Practical AI for Architecture: Automating Site Feasibility Reports

The Core Problem: Billable Hours vs. Speculative Work

Every architect knows the grind of the pre-design phase. A potential client has a parcel of land and a dream. Before you can even sketch a floor plan, you’re on the hook for hours of unbilled research. You’re digging through arcane municipal websites, deciphering 200-page zoning PDFs, and calculating floor area ratios (FAR). This is speculative work, and if the project doesn’t move forward, you just ate that cost. This article shows you how to build a semi-automated system that uses AI for architecture to produce a comprehensive site feasibility report in under an hour.

This isn’t about replacing architects. It’s about front-loading intelligence and killing bad ideas faster, so you can focus your expensive human time on the projects that actually have legs.

What Most “AI for Architecture” Guides Get Wrong

Most of the discussion around AI in this field is, frankly, a distraction. It’s dominated by stunning, impossible renderings from tools like Midjourney. These images are great for mood boards but useless for actual construction. They show you a fantasy, not a workflow. They can’t tell you the required side-yard setback, the parking space minimums, or the local stormwater management requirements.

My gripe is that this focus on eye-candy completely misses the point. The real value isn’t in generating a pretty picture of a building that can’t be built. It’s in automating the tedious, error-prone, and non-creative work that precedes design. It’s about using language models to do what they’re good at: reading, summarizing, and structuring vast amounts of text-based information. The flashy images are a fun party trick; automated document analysis is a business tool.

The Stack: Building Your Feasibility Agent

You don’t need to be a coder to build this. You just need to connect a few powerful tools. This isn’t a single push-button solution, but an assembly of parts that work together. Think of it as an operator playbook for your firm.

Here’s the stack I use:

• Web Scraper: You need to pull data from the web. Specifically, the local zoning authority’s website. I use Phantombuster for this because it has pre-built ‘Phantoms’ that can handle simple scraping tasks without code. You point it at a URL, tell it what text to look for, and it extracts it. Browse AI is another solid option if you prefer a record-and-play interface.
• Document Analyzer (LLM): This is the core of the system. You need a large language model with a massive context window and strong reasoning capabilities to digest the zoning PDFs. As of 2026, Claude 3 Opus is the best tool for this job. Its 200K token context window means you can drop an entire municipal code into the prompt without it breaking a sweat. GPT-4 is a close second, but I find Claude makes fewer errors on dense, legalistic text.
• Report Generator (LLM): You can use the same LLM (Claude or GPT-4) for this. After extracting the raw data, a second prompt will format it into a clean, client-facing report.
• Conceptual Massing (Optional): Once you have the constraints (e.g., max height, setbacks, FAR), you can feed them back into an image model like Midjourney to create a visual representation. This is the last step, not the first.

Is the Free Tier Actually Usable?

This is a common question, and the answer is a hard no. The free tiers of these services are for testing and hobbyist use, not for running a business process. Trying to run this workflow on free credits is a recipe for frustration.

Let’s talk real costs. Phantombuster starts at $69/month, which gives you enough credits for dozens of scrapes. The real cost is the LLM. You’ll be using the API, not the public ChatGPT interface. A single, complex analysis of a 150-page zoning document with Claude 3 Opus might cost around $2-$4 in API tokens. If you run 10 feasibility studies a month, you’re looking at maybe $30-$40 in API fees. So, all-in, you should budget around $100/month to run this system reliably. Honestly, considering it can save 5-10 hours of an architect’s time *per project*, $100/mo is an absolute steal. It pays for itself the first time you use it.

A Step-by-Step Workflow: From Blank Page to Preliminary Report

This process takes what was once a multi-day research project and compresses it into about 45 minutes of active work. The key is structuring your prompts correctly and verifying the output.

Here are the steps:

1. Data Acquisition: Go to the city or county’s planning and zoning website. Find the page for the specific parcel address. Use Phantombuster to scrape the key text on that page: Zoning Designation (e.g., R-1, C-3), Parcel ID, and any direct links to PDF documents for that zoning code.
2. Document Ingestion: Download the relevant zoning code PDF. It will be dense and terribly formatted. That’s fine.
3. Constraint Extraction Prompt: This is the most important step. You’ll feed the entire text of the PDF into Claude 3 Opus. Don’t just ask it to “summarize the zoning.” You need to be specific. Here’s a prompt structure I’ve had success with:
   "You are an expert zoning analyst assisting an architecture firm. I have provided the full text of the zoning code for [City Name], specifically for the [Zoning Designation, e.g., 'C-2 General Commercial'] district. Your task is to analyze this document and extract the following specific constraints for a new construction project. Present the output in a clear, structured format.

   1. **Permitted Uses:** List the primary permitted uses for this zone.
   2. **Building Height:** What is the maximum allowable building height in feet and/or stories?
   3. **Setbacks:** What are the minimum required setbacks for front, side, and rear yards? Note any special conditions for corner lots.
   4. **Lot Coverage:** What is the maximum percentage of the lot that can be covered by buildings?
   5. **Floor Area Ratio (FAR):** What is the maximum FAR?
   6. **Parking Requirements:** Summarize the off-street parking requirements (e.g., spaces per 1000 sq ft of commercial space).
   7. **Landscaping:** Are there any specific landscaping or open space requirements?

   For every piece of data you extract, you MUST cite the page and section number from the source document. Format citations like this: [Page 45, Section 11.2A]. Do not invent information. If a constraint is not mentioned, state 'Not specified in document'."


4. Verification: The AI will give you a beautifully formatted list. Your job is to spot-check it. Because you forced it to include citations (which, yes, is the only way to make this trustworthy), you can quickly jump to Page 45, Section 11.2A and confirm the number. This turns an hour of reading into five minutes of verification.
5. Client Report Generation: Take the verified output from the previous step. Feed it into a new chat with your LLM. Prompt it: “Using the following verified zoning data, write a concise, one-page summary for a non-technical client. Explain what each constraint means in simple terms. Adopt a professional but clear tone.”

The output is a clean, professional-looking document you can send to the client, positioning you as efficient and thorough from day one. My favorite part is seeing the client’s reaction. They get a detailed report in a day, not a week, which builds immense trust right away.

How to Debug When It Hallucinates Zoning Codes

It will happen. At some point, the AI will confidently state a setback is 10 feet when the document says 20. This is especially common with complex tables or nested regulations. Without a debugging process, this entire workflow is dangerously unreliable.

This is why the citation prompt is non-negotiable.

When an output looks suspicious, or for any critical number like height or FAR, you must check the source. The citation prompt `[Page X, Section Y.Z]` is your debugging tool. It forces the model to show its work. The most common failure mode is the AI misinterpreting a table column or missing a footnote that modifies a rule (e.g., “…setback increases to 25 feet on lots adjacent to residential zones”). If you see a number that feels wrong, go directly to that page in the PDF. Nine times out of ten, you’ll find the nuance the AI missed. You can then correct that one data point manually or re-prompt the AI with a more specific query about that section, asking it to pay attention to the footnotes. This verification layer is what separates a professional tool from a toy.

If you skip the verification step, you will eventually give a client wrong information, and that mistake will be on you, not the algorithm.

You can build this entire automation yourself following the steps above. It will take some trial and error, particularly in refining your prompts for the specific language your local municipality uses in its codes.

We cover this in more depth elsewhere — deeper coverage of AI agent platforms.

If you’d rather skip the build and deploy a working version in an afternoon, we’ve packaged this entire workflow, complete with tested prompts and tool recommendations, as a deployable blueprint. You can find it at deepusecase.com/vault.