MeetGeek’s Dan Huru: “We’ve assumed transcription goes to zero for years”
The MeetGeek co-founder on why payment is the only real trust signal, the two questions that decide what his AI does without asking, and the products he refuses to build.
Key takeaways
- Payment — not surveys — was the signal that customers trusted an AI assistant inside private meetings.
- Reversibility and blast radius decide what the AI does on its own: internal and undoable actions run automatically, anything leaving the company needs human approval.
- Transcription is treated as a commodity heading to zero; the product bet is on executing the work a meeting creates.
- Hard limits: no monetizing customer meeting content, no individual surveillance scoring, nothing built for recording people who don't know they're recorded.
- ARR tripled past $2M by moving beyond the transcript into the tools teams already use — shifting from individual sign-ups to team rollouts.
Dan Huru spent years inside Accenture, IBM and Ericsson building automation products — work that, as he tells it, meant constantly measuring where time disappeared inside a company. Meetings were always the largest number nobody owned.
In 2020 he left with two Ericsson colleagues to build MeetGeek, now a Bucharest-founded AI meeting assistant that records, transcribes and increasingly executes the work that meetings create. The company raised €1.6 million in September 2025, led by Early Game Ventures with Inspire Capital participating, after annual recurring revenue tripled past $2 million.
We asked him about trust, what he refuses to build, and why he has assumed for years that transcription is worth nothing.
From Ericsson to founding MeetGeek
Before we get into MeetGeek, could you introduce yourself and tell us how your journey from software engineer and university teacher to startup founder unfolded?
I'm Dan Huru, co-founder and CEO of MeetGeek. Engineering background: Electronics and Communications at Politehnica Bucharest, then a Master's in Computer Science, then a PhD in AI. I taught there while finishing it. Worked as a software engineer, later an engineering manager, at Accenture, IBM and Ericsson.
My last years in corporate were on automation products, which means you end up measuring where time goes inside a company. Meetings were always the biggest number nobody owned. I did the Founder Institute in 2020, pulled my co-founders in, and quit to do this properly.
You spent years building complex systems and managing engineering teams at Ericsson. Was there one particularly frustrating meeting that made you think, “There has to be a better way to capture what happens here”?
Not one meeting. At Ericsson we were building task automation, so we spent a lot of time analysing what people did with their tasks. The analysis meetings were repetitive and usually took longer than the task we were automating. That's what stuck with me: we were burning hours in meetings to understand work that took less time than the meeting did.
We were burning hours in meetings to understand work that took less time than the meeting did.
MeetGeek was started by three technical co-founders who had previously worked together. What convinced you that your former colleagues were the right people to build this company with, and how did you divide responsibilities during the earliest stage?
We'd already worked together at Ericsson on hard problems. I'd seen how they behave when something breaks at 2am and when they turn out to be wrong. You don't get that from interviewing someone.
All three of us being technical had one obvious benefit: no arguing about whether something is hard. The downside is we all wanted to build instead of sell. Took a while to fix.
The split was rough at the start. They kept the product moving, I took the rest — pricing, support, sales calls, contracts.
Early product and earning trust
What did the earliest version of MeetGeek look like, and which assumption about meetings or user behaviour did real customers prove completely wrong?
Recording, transcript, highlights. That's all it did.
We didn't walk in with big assumptions. We ran over 300 discovery calls, so most of what we learned came from asking rather than from shipping and failing.
What surprised me was how little people wanted to open the product. If they never log in and still get what they need, that's a win for them, so we built around that.
Inviting an AI assistant into private meetings requires a high level of trust. What was the first moment when you felt customers were not merely trying MeetGeek, but were comfortable making it part of their daily work?
When we turned on payments. Free users try anything. Someone putting a card down for a tool that sits in their private meetings answers the trust question better than a survey does.
Then we saw where it was being used. We built it for sales calls and interviews. It showed up in 1:1s and leadership meetings.
Someone putting a card down for a tool that sits in their private meetings answers the trust question better than a survey does.
Beyond transcription
Transcription and meeting summaries are quickly becoming standard features inside Zoom, Microsoft Teams and Google Meet. If those features become commodities, what is the deeper problem MeetGeek is actually building to solve?
We've assumed transcription goes to zero for years. It's the cheapest thing we do.
What's left is task execution. A meeting creates work and someone still does that work by hand: updating the CRM, opening the ticket, chasing the follow-up, flagging the risk. To do that for them you need more than a conversation’s summary. Then there's the split across tools. Zoom for clients, Teams because IT picked it, Meet for partners, plus calls and rooms. Native notes leave your history in five places.
We've assumed transcription goes to zero for years. It's the cheapest thing we do.
MeetGeek is moving from a passive note-taking assistant towards agentic AI that can update CRMs, create tasks and execute follow-ups. How do you decide which actions AI should perform automatically and which should always require human approval?
Two questions: can it be undone, and how far does it reach.
If it runs inside your own systems and is reversible, we just do it: CRM fields, tasks, alerts, flags. Undoing any of those takes seconds.
If it leaves the company with someone's name on it, a human approves: Customer emails, prices, dates, anything that commits you.
The tricky part isn't accuracy on the transcript, because a model can be right that a sentence was said and wrong about what it meant. People hedge, think out loud, change their mind four minutes later. That's why we ship narrow Skills tied to specific jobs instead of one agent with broad permissions, and why every action points back to the moment in the call it came from.
A model can be right that a sentence was said and wrong about what it meant.
Lines we won’t cross
Meeting data can include confidential deals, employee conversations and sensitive customer information. What privacy or ethical line would you refuse to cross, even if crossing it could produce a more powerful product?
We don't make money from customer meeting content. Not selling it, not packaging it, not training models on it unless someone explicitly opts in.
No individual surveillance either. People ask for it — e.g., take the talk ratios and participation data, turn it into scoring for managers — but that data exists so a team can fix its own meetings, not so someone can build a case against an employee.
And nothing built for recording people who don't know they're recorded. If the product only works when participants are unaware, we don't build it.
If the product only works when participants are unaware, we don't build it.
Growth, funding and leadership
MeetGeek’s annual recurring revenue tripled to more than $2 million before the company raised its €1.6 million round. What changed inside the business to produce that growth, and what can the company now pursue that was previously out of reach?
Growth came from getting past the transcript. Once meetings started feeding the tools teams already used, we stopped being something individuals tried and became something teams rolled out. Different buyer, different expansion. The funding pays for the automation layer. Integrations and agents are slow and expensive to build, and hard to justify when every quarter has to pay for itself.
As a technical founder, which transition has been harder: moving from building software to selling a vision, or moving from solving problems yourself to trusting a growing team to solve them?
Both, at different points. I wouldn't rank them. Selling was the early one. Engineers open with caveats, which kills a sales conversation. Delegating came later and takes longer to learn because you find out slowly. Bad code tells you in minutes. A bad call on a person takes months.
Bad code tells you in minutes. A bad call on a person takes months.
What comes next
Imagine MeetGeek succeeds exactly as you envision it. Five years from now, what will happen automatically after someone says, “Let’s move forward,” and which familiar post-meeting task will have disappeared entirely?
Deal stage updated, tasks assigned, contract draft generated, next meeting booked with an agenda based on what was left open, briefing sent to everyone who wasn't there… and the list goes on. Much of this we already do.
Five years from now, AI won't just handle what happens after meetings. Voice Agents will be taking certain calls autonomously, ask the right questions, gather what they need, and trigger the workflows that follow.
We work with teams to automate the workflows their meetings drive, rather than just handing over a tool. What disappears is the follow-up email, the CRM update everyone does two days late, and many of the status meetings people attend just to stay aligned.