Early ML stack and techniques

Watch on YouTube • Listen on Spotify • Listen on Apple

SaaS as we know it is dying and giants are being felled. Today, I want to talk about EliseAI’s playbook for murdering such giants in the world of property management software. The company’s displacement strategy is elegant and brutal: Start with conversational AI that solves an obvious pain point (answering repetitive tenant inquiries), use that wedge to integrate your product into every system used by a customer, expand into workflow automation that makes human interfaces less necessary, then give away the CRM and other legacy software for free. You can afford this last part because you’re not selling software subscriptions anymore—you’re selling AI that replaces labor.

This pattern has been proposed by multiple thinkers over the last few years in tech. However, most of these proposals have been theories, ideas, and Substack posts without tangible examples. EliseAI makes their words come true. The company recently raised $250 million at a $2.2 billion valuation and has been at this strategy since 2017, with the sort of scale and execution that were previously just fantasies. The firm has done very, very little press and just been quietly building. Today, that changed, and I sat down to interview co-founder and CEO Minna Song.

EliseAI is particularly instructive because the company didn’t start by trying to replace property management systems; they started where there was no software, automating the thousands of manual tasks that humans suffered through daily. Answering tedious questions like “What’s the pet policy” or “Can we schedule a tour?” Thus came the beauty of automation: Every automated workflow created more integration points, more data capture, more surface area to expand. Seven years in, EliseAI is now giving away tools like CRMs that incumbents charge for while making money from the actual value-generating work of AI agents. The traditional software becomes a dumb database, which is a short walk to obsolescence.

This interview is a decision tree for the next five years. It will tell you what each of your decisions will lead to if you’re an investor or an operator. If you’re an incumbent vertical SaaS company and someone offers conversational AI as a “partnership” or “integration,” you’re opening the door to your eventual replacement. If you’re building in vertical AI, the question isn’t whether or not to give away traditional SaaS features to beat your competition: It might be the only way you survive. And if you’re allocating capital, you’re competing against companies that got lucky on timing (founded early enough to build pre-GPT infrastructure but late enough to benefit from transformers), are disciplined on spending (EliseAI bootstrapped 2.5 years before raising a $1.9M seed), and are ruthless about expansion once the wedge worked. You’ll need to know what you’re up against if you’re operating in this space. Which is why I think this conversation with EliseAI is incredibly valuable.

Below are my notes and primary takeaways from the conversation.

Ideas & Analysis

The Wedge is Where Software Never Existed

Thesis: Elise entered property management by automating human workflows nobody had bothered to software-ize, not by competing with existing tools. A classic AI displacement entry point.

Pull Quotes:

“I took a job working at a real estate firm in New York City and that was really where I got a lot of exposure. So I was a front desk admin, so I met a bunch of people, I greeted everyone and I learned everything about what we built here at the beginning from that role…People, they just had loads and loads of people doing really tedious tasks on site, answering the same email 50 times a day. Here’s how much a one bedroom costs. Here’s our pet policy. I’ll schedule you an appointment, right? All manually.”

“Every time we think about what do we build next, we’re not there to just try to capture zero sum part market share. We’re trying to capture the value that we’re trying to create.”

Analysis:

Co-founder Minna Song started off the company by working for a three-month stint as a front desk admin. The insight was simple: property management software handled databases and accounting, but humans still manually answered “how much is a one bedroom” fifty times a day because nobody had bothered to automate that layer. It’s too conversational, too variable, too low-margin to justify traditional software development. But for an AI company in 2017 with just-released transformer models, it’s the perfect entry point. You’re not displacing anyone because there’s nothing to displace. You’re just removing tedium, which makes you a hero rather than a threat. This is the wedge strategy at its purest: find where humans do repetitive work that looks too “soft” for traditional software, automate it with AI, then use that foothold to expand.

The framing about “creating value not capturing market share” is both true and strategic misdirection. Yes, automating unautomated work creates new value. But once you’re integrated into the leasing workflow—answering inquiries, scheduling tours, qualifying prospects—you’ve got data pipes into every other system the property uses. The CRM, the maintenance system, the payment processor, the smart locks. Each integration point is a future expansion vector. The reason this works as displacement is that incumbents can’t defend against it. How do you compete with someone giving away conversational AI when your business model is selling seat licenses? You can bolt on chatbots, but you don’t have the AI-first architecture to make it actually good, and you can’t afford to give it away.

From Conversations to Workflows to Systems

Thesis: Elise’s expansion from answering questions to executing tasks to replacing entire software categories follows the natural gravity of AI integration—each step makes the next inevitable.

Pull Quotes:

“It’s not just that conversational component. It’s not that interface with the resident, but it’s actually executing on all the tasks. And then yes, we have a CRM so agents can interact with the AI and get all the information that they need.”

Analysis:

The progression from conversational AI to full workflow automation reveals the displacement mechanics. Stage one: answer tenant questions (the wedge—low risk, obvious value). Stage two: take actions based on those conversations (schedule tours, process applications—now you need write access to their systems). Stage three: execute the entire workflow without human intervention (generate leases, route maintenance requests, manage contractors) and now you’re the orchestration layer. Each stage requires deeper integration with the customer’s tech stack, and each integration creates more dependency. By the time you’re routing maintenance requests to the right contractor with the right priority, you’re running their entire operations.

The sneaky part is how natural this expansion feels to customers. They’re not adopting a new category; they’re just saying “hey, since you’re already handling inquiries, can you also handle applications?” Then renewals. Then maintenance. Then payments. Each request sounds incremental but collectively they add up to Elise becoming the operating system for the property.

Traditional property management software can’t replicate this path because they’re database-first, not AI-first. They can add chatbots, but those chatbots can’t actually do anything without breaking their existing architecture. Elise built the opposite way: AI that takes actions, with software scaffolding to support it. The counter-argument is that full workflow automation requires reliability that current AI can’t guarantee—one bad maintenance routing decision costs thousands of dollars. But Song’s bet is that as models improve, the reliability gap closes, and by then Elise owns the entire workflow layer.

Free CRMs Aren’t Charity, They’re Strategy

Thesis: Giving away the CRM for free isn’t just land-and-expand pricing.

Pull Quotes:

“Our CRM is free because people are sort of used to having that be the tool. Whereas we see that as more, the AI is actually adding the value and the CRM is over time. Hopefully the CRM goes away actually because the AI is doing all that work.”

Analysis:

Song’s casual mention that the CRM will “hopefully go away” is the entire vertical SaaS displacement thesis in one sentence. Traditional software companies charge for CRMs because that’s their product—the interface where humans do work. Elise gives it away because in their model, the CRM is an artifact of insufficient automation. If the AI handles inquiries, schedules tours, processes applications, routes maintenance, and manages renewals, what exactly do humans need a CRM for? Just monitoring exceptions and handling edge cases that get escalated. That interface becomes simpler over time, not more complex, as the AI gets better. Charging for it would be charging for your own obsolescence, which is terrible unit economics.

This is the nightmare scenario for incumbent vertical SaaS: a competitor enters via AI features, integrates with your system, then starts giving away your core product for free because they’re monetizing a different layer. You can’t match their pricing without destroying your business model, and you can’t match their AI capabilities without rebuilding your entire architecture. Song’s framing about “not charging for something that will disappear” sounds principled, but it’s also predatory—she’s explicitly designing a product where the traditional software layer gets thinner every quarter. For the vertical SaaS incumbents reading this, if your revenue comes from human interface software (CRMs, dashboards, workflow tools), you’re in the crosshairs. The only defense is to become the AI layer yourself, which requires burning down your current business model and rebuilding. Most companies won’t do it, which is why they’ll get displaced.

Bootstrapping Bought Them Perfect Timing

Thesis: Elise’s tiny seed and 2.5-year bootstrap looks like discipline, but Song admits they’d have died with a big round—spending on the wrong stack before technology was ready.

Quotes:

“I don’t think we would have survived actually because if you think about the timeline of AI, you think about a bunch of different, the timeline of real estate adoption, I think we would have been spending in the wrong places.”

“Money doesn’t solve a lot of problems either. So even for the seed company, you know, I don’t know which company you’re talking about and maybe it’s a very capital intensive company and you actually just need that to get it off the ground. But for company like ours, you didn’t need that.”

Analysis:

The key thing that made their company work is that they bet early on tech that no one cared about in a market most people were ignoring.

Song’s reflection on their $1.9M seed contains a brutal truth about AI timing that today’s $100M+ seed rounds are learning the hard way. Elise got lucky by being poor—if they’d raised huge in 2017, they’d have spent 2018-2020 building the wrong architecture (pre-Transformer models weren’t good enough), hiring too fast for a market that didn’t understand AI yet, and burning out before AI agents started working. The forced frugality from bootstrapping meant they iterated slowly with customers, rebuilt when models improved, and only scaled after proving enterprise fit. By the time they raised serious money (Series C onward), they knew exactly what to build. Capital can’t compress time when you’re waiting for foundational technology to mature or customer behavior to shift, and dumping money into the wrong hypotheses just kills you faster.

This creates an interesting filter for which AI companies will execute the vertical SaaS displacement playbook successfully. The ones that raised huge early are now stuck: they built on old models, have expensive teams, and can’t easily rebuild without admitting their Series A/B was wasted. The ones that bootstrapped or raised small through 2017-2021 can now ride the GPT wave with validated use cases and efficient teams. Song’s three-week $250M Series E while still sitting on Series D cash shows the flip side—once you’ve proven the model works at scale, money chases you because it’s mostly execution risk (hire more engineers, scale GTM), which capital can actually solve. The decision rule: early funding should match the speed limit of your category’s readiness (technology maturity, customer behavior, regulatory environment), not your ambition. If you’re building on bleeding-edge tech that isn’t proven yet, stay lean until it works. Then raise huge and scale. Otherwise you’re just burning runway on bad assumptions at venture scale.