EliseAI's Long Bet: Building Conversational AI for Housing Before the Models Existed

Starting with AI in 2017, before there were models worth using

EliseAI's origin story, as Song tells it, is unusual because there was no single technology moment that triggered the bet. The company committed to AI from the start in 2017, when the useful generative models it now depends on did not yet exist.

What made it work early was picking a target that was almost untouched. Song calls housing a greenfield where even traditional technology was outdated, which meant that older techniques added value before generative AI arrived.

We really focused on solving problems for our customers. Once we understood the problems in the housing industry, it was very clear that AI was the only way that we could solve them.Montana Labs

In practice, that meant shipping with what was available. Before strong generative models existed, EliseAI used models like BERT to build a traditional conversational experience, then upgraded as new capabilities dropped.

GPT-4 and Whisper as concrete turning points

Song is specific about which advances mattered. GPT-3.5 was a meaningful improvement over GPT-3, but the jump to GPT-4 was where the team realized how much of the industry's problems it could actually solve.

Whisper unlocked a different door entirely. EliseAI had built text-based products for housing but treated phone calls as a clear gap. Song says the earlier voice technology 'wasn't even close' to usable.

That gap was existential for healthcare specifically. Song says almost all communication in healthcare happens over the phone, and that without workable voice AI the company could not have entered the industry at all. Voice was not a feature request; it was the entry ticket.

Designing AI to imitate the workflow it replaces

For non-technical industries, EliseAI's adoption strategy was deliberately unambitious about novelty. The team replicated existing workflows so the tool felt familiar, aiming for users to feel the AI was doing a task exactly as they used to, just faster.

Song notes this has shifted over time. As customers grow more comfortable with AI and recognize the brand, EliseAI now leans toward promising to rethink day-to-day processes rather than mirroring them step by step.

Success is measured against the customer's own numbers. In housing, Song points to occupancy rates, service quality, maintenance resolution times, and resident satisfaction. The internal metric she highlights is the percentage of a workflow the AI can automate, benchmarked against the best human agents on both effectiveness and reliability.

The harder problem is maintaining old products, not shipping new ones

Song's most useful observation for teams shipping AI is counterintuitive. The difficulty isn't only planning for a future where the model landscape may change before you finish building. It's 'planning for the past'—keeping existing products current for a large installed base while new tools keep arriving.

That reframes the fast-moving-model problem as a maintenance and architecture question. Every time EliseAI adopts a new tool for a new product, Song says the team asks whether it applies to already-solved problems in a better way and whether the architecture should be reconsidered.

The specific lesson from EliseAI is that betting on AI in an untouched, non-technical vertical means the model upgrades are the easy part; the durable competitive work is continuously re-evaluating a live product against capabilities that didn't exist when you built it. As Song puts it, if you don't keep improving existing products, someone else will.