Fireflies.ai is an AI meeting assistant that joins calls, transcribes them, and automatically generates notes, summaries, action items, and follow-ups. Founded in 2016 by Chris Roman (who was 21 at the time) and a co-founder fresh out of college, it now serves over 300,000 customers and 16 million users, including 75% of the Fortune 500. Ramp ranked it the 4th highest AI platform by corporate spend. The company is profitable, has raised only about $2 million total, and has been largely bootstrapped in mindset.
How meetings will evolve with AI
Before meetings: AI will brief you on who you’re meeting, what you discussed last time, and what topics to cover — essentially replacing the human EA prep.
During meetings: An AI assistant listens in real time, turning conversation into action items, CRM updates, and documentation as you talk, rather than requiring you to do that work after the call ends.
After meetings: AI nudges you about priorities, reminds you of commitments, and surfaces what you promised to follow up on — holding you accountable.
10 years out: Chris envisions a world where your AI agent talks to the other person’s agent and they figure things out on your behalf, rather than humans needing to meet at all.
What stays human: Meetings should be reserved for deciding, debating, and discussing — not for information transfer, which can happen asynchronously beforehand via Loom, writing, etc.
Fireflies’ current capabilities
Core product: An AI notetaker that joins Zoom (and other platforms), transcribes meetings, and generates notes, summaries, and action items. Word-of-mouth has been the primary growth driver.
The Feed: A self-updating news feed that surfaces important decisions and discussions from meetings you didn’t attend — e.g., if your team is debating which LLM to use, Fireflies pushes that to your feed.
Pre-meeting debriefs: If you’re meeting someone again after two months, Fireflies reminds you what you discussed last time and what you promised to follow up on.
Task management: Automatically creates tasks from action items across all your meetings, with a ready-made task management system. Early NPS feedback has been very strong, with some users saying they no longer need Asana.
Sound bites / highlight reels: AI automatically clips the most action-packed moments of meetings into shareable highlight reels.
Ask Fred: An assistant you can query about your meeting data — e.g., “What were the most common feature requests from sales calls this week?” — which would take a human researcher ~10 hours but takes minutes.
How new models have impacted Fireflies
2016–2020 was “chewing glass”: LLMs didn’t exist at usable quality. The company bet on transcription costs falling and accuracy improving, which happened. Early customers bought Fireflies purely for transcription because hiring a human for a 2-hour board meeting transcription cost more than a full year of Fireflies.
GPT-2/GPT-3 unlocked summarization: Before LLMs, they used basic NLP that just sliced up text. GPT-3 enabled human-level paraphrasing and true summarization rather than extraction.
Today’s models are “good enough” for summarization: The real unlock Chris is waiting for is models smart enough to take actions and recommendations based on meeting content — not just summarize it.
Multimodality is the next frontier: Chris is excited about models that can see what’s on screen (e.g., a Wi-Fi password), run background checks on candidates in real time, or fact-check statements during a meeting by querying the web.
Shortcomings of current models and how Fireflies handles them
Inconsistency is the biggest problem: Running the same query eight times can yield different answers. Sierra (customer support AI) published data showing that even if a model is accurate 80% of the time, the odds of it being correct eight times in a row are ~20%.
Fireflies’ solution: They built their own A/B experimentation platform. They don’t believe in fine-tuning (more on that below). Instead they use heavy prompt engineering, constrain output formats (e.g., JSON), and test across every model vendor. They found different models excel at different parts of a summary — one does overviews well, another does action items — so they flexibly combine them.
Customer rating as ground truth: They let users rate responses and use a stack-ranking algorithm. With millions of users, they get strong signal quickly.
Why Chris doesn’t believe in fine-tuning
Fireflies does not train on customer data by default — CIOs appreciate this.
Fine-tuning is expensive and returns diminish as base models improve weekly. What made sense at GPT-3 makes less sense at GPT-4 and will make even less sense at GPT-5.
Prompt engineering + meeting context is sufficient: You can give the model enough context from a customer’s meetings to get great results without fine-tuning.
Speed of market change: Fine-tuning slows you down. The market shifts weekly, and you need to be able to throw assumptions out the window and adopt new models quickly.
Analogy: When GPT-4’s low-latency voice interaction arrives, contact center companies that spent 2 years and billions building custom solutions will likely just adopt the new model and throw their old work away.
How to build around models that keep improving
Don’t build what the base model will soon do for free: Fireflies experimented with building their own ASR engine and open-source LLM but realized as a startup with limited resources, focus was critical.
The defensible moat is workflow depth, not model capability: Even if Zoom and Teams offer basic transcription and note-taking, Fireflies competes by going deep — helping you make a hiring decision, close a big deal, or fill out ERP systems based on conversation data.
Commoditize yourself first: As models get cheaper, Fireflies’ philosophy is to pass those savings to customers immediately rather than trying to maintain premium pricing on capabilities that are becoming commoditized.
Pricing model: Hybrid — seat-based pricing for base features (unlimited transcription, notes, Ask Fred), with value/utility-based pricing for complex tasks that require significant compute.
Horizontal vs. vertical SaaS in the AI era
Chris’s controversial take: Vertical SaaS is extremely difficult and often overpriced. In a world where general intelligence keeps improving, a horizontal tool that lets you customize for your vertical (like Monday.com or Notion) is more defensible than a vertical-specific product.
Fireflies’ approach: Notes, tasks, and contacts are 70–80% of what every knowledge worker needs regardless of industry. They build those as universal building blocks, then let users tell Fred “I’m in pharma, surface these insights” — or let third-party developers build industry-specific AI apps on top of Fireflies’ platform.
The agent ecosystem future: Rather than one agent ruling everything, Chris envisions specialized agents (Fireflies for meetings, Harvey for legal, Perplexity for search) talking to each other via APIs, similar to how SaaS integrations work today.
Competing with incumbents (Microsoft, Google, Zoom)
Incumbents have distribution but different incentives: AI is often a checklist feature for them — it looks good on earnings reports. For a startup, AI is the entire product, so you can go deeper on specific workflows.
Corporate bureaucracy is a real advantage for startups: Fireflies can be lean, AI-first, and out-innovate on specific problems.
Trust with sensitive data: Meeting data is as sensitive as it gets. Chris credits incumbents like Microsoft for moving fast on AI (not repeating their mobile mistakes). Fireflies earned trust by not training on customer data and building security infrastructure to handle 75% of the Fortune 500’s conversational volume.
Chris’s personal arc: He wrote a letter to Bill Gates as a kid about personal computing, later worked at Microsoft as a PM, and recently sat across from Gates presenting Fireflies — while knowing Microsoft was building Copilot. He sees Fireflies not as competing with Copilot but as building the conversational infrastructure layer across all meeting platforms.
What metrics matter most
Downstream value: The key metric is how much Fireflies improves downstream systems — is Notion better because you use Fireflies? How much data routing are they doing?
Conversational infrastructure: Inspired by Segment and Zapier, Fireflies aims to be the piping that runs conversational data through your company — hard to rip and replace once embedded.
The Northstar is integration depth: Not just “cool meeting notes” but whether the output actually flows into Salesforce, Slack, Asana, and other tools in the organization.
Teaching users to use AI products
The blank canvas problem is real: Users don’t know how to talk to AI. Fireflies learned to start with recommendations and branch out like a decision tree (similar to Perplexity’s “related questions” approach).
Nudge, don’t bombard: Get users comfortable with universal needs first (notes, search, tasks), then surface industry-specific applications once they’re hooked.
Simplicity as defense against disruption: Every product gets feature creep. Fireflies stays focused on a clean first-time user experience so a “simpler version of us” doesn’t eat their lunch.
Model evaluation at scale
Internal A/B testing framework: Fireflies built their own toolkit for rolling out different models, A/B testing them, and measuring results. They test granular changes (e.g., removing 2–3 words from a prompt).
Customers are the ultimate judge: With millions of users, they get strong positive/negative signal quickly. Usage data tells them whether a model is delivering value.
Multi-model strategy: They’ve tested Llama, Grok, Mistral, Anthropic, OpenAI — and use different models for different tasks based on performance.
How Fireflies is so fast
From 30 minutes to near real-time: When they started, it took ~30 minutes to process a meeting. They got it down to 15, then 10, and continue optimizing. Speed creates a positive feedback loop — faster notes mean more people look at them, forward them, and find them useful.
The hardest challenge isn’t AI — it’s infrastructure: Joining millions of meetings simultaneously (peak at 10am PST), managing rate limits (they’ve exceeded OpenAI rate limits 8 times in 3 months), and building the security and reliability to handle 75% of Fortune 500 volume.
They’ve exceeded OpenAI rate limits 8 times in 3 months — Chris says OpenAI folks told him they’ve never seen anyone need that many tokens per minute.
Hardware and adjacent spaces
No hardware plans: Chris believes hardware is insanely difficult and that people won’t adopt something meaningfully different from their phone. Fireflies partners with hardware companies rather than building their own.
Adjacent spaces Chris is watching: AI design tools (prompt-to-UI), coding copilots, and visual content generation (Sora, Runway) — because visual storytelling is powerful for enterprise use cases like investor decks and sales presentations.
Over-hyped and under-hyped
Over-hyped: Fundraising in AI is insanely overhyped. Startups chasing unjustifiable valuations. Also, fine-tuning and cost-reduction services are overhyped — you’re competing with the big players who are already doing that.
Under-hyped (implied): The value of going deep on customer workflow and the thousand small product decisions that compound into a magical experience.
Biggest surprises in building Fireflies
Ask Fred took 6 months of iteration: They assumed users would be smart about querying, but many customers don’t even know what an LLM is. They had to do extensive handholding, nudges, and suggestions before it became a breakout feature.
Tasks was an unexpected hit: Early data and NPS are very strong. The concept originally existed as a Chrome extension pre-LLM era (tracking commitments across email, LinkedIn, Slack) and has now come full circle.
What Chris changed his mind on
He initially thought he needed more corporate experience before building enterprise SaaS, but in retrospect he’d do it the same way. The first 3–4 years of struggle build character. He now values grit and persistence over industry experience in hiring.
Other AI startups Chris is excited about
Perplexity: Chris is a regular user. He sees them as proof that if you do the core thing really well, you can take on incumbents — even Google. The quality of answers has been impressive.
