Delight Drives Outcomes: What Voice AI's Best Conversations Have in Common

This isn't about pickup rates anymore

Three years ago, a successful voice AI call came down to one number: pickup rates. Every product in the category had to clear the same bar: answer every call, 24 hours a day. For the first time in most businesses, a caller had a 100% chance of reaching someone, and that was better than nothing.

It was a real win for every industry that bought into it. Teams that wanted to provide excellent service but couldn't afford an answering service had the first opportunity to attempt to provide more coverage to leads, answer questions faster, etc.

As models and AI advancements progressed (have you seen Jev?), voice AI has been largely treated like a solved problem. Just hook up an ElevenLabs voice to Twilio, and you've got your own AI receptionist, right?

Here's the issue though: we as an industry have been overindexing on volume, rather than outcomes and experience. As models and architecture progress, a 100% pickup rate is the expectation. Less than that is a bug.

People are already jumping for the AI to make outbound calls, but have we really solved inbound yet?

Haven's agents have been through many iterations since we started in 2023. Starting with maintenance intake, adding leasing nurturing, and of course our receptionist, we've had hundreds of thousands of calls route through our system.

The real question is what happens once the call connects.

What Susan at the front desk already knew

Human-to-human phone calls have been measured for a long time, through surveys, call outcomes, and transcript review. When an employee wasn't handling calls well, you coached them:

• Introduce yourself the same way each time.
• Pause when reciting a number so the person on the other end can write it down.
• Hang up when the call reaches its natural end.

The goal of the call wasn't just the transaction, it was the delight that came along with interfacing with someone from your team.

Voice AI should be held to that same standard of delight, and any agent you present as a part of your team should come with the skills necessary to reflect your brand well. It should create the same delight Susan at the front desk does.

Answering the call is the price of entry in the Voice AI space. Delight is what wins or loses the lead, gets the review, causes the tenant to refer a friend.

Since May of this year, our AI engineering team formed a hypothesis:

Can we superpower the rest of our leasing agent conversions by strengthening their first impression of the agent via phone calls?

You see, we've had a leasing agent for years now. And it got the job done. It answered property info, routed leads to team members, helped book showings, etc.

We have a roadmap of features we plan to add to our leasing product, but before we could do any of it, we had to increase the user's trust of the agent's ability to help.

Voice AI was the first impression for most leads, whether through our Receptionist or Leasing products, so that's what we focused on, for 5 months.

Making Voice AI delightful

Delightful voice agents don't come out of the box, and why would they? At the end of the day, they do not innately understand turn-taking, the ability to be interrupted, or things like sarcasm or tone of voice and its impact on a conversation.

We wanted to make sure our agents acted the way a normal person would expect another normal person to act.

A few takeaways from this sprint

We learned that a leaner agent prompt beats a bigger one. The leasing instructions grew to four times their original size as we added capability, then we cut them in half and the agent got better, not worse. Two phases:

• First we added. New rules, new skills, and a lot of scripted example phrases, until the instructions were long and heavy.
• Then we simplified. We stripped scripted examples or templated messages, merged rules that could overlap with one another, and moved situation-specific facts so the agent assembles them fresh on each call instead of carrying every possibility in its head.

These are some of the skills we gave the agents:

• A scripted line became a plain rule to follow, rather than "say exactly this."
• A hardcoded assumption became a live check. Before: a rule that guessed whether a transfer was possible. After, the agent is told per call whether a live transfer route is available, and never offers one that is not.
• Four overlapping rules became one. Before, four near-identical warnings about promising a callback. After, a single rule: "Never invent a callback promise, promise a specific person, or guarantee when they will respond."
• Scattered tone notes became one personality given to the agent each call.

One thing we tested specifically was our greetings: we changed the default greeting from a menu-style opening to a warm, named one: "Hi, this is [name] from [firm]." Holding the same firms and the same instructions constant, call success rose from about 56% to 64%.

As our September iteration peaked, the share of leasing callers asking for a human slid from 33.1% to 24.0%, week over week. The receptionist, whose job is to route people to a person, stayed flat over the same period, which tells us the leasing change is real and specific, not a quirk of how we measure. To be sure the drop was the agent improving and not just a changing mix of customers, we held one high-volume firm fixed across the whole window and saw the same decline.

The new leasing voice agent went live on August 20, 2026. In its first seven weeks it handled 5,262 calls across 56 property management firms. Our receptionist agent, live since mid-July, handled 36,000 more in the same window.

Weekly call volume grew roughly 40 times over the window as we added firms, from 4 firms to 56. Through that growth, the outcome rate held steady around 60% and the share of calls that turned negative fell by more than half.

Everything below comes from those real calls.

Finding #1: The longer a caller stays engaged, the more often the call ends in a real result.

The job of the agent is to earn every second of the conversation. The calls it keeps delightful are the calls that convert. That is delight doing its work: a caller stays because the conversation is easy, and the outcome follows.

Finding #2: Allowing people to speak to people is a feature, even for delightful agents. The leasing agent hands a caller to a person on about 30% of calls, and the receptionist does so on 62%. Those handed-off calls succeed far more often than the ones the agent keeps to itself. For the receptionist, calls that reach a person succeed 95% of the time, against 40% when they do not. The agent earns its keep by resolving what it can and routing the rest, never by walling callers off. Getting someone to the right person is itself a form of delight.

Finding #3: Delightful calls lead to fast outcomes. On the firms where we coordinate leasing showings, half of booked tours are scheduled within about 14 minutes of the lead's first call. A third happen within five minutes of hanging up. A prospect who calls is booked while they are still deciding, not the next day when a human team would have called back. Delightful.

What we have learned

• Delight is the advantage, and it compounds. The polish that looks cosmetic like speaking a price clearly, dropping a clunky menu, confirming a detail once instead of three times, is what buys the minutes where outcomes happen.
• A warm, human opening changes the whole call. One greeting change moved outcomes by eight points.
• Smaller, sharper prompts beat large, scripted ones.
• Callers are starting to trust the agent. Fewer of them reach for a human the better it gets.

Volume is a vanity output. Growing 40 times only matters because delight held while it happened. If it had not, we would have 40 times more bad calls.

What we are still learning, and what we will measure next

We are clear-eyed about the limits of what we can prove today, and each limit is a measurement we are building toward.

• Tie calls to signed leases. Today we can see a call go well and a tour get booked. Within the next 45 days, we're committing to following prospects through the entire leasing process to track full conversions.
• See every tour, not just ours. About half of our firms book tours through outside scheduling tools we cannot see. We're exploring integrations with other showing coordination tools, so that users can use Haven's delightful agents without having to change their tech stack.
• Close the handoff gap. Our clearest losses are calls where the agent tried to reach a person and no one was available. We'll be showing per-team-member availability rates so that your team can improve your own transfer availability.
• Test how hard to qualify. Some firms have the agent qualify a caller before offering a tour. We want to measure cleanly whether that trades volume for better leads, and help tune it per firm.
• Keep watching what callers do. Reading caller behavior over time, who asks for a human, who stays engaged, who books, is how we tell whether each change actually landed. We will keep doing it on every meaningful release.

The through line is simple. Answering the call is a low bar. Delight is what we compete on, and we get there by building deliberately and measuring what real callers do. Seven weeks of our own production calls say the bet is working, and we know exactly what we want to prove next.