Collecting Keys - Real Estate Investing Podcast

AI & Tech for Real Estate Investors: What Actually Works

Mike DeHaan, Dan Austin and Dylan Koch have tested AI tools, CRMs, property management software and virtual assistants across years of wholesaling, flipping and rentals. This guide collects what they've found genuinely useful, what they've watched cost people money, and how their thinking shifted from early enthusiasm to hands-on skepticism.

Start with these episodes

Where does AI actually help in a real estate business?

On EP 99 in early 2023, Mike walked through the first three jobs he handed to ChatGPT: writing job listings tailored to a specific role and industry, producing marketing copy for ads, social posts and direct mail, and drafting standard operating procedures. He prompted it for a 300-word letter to a motivated seller and got multiple versions to split test. The online copy came back SEO-checked and formatted with headers that dropped straight into their WordPress plugins.

The pattern the hosts keep returning to is that AI is good at structure and bad at knowing things. On EP 99, Mike said AI got their SOPs roughly 85% of the way there — the missing piece was the internal detail unique to their business. On EP 479, his hands-on testing found it worked well for reformatting structured data like marketing lists, but returned inconsistent and sometimes wrong results when he fed it the same documents and instructions twice.

The clearest win came on EP 488: over a weekend Mike used Claude to combine investor rate sheets into a single embeddable loan sizer, so a new loan officer can quote a fix-and-flip loan in about ten seconds. He estimated four to five hours of work — and said the tool confidently returned wrong numbers until he clicked through every setting to verify the math.

From: EP 99 · EP 488 · EP 479

Where does AI cost you deals?

EP 485 is the bluntest episode on this. The hosts cite a reported case where Ryan Serhant lost a $50 million deal — roughly $1.5 million in commission — after both buyer and seller asked AI whether they were getting a good deal and got opposite answers from the same information. The point isn't that AI is evil; it's that a confident wrong answer handed to either side can blow up an agreed transaction.

On EP 462, Mike and Dylan tested ChatGPT on market data live. It produced a market map with cities in the wrong states and the colors reversed, and when asked about Cincinnati it named what the hosts considered the worst possible ZIP codes to invest in. Their summary: LLMs predict a satisfying answer, not a correct one.

The second failure mode is opportunity cost. On EP 485 the hosts argue that at small scale, more sales beats more automation, and that most AI use cases investors brag about already existed — connecting a lead form to a CRM is Zapier, which has been around for over a decade. EP 488 goes further: building your own AI agents is usually a deflection from the marketing and sales work that actually grows revenue.

From: EP 485 · EP 462 · EP 488

Should you build your own software or just buy it?

The hosts' view has genuinely evolved here. On EP 480 they argued the durable value from AI will sit with software companies that embed it, not with individuals building their own tools — the same way almost nobody builds their own website or cloud storage anymore. EP 488 reinforced it: someone working full time in your industry will probably ship something better and cheap within a year.

By EP 510, after RESimpli's 6.0 rollout reportedly broke core functions for users, Mike had changed his tune on the narrow stuff. He built his own e-sign tool with Claude in about an hour after PandaDoc quoted $1,000 per seat per year plus $2 per document, and rebuilt his loan origination system the same way. His method: point Claude at the platform's tutorial videos, tell it to build the same thing and get as far as it can without asking questions, then host it somewhere like Netlify.

Both things can be true. Simple internal tools are now cheap to clone. Anything touching texting, calling or money is still hard — EP 510 flags 10DLC rules as the main barrier, with services like Batch or JustCall as a bridge. And EP 488 raises platform risk: an Anthropic billing change reportedly took some Open Claw setups from about $200 a month to $200 a day.

From: EP 510 · EP 488 · EP 480

Do virtual assistants beat automation for a small operator?

Long before AI, the hosts' answer to repetitive work was people. On EP 39 they described using overseas VAs for lead management, marketing, bookkeeping and return-to-sender skip tracing, hiring directly through Upwork, Fiverr and OnlineJobs.ph rather than paying an agency — Mike was quoted $2,500 just to source an admin VA. They pay their Philippines-based VAs about $5 an hour. Their screening trick: build an odd request into the job post, like a voice recording naming a favorite food, to confirm the applicant read it and speaks English well.

EP 84 lays out Dan's three-layer self-management stack: property management software (DoorLoop, after trying Hemlane, AppFolio and RentReady), Asana for recurring tasks with Loom videos attached, and a VA acting as remote property manager at roughly $200 a month for about ten hours a week. Digital locks issuing hour-long expiring codes handle most showings.

On SOPs, EP 92 explains why they record a short screen video and have a VA turn it into a document — when the software changes you re-record five minutes instead of rewriting everything. On EP 358, Dylan's first executive assistant hire worked because he already knew the person, still ran DISC assessments and multiple interviews, and handed off tasks via Loom, replacing a three-hour weekly foreclosure data pull.

From: EP 39 · EP 84 · EP 92 · EP 358

Which parts of the transaction is tech most likely to disrupt?

On EP 479, Dan argues title, escrow and county recording are the functions most likely to be reshaped by technology, but expects slow adoption because regulation is jurisdiction-by-jurisdiction and the industry moves at its own pace.

EP 493 goes further on valuation. The hosts say large lenders and capital sources are building internal AI valuation tools that compare photos of a subject property against comps, and expect that to make paid appraisals — and the appraisal management companies in between — unnecessary over time. They also mention a broker friend with 2,800 agents who rebuilt a private, searchable MLS for his brokerage in a single weekend, which they read as an early crack in the MLS's monopoly on listing data.

That same episode notes one of their capital partners has flagged the Cincinnati and Northern Kentucky market for wildly inconsistent appraisals and is tightening scrutiny on 30-year debt there. The hosts' explanation is that appraisal problems tend to come from small networks of wholesalers, title companies, agents and appraisers bending rules for each other.

From: EP 479 · EP 493 · EP 485

How do you tell good real estate software from a sales pitch?

EP 374 is the practical filter. Mike explains that many online real estate gurus no longer do deals and rotate from product to product chasing affiliate commissions — a $5,000-a-year product can pay a $1,000 referral fee. He lists the categories that have cycled through: CRMs, skip tracing services, data products, and more recently AI tools. His checks are to look up the company's affiliate program and the promoter's relationship to it, verify the person still actively does deals, and ask active local flippers and wholesalers what they think.

Mike and Dan's own test, described on EP 85 with TwnSqr founder Paul Wakim, is to ask where the value lands. If the only beneficiary is the platform owner, it's a churn business — which is why they called InvestorLift's roughly $6,000 upfront fee and mass-email blasts a waste of money in that conversation.

EP 240 with DealMachine founder David Lecko shows the other side: a tool built from a personal widget for pinning run-down houses that has done roughly $60 million in revenue since 2017. Lecko also shared a cautionary tale — he over-hired engineers as copycat apps appeared and watched a $3 million profit business go to months of zero profit.

From: EP 374 · EP 85 · EP 240

What about crypto and the broader AI trade?

The hosts have been around this block since EP 11 in 2021, when they detoured into metaverse land and Ethereum gas fees — Mike skipped gifting an NFT because transferring it would have cost about $300. On EP 485 they push back on the 'Fannie Mae accepts crypto mortgages' headline, noting it's limited to covering up to $10,000 in closing costs, likely over-collateralized, and reads to them mostly as a lead-generation play by one lender.

On the macro AI trade, the hosts are skeptical but not unanimous. On EP 477, Mike reported from a GoBundance event that sentiment split into extreme bulls and bears with almost no middle, with bulls leaning on AI capex as proof the economy is fine. His bear case: that money flows to shareholders, CEOs and data centers, not to the people who buy houses, and data centers can raise electricity prices in surrounding residential areas. Dylan, citing Luke Roman, argues the US lacks the grid, natural gas, rare earths and skilled labor to deliver the buildout on a one-to-two-year timeline.

Dan disagrees. On EP 496 he argues data centers are a net local positive because they must pay to build generation and transmission infrastructure, which then benefits nearby development and grid resiliency. That same episode flags distorted signals — SpaceX's IPO at $135 a share implying roughly a $1.77 trillion valuation against $18.7 billion in revenue and a $4.2 billion operating loss, with Dylan betting it trades under $135 five years out.

From: EP 11 · EP 477 · EP 496 · EP 488

Frequently asked questions

Is AI reliable for pulling real estate market data or comps?

Not on its own, according to the hosts. On EP 462 they showed ChatGPT producing a market map with cities in the wrong states and colors reversed, and naming what they considered the worst Cincinnati ZIP codes to invest in. Mike's take on EP 479 is that LLMs predict a satisfying answer rather than a correct one.

Should a small investor build their own AI tools?

Mostly no, with one exception. On EP 488 the hosts argue building your own agents is usually procrastination on marketing and sales, and a real vendor will likely ship something better within a year. But on EP 510 Mike cloned an e-sign tool with Claude in about an hour after PandaDoc quoted $1,000 per seat per year.

Do the hosts still recommend virtual assistants now that AI exists?

Yes. Their systems in EP 39 and EP 84 run on VAs at around $5 an hour handling lead management, bookkeeping and remote property management, trained with short Loom videos. EP 488 explicitly argues using AI to cut payroll is usually a smaller win than using it to raise revenue.

Which real estate jobs do the hosts think AI will replace?

On EP 493 they expect AI valuation tools built by large lenders to make paid appraisals and appraisal management companies unnecessary over time, and see private brokerage MLS clones as a threat to the MLS. On EP 479 Dan named title, escrow and recording as most likely to be reshaped, but expects slow adoption.

How do I check whether a promoted software or AI tool is worth buying?

EP 374 lays out Mike's checks: look up the company's affiliate program and the promoter's relationship to it, confirm the person still actively does deals, and ask active local flippers and wholesalers what they think. He also warns about promoters whose 'secret weapon' product changes every month or two.

All 21 episodes on ai & tech