Aditya Sanghvi:
It's wonderful to be here with you all today. My name's Aditya Sanghvi. I'm a senior partner with McKinsey. I'm with QuantumBlack, which is our AI unit. I'm a computer scientist by background, but very incidentally for this conversation, I also lead McKinsey's global real estate practice. Founded about 10 years ago, we have about 80 partners globally, and we serve many of the top real estate institutions out there. And very germane for the conversation today, we have a team of about 100 data scientists and engineers and developers who focus on building AI solutions for our clients.
The goal today is to bring to life the opportunity AI has in rental housing, including home building, basically the residential real estate ecosystem. And I want to talk about what's blocking the opportunity and what you all as leaders should do in order to unlock that opportunity.
But let's just take an example situation, one that happens all the time in the buildings that you all own, operate, or service. It's 6:12 a.m. and a pipe lets go. Water pools across the floor. And when that leak happens, even 10 years from now, an AI model is not what's showing up at the door. What's going to show up at the door is a person, but that person will be AI-powered. And there's a lot of fear in this room and a lot of hype outside of it, but I want to be very clear that what AI is going to do is going to change everything between when the leak happens and when the plumber is there to fix it. And this is what we'll bring to life today.
Just in a very standard flow of what happens today. Today a resident calls, a manager logs in the service request, they assign a plumber. After a couple of days, the resident might escalate if it hasn't been fixed yet. A manager will then call to follow up with the plumber and say, why isn't it done? Get over there. The plumber then may fix the leak. Every single step in this value chain is very simple and quick. Very simple, very quick. But yet there's a lot of elapsed time. And so where does the time go? The time happens in the coordination, not the work. It's the ticket and inbox, the call that's not returned, the follow-up that no one zones. There are dead zones that happen in almost every single one of the real estate processes. And those dead zones are ones that the residents feel.
Now let's think about what this could look like. And this is not in 2030; this is now. And we have done it with clients already. You could have a team of agents and humans that move way faster and produce better outcomes. For the intake, you can have an agent where the resident sends a picture, the agent actually takes all the information and makes sure it's stored in the right place with the right information in order for the system to be able to act in the proper way. You have an agent that can triage, that can figure out how severe of a request this is, figure out the right vendor, and automatically follow up in case the vendors didn't show up. You have a human who then comes and fixes it. After, an agent can practically follow up with the resident, and there's an agent that can learn. How do we do this better the next time? An AI doesn't replace the maintenance, but the AI actually removes the dead zones to make this faster and better.
I wanted to start with this example just so that it'd become a little bit clear what we're talking about. But really the idea of today, this session, is to talk about four things. One, what is this AI moment? To sort of separate signals from the noise and help you all understand what the true opportunity actually is. Second, the great paradox, and I'm sure all of you are experiencing this already. There's been an incredible amount of adoption by companies of AI, but very few companies can point to value from AI. What's the reason that's happening? Third, we'll talk about the re-imagination of rental housing itself.
What does this mean for you and your business? And then fourth, what does that mean that you should do as business leaders, as CEOs, in order to unlock the opportunity?
But let's put AI into context. One could easily argue that the history of the last 50 years of business has been giant leaps in technology. And each of these giant leaps in technology have made something that was scarce into something that is abundant. In the late 70s, we had computers. They took up entire rooms, and very few people had them. They were almost all by businesses or the government. But there was computation, and personal computers made computation available for everybody. That was the first great leap of technology.
The second one was the internet. The internet connected all of these computers. It made it available to see what was on someone else's computer in countries away. It made distribution from scarce to abundant. And then the mobile and cloud, which I learned yesterday, were the source of all my unhappiness. They made access available. We could suddenly have what was on our computer with us all the time, everywhere that we were. And now AI has made intelligence from scarce to abundant, which is that we can now have intelligence at our fingertips at all times.
Now, some of you might say, well, AI isn't actually new. And in fact, it's what I majored in college two decades ago. But what happened in 2022 was the emergence of large language models. And those rapidly and exponentially changed the game. I want to use one little example for that. There was a test that was created. It's called Humanity's Last Exam. Humanity's Last Exam has 2,500 questions that only a PhD in the discipline of which the question is in could answer. I give you one example up here on the screen, which is about hummingbirds and some bones they have and how many tendons are connected to that very unique bone.
In January 2025, the best AI model could only get it right, get 3% of all those 2,500 questions. To be very clear, I would get zero. Most people in this room would get zero. These are very difficult questions. As of July of 2026, Gemini's models are now at 50%. This test that was designed specifically in 2024 to be a proof that AI can't get to where we are is already getting to where we are and moving very fast. But AI is not perfect. A University of Edinburgh study last year asked AI models to read clocks, regular analog clocks. The models got the time right 13% of the time. So the same system that can answer this doctoral question on hummingbirds can't tell you if it's a quarter past three. AI is powerful but uneven. And it's actually very hard to predict where it'll be brilliant or where it's going to fail. And that's why, exactly, you'll see that this is about agents and humans together that is a true power.
The AI revolution, I mentioned earlier we had AI before. There was robotic process automation, there was machine learning. Those two things that were for 2022 had actually already created a lot of automation potential. 20% of work by McKinsey Research could be automated based on machine learning that we already had before 2022. But then we had three innings of changes, and there's already been three innings in just a few years. The first inning was generative AI. This is probably what most of you use AI for. This is using ChatGPT or Anthropic, in which you put in a prompt and you get a response. And I'm sure all of you used it to create birthday raps for your eight-year-old granddaughter to talk about them to the tune of Ice Ice Baby. All sorts of crazy stuff that we've done with it. But the automation potential actually didn't go up very much. You can see it went from 20% to 25% from gen AI. And the reason is that it always relied on human prompting.
But then in 2023, agentic AI started. I'm sure you've all heard that word, and I'm going to explain what that is because it's actually the crux of the whole opportunity here. Agentic AI is an AI that doesn't wait for a prompt, but you give it a task, you give it a goal, and it can then plan and execute around that goal. For example, you could have an agent that does scheduling, and all it does is schedule. And it gets better and better over time, but it does scheduling. We moved to do a task for me with a single agent. But the problem is that each of these agents operates in silos, and they can only do very narrow and specific things. Now, the great revolution that's happened and it was really unlocked last year by something called model context protocols that were developed that allows the agents to actually talk to one another and work together is multi-agent meshes. And that's where you go from doing a task for me to doing an entire process for me. And there you can see there's a big leap in the automation potential from 35% with single-agentic AI to 55% with multi-agent meshes.
And these multi-agent networks are powerful not just because of what the agents can do, but because explicitly humans are in the loop. Now, if I asked almost every single one of you, would you allow an AI to read and respond to your emails without you looking at their response? I'm sure the answer would be absolutely not. But what's the reason for that? The reason isn't that AI doesn't know how to write emails. It's that there's context that AI doesn't have.
It may not know the personality dynamics of the people on the email thread. Also, the AI hasn't been trained. It doesn't know how to write like you. It doesn't know what you like to say, how you like to come across.
Humans are essential to be in the loop of almost every single agentic AI process because they have context, because models haven't been trained, because humans still outperform agents on many tasks. Also, agents can't be held accountable. You can't fire, you can't sue an agent. And agents can sometimes have unexpected second-order effects. A client that we did work on as McKinsey had a customer service agent. That customer service agent was told and given the goal of making customers as happy as you can as fast as you can. What did that agent do? Every time a customer called, they refunded the entire purchase. It worked, but it may not be what you want.
The other thing about humans in the loop is that this makes it so powerful because it means that right now you can actually get the value from multi-agent networks. If we had to wait for AI to be able to fix all those things on the right-hand side of this page, we'd be waiting for a long time. But with the right redesign of work where we have the human intervention at the right points, we can use it now and take away the low-value stuff to agents and put the high-value stuff in the hands of humans.
Agentic AI. It has already driven significant value across industries, and we're going to talk about real estate and rental housing in particular in a few minutes. The most common use cases have been in productivity, product, and some in disruptions. But on productivity, the ones that you'll hear about all the time when you dig in, it's really about four things. One, customer service. Two, coding, instantly, at scale, doing years of coding work in just a matter of hours or days. Salespeople are empowered with co-pilots, and process and claim transactions that are done at heavy volume. These are things that are delivering value for companies right now. But then the more interesting stuff on top of that also becomes in product, what you actually offer to customers. There are a lot of companies that are changing by adding AI into the product themselves. They're also able to do personalization at scale. And then also, there are some businesses like accounting businesses that are starting to sell outcomes, not just the software, because they are made so much more powerful based on the agents that they have. But then we also see that agentic AI is starting to disrupt industries overall. There are many things that could happen, but let's take the software industry, for example. In what feels like almost overnight to them a few months ago, all their valuations radically changed. Nothing about the businesses changed, except AI created this uncertainty that maybe their businesses could be replicated by agents very quickly, and they no longer have the competitive advantage that they thought they had. This might happen in our industry too, which we'll talk about.
The net of all of this is that AI is creating incredible opportunity, but also incredible uncertainty at incredible speed. When we work with most CEOs, there's just incredible whiplash. One day they had conviction to roll out a tool across the entire enterprise. For example, they might build their own large language model that they think represents their entire firm's intelligence. But then they realized that that large language model was built on a tool that suddenly can't get the benefit of what the latest models have these days, and their tool is now obsolete. But then also, people are asking questions like, if I keep using these LMs, am I training my replacement? Do I have to become a technology company? Do I have to be a first mover, or can I wait? We're not going to answer every one of these questions, but these are the questions that we see on the minds of most CEOs.
Okay, so I alluded to this earlier: the paradox. AI adoption is incredibly high. This stat was as of last year. 80% of companies have used AI in at least one business function. If you look at the data now, I'm sure it'd be way higher. But very few get the value. Only 6% of companies report material financial impact from AI deployment. And most companies will tell you that it's a waste of time from what they've seen. What's the answer? The gap is not technology. Everyone's using the same models. The gap is actually the approach.
Here are the six reasons why AI is not delivering value. First and foremost, and this is the one about the top left here, when people deploy AI, they're largely deploying generative AI, not agentic AI. Less than 10% of deployments are agentic AI. And generative AI, as I'm sure all of you have experienced in your own lives, can sometimes be just a better form of search. And it improves individual productivity, but it doesn't improve enterprise productivity. That's the first reason.
The second reason, and I would say in real estate this is the most important reason, is data. Agents are not at the place yet where they can take data that's not clean and make it clean. They are not there. And I don't think they will be there for a while. They operate on shared context and data that is accurate and usable. And the problem is that if you have data that's 90% accurate, it's actually 0% useful. And I'm sure all of you have seen this. A Gartner report found that 60% of AI efforts were abandoned because of data quality. It wasn't the problem with the models; it was the problem with the data.
Third, a thousand flowers bloom. Most CEOs say to us all the time, with great pride, that they have thousands of experiments running across their organization, and they have all these young people that are doing such cool things with AI. The problem is that in every single one of these little use cases that people do, you can't actually move the needle on outcomes. What ends up happening is that people do something that's cool but small. And these things don't add up into anything that you can see in a P&L. And they're often done outside the core systems, so they're not really scalable in what's done.
Fourth, over-delegation by CEOs. I'm sure it may or may not be true in this room, but broadly speaking, most CEOs are not taking enough ownership over AI. They delegate it. If a head of IT is leading an AI initiative, you're already dead in the water. This is a business transformation; it has to be led by the business, and it has to be done with the same rigor of any other transformation that a CEO would lead.
Fifth, living on the extremes. Companies try to do one of two things. Either they take a process, and they just add an AI to it. Instead of somebody going through a lease, they add a lease abstraction tool, and then they try to see if it gets better. It doesn't get better. But on the other extreme, some people try to automate processes entirely. We have a client that actually tried to automate, which we'll get to in a little bit, their entire investment committee underwriting process. But then the AI would return answers that nobody could explain, and then nobody trusted it. Neither extreme works; you have to reimagine the work in order to get there.
Finally, change management. Just with anything, it is always underappreciated. When you roll out new tools and new processes, unless people know how to use them, unless they see other people using them, unless their rewards and consequences are set up against it, they won't use it. They just won't. Change management is usually considered an afterthought, and it has to be fundamental to the whole process. Those are the reasons why, generally, across industries, AI has not delivered value. I want to just hit on a couple that are very specific to our industry.
First, data. The data in real estate is terrible, just terrible. Because there are all these different property management systems that don't talk to each other.
People have offline Excels, you have ledgers and CRMs, and nothing connects. The tenant names from the CRM don't match to the tenant names in the PMS. And what that means is that the data isn't usable. And unlike other industries, real estate does not spend the time to put the right data governance in place, little boring things like data definitions, accuracy measures, in order to get data that's usable.
Second, the industry is very fragmented, and the incentives are a bit out of whack for AI. As I'm sure you all think about your place in the value chain, for one of our clients, for example, they're a third-party property manager. And they get paid as a percentage of revenue on the buildings; if they do all this work to basically create agentic solutions that deliver better cost outcomes to the property NOI, they don't get compensated for any of that. Now, they could win a greater share of business, and there are, of course, benefits to it, but not enough to make the ROI case work. And that's one of the big challenges that we have up and down the entire value chain.
And finally, the tech talent and tech maturity of business leaders in this industry is worse than other industries. It is worse. And that's not meant to be an attack per se. It's because over the last 10 years, if I asked almost every one of you and every one of your firms, you would not say that tech has been the differentiator for you. But in other industries, it is the case. And so our industry is behind in terms of maturity on some of these things. And also, a lot of CEOs and a lot of people still think about tech as a cost to be managed rather than a value driver for the business.
The opportunity. I would actually put out there that rental housing is perfect for AI, perfect. Why? Number one, coordination is the work. We used the maintenance example earlier. Let me use another one: investor reporting. What happens if you need to report up to investors metrics about your property is you take data from the property manager, who then sends it up to the asset manager. The asset manager then sends it to investor relations. The investor relations team then sends it to the LP. Four different steps where everybody on each of those steps is recutting and curating the data in their own way. And if you look at almost every process in the industry, it's basically the same thing. It is just a whole bunch of relatively simple tasks that involve complex coordination. And this is where agentic AI thrives more than any other way.
Two, every decision is a lesson. In industries in which you have millions of repeated decisions, like what is the retention offer that you provide to a resident? How do you resolve a maintenance request from a resident? How do you underwrite a deal? And then you can see the impact of that. That becomes an opportunity for learning. Your portfolio can actually become a teacher. Static assets can become dynamic self-learning systems. When your businesses are only a couple of things that happen in your matter, you don't have the kind of scale to get that learning.
And finally, propagation, which is an improvement scale. What I mean by this is, right now in the industry, when someone figures out something great to do, it stays local. If somebody figures out a new way of doing things in a property, it benefits that property. In the best case, maybe somebody creates a playbook, and that playbook can then be used across a few more properties. But with agentic AI, when someone creates something different, it actually can be spread throughout the entire portfolio automatically. And there have been proofreadings around the value chain already. For those who are wondering where values come from, we at McKinsey estimate there's about 430 to 550 billion in total value from real estate AI, and that's across all real estate. It's across the entire value chain. But the biggest ones to point out are, one, lease conversion. Two, renewals. Three, maintenance. Four, collections. Five, continuous IC input, which we'll talk about. And finally, generative scheduling and development.
These are things that have already happened, have had value, move KPIs that matter for the business, and are areas where most of you don't need to reinvent. You don't need to come up with the areas that will have value, you just need to start doing and acting on the areas in which it's already been proven.
I want to very quickly bring to life how, actually, these whole agents and people work. And we're going to do it with an example in maintenance. The same example I had earlier about the leak. What we did is we created four different agents, and then there were two humans. Just to explain it, the first agent was called the inspector agent. The inspector agent takes new tickets and triages them. Figures out what's the priority level of it if it was put in as a low priority, but it seems like a high priority, actually changes it into a high priority and takes action there.
The notifier agent, and to me this is the most important of all the agents, is actually the one that continues to do the follow-up, does the coordination. Updates everybody in the value chain, sends ETAs, escalates when thresholds are hit. They are the ones that make sure that all the dead zones that we talked about earlier are gone. Then there's the actor agent. The actor agent takes action. The actor agent actually is like five agents, but keep it simple. The actor agent routes the right crew, checks certificates of insurance, schedules them to vendors.
And then finally there's a pattern agent. The pattern agent is the one that learns. Two examples of this. A pattern agent in one of the deployments that we did found that there were certain technicians that just took a lot longer than they should. And that wasn't to blame those technicians; it was actually to incorporate that knowledge into how we actually deployed those technicians, and when we deployed them. Second, there was a building in which the frequency of HVAC breakdowns were happening so quickly that instead of fixing the next one, it was determined that actually you should replace all of the HVACs in that building. That's how the pattern agent is not just within one of these tasks or workflows, but looks across workflows to figure out what to do.
Then there are two humans involved. One is the actual plumber who goes and fixes the leak. The second is the facilities manager. And the facilities manager basically is the one who gets the approval rights for all the money moves. In this deployment, anytime money was being spent, the facility manager was there to make sure that the right decision was happening. The impact was substantial. It gave about 50% time back for facility management teams. It brought the average time of resolution from weeks to days. All because of this waste in the system was basically removed.
We're going to talk about another example here, which is renewals. And instead of me describing it, I'll show you a quick video on it.
“Lease renewals. High stakes, high volume, and historically very manual. Not anymore. The moment a lease approaches expiration, the workflow kicks in. A lease expiration agent surfaces the case. A churn analysis agent digs into payment history, service requests, past behavior, and predicts renewal risk.
An incentives agent reasons over that tenant's preferences and recommends the right offer. Then the renewal offer agent writes a personalized letter ready to send. On the manager's screen, one click to approve, adjust, or add a note.
That's it. Hundreds of tenants proactively engaged without a single manual touch point. Higher retention. Optimized revenue. Stronger relationships.”
Aditya Sanghvi:
Great, so for those of you who couldn't follow that quick video, just to explain real fast. Basically today, the way that most renewal happens in rental housing is a very reactive process. A generic offer gets sent; the same thing gets sent to almost every resident. If there's a resident that needs to be engaged, people look up the context and try to figure out what the resident might like. Often the renewal offer is too little, or sometimes it's too late, and you don't get the renewals that you want. What agentic allows you to do is you have all these different agents that then can do very specific things.
One agent will actually build a Resident 360, which is an entire profile of that person. Another agent will figure out what is the churn risk of that person based on information about things like amenity usage, number of kids that they have, other information we have in Resident 360. And then it figures out the right move. For example, we often found in one building that for people who use the gym frequently, there was a gym attached to the building that was third-party managed. Offering one year free gym membership was way more effective for retention than offering one month of free rent. And way cheaper, by the way. It then prepares the human for the actual negotiation moment, and then you have a learning algorithm that learns. What this basically does is it's super intelligence for the renewal process, and its idea is to make the right offer for the right unit and the right person at the right time, every single time. This is only possible with self-learning and with agentic AI.
Final example I'll show: investment committee. At many firms, the investment committee is a few people who have incredible knowledge and insight about investing in real estate. They meet every once in a while. They're pretty hard to get time with. What we did with one client is we actually built a multi-agent agentic system in which we had about dozens of agents that did very specific things that were trained on all the investment committee members and investment committee decisions that have ever been done by that firm. They did things like, there's one agent that was a risk agent, another that would come up with market scenarios, another agent that was deal structuring, things like that. And what this firm then did is it made it available to every single deal professional at every stage of the process to have a chat with. We actually created a war room where you can bring your deal to the investment committee at any stage of the process, and you actually see the different agents arguing. There are actually agents that were designed in the mold of specific people on the investment committee. You can say, what would our CEO say about this specific deal? And therefore, it wasn't just about getting input from the IC at the moments in which the IC could meet; it was getting continuous inputs for deal teams to make them make better decisions and save time and triage deals that shouldn't happen.
Those are a few examples of what agentic deployments have already been in real estate that have made a significant impact. But then you start to think about: what does all this mean about where things will go? And specifically on differentiators. What I lay out here is the differentiators that many firms say today and what tomorrow's differentiators could be. For example, on investment experience, most people talk about how great they are at reading markets, that they have local knowledge, that they're able to figure out exactly what are the right assets to buy. Maybe everyone will have that soon. And maybe tomorrow will be more about, well, what can you do when you actually own the assets in terms of using agentic AI to lead to operational value creation in the assets themselves?
Second, the operating platform. Most people talk about how their operating capabilities allow them to consistently deliver on plan, whether the asset management team or the property manager itself. But with agentic, you can get better outcomes for lower cost. If you're not doing this, do you start to fall behind? On the resident side, people are usually proud that they have great teams that treat the residents with respect and have good retention. But maybe you can start to do personalized experiences for them, make people's lives actually better, and get to outcomes and get people to want to stay with your properties, even if they have options elsewhere. And finally, proprietary data. A lot of people talk about how they have great experiences. Experience is static. What's not static is memory and learning loops. And that's what AI can give you. Differentiators will change. And then with that, the industry structure itself might change. And these are questions. These are not forecasts of what will happen. But one, vertical integration, has already been a trend for the last 10 years. But if a lot of these agentic deployments happen across all aspects of the value chain, are those who are owner-operators going to win out even more?
Second, a lot of what third-party property managers do and what the prop tech firms, PMSs especially, are trying to do with AI feel very similar. Does that stay as two separate, or does that start to merge too? Second question. A lot of the firms in the ecosystem get paid for coordination. Agentic AI makes coordination easy. What happens to those firms whose real value proposition, when you think about it, is coordination? Third, owners. I believe very strongly that benefits go to scale. Those who have memory and learning loops will win out. Let's say the big ones, the giants, if they're able to do AI properly, will win. Small firms actually can start to get really sophisticated tools that can help them very cheaply. Does that lead to the midsize owners getting caught in between? And finally, I think the most important question is who owns the memory? The most important differentiator for the next 10 years will be the data, the memory, and the learning loops. Is that owned by the owners, the operators, the PMSs? How does that shake out?
What does this mean that you should do as CEOs and business leaders? This is McKinsey's rewired approach. What it basically says is you have to do three things together, and there's no silver bullet. One is you have to have clear alignment on value. You need to have two or three bets that you really focus on in which you can get value from agentic AI deployments. Second, you have to reimagine work, and you reimagine work through talent, through operating models, through technology and data all together. But basically, instead of trying to change your entire tech stack, change it just for the work that you're trying to reimagine, just for those individual bets. And then you have to focus heavily on change management. If you miss any one of these three components, the value will leak away. It has to all be done in concert.
But very specifically, I think CEOs need to do six things. Number one, they have to re-underwrite their advantage. Basically, try to answer all the questions that I just laid out, but for them specifically. Two, they have to pick their domains. What are the places in which they are going to focus to get value? Three, redesign work, truly clean, cheap in a business-led way the way that work happens today for those specific areas. Four, curate the stack. Every single one of these conversations I have, people ask me about buy, build, partner.
The answer is all. You should be doing all three of them. But the real value is in making sure that you are making sure there's connectivity across the stack, which we'll talk about in a second. Fifth is scale trust, and sixth is drive adoption. Let me talk about them each in turn real fast.
I want to introduce a concept which is called domains. In AI, there's a Goldilocks-type dynamic. Many people try to approach AI as the entire enterprise. Too big, too shallow. Usually what this boils down to is rolling out one of the large language models across your enterprise. Doesn't leave a lot of value. But too often, as I said earlier, people try to go too small. I want to automate a first draft of a lease contract, or I want to automate billing for a specific item. There's a middle ground, which is a domain, which is an area of value in which you can have clear KPIs, you can have a business owner in charge of it, and you can start to shift the conversation from cost per task to cost per outcome. That's the right level. Those are the two to three sponsored bets that should be at that level. Renewals, maintenance, collections, not little experiments.
Second, I want to tell you the story of Marcus. Marcus is completely made up, but based on a real example. Marcus is a head of asset management, a large multifamily investor, and he did, let's say, what we talked about earlier in terms of the renewals. But just think about what his job is that is so different from what he used to do before. The head of asset management before would manage teams; he would set policies; he would get feedback; he would act and execute. Now what Marcus is doing is he's reimagining it. He's creating a team of people within the business and tech leaders to actually figure out what agents should do, what humans should do, and how the process should be done. He's leaning into the tech stack to understand the differences in what he's able to do and what he's not able to do based on the tech stack that he's chosen. And he's building a tech muscle for himself. This is what business leaders have to do now in an agentic AI, not just execute themselves.
Data governance. This is the most boring, but also the most important slide that I have in my entire presentation. Every differentiator, every agent, every dollar of value that I'm talking about depends on data. And almost no one has solved it in our industry. What you need to do is you need to focus on having a business-led approach to data, not IT. You need to have data accountability, data definitions, data quality measures, and focus on data as the most important thing because it is the most important thing. However, the one big caveat is what everyone tries to do is they try to fix their data first. They do it in isolation. They have these large 5, 10-year data lake projects that haven't led to any value. And the reason they don't lead to value is because they're tech-led. What you need to do instead is to use a domain, and actually clean the data in the context of that domain. When you start working on renewals, you take your renewal data, you clean that data while you start to generate the outputs of that renewal algorithm that you create. And that allows you to actually, in real time, validate the quality. Because otherwise, you do these large data lake exercises, but no one's there to actually validate the quality. And then people find out later on that quality isn't actually there.
I mentioned this before, but the agentic stack. There are many different layers of the agentic stack. There are experiences, which is what people actually use; apps, think about it like that. There are the agents; there's the trust and control, intelligence, memory and data, and internal systems. People always want a very clear answer by build and partner. I'll tell you two things that are very clear. Number one, you should own your data. You must own your data. And number two, do not build a large language model. You will not do well at it; you just won't. In between that, your job is to curate the stack, to make sure there's connectivity across all elements of it, to put the governance in place. And I know this is a little complicated, but if you're expecting someone to give you an answer and give you a tech stack that will solve it, you'll be waiting a really long time. This is something that you must curate and must be business-led. You're the conductor of the orchestra.
Okay, last two things and then we're going to open up Q&A. Trust, so we talked at the very beginning about how agents can't do everything yet, and you need to have humans in the loop, but how do you actually do that? One of the things that you should think about as leaders is lean into AI taking action, but put very strict guardrails in place. You can have three different levels, which is what this shows, but in very simple terms, set boundaries and guardrails for what the AI can do. And then get to a place in which your team has to approve what the AI does. And then when your team hits 99% of the time, they will ask you to then automate the next step. Trust is something that you gain by your humans actually seeing over and over again that the agents are getting to the right place and training the agents rather than something you should just assume from day one.
Finally, change management. Change management is about role modeling. It's about making sure people understand, developing talents and skills, and rewards and consequences. This is what people usually forget about. They assume that when you roll out a tool, people are going to use it. As I mentioned earlier, about half of your effort should focus on this. With that, and before we just jump to Q&A, the big takeaways from this conversation: number one, agentic AI is the unlock. If you're using generative AI, you're not going to get a place of value. Agentic AI is where you have to go. Second, rental housing is built for this. The opportunity is ripe. Third, those who win will be the ones who redesign work the most. And who have the data and then start these learning loops that make every decision better. And stands on the shoulders of all the decisions that came before it. And finally, this is the CEO's job, so start now.
All right. Thank you all.
