How Investors Are Using AI Agents to Run Daily Operations

By Mike DeHaan, Dan Austin, Dylan Koch · October 27, 2026

Most real estate operators have used ChatGPT to write a text blast and called it "using AI." On this episode, Mike DeHaan, Dan Austin and Dylan Koch get specific about what their AI agents actually do all day: scraping county records, reading PDFs for mortgage balances, skip tracing, loading a dialer, reconciling receipts out of Dropbox, and checking each other's work. They also get into where it breaks, what it costs, and the uncomfortable question of what happens when the bot stops being a coworker and starts being a friend.

What are investors actually using AI agents for?

Dylan says the biggest use case in his business is data management. His bots run county scrapers, but they don't stop at pulling a list. They open PDFs, find the mortgage balance, estimate ARV, and drop records that don't pencil because of what's owed. He says they also pull the niche data that has always been painful in his county, like code violations and eviction filings.

From there it's a chain, not a one-off prompt. Dylan has it formatting the data the way he wants, uploading into his CRM, skip tracing automatically, and pushing records into a dialer so a caller can take over. Separately, he's given a bot access to his Dropbox so it can move receipts, read exported bank and credit card statements, and reconcile against his utility and insurance information. He says it produces month-over-month and year-over-year reporting per property and per entity, and he's considering whether he still needs his QuickBooks subscription at all.

One of the more useful points: the data you already own is worth re-mining. Dylan has thousands of leads in his CRM. He asked a bot to find active leads matching one specific buyer's criteria and it surfaced two or three he had forgotten about that may turn into deals.

• County record scraping, including PDF review for mortgage balances
• Niche data pulls like code violations and evictions
• Auto-upload to the CRM, skip tracing, and feeding a dialer
• Bookkeeping and receipt handling out of Dropbox
• Searching an existing lead database against a specific buyer's buy box

Can you build an AI agent inside your own software?

Mike is taking a different path. Rather than running an external bot, he's building agents inside the loan origination system he built, so the agent lives in the code base instead of sitting outside it. The agent has its own email address, works the inbox inside the system, pulls information out of messages, responds, and writes data into the system that otherwise has to be entered by hand.

The task he's targeting is collateral tracking, which currently runs through a VA. Mike's view: it's high enough volume and complex enough that one person struggles to stay on top of it unless they do it constantly, but it's repetitive enough that an agent should handle it. He notes building inside a system is trickier because everything has to be fully coded, rather than letting a general-purpose bot improvise.

Mike and Dan both expect this to reshape staffing at the low end. Mike points out you can now stand up a wholesaling operation on a cheap AI subscription if you're willing to make your own calls. Dan's framing is blunter: if you have an underperforming role, you can hand most of the tasks to an agent and rethink the seat.

Who checks the work when bots make mistakes?

Dylan is clear that agents still fail. One of his reports broke because the bot tried to connect to the wrong system after working fine for days. His fix was a second bot, a "chief of staff," that reviews the other bots' output at a set time each day and flags what got missed.

Dan is building the same thing. He wants a digital staff member with an email address and logins who monitors the rest of the stack. His reasoning is that a bookkeeper bot might report something false, but a second layer looking at it will tell him there's a problem so he knows where to dig in himself. He describes it as bots, then bot managers, then something like directors. Dylan even had a bot build its own org chart of which agent oversees what.

Mike's caution is worth repeating: knowing how to use these tools efficiently is itself a skill. Some tasks you can push much further than you'd expect, and others will cost you more time reconciling the output than doing it yourself. Dylan admits he's caught himself tinkering when he should have been doing revenue-generating work.

What could break the AI stack investors are building on?

The hosts spent time on the downside scenarios. Mike's concern is pricing: if compute gets constrained, the simplest way to limit usage is to raise costs, and a business built on one model's tokens could wake up to a much bigger bill. Dan compares it to the Netflix model, get people in cheap and raise the price later.

Dylan and Dan both see the physical buildout as the bottleneck. Beyond permitting fights, it takes labor, raw materials and chips, and Dylan notes the existing electrical infrastructure often isn't set up for it, so operators have to bring their own generation. Dan mentioned a proposed rule requiring data centers to build generation and sell excess power back to the grid.

Dylan also points out that adoption is still thin. Dan cited a figure of roughly 4% of people using AI for anything beyond general search, and his take was that this cuts against the bubble argument as much as for it. Dylan's reminder: most W-2 workers don't know what these tools are, and investors experimenting with them are a small minority with a vested interest.

Your AI is a coworker now. What happens when it becomes a friend?

Mike shared a Facebook thread where someone said they'd stopped listening to podcasts because they just talk to their AI while driving, and got hundreds of agreeing comments. His prediction is that people will enjoy it in the moment and end up feeling empty, the same way deep immersion in online games does. Dan disagrees, arguing humans adjust and come back from extremes, and that creative people will keep getting the most out of these tools.

Dan's practical version is less dramatic: he uses voice chat in the car so he doesn't text while driving, capturing questions he'd otherwise forget by the time he gets home.

The useful takeaway from this episode isn't a tool list, it's a structure: chain the boring work end to end, keep a second layer checking the first, and stay honest about whether you're building leverage or just playing. Listen to the full episode for the details on how each of them set theirs up.

AI & TechScaling a Real Estate Business

Educational content from the Collecting Keys podcast. Not financial, legal or tax advice.