Will AI Replace Urban Planners?

It is 10:40 on a Tuesday night and you are on submission 212 of 400. Your glasses are on the arm of the couch. The hearing is Thursday.

Somewhere around comment 180 you stopped reading and started scanning, and you know it.

That is usually when the thought shows up. Not a headline about robots taking over. Something quieter, which is that a machine could do this in four minutes, and you are not sure whether that is a relief or a warning.

It is both, and the reason is specific to planning in a way that most of the coverage misses entirely.

AI arrived in planning through the permit counter, not through the profession

Nobody in your department went looking for this. It came in through the city manager’s office, attached to a backlog.

Austin brought in automated plan review in late 2024. Los Angeles launched a version of it in April 2025, after the wildfires. Honolulu, where Smart Cities Dive reported a median wait of 393 days for a commercial permit in early 2025, went to two vendors at once. These tools read a submitted plan set against the zoning and building code and produce a compliance report in hours.

That is the ministerial work. Measuring a setback against a standard was never the interesting part of anyone’s job, and handing it to software is not a loss.

Here is what it does instead. A faster intake does not shrink the department, it moves the pressure downstream, toward conditional use permits and variances and rezonings and everything else where a human body has to make a judgment in public. The queue was never caused by planners typing slowly. It was caused by the fact that discretionary decisions take a hearing, a notice period, a staff report, and a vote.

So the volume arrives faster at exactly the place it was always slowest.

What AI cannot do in a land use decision

Three things sit outside this, and they are not sentimental ones.

The first is discretion itself. When an adopted housing policy pushes toward density on a corridor and an adopted neighborhood policy protects established character, somebody has to decide which one governs this parcel on these facts. That authority was delegated by a legislative body to a commission and to staff, and it has never been delegated to a vendor. A tool can lay both policies side by side, which is genuinely useful. It cannot weigh them, and if it produced a weighting you would still have to own it.

The second is the finding. A staff report is not an essay about a project, it is a chain that runs from a quoted criterion, through named evidence in the record, to a conclusion about whether the criterion is met. Each link gets tested separately on appeal. A finding you edited into agreement reads identically to one you reached, right up until somebody asks you to explain the reasoning without the document in front of you.

The third is the room. At the Workplace AI Institute we work with a lot of public sector professionals, and planners are unusual even among them, because the job routinely ends with one person standing in front of people who are upset and entitled to question them. That is not a task with an output. It is an accountability, and accountability does not transfer.

If your own week is heavier on the first two than you would like, the 3-minute readiness check is built around a planner’s actual docket, and it is blunter about where the reading is going than most people expect.

The risk that is specific to this profession

Every profession has an AI accuracy problem. Planning has a version of it with teeth, and this is the part worth taking seriously.

Ask a general AI tool what the accessory structure setback is in your single-family zone and it will answer with a section number. It will look exactly like your citations look, because that is the pattern it learned. It has never seen your ordinance.

In marketing, that is an embarrassing error. In planning, it is a defect in a record that an appellant’s attorney is reading specifically to find. Courts have been dealing with the professional version of this for a while now, and there is a running public tally of decisions where a filing turned out to contain citations that were generated rather than retrieved. Those were filed by attorneys who believed they had checked.

The fix is not complicated, and it is the whole difference between a planner who uses these tools well and one who gets caught. You never ask what the code says. You paste the code and ask which of the pasted provisions govern your facts.

I am a planner answering a question about what our code requires. Below is the text from our zoning ordinance. Work only from this text, and if something is not in it, say so rather than supplying it. The fact pattern is [WHAT IS PROPOSED, IN WHICH ZONE, AT WHAT SCALE]. Tell me which provisions govern, quoting the exact sentence from each with the section number as it appears, and name any provision that could reasonably be read two ways on these facts.

Ninety seconds, and everything it hands back is checkable against text on your own screen.

What to change before your next packet

Two things, and neither of them is a tool purchase.

Start by moving your reading. The ordinance extraction above, a study summary that asks for the assumptions rather than the conclusions, a policy comparison that quotes both sides of a tension. That is where the hours are, and none of it touches your judgment.

Then fix your submission process, because this is where planners lose the most and notice the least. Theming four hundred comments is a frequency operation, and the comment that decides a hearing is very often the one nobody else made. Run a second pass whose only question is which submissions raise something no other submission raises. We have watched planners find a conflict with an adopted subarea plan sitting in comment 311, on the Monday rather than at the podium on the Thursday, and it changes the entire meeting.

For a structured walk through both, with the prompts and the sorting rule that keeps a commenter’s home address off a vendor’s servers, the AI for Urban Planners course takes code interpretation, staff reports, submission analysis, and stakeholder correspondence one task at a time.

So will AI replace urban planners?

No. The Bureau of Labor Statistics has urban and regional planners growing about 4% between 2025 and 2035, on a base of 45,800 jobs, with a master’s degree as the typical entry point. That is slow growth, not decline, and slow growth with rising caseloads has a predictable effect. More of the docket per person, and more of it discretionary.

Adoption is already well past the experimental stage around you. An Arup survey of 5,000 planners, architects, and engineers across ten countries found 42% of the United States respondents using AI tools daily, ahead of the global figure of roughly a third. Something we keep noticing is that the planners furthest along are also the ones with the sharpest answer to where they stop, and those two facts are connected rather than coincidental.

Could you state your own line out loud if a council member asked you at a study session? The readiness check walks a planner’s docket and hands the answer back written down.

Write down the five sentences before somebody asks you for them. What you use it for, what you never use it for, how you verify anything entering a record, how you handle a commenter’s address, and what you would tell a colleague. The planner who can say that out loud is the one who ends up drafting the department’s policy instead of complying with someone else’s.