The voiceover rebuild
Lumen5's signature video format was born in 2018 and starting to show it. Kaegan led the rebuild around AI voiceovers, and paid purchases ran up roughly 40% compared to text-only videos.
The setup
Lumen5 made its name on a style the team called text-on-media: stock footage with animated captions, built straight from a blog post. Groundbreaking in 2018. By 2023 it had grown stale, and everyone could feel it. Social feeds had moved to voice-led video, and the output looked like what it was: a format from five years ago.
Context for why that was even possible: Lumen5 was founded as an AI-first product, so the team had spent years figuring out how to solve problems with AI that feel fundamental to the product experience, rather than bolted on. By 2023 the models had finally caught up to the ambition.
The bet was to rebuild the product around voiceover-driven videos, with ElevenLabs providing the voices. Not a feature bolted onto the side. A new spine for the whole creation flow. The strategy holding it together was Kaegan's own, and he wrote it up at length:
The rebuild started as a strategy I came up with and wrote extensively about called "the narrow path". It brought together three separate projects (LLM-powered script building, improved video recommendations, AI voiceover) into a simple strategy: build the most barebones flow imaginable to get a user from idea to an AI voiceover video quickly, and then "expand the path" by adding customization options and solving other pain points along the way. This provided the foundation for well over a year of product work across three teams.
The writing was the artefact, and not by accident:
I created and grew the documentation-first culture at Lumen5, and everything I worked on existed in writing.
It is also what one of the founders singled out:
Your written documentation is top notch and almost certainly the best in the company.
What he owned
- The voice-provider evaluation and the internal case for ElevenLabs over the alternatives. Lumen5 was an early Enterprise customer (ElevenLabs didn't have a sales team yet), and Kaegan owned the relationship and negotiated the deal.
- The transcript-based editing experience: scenes timed to words, click a sentence to seek, edit the script and the video follows.
- Script Composer, the LLM-powered flow that turns a prompt into a voiceover-ready script. It cut video-creation time by about 30% versus starting from scratch.
- An AI onboarding flow that scanned a new user's website, imported their brand colors and fonts, and matched them to a theme, so their first-ever video came out on-brand automatically.
- The pricing and paywalls for the new format: voiceover minute limits, plan tiers, and the paywall experiments that drove free-to-paid conversion.
Two projects, one new front door
The rebuild did not arrive as a single launch. It arrived as two projects back to back, and the measurement discipline between them is the part he tends to lead with now:
Two back-to-back projects completely replaced the new user experience: an LLM-powered script builder, and then a rebuild of our video creation flow around AI voiceover. I ran dozens of A/B tests along the way, first to get to parity in activation between the old and new flows, and then later to decrease time to first value by 30%.
Parity first is the unglamorous half. Replacing the front door of a product with tens of thousands of monthly signups means the new flow has to be provably no worse before anyone gets to argue it is better.
The courage to sunset
For most of Lumen5's life, AI did the narrow, repetitive jobs: media selection, text positioning. The voiceover era was different, because the models were now good enough to collapse the editing work itself. That created a problem nobody puts on a roadmap.
We had to have the courage to demote or even sunset core parts of our app we'd spent years building, because the need for them had been negated by new AI-powered workflows that solved those same problems more simply.
The habit underneath it: regularly collapsing complex creation workflows into shorter, step-by-step AI-powered ones, to save users time and deliver a better output. Every time that worked, something the team was proud of got smaller.
How it played out
The product-sense signal came first, and it is how Kaegan still describes the bet:
From a product-sense perspective, it was clear that building videos around ElevenLabs voiceovers was a win. Videos were no longer text-heavy, meaning we could focus on the visuals or use text for strategic emphasis (which we called "animated highlights"). The voiceovers also gave the videos a very human quality, and our customers on the Enterprise side of the business told us the videos performed a lot better.
Then the numbers agreed. He made monthly active paid users the north-star metric for the push, and usage grew from 0 to 1,000 MAPU, 2X the original goal. Paid purchases ran roughly 40% higher than text-only videos, a comparison across the transition rather than a controlled test. The format hit its adoption goal in about three months and doubled it within the year, with a large share of new customers choosing it. Script Composer beat its own adoption target by 1.7X. Within a year, voiceover videos had overtaken the format that built the company, on both the self-serve and Enterprise sides, and deals were won specifically because of the capability.
Transcript-based editing began as an idea he validated himself, and became the foundation of Lumen5's video creator UX for years to come. He rebuilt it as a live demo: scenes timed to words, click-to-seek, and an AI voiceover button.