Urban intelligence

Overview

Urban intelligence is the use of data and AI to understand cities at the level of individual parcels, buildings and rules. It brings together maps, property records, zoning and building regulations, permits, transport access, demographics and market data to answer practical questions. What could legally be built on this lot? How much floor area is still unused? Which streets are likely to change in the next five years? Where would new housing have the best access to transit and schools?

Where a smart city programme focuses on running services such as traffic, water and waste, urban intelligence focuses on land, buildings and development. Its users include planners, architects, developers, lenders, insurers, investors, community groups and residents who want to understand what is possible near them.

An everyday picture is an experienced planning consultant. Asked about a site, the consultant pulls the map, reads the zoning text, checks overlays and past permits, looks at nearby sales and gives a reasoned opinion. Urban intelligence systems do that reading and cross-checking at city scale, in seconds, and show their sources so a professional can verify the answer.

What it is

The core analytical object is the parcel joined to everything that applies to it. Once each lot carries its zoning district, overlays, existing building and surroundings, many questions become computations: subtract built floor area from permitted floor area to estimate unused development capacity, or filter all lots where a given use is allowed within a ten-minute walk of a station.

Where it came from

Land-use regulation is long established. New York City's 1916 Zoning Resolution is often described as the first comprehensive zoning code in the United States, and the US Supreme Court upheld municipal zoning in Village of Euclid v. Ambler Realty Co. in 1926. Many other countries developed their own planning and building codes through the twentieth century, so the rules that urban intelligence systems read are often decades old, heavily amended and written in dense legal language.

On the data side, geographic information systems became standard in planning departments from the 1980s and 1990s. In the 2000s and 2010s many cities published parcel maps, zoning layers, building footprints and permit records as open data, and property-technology companies began combining them with market data.

Since the early 2020s, language models have added a new ingredient: the ability to read regulatory text and connect it to spatial data. Retrieval methods such as agentic RAG and GraphRAG, and text-to-SQL for querying property databases, let systems answer questions in plain language while citing the rules and records they used.

How it works

Step 1: assemble and align spatial data

Load parcels, zoning districts, overlays, building footprints and other layers into a spatial database, such as PostgreSQL with PostGIS. Reproject them to one coordinate system, fix invalid geometries and decide how to treat lots that straddle district boundaries.

Step 2: join records to parcels

Link property records, permits and sales to parcels through identifiers or addresses. Address matching is messy; expect to spend real effort on normalisation, geocoding and resolving duplicates.

Step 3: encode the rules

Translate zoning text into structured parameters: permitted uses, maximum floor area ratio, height limits, setbacks, lot coverage and parking requirements, per district and overlay. Some rules are simple tables; others have conditions, exceptions and bonuses. Language models can draft this extraction, but experts must review it, and every parameter should link back to the clause it came from.

Step 4: compute derived measures

For each parcel, compute permitted floor area, existing floor area, unused capacity, distances to transit and amenities, and flags for constraints such as flood zones or historic status.

Step 5: model and predict

Use machine learning for questions that rules alone cannot answer: predicting which parcels are likely to be redeveloped, estimating values with gradient boosting on property and location features, or detecting new construction from imagery with computer vision.

Step 6: answer questions with evidence

Provide map views, parcel reports and natural-language answers. A good answer cites the zoning clause, the data source and its date, and states uncertainty, for example where a rule needs professional interpretation.

Key concepts and terms

A worked example

This scenario is illustrative. A housing nonprofit wants to find sites in one city where small apartment buildings could be built as-of-right near transit.

The team loads 180,000 parcels, the zoning map, overlays, building footprints and station locations into a spatial database. It encodes the residential districts' floor area ratios and height limits from the zoning text, with each value linked to its clause. For each parcel it computes permitted floor area as lot area times FAR, subtracts existing floor area from building records, and measures walking distance to the nearest station.

Filtering to parcels with at least 1,500 square metres of unused residential capacity, no historic or flood overlay and a station within 800 metres leaves 2,140 candidates. A redevelopment model trained on ten years of permits ranks them by likelihood of change, which highlights single-storey commercial lots with large car parks. Staff review the top 100 by hand and find that 14 have rule interpretations that need a planner's opinion, which the report flags rather than guesses. The shortlist goes to site visits and owner outreach.

Strengths and limitations

Strengths:

• Turns scattered maps, records and rules into consistent parcel-level answers.
• Saves hours of manual research per site.
• Makes city-wide screening of opportunities and risks possible.
• Can make planning information more accessible to residents.
• Supports evidence-based policy, such as testing rezoning scenarios.

Limitations:

• Zoning text is complex, ambiguous and frequently amended.
• Public data can be outdated, incomplete or inconsistent between departments.
• Address and parcel matching errors propagate through every result.
• Models of future change can reflect past patterns, including inequitable ones.
• Answers support, but do not replace, professional and legal review.

How it compares with alternatives

Practical guidance and best practices

1. Make the parcel, with a stable identifier, the centre of the data model.
2. Record the source and date of every dataset and rule.
3. Link every encoded rule to the exact clause it came from.
4. Have planning experts review rule extraction, especially conditions and exceptions.
5. Track zoning amendments and re-run affected parcels when rules change.
6. Show uncertainty and flag cases needing professional interpretation.
7. Validate predictions against later permits and outcomes.
8. Check results for uneven quality or bias across neighbourhoods.

Common mistakes

• Treating a model's reading of zoning text as authoritative without review.
• Mixing datasets from different years without noting it.
• Ignoring overlays and special districts that override base zoning.
• Assuming permitted capacity will actually be built; markets and ownership matter.
• Presenting answers without citations, so users cannot check them.

Frequently asked questions

Is urban intelligence the same as GIS?

GIS is a core tool, but urban intelligence goes further by encoding regulations, joining many records to each parcel and adding prediction and natural-language interfaces. It is an application area built on GIS, databases and AI.

Can AI read zoning codes reliably?

Language models can extract many rules well and greatly speed up the work, but zoning texts contain conditions, cross-references and exceptions that are easy to misread. Reliable systems combine extraction with expert review, link every value to its source and flag uncertain cases.

Who uses urban intelligence?

Planners and policy teams, architects, developers, lenders and insurers, investors, nonprofits and community groups. Residents increasingly use it to understand what might be built near them and how to take part in planning decisions.

How accurate is unused development capacity?

It is only as accurate as the lot areas, existing floor areas and encoded rules. It is a screening measure that shows where to look more closely, not a guarantee of what can be built on a given site.

Does this replace planning professionals?

No. It removes much repetitive research and makes information more accessible, but interpreting complex rules, negotiating approvals and judging design and community impacts remain professional tasks.

Glossary

• Urban intelligence: data and AI applied to land, buildings and urban rules.
• Parcel: a legally defined lot of land.
• Zoning: rules that regulate land use and building form by area.
• FAR: floor area ratio, a common density limit.
• Overlay district: an extra set of rules on top of base zoning.
• PostGIS: a spatial extension for the PostgreSQL database.
• Geocoding: turning addresses into coordinates.
• As-of-right: development allowed without discretionary approval.

Summary

Urban intelligence joins parcels, zoning rules, building records, permits, market data and spatial context so that questions about land and buildings can be answered quickly and with evidence. It builds on a century of land-use regulation and decades of GIS, and language models now help read regulatory text and answer questions in plain language. Its value depends on clean parcel-level data, rules linked to their sources, expert review and honest uncertainty. This page is general information, not legal advice.

Related

It belongs to Industries and applications in my stack, under Machine learning. Others in the same group: