A developer we know tracks time by the hour for every active deal. On a typical pre-acquisition week, he spends 32 hours gathering information. Fourteen of those hours are phone calls to planning departments. Six are spent reading municipal code PDFs. The rest is a mix of GIS screen grabs, soils reports, and the occasional call to a zoning attorney who charges $650 an hour to explain whether a special use permit triggers a new traffic study.
At the end of that week, he has a recommendation memo that is maybe 1,400 words long. The ratio of research hours to deliverable words is worse than it sounds, because two thirds of what he finds is ruled out.
This is the math that AI changes. Not the writing. The gathering.
What entitlement research actually is
Entitlement due diligence breaks down into a predictable set of checks. What is the base zoning and what are its dimensional standards. What overlays apply. What is the plat history and are there lot-of-record issues. What permits are triggered by the intended use. What is the historic pattern of approvals and denials in that jurisdiction. What is the political temperature.
The first five of those are, at root, lookup problems. They are not judgment problems. A junior planner and a senior planner, handed the same ordinance, will arrive at the same setback number. They will disagree on whether a variance is achievable, but they will agree on what the setback is.
The lookup hours dominate the timeline. The judgment hours do not. This is the place where compression is available.
The agent loop, specifically
When the research agent handles a setback question, it does three things in sequence. It identifies the zoning classification from the parcel's location. It pulls the relevant ordinance chapter from the municipal code. It returns the dimensional standards with a citation back to the ordinance section.
On a clean case, that loop takes eight seconds. On a complicated case with overlays, it takes forty. The slowest version is still orders of magnitude faster than the phone-call alternative.
The time saved is not the point, exactly. The point is what it unlocks. A developer who can answer "what is the setback" in under a minute can ask a second question. And a third. And a fourth. The speed of iteration is the speed of thinking.
What does not compress
Political temperature does not compress. A planner who is hostile to a specific use will not become less hostile because the applicant read the code faster. A neighborhood that objected to the last six mixed-use projects will object to the seventh. These are the hours that need to be paid in person, on the phone, over coffee.
The agent does not try to replace those hours. It tries to move them later in the timeline. Gather the factual record first. Make the political calls once the facts are clear.
What the compressed timeline looks like
A pre-acquisition screening that used to take five days now takes one. The first three hours cover zoning, flood, wetlands, soils, topography, and a political scan based on historic approvals. The next three hours cover utility availability and a preliminary site plan fit. The last two are reserved for the planner call and a conversation with counsel.
The five-day version and the one-day version produce the same memorandum. The one-day version is written while the information is still in the developer's head, which is worth more than the four days saved.
The friction that remains
The friction that remains is mostly at the boundary between public sources and paid sources. Soil data is free, if you know USDA Web Soil Survey. Flood data is free, via FEMA MSC. Title work is not free, and the gap between a preliminary title commitment and a final policy is a legal question the agent cannot answer.
For that boundary, the agent is a research partner, not a replacement. It identifies where title work is likely to matter, flags easements of concern from the county recorder's office, and hands the rest to the attorney. That handoff, done well, is the difference between a defensible file and an expensive surprise.