How I Used OpenAI Agent Mode to Automate My Real Estate Market Reports

Every month I spend over two hours doing the same thing: pulling market stats from my MLS PDFs, extracting the numbers city by city, and reformatting everything into graphics for my market update videos. It’s tedious, error-prone, and the kind of work that adds up fast when you’re covering not just Austin but all the satellite cities — Bastrop, Pflugerville, Cedar Park, Dripping Springs, and more.

My MLS only provides aggregated data for Austin. The satellite cities are buried inside a PDF, each on a different page, and each one requires me to track down the right row: combined residential, not single-family only, not condos — the right number. Get one wrong and the whole graphic is off.

That’s why I decided to test OpenAI’s new agent mode to see if it could do this work for me.

What OpenAI Agent Mode Actually Does

Agent mode is available on the ChatGPT Plus plan ($20/month). It works differently from a normal ChatGPT conversation. Instead of responding to each message, the agent takes a set of instructions and works independently — opening pages, reading through documents, extracting data, and completing tasks without you watching every step.

The key distinction is that it operates more like an employee you’ve delegated a task to, not a chatbot you’re prompting back and forth. You write your instructions once, attach your files, and walk away.

One thing to know upfront: the credit system matters. Each new instruction or follow-up question counts as a credit, and Plus members get roughly 35 credits per month. That means you want to front-load your prompt with as much specificity as possible. If the agent has to ask clarifying questions, those eat into your credits.

The Setup: What I Gave the Agent

My workflow starts with the monthly market stats PDF my MLS publishes. It’s a large document, and PDF is not a great format for AI to read — the extraction often goes sideways because the model has to convert it to text first, and that conversion introduces errors.

Knowing this, I gave the agent very specific instructions upfront:

Which cities to extract data for
The exact page numbers for each city in the PDF
Which row to pull: combined residential only, not segmented property types
Which data points to capture: median sales price with percentage change, closed sales with percentage change, sales dollar volume with percentage change

For Bastrop, for example, the data was on page 31. The month showed 10 closed sales, down 37.5%. Median price: $405,635, up 28.7%. Days on market: 43. I included an example like this in the prompt so the agent knew exactly what format I was looking for.

The final instruction: compile everything into a CSV file.

What Happened During the Run

Once I submitted the prompt and attached the PDF, the agent turned on a reading mode — something different from how a standard ChatGPT conversation handles documents. I could watch the activity log in real time. It worked through the pages, flagging when it couldn’t find a city (Westlake Hills wasn’t in the document) and pushing forward on the others.

When it hit something uncertain, it made a judgment call rather than stopping to ask me. That’s the point of agent mode — it’s built to operate with less hand-holding.

The output was a CSV file with all the cities listed, each with their current month’s closed sales, median price, dollar volume, and percentage changes. Exactly what I needed.

I did spot-checks against the PDF. The data came back accurate for the cities that were present in the document. Cities that were missing got flagged correctly.

What Didn’t Work (And Why)

The original plan was to take it one step further: give the agent access to my Canva account and have it update the infographic templates directly. I tried it. Canva blocked the login attempt because the agent operates through what appears to be a virtual machine or proxy, and Canva’s security flagged it.

So the Canva automation didn’t work. At least not yet.

That said, I still cut the hardest part of the process — the data extraction — down from a multi-hour manual exercise to a task I can hand off entirely. I take the CSV, upload it to Google Sheets, and copy the numbers into my templates. That step takes minutes, not hours.

Why This Matters for Your Marketing

The reason I put this much effort into market data isn’t the data itself — it’s the lead generation engine behind it.

When I create a market update video for my seller-focused YouTube channel, I offer the full market stats graphic as a download. Viewers click the link, enter their email, and get the PDF. That’s a lead. I put them into my CRM and start sending them market update newsletters.

The workflow is: AI agent extracts data → I format the graphic → I shoot the video → the video drives email captures → the email list becomes my lead pipeline.

The data collection was always the bottleneck in that process. Now it’s not.

If I’m being realistic, the agent got around 80-90% of the data right on the first pass. For a task this tedious, that’s already a significant win. And the more specific your prompt, the better the accuracy gets over time.

What This Version of AI Agents Can and Can’t Do

Agent mode is genuinely new. It can browse documents, extract structured data, visit websites, and take action on platforms where it can successfully authenticate. It works best when you give it extremely clear, specific instructions and when the data source is reasonably well-structured.

It doesn’t handle ambiguous instructions well. It can’t break through security walls like Canva’s VPN detection. And the credit limits on the Plus plan mean you need to be thoughtful about how you write your prompts — ask everything upfront rather than iterating.

What it does well: repetitive, structured data work. Anything where you’re doing the same extraction task every month from the same kind of source document. That’s exactly the kind of work where delegating to an agent makes sense.

The Bigger Picture

This is the beginning of a workflow I plan to build out further. Right now, the agent extracts the data and hands it to me. The next version would have it push directly to a spreadsheet, trigger a Canva API call to update the template, and queue the graphic for review. The Canva step is currently the blocker — but that will change.

For now, if you spend any significant time on repetitive data collection for your real estate marketing, agent mode is worth testing. You’ll need a Plus subscription and patience to write a solid initial prompt. But the payoff — getting hours of manual work done in the background while you’re doing something else — is real.

You can see the full workflow demo in the video above.

If you want to build out similar automations for your own business, subscribe to the newsletter — I document these workflows there. And if you’re curious about the broader set of AI tools I use in my real estate practice, the tools page has my current stack.