In one session in October, I had an AI coding tool download and join three public data portals for my own region: the city’s, the regional government’s and the police service’s. Collisions, break-ins, zoning, census areas, a business directory, traffic counts. The question was simple. What could a local business buy from this, every month, that it can’t get today?
The joining was fast. That’s the part AI has made cheap. What decided the answer was four checks afterwards, and two of my favourite ideas didn’t survive them.
1. Check what the data holds, one category at a time
The city publishes every service request it receives. The five biggest enforcement categories have no location. Parking, property standards, animals, winter and waste: no coordinates and the address blanked. Infrastructure requests (permits, trees, streetlights, sidewalks) are 70 to 100 percent mapped.
Cities locate what they fix and blank what they enforce. That one finding killed two ideas: parking-ticket hotspots for fleets and complaint load per block for landlords. Both need a freedom-of-information request or a deal with the city, not open data.
It wasn’t the only gap. The rental licence list has no download, so landlord risk can’t be tied to licensed units. And the regional “business count” column is empty in every row. I’d have designed around it if I hadn’t looked.
2. Divide by exposure before you call it a hotspot
The top ten roads by number of crashes and the top ten by crash rate are different roads. The data covered 10,318 collisions from 2022 to 2024. Divided by traffic (crashes per million vehicle-kilometres), the median road segment scored 2.5 and the worst between 40 and 90. One road that ranked fourth by raw count came out at 7.9, close to normal. It’s busy, not dangerous.
That gap is the sellable part. Anyone can count crashes on a map. The raw list is what your buyer already has.
There was a trap inside the fix, too. Crashes at intersections snap to very short road stubs, which inflates their rate. So the unit you sell has to be the intersection, not the segment.
3. Find the incumbent before you believe “nobody does this here”
I started out sure nobody covered my region. A desk check found three who did. A planning-data service at $599 a month that covers 54 municipalities, including mine. A weekly planning newsletter at $468 a year that has covered the region since 1998. A free app, launched in September, that maps permits and applications from the same open data.
The room was narrower than I thought, and more specific. Nobody turns the smaller towns’ council agendas into a list, because those towns don’t publish their applications. Nobody compares the zoning maps week to week. Nobody sells at a local price. And nobody lets you pin a parcel and get told when something near it changes.
That’s a product. “A map of everything” wasn’t.
4. Ask whether they pay once or every month
The first thing I built was a report on one property: zoning, hazards, neighbours, safety, and what the data couldn’t answer. It held up against a real development file. It’s also something a buyer pays for once.
I want recurring revenue, so three days later I scored six recurring versions against what competitors charge and what buyers already pay for. The winner was a weekly digest of planning activity plus alerts on the parcels you’re watching. The one-off report became its paid front door at $149, not the product.
Build it wide, sell it narrow
Every police, census, business and traffic layer was already region-wide, so I’d build for all eight municipalities from day one. Only zoning, planning and parcels are per city. Building for one city and widening later is the same work plus a rebuild.
But I’d sell in one city first, because a pilot needs a named place and people you know. The price and the name cover the whole region, so the second customer, one town over, is a settings change.
Then test before you build the rest
The next step isn’t more software. It’s one real first issue, made from data already pulled, sent by hand to twenty people (planning consultants, commercial brokers, trades and suppliers) with a founding price of $29 a month. Three sign-ups in fourteen days is a go. Fewer, and I’ve lost two weeks, not a quarter.
That’s the plan part. AI made the build cheap enough to start on a whim. It didn’t make the checks optional. If you’re about to build on someone else’s data, run all four first: what it holds, what it means per unit of exposure, who already sells it, and how often the buyer pays.