AI Voice Agents

The insight that started it all

In early 2025, voice AI was getting genuinely good. But most products were still treating it as a novelty.

We saw something different: what if voice could eliminate entire categories of repetitive human work?

We started with recruiting. Screening calls were eating hours of recruiter time every week.

After 40+ recruiter interviews, the answer was consistent: it's not that the calls are hard. It's that there are just so many of them.

We dogfooded it first

Before pitching a single customer, we handed it to Fireflies' own recruiting team for one real role.

Week one, they were skeptical.

Week six, five roles were running on Voice Agents.

Today, all pre-screening at Fireflies is handled by AI, saving the team 15+ hours every week.

That one-role-to-all-roles pattern later repeated itself across hundreds of customer accounts.

The first build humbled us

For v1, we integrated with a Google Meet to move fast. Smart on paper, painful in practice. The latency killed the experience.

A 400ms pause doesn't sound like lag. It sounds like disinterest. Candidates hated the expeirence.

We dropped the idea of hosting the experience on a video conferencing platform entirely, rebuilt on audio-only on our in-house platform, cut latency by 60%, and went from pilot to production in under six weeks.

Activation was a product problem, not a growth problem

Voice Agents run on credits. The challenge: how do you get users to activate something they haven't experienced yet?

We gave 100 free credits when users created their first voice agent - enough to run a full week of screening calls.

Users who completed their first call converted to paid at 3x the rate of those who never ran one. 62% purchased additional credits within 30 days.

Then we found the real problem

58% of users who started agent setup never finished it. They'd see the configuration screen and leave. Going back to users, the answer was humbling:

"I didn't know what questions to put in. I just wanted to hire someone."

So we asked ourselves: what if they didn't have to fill in anything at all?

We used their email domain to identify their company, scraped their careers page in real time, and pre-populated their live job openings the moment they landed on setup. We made the entire voice agent creation process one click.

Select a role and Fireflies auto-generates the questions, instructions, and knowledge base from the job description.

Users in testing literally said "wait, that's it?"

Setup completion jumped from 42% to 64%. Time to first agent dropped from 11 minutes to under 90 seconds. Agent creation volume tripled in the following month.

What started as a recruiting tool became a platform

Customers kept asking - can it do this for sales? Discovery? Research? We said yes, and meant it.

Recruiting was just the entry point. We expanded into sales outbound and lead qualification, marketing feedback calls and NPS conversations, structured user research interviews at scale, and fully custom agents that teams could build for their own workflows.

To drive adoption across these new use cases, we ran a webinar series - not as a marketing play, but as onboarding at scale. Average attendance was 400 users per session.

Every webinar drove a measurable spike in new agent creation, typically 35% in the week following each session. When users saw real examples of agents running real calls, they stopped imagining it and started building.

Distribution built into the product

Creating the agent was only half the job. We needed users to actually deploy and share them. So we built distribution directly into the product.

Embed on a website - drop a voice agent onto any careers page, landing page, or support page. Share to socials - one-click shareable link for LinkedIn, or socials. Add to calendar - candidates or leads can self-book a voice agent call like any other meeting.

Each of these removed a deployment barrier. The agents weren't just powerful. They were spreadable.

We also shipped voice cloning so teams could run calls in a voice that felt like theirs - consistent, on-brand, and human enough that candidates stopped noticing they were talking to an agent.

Voice-cloned agents had a 22% higher call completion rate compared to default voices. People stayed on the line longer when it felt personal.

Where it stands today

10,000+ MAU. 90,000 calls automated per month. Customer NPS of 71. Recruiters saving 12 hours a week. Voice Agents now referenced in over 40% of new Fireflies sales conversations. The average customer goes from 1.3 active agents at month one to 3.6 by month three.

What this taught me as a PM

Dogfood with real stakes. A sandbox pilot tells you nothing. Putting your own team's hiring pipeline on the line tells you everything.
Activation is a product problem, not a growth problem. The 100-credit give wasn't a discount. It was the product doing its job of earning trust before asking for commitment.
Drop-off data is a gift. The 58% abandonment rate was uncomfortable to look at, but it pointed directly to the insight that led to our best feature.
Non-technical users don't want configuration. They want outcomes. The one-tap setup wasn't a UX improvement. It was a product philosophy shift.
Distribution is part of the product. Embed, share, calendar - none of that was an afterthought. Getting the agent in front of the right person is as important as building the agent itself.
Voice UX has no visual safety net. No spinners, no skeleton screens - just the conversation. It forced us to get the core experience right, not good enough.

Customer Love for Voice Agents

"That was so dope, felt like I was talking to a real person." — Carlexus, 5/5
"Beyond imagination — I liked it!" — harsh D., 5/5
"Seems like a real person." — Chris A., 5/5
"INTERATIVIDADE PERFEITA, TEMPO PERFEITO, DIDATICA PERFEITA" — Vanderson, 5/5
"The session was great — felt like speaking with a real human. She showed me where I was having problems." — Helena, 4.5/5
"Wow this blew me away!" — Sally H., 4.5/5
"I enjoyed the interview. It was as if I was speaking to a human." — Shaberah C., 5/5
"Very different and unique first round experience." — Pavithra, 4/5
"Superb. This is exactly what I need for my business." — Nosa, 5/5