." Everyone expects the analysis and nobody expects that you can just say what you want the map to look like. It turns out to be what people use most.

Real analysis. "Buffer the wells by 500 metres and tell me how many parcels fall inside." "Join the soil types onto my paddock boundaries." "How far is each school from the nearest fire station?" "Clip the vegetation layer to my property."

Chains. This is where it earns its keep:

1. "Buffer the streams by 20 metres."
2. "Select the parcels that intersect that buffer."
3. "Save those as a layer and colour them red."
4. "Export it as a shapefile."

Four sentences. In a desktop GIS that is four dialogs, each with parameters you need to know the meaning of.

Things with no buttons. Geo Hound's assistant can clip, erase, and run the field calculator, none of which have an interface in the workbench. Asking is the only way to reach them.

What it cannot do

It cannot fix bad column names. If your fields are F1, F2, and VAL_3, the assistant has the same problem you do. It will read the schema and sample values and sometimes work it out, but ask it about "the depth field" when there is no depth field and it will pick something. Look at what it chose.

It cannot do rasters, or scale. Same limits as any browser GIS: no elevation analysis, no imagery, not hundreds of thousands of features.

It cannot know your intent. "Near the river" is not a distance. It will pick one. If it matters, say the number.

It cannot be trusted blind. It is a competent operator who has never seen your data before and cannot ask a colleague. Look at the result on the map. If a buffer count looks wrong, it probably is, and usually because it grabbed the wrong layer or your layer only had the features in the current view loaded.

What leaves your browser

Worth being blunt, since this is the part people should ask about.

The model gets: your layer names, geometry types as words, feature counts, field names and types, and the results of the operations it runs. Those results are capped hard, at most 20 attribute rows from a query, 10 sample values per field, values truncated at 200 characters.

The model never gets: geometry, or whole datasets. It cannot, because the tools run in your browser and only their summaries go back.

In Geo Hound's case the route is your browser, to our server, to OpenRouter, which routes on to a model provider. Conversations are not stored on our servers. The privacy policy has the full detail, including what is stored and what is not.

If you are evaluating any other tool that offers this, that is the question to ask: does my data go to the model, or do tool results go to the model? The answer tells you both how private it is and how accurate it can be, because they are the same design decision.

Try it

The assistant lives in the Geo Hound workbench, on top of whatever public map data you were already looking at. Up to ten free questions a month, then credits you buy once that never expire. One credit per question you ask, however many operations it takes to answer.