Ep 5 | AI Breaks Where Liability Lives w/ Daniel Cunningham

AI is getting easier to build and as a result we already see the business community losing sight of what's really important, what's difficult.

Every day there are
New tools.
More automation.
Faster, more human-sounding models.

That’s where most AI conversations go.

And it’s the wrong place to focus.

In this solo episode of The Multifamily Property Management AI Playbook, Daniel Cunningham breaks down where AI actually breaks when used in real life to automate property management operations.

Not in the 80–90% of routine work, the every day issues
but in the moments where things don’t go as expected.

This is where:

- Residents are stressed
- Legal and health issues are at stake
- Policies conflict with reality
- And simple work orders turn into liability

Most AI solutions can handle the obvious…
but they need to hand the problem back when it actually matters.

That’s what's hard these days. That's what vendors selling you something out-of-the-box don't care about or want you to recognize.

You’ll learn:

- Why AI doesn’t break in the majority of work—it breaks in the moments that matter
- What actually happens when workflows hit edge cases and gray areas
- Why “fully automated” AI falls apart in real property operations
- The difference between automation and true operational coverage
- Why human-in-the-loop is the architecture of trust—not a backup plan

🎧 Listen now to understand why the last 10% of operations is where AI wins or loses.

Continue the Conversation

🔗 Connect with Daniel Cunningham on LinkedIn
🏢 Follow Vendoroo on LinkedIn
🌐 Visit Vendoroo's Website

📬 Get insights from the Vendoroo Weekly Newsletter

Chapters:

00:00 AI Automation Fantasy
01:09 When AI Gets It Wrong
03:32 Why Humans Still Matter
05:48 Demo Versus Real Ops
07:21 Designing True Coverage
09:34 Maintenance Edge Case Maze
12:03 Red Flags and Real Limits
13:51 Robotaxis and Audit Lessons
15:55 Vendoroo Trust Architecture
19:54 100 Percent Coverage Promise
21:20 Trust Wins the Future
22:27 Closing and Next Episode

Transcript

SPEAKER_00 0:00

Welcome back to the Multifamily AI Playbook Podcast. Everybody wants the AI story that sounds cleanest. You flip the switch, the software takes over, payroll stays flat, customers are delighted, and somehow the messy parts of business just disappear. That is a beautiful story. It is also, in most real operations, not how trust actually works. Welcome to the Multifamily Property Management AI Playbook. I'm your host, Daniel Cunningham. If you're responsible for running properties and trying to make sense of where AI fits, no hype, just insight, you're in the right place. Presented by Vendoroo, your all-in-one AI solution for resolving maintenance needs. Let's dive into the playbook. The truth is, when AI touches the real world, when it's not just summarizing notes or drafting an email, but it's it's actually taking action across workflows, making judgment calls, routing decisions, communicating with customers, now spending money or handling anything urgent, then the question changes. The question is no longer can AI do most of the work? In many cases, yes. Yes, it can. But the real question becomes: what happens? What happens when it's wrong? What happens when the situation is weird? What happens when policy conflicts with reality? What happens when a resident is upset? A vendor goes dark, an owner needs clarity, or the issue lands in that uncomfortable gray zone that every experienced operator recognizes instantly. You know what I'm talking about. This is where uh the fantasy of full automation starts to wobble. Okay. And that is why I think the real future of AI in operations, at least, is not AI alone. It is AI with human coverage. Not because AI is useless, quite the opposite. Because AI is right now, the AI is powerful enough to run most of the playbook. And that makes it even more important to design for the moments when the playbook is not enough, when the playbook's not right, when the playbook has some gaps in it. And the evidence is piling up. In just the last few months, Reuters reported on new research that's suggesting hallucinations may be baked into large language models more than many buyers want to admit, especially as context gets longer and the tasks get more complex. Thompson Reuters was blunt in its its recent legal coverage, too, that agentic AI is gaining momentum, but adoption depends on clear human oversight. And in courts, AI should remain under human supervision rather than act as an autonomous decision maker. Why? Why are they saying that? Well, Anthropic's own recent guidance for trustworthy agents is built around the idea of keeping humans in control, requiring human approvals in certain instances, and teaching agents to stop and ask when they are uncertain. I'm going to come back to that because that's not as easy as it sounds. But it matters because there has been kind of this, especially now, that that kind of homebrewing your own agent is getting easier and easier. There's like this lazy argument floating around in tech that humans in the loop are just a temporary crutch. You know, it's the models are getting so good, so easy that humans are just going to disappear. And and maybe in some narrow, low-risk use cases that that's true. In fact, there's probably a lot of those in the world where it really doesn't matter that much if the AI is wrong or you know confuses policies and doesn't get it exactly right. There's a lot of use cases where that's not the end of the world. I do not think property management is one of them. Not when we have we are stewards of the health and welfare of human beings. So when it comes to property operations, I think that view that humans are not necessary, I think it misses the point. The humans are not there just to patch technical weakness. I think actually, right now, there's a lot of AI systems that are still kind of mechanical turks. There's humans in the background really doing the work. Vendor is not one of those. Many of our competitors are not those. However, the humans are still there, even when there's not like problems architecturally. They're there to provide judgment, accountability, trust, sometimes empathy, and last mile end-to-end edge case reliability. Those things don't vanish just because the model got better at pattern matching, just because you can download Claude code and brew your own AI agent that responds to maintenance, it doesn't mean that it inherits sudden judgment, accountability, trust, all those things. So here's the distinction I think buyers need to understand. A lot of AI companies are still basically selling a demo, a very impressive demo sometimes. AI can look fantastic for five minutes. For any five minutes, AI, anybody's AI, something out of the box, untrained in maintenance or leasing, can blow your socks off. It's amazing. It's an amazing time. It's a these things are amazing what they can do. But it's still essentially a demo. The AI can answer. It's easy. We will answer. It can classify, it can draft, it can maybe route, it can maybe take a first pass. But if the system gets confused or the customer says something unexpected, or two rules collide, two policies you've established collide, or a decision has legal implications, financial implications, or emotional health, safety consequences, suddenly the software needs to hand you back the problem, right? Because you don't have end-to-end automation in those cases. That's assisted software with a trapdoor. And sometimes that's fine. But but if you are buying an AI solution for workflow, for a workflow that matters, that that residents, owners, field teams, and vendors actually depend on, well, then assisted software is not enough. Okay, you need coverage. Okay, you need a system where the AI handles the repetitive 80 to 90 percent, okay, which which I again going back to this, I think that's what a lot of out-of-the-box homebrew stuff is doing. 89% coverage. It's not bad. But the remaining 10 to 20%, that needs to be covered so it doesn't become your surprise in an emergency. It doesn't become your liability in the wrong situation. That remaining slice has to be designed for. And that's what human in the loop should mean in practice. Okay. It's not a marketing phrase. It's not, as I was saying earlier, it's not a defect of immature AI. It's not a vague promise that, you know, a person's available somewhere. You can they are stitched into the system with human in the loop. They are part, an integral part that ensures end-to-end coverage in that last 10 to 20% where the liability lives. A properly designed AI system, integrating human in the loop, should know when to escalate. It should raise its hand and say something's different here or something doesn't sound right. And, you know, we have to bring a human. And the humans have to be trained for for this kind of every class of exception, right? And the handoff has to be fast enough that the resident or the customer never feels the floor drop out beneath them. And this is what I said, or I wanted to come back to this. That is really hard. That's the hardest thing, maybe, that that vendor has spent time building. We have great data labels for sure. Maybe the best. You get the best data labels, the best. We we might have the best data labels. Uh, we might have the deepest maintenance brain. But what for sure we have is the most well-developed, seamless handoff to our human in-the-loop team to make sure that we don't drop the ball at that critical moment and that the that your resident doesn't feel that that transition. And that is hard to do. And I think nowhere else is this more obvious than in maintenance, because maintenance looks simple from far away. I mean, ask ChatGPT, how do I clog a toilet? It gives you a reasonable answer. Ask ChatGPT, how do I change the filter in my HVC system? It'll ask you the type of HVC model you're using and give you step-by-step instructions. It looks easy, except to anyone who's actually run maintenance. They know that that's not all what maintenance is. It's some of it, maybe even a lot of it. But in reality, maintenance is conditions stacked on conditions stacked on top of conditions. Is it an emergency or is it just urgent? Is this resident caused or is it owner responsibility? Is there a warranty? Is there a preferred vendor to do this work? Does the property have some special rule about this? Was it built before 1970? And there's there could be a specific, and we need to we need to check some things before we actually authorize any work. Is a resident themselves, are they elderly? Are they vulnerable? Are they known to be unstable, furious at the moment? Or are they unreachable? Can you make a decision if they don't say anything at all? Did the vendor accept the the work and then ghost? Is the technician sick and can't respond to an emergency? Does the owner want repair or do they want replacement? Does the issue have a fair housing angle, a safety angle, a liability angle? In other words, maintenance is not one workflow. Maintenance is not handled. Everything I just mentioned to you, that is not handled out of the box with your Chat GPT subscription. Okay? It's not. And it can try, but if you can't sleep at night knowing that the answers to all of those situations I just mentioned to you are handled right 100% of the time, then you have not unloaded your work to an AI. You have simply added liability into your company because it might make decisions that you don't want. It is a thousand tiny judgment calls, all pretending to be one workflow. That's what it is. This is exactly the kind of environment where AI can be incredible at the standard playbook. And it still needs humans for the edge cases. It still needs humans for those moments that I was just mentioning. And frankly, if any AI vendor says otherwise to you, if they say we're 100% automated or you can trust us 100% all the time, I would not, I wouldn't treat that as confidence from that vendor. I would treat it as a red flag. Because the dangerous phrase in AI right now is not human in the loop. It is don't worry, it handles everything. What might be more dangerous right now is don't worry, I built this myself, but that's another show. It does not yet handle everything. And I think the case I'm making here is it probably should never handle everything. It can handle a lot. It probably, in certain categories, will never handle and should never handle everything. Reuters recently reported that even in legal work, I think anybody in AI has heard this anecdote, a lawyer was sanctioned after using an AI-generated material that contained like 20 some fabricated citations or misrepresentations. If lawyers are being reminded not to trust very fluent-looking output without review, then property operators should be very cautious about any claim that AI can just freewheel through real-world exceptions. It's not the case. And it is worth saying out loud that this is not just the view of Venderoo or a few cautious operators, but Frontier Labs themselves are signaling it. Anthropic CEO Dario Amoday has said that the current generation of AI is not accurate enough for AI-driven weapons, right? That's a big controversy right now with them refusing to let the government use their AI to operate within the defense parameters. You know, the best analogy I have for all of this is robotaxis. Robotaxis are probably the cleanest public story of automation that we have right now. There's cameras, sensors, software. There's what appears to be autonomy and machine decision making that has to react in real time to uh road conditions that can change from time to time, right? It's our future on wheels. And yet, it was reported recently that Waymo defended its use of remote assistance personnel, saying that those humans provide advice and support when the vehicles need help. They do not directly drive the car on the road, but they are still part of the system. That's the point. Even one of the world's most advanced autonomous businesses still keeps a human layer in the loop for last mile, the pun intended, trust problems. That's not a failure of autonomy. That is mature autonomy. It's a recognition that edge cases exist and the trust is not built by pretending they do not. There's a similar story playing out in the world of accounting surrounding audits. There's a this report I read that says, you know, firms are using AI more aggressively for the repetitive work in audits, but they're still drawing a line around certain items that require professional judgment. There's one executive who said something akin to, you know, there's all this fear about replacing humans and that they need to keep humans in the loop because AI cannot replace professional judgment. So that's another industry learning the same lesson. Let it chew through the repetitive stuff. Keep the human judgment where the cost of being wrong is too high. So that's the broader pattern we're seeing. The future is not machine only. It's maybe majority machined, or a lot more machine than it used to be, or is now even for sure than it is now. But the future is machine first and then human-backed. Now, let me bring this back to Vendoroo. The strongest argument for Vendoroo is not just that it has AI. Everybody in the world has AI in the deck at this point. That's table stakes. As I was saying earlier, more and more people feel like they are getting confidence that they can roll their own with these systems that are becoming easier and easier to create agentix systems. And they have a lot of use. In fact, we'll teach you. You can build it, but architecturally, you need to be for to for you to have trust in this industry, it needs to be built around the idea that AI can run the playbook, but humans should remain available for the moments that require judgment, for the moments that require delicate communication or exception handling. When we talk to prospects, we explicitly frame Vendru as a human in the loop for edge cases. And, you know, we have these AI teammates who rely on real maintenance experts who we have that have been trained for two years, who step in when deeper judgment, sensitive communication is necessary. They're there for the failovers, when human touch is needed. Like that is part of our system. There's a data part, there's a data labeling, there's the AI brain, and then there's our team of trained maintenance experts for all of the exception handling. And that's something you're not going to get out of the box, rolling around. You will have to rely on AI to get it all right. And I think I've made the case that that's foolhardy. Because there is a much more serious reason to consider Vendoro other than our agents are really smart. It's really hard. It's really hard for most people to even tell the difference between one agent who's good at something and somebody else's agent who claims to do the same thing. That's really hard. And what I'm saying is the real buying question is not whether or not our AI can handle the easy stuff. It can. So can a lot of other folks that are out there talking about their AI maintenance solution or their leasing solution or their resident communication solution. It's very good at handling the easy stuff. Of course it can. There's a million people I see on LinkedIn every day talking about how they're building their own bots that do that successfully. And good on them. You should all learn how to do that because it can make your life a lot easier in a bunch of other ways. But the real question is what happens when the resident sends a blurry photo? And the issue could be minor or catastrophic. What's AI going to do with that? What happens when the policy says one thing, but the owner has acted in a different way and AI has learned something contrary to that policy? What happens when a vendor accepts and then disappears? What happens when a resident is panicking after hours and the issue has to be triaged now? Not whenever, not when the office opens. Now. What about that? The issue needs to be triaged now, or you get a million dollars in water damage by the time you guys wake up in the morning. What happens when a husband on the lease but with a new restraining order is asking for a rekey? In that world, the best system is not the one that boasts the highest autonomy score. It is the one that gives you confidence that nothing falls on the floor. And that's why I think the human layer matters so much in Venduru's story. It's not just we have people because the AI is not ready. That's just the point I'm trying to make here, is that it's that we have people because trust at scale requires judgment at the edges. That's a very different story. And it's a much better one. So if I could boil all this down, I would say this. What I tell my sales team members is do not promise 100% automation. Promise 100% coverage. That line matters because buyers are getting smarter. They're getting firsthand experience with AI themselves, and they they'll ask you the questions 100 times, and then after the 99th time, it says something wrong. They're smart enough to extrapolate what that could mean in a critical moment. They know buyers are learning that perfect automation is fiction. What they actually want, they need to be worth investing in an AI platform is confidence. Confidence that the routine work gets handled instantly, yes. But confidence that the weird stuff gets escalated fast and handled correctly. Confidence that the residents are not left in limbo or misguided. Confidence that owners or your clients are not surprised by what your AI platform is doing at their property. Confidence that vendors do not drift into a communications black hole, or even residents for that matter. Confidence that the operations keep moving, even after hours, even on weekends, even when something does not fit the script. That is what trust looks like operationally. It's not magic, it's coverage. So here's the bottom line. The future belongs to AI systems that do real work. Yes. But the winners in serious operational categories like property management, where literal lives are on the line, the winners in these categories will be the ones that respect the boundary between repeatable execution and human judgment. That is why I think human in the loop is not a temporary patch. It is the trust architecture. It is the architecture of trust. And for maintenance specifically, that trust architecture is the difference between an impressive demo and an Actual operating system that you can rely on. An actual replacement that in the real world can take the burden of human beings and do not just as good of a job, but better. AI alone can look smart. AI with the right human backup can be trusted. And in the end, trusted is what gets bought. Trusted is what keeps you out of litigation. Thanks for listening. We'll be back again next week with another interview, and we'll dive deep into the world of AI and property management. Talk to you then. Thanks for listening to the Multifamily Property Management AI Playbook, where AI stops being theoretical and starts improving how multifamily operations actually run from processes to people to performance. Learn more at vendoroo.ai. Until next time.