Minna Song

The through-line

Minna Song's consistent position is that the most valuable AI work sits in the least glamorous places: the administrative grind of industries that serve fundamental human needs. She and co-founder Tony Stoyanov set out in 2017 to "build technology for industries that serve fundamental human needs," starting with housing "because it is our largest cost of living, and everyone needs to have housing," and later extending the same mission into healthcare [1]. The two sectors together represent close to half of household spending, and both are burdened by inefficiency and rising costs, which is the affordability case she and Stoyanov make for automating them [6]. What EliseAI sells is the removal of manual work: "all the manual things that happen on site at a property or at a a healthcare provider's office, and these teams are staffing up tons of people for work that can really be automated using AI" [1].

The second, equally durable idea is that technical founders should resist the urge to build first. "I think the inclination from engineers is to start building, not necessarily start with understanding the problem you were actually trying to solve" [1]. That belief produced the founding decision she is best known for and still shapes how the company operates a decade in [1][2][4].

On learning the industry from the inside

Rather than research housing from the outside, Song took a front desk admin job at a Manhattan real estate firm. The insight arrived within weeks, and not as a revelation: "It was not some aha moment. It was this continuous thing that you kept hearing, this continuous pain point. Where every day someone was trying to call a building and they couldn't reach that building" [1]. Communication was obviously the bottleneck, and slow response times were the wedge AI could attack. She finds the surprise this provokes odd: "people are are shocked that I went and took this this job, which is strange to me because there's nothing that more obviously tells you than being in the work environment of your customer" [1].

On being early, and on rebuilding everything

EliseAI was doing conversational AI more than five years before ChatGPT, which meant building with tools that now look primitive. The company collected its own data, tagged test and training sets, and trained in-house classification models to determine intent, whether a message was about pet policy or application requirements, then answered with templates. "If you make the templates kind of fine-grained enough, people did actually think that they were speaking with a human, but not all of the time. Sometimes it was a little bit robotic" [1]. Song credits luck as much as foresight: the transformer was invented in 2017, the year they started, and what they were doing was state of the art at the time even if "now it seems so so silly" [1].

GPT-3 in 2020 was not commercially viable in the product, but it was a signal that something had changed, and the company rebuilt its entire system between 2021 and 2022. She does not romanticise it: "Extremely painful time in the company. Doing a full system rebuild is nothing to be excited about. It was a year-long process. Hard on the company, hard on our customers," and many companies attempt the same thing and fail [1]. The payoff was architectural. The rebuilt system lets EliseAI plug in and evaluate new models quickly, which is what keeps product velocity high. "If we hadn't done that we would be in a very different spot and we might we might not be here today" [1]. Some in-house conversation models survive, used mainly for reliability, stability and compliance in highly regulated industries [1].

On why there are no moats left, only execution

Song is blunt that generative AI has flattened the barrier she once had. Conversation gets cheaper and easier over time, models follow instructions better, and "if we could build it, of course someone else could also build it" [1]. Her answer is not a technical secret but compounding execution and product depth. "Everything we think about culturally, internally, is how do we get our product velocity higher and higher all the time," and she argues the gap between EliseAI and competitors has actually widened as a result [1]. Scale supplies the raw material: more data, more edge cases seen, and a customer base large enough to test and evaluate quickly. A newer entrant lacks the volume of cases, "and when something goes wrong, you can really hurt the trust that a customer has in you" [1]. Ultimately she locates defensibility in the relationship rather than the model: "the best moat for anybody is is having a really strong customer base that that loves you, that is loyal to you, and having a really strong brand," which takes a good, robust product plus great customer service, and comes down to "executing a thousand things every day" [1].

On the move from conversational AI to agentic workflows

Leasing was the wedge, not the destination [3]. The original assistant answered emails and texts about availability, pricing, subletting policy and application paperwork, and integrated with customer systems to schedule and reschedule appointments [1]. Song has since pushed the product across the whole renter journey, so the assistant continues after move-in, and from responding to requests toward executing them. Maintenance is her worked example: not just intaking a resident's AC complaint but resolving it, which means understanding context such as the outside temperature, since "if it's 105° in Dallas, or if it's 72°, then you triage that differently" and escalation changes accordingly [1]. Beyond triage, the system maps the customer's staffing model, matches skill sets to jobs so the right technician is dispatched, and handles backlogs, scheduling and resource optimisation [1]. That operational depth is precisely where she says the defensibility comes from. The company now has roughly 15 products, after several years of building only the leasing product while investing heavily in the underlying platform [1].

On why proptech was starved of capital

Housing is around 20% of GDP while proptech has drawn about 1% of venture dollars, a gap Song attributes to real structural difficulty rather than investor blindness. The sector was once dismissed as "floptech," which she says made her first several rounds hard to raise [1]; fundraising for conversational AI in 2017 was difficult enough that the company bootstrapped early on [5]. Budgets are small and margins low, so buyers historically spent little on software and demanded long trials to prove ROI, since they have no room for something that does not work. That forces startups to spend heavily serving a customer long before any revenue arrives, making escape velocity hard [1]. Compounding it, the end users are non-technical: "Even if something provides value in real estate, if it's not extremely easy, the customers just won't adopt it," because unlike developers they will not invent workarounds to capture value from an awkward product [1].

On centralization and what the savings actually look like

Song frames the industry's adoption of AI through centralization, taking staff who traditionally worked on site at every community, automating roughly 95% of their work, and having one person serve 20 buildings while handling only what cannot be automated [1]. She points to disclosed customer numbers rather than generalities: Equity Residential saw $15 million in savings from the leasing product alone, amounting to about 15 to 20% of on-site staff, and Villa Serena reported a 43% increase in NOI, roughly $731 per unit, driven by payroll savings [1]. These are the operational shifts she describes as transforming multifamily operations [2][4].

On expanding into healthcare from the same platform

Healthcare was always part of the mission, and Song entered it only when she believed there was a genuine advantage: the platform built over years of housing work [1]. Two capabilities transfer directly. The first is reliable, enterprise-grade conversation with the customer's customer. The second is deep integration into systems of record so the AI can take actions, including experience with legacy technology, which dominates both housing and healthcare and which she calls "a different skill set, different muscle than I think most companies have already built" [1]. Because both industries run on the same platform, "they both get better as we develop in each industry" [1]. What did not transfer was go-to-market, operations and customer support, which required building a separate team and learning from scratch [1]. The four core specialties served so far are women's health, dermatology, orthopedics and ophthalmology, with ambitions to add more [1].

On growth capital and where it goes

After crossing $100 million in ARR and closing a $250 million Series E at a $2 billion valuation [1][5], Song frames the raise as getting ahead of exponential growth in both industries rather than funding a new strategy. Most of it goes to headcount: new R&D, customer success capacity to support the clients being added, and building out the healthcare organisation, including salespeople needed to penetrate additional specialties [1].

On building customer obsession without industry DNA

EliseAI's offices are full of young post-grad AI engineers rather than real estate veterans, and Song sees that as a net advantage: "because we're technologists, because we're not from the industry, we've been able to sort of think from first principles and think in an automation-first way about how to solve the problems in the industry, which is why I think our technology is quite transformative" [1]. Only about 1 to 2% of the company comes from real estate, enough to give teams someone accessible to run to for domain judgement [1].

The compensating mechanism is structural dependence on customers. "Because we knew we came from outside the industry, we didn't know anything about the about the actual operations, we knew right from the beginning that we needed to depend heavily on our customers, and that really did become part of our DNA" [1]. Engineers go on site, visit customers, and watch how a maintenance worker schedules a day and what goes wrong, so the team can replicate the decision-making and then optimise it with AI. Internal rules track how many customers someone has spoken to and how many on-site visits they have done. Song ties this straight back to speed: it "gives us the most product velocity, cuz we waste less time building the wrong products, which I think is really really easy to do as a company" [1].

Takeaways

• Before writing code, work the customer's job. Song took a front desk admin role at a Manhattan real estate firm and found the core pain point, unanswered building calls, within a couple of weeks [1].
• A full system rebuild is worth the pain if it buys architectural flexibility. The 2021 to 2022 rebuild lets EliseAI plug in and evaluate new models fast, which Song says is why the company still exists [1].
• Traditional moats are gone in generative AI. Defensibility now comes from product depth, accumulated edge cases, a loyal customer base and brand, sustained by "executing a thousand things every day" [1].
• The real value is in agentic execution, not conversation. Maintenance means triaging by outside temperature, matching technician skill sets, and optimising backlogs and scheduling, not just logging a request [1].
• Proptech was underfunded because low margins, long ROI trials and non-technical users make it structurally hard, and products must work perfectly out of the box or go unadopted [1].
• Centralization is the buyer's business case: automate roughly 95% of on-site work so one person can serve 20 buildings. Equity Residential saw $15 million in savings and Villa Serena a 43% NOI increase [1].
• Entering a second vertical works when the platform transfers. Reliable consumer conversation and deep integration with legacy systems of record carried into healthcare, while go-to-market and support had to be rebuilt from scratch [1].
• Hiring outside the industry enables first-principles, automation-first thinking, provided you offset it with enforced customer contact and on-site visits [1].