The Meeting Is the Moat: Fireflies.ai’s Path to 16 Million Users

Jacob Effron introduces Krish Ramani, co-founder and CEO of Fireflies.ai, describing Fireflies as an AI-powered voice assistant that records, transcribes, and analyzes meetings. He notes that Fireflies has more than 300,000 customers, over 16 million users, and 75% of the Fortune 500 using it, and that Ramp named it a top AI vendor by corporate spending. Effron frames the conversation around operating at true scale, Fireflies’ evolution since 2016, and how an early-stage AI company can win against Microsoft Teams, Zoom, and Google Meet.

Krish describes a future in which a pre-meeting debrief tells him who he is talking to and what was discussed last time, while the meeting itself becomes a place where his voice is converted into action. After the meeting, an AI assistant would handle the downstream work—updating a CRM, creating tasks, writing documentation—and nudge him on priorities he might forget. Ten years out, he imagines sending his agent to talk to another person’s Fireflies agent so the agents can figure things out together.

Krish still believes meetings should be used to decide, debate, and discuss rather than to transfer knowledge. He says presentations and debriefs should happen before the meeting, often through a Loom or written memo, so the live time is reserved for important decisions and minute details. Human creativity, decision-making, and the human-in-the-loop process remain central; he is not arguing for an AI CEO that makes every call.

Fireflies today is an AI meeting assistant that joins Zoom calls, takes notes, and does work afterward for millions of people. Krish says the product has generated roughly 5,000 years of notes in the last year and is evolving from a note-taker into a work chief of staff. The company’s thesis is that conversations are where work happens and are the most important source of data inside an organization.

A recent Fireflies feature called the Feed surfaces important chatter from meetings a user did not attend, such as a decision about which LLM to use, and presents it as a self-updating newsfeed. Krish says the hard part is prioritization: what matters differs by person, role, and company. Fireflies starts from the data source, then studies each customer’s meeting workflow and integrates with downstream systems such as Salesforce, Asana, and Slack.

Krish argues AI is most valuable when it tells users things they do not know and helps with the minute, annoying tasks of every day. Fireflies prioritizes “automagical” features: a feed that appears without configuration, action items turned into a ready-made task system, and a pre-meeting debrief that reminds users what they promised to follow up on. The goal is to give every knowledge worker the kind of support that previously only an executive assistant could provide.

Krish calls 2016 to 2020 a period of “looking out into the abyss, chewing glass, and hoping things will get better.” He was 21 when he started Fireflies, with about a year of PM experience at Microsoft, and his co-founder had just graduated college. They had to learn enterprise software, use cases, and customers from scratch; he recalls that his co-founder did not know what Salesforce was when they started.

When Fireflies raised its seed round in 2019, investors asked whether transcription would ever be accurate and cheap enough. Krish and his co-founder bet that the cost of transcription would fall and accuracy would reach human level, even though few people were confident of that at the time. Customers initially bought Fireflies for transcription because hiring a human to transcribe a two-hour board meeting cost more than an annual Fireflies subscription.

Transcription is now a commodity, and users expect it to be perfect, so Fireflies moved up the stack to summaries, notes, and action items. Before LLMs, the company used NLP that sliced text into chunks; Krish remembers Salesforce’s 2018 sequence-to-sequence summarization paper as an early signal. GPT-2 and GPT-3 raised the possibility of human-level paraphrasing rather than mere extraction, and Krish admits Fireflies was a couple of years early to the market.

Krish describes model progress in terms of intelligence levels: GPT-4o and Claude 3.5 are like high school or college students, while GPT-3 was more like a 10-year-old, and future models may be PhD-level. But for any founder, the key issue is whether the model’s output is repeatable and consistent. Fireflies saw that the same input could produce a completely different answer, so controlling variance became a major challenge.

To manage that variance, Fireflies built its own AB experimentation platform and tested models from every major vendor. Krish says prompt engineering matters enormously: constraints, JSON formatting, and query parameters can break results, so the team tells the AI to work within the information it has and not get too creative. Fireflies lets customers rate responses and uses a stack-ranking algorithm to match different models to different parts of a summary.

Different models are better at different jobs—one may write a good overview, another good shorthand notes, another good action items—so Fireflies stays flexible across providers. Krish believes current models are sufficient for summarization, and he is more excited about taking meeting outputs and turning them into actions, recommendations, and decisions. He also points to multimodality: a model that can recognize a Wi-Fi password on screen or run a background check on a candidate while a conversation is happening.

Krish expects a future with many specialized agents rather than one agent that rules them all. He imagines the Fireflies Fred agent talking to Harvey.ai’s legal agent to draft something discussed in a meeting, or Perplexity fact-checking a statistic someone cited. In that world, agents are good at specific tasks—search, legal work, meeting context—and work together much like APIs.

Krish says Fireflies does not train on customer data by default, a policy CIOs appreciate. He argues that fine-tuning is expensive, has diminishing returns as base models improve, and is a poor fit for a market that changes weekly rather than yearly or monthly. Fine-tuning made more sense in the GPT-3 era; by the GPT-4 era, it was slowing the company down.

Instead, Fireflies uses prompt engineering and context from meetings to help the AI perform well for a customer. Krish says the company has to be willing to let its “Legos” go when a new model arrives, as GPT-4o’s low-latency voice demos may force contact-center companies to throw out years of latency work and R&D. The broader lesson is to ride the technology wave and not over-invest in assumptions that a new model release can erase.

Krish says Fireflies experimented with building its own ASR engine and an open-source LLM initiative using Llama, but abandoned those efforts because a startup has limited resources and must focus. He recalls that some companies convinced investors they needed to build an LLM from scratch, poured in billions, hired expensive AI engineers, struggled for traction, and eventually crashed. Fireflies’ CTO was bullish on riding the technology wave instead.

Three transformations shaped Fireflies: transcription got cheaper and more accurate, COVID drove video-conferencing adoption, and LLMs brought intelligence at falling cost. Krish argues the only defensible position for an application company is to solve an end-to-end customer problem and go deep into that workflow. Teams and Zoom can offer transcription or general note-taking, but Fireflies aims to help with a hiring decision, closing a deal, or filling out ERP systems.

On pricing, Krish expects a hybrid: a seat-based base plan with unlimited transcription, note-taking, and Ask Fred, plus value- or utility-based pricing for complex, compute-heavy work. Fireflies is self-service and product-led, and if commoditization happens, Krish would rather be the first to commoditize and pass benefits to users. He warns founders not to innovate in a race to the bottom, because model costs will keep falling.

Krish compares the base subscription to a Costco or Sam’s Club model: give as much value as possible, then charge more for costly tasks. Fireflies has been run like a bootstrapped company, is profitable today, and has raised only about $20 million total, with its last round in late 2020 or early 2021. He gives the example of asking Fireflies to comb through a week of sales calls and identify common feature requests; it may cost a couple of dollars today but saves a PM or user researcher roughly 10 hours.

Fireflies can automatically create sound bites and highlight reels from the most action-packed moments of a meeting, an example of using LLMs to automate manual work. Krish is skeptical of vertical SaaS, arguing it is often overpriced, sales-led, and less defensible when general intelligence improves; he prefers horizontal products that users can customize, like monday.com or Notion. Fireflies wants an app ecosystem on top of its meeting data, and Krish says notes, tasks, and contacts make up 70% to 80% of most knowledge workers’ day-to-day work.

When investors ask about Zoom, Google Meet, or Microsoft Teams, Krish acknowledges that incumbents have incredible distribution. But he argues startups can execute better on the core product and go deeper on things that are checklist features for larger companies, where AI may be valued partly for its effect on earnings reports and stock price. Fireflies raised only about $20 million, became profitable, and tries to control its own destiny; Krish also warns that AI fundraising is overhyped and some startups are chasing valuations they cannot justify.

Krish says incumbents deserve credit for moving fast in AI, especially Microsoft, which does not want to repeat its mobile mistake. He recalls writing a school report on personal computing and a letter to Bill Gates, getting a form reply from Microsoft, later working at Microsoft, and eventually sitting in front of Bill Gates to explain how Fireflies pulls data from many meeting platforms. He does not see Fireflies as competing directly with Copilot; instead, every company will have AI features, and Fireflies wants to build the platform and ecosystem rather than remain a feature.

Fireflies’ key metric is the downstream value it creates when conversational data flows into systems of record and makes tools like Notion better for the user. Krish measures the amount of data routing Fireflies does and compares the ambition to Segment or Zapier: a connector that becomes hard to rip out once it is embedded in a company’s workflow. He calls this “conversational infrastructure,” taking knowledge from meeting systems and improving everything downstream.

Krish acknowledges the blank-canvas problem: a powerful AI tool can confuse users who do not know what to ask. Fireflies starts with recommendations and decision-tree-style branching, similar to how Perplexity suggests related questions after an initial query. The company focuses first on universal needs—meetings, notes, search, action items—then teaches users role-specific applications without bombarding them; it also tries to avoid feature creep that would let a simpler competitor take the market.

Fireflies built an internal toolkit for rolling out models, AB testing, and measuring customer response, and Krish believes customers should be the final judge; the company can get a strong signal quickly because of its user base. It has tested granular prompt changes, such as removing or including two or three words, and has optimized processing time from about 30 minutes to 15 minutes to 10 minutes, because faster notes are more likely to be read and forwarded. Krish says the hardest challenge has been infrastructure scale—joining millions of meetings, breaking down a monolith, measuring latency, and handling email rate limits—not the AI or UI.

Fireflies has exceeded its AI rate limits eight times in three months, and Krish says OpenAI employees have told him they have rarely seen such high tokens-per-minute demand, with 10:00 a.m. PST as a peak meeting time. The company works with OpenAI, Anthropic, Llama, Grok, Mistral, and other providers because the field moves too fast to lock into one vendor. Krish is wary of hardware, preferring partnerships because it is hard to get people to use a device that is less convenient than a phone, and he is bullish on adjacent AI areas such as design generation, coding agents, and visual content from Sora and Runway.

Krish calls fundraising and fine-tuning overhyped, and he also warns against focusing too much on cost reduction when the big model providers are already driving costs down. He says founders should ask whether what they are building will still matter two years from now. The biggest surprise for Fireflies was Ask Fred: the team expected users to ask thoughtful questions, but many customers did not know what an LLM was or how to talk to AI, so the feature needed hand-holding, nudges, and about six months of iteration.

Tasks has been a breakout success: Fireflies automatically assigns tasks discussed in meetings, and users report that it brings structure to their work and sometimes replaces a trip back to Asana. Krish notes that before Fireflies did meetings, it had a Chrome extension that followed users across email, LinkedIn, Slack, and messages to create tasks when they promised to send something. He says he would still start Fireflies the same way, and that the first three or four years of struggle built character; the company now values grit and ability to learn over a decade of industry experience.

Krish says he is a regular Perplexity user and is excited by how the company has proven that doing the core search and answer experience really well can challenge Google. He also points to Nvidia as an example of a company that started with dependencies on a larger incumbent and eventually became dominant. Fireflies can be reached at fireflies.ai and Krish is on LinkedIn; he says the company has been low-key on media and PR but has more product updates coming.

In the debrief, the hosts emphasize that Fireflies is profitable, capital-efficient, and one of the most-used AI apps in the enterprise despite not having splashy fundraising headlines. They see Krish’s story as proof that surviving long enough for model quality to arrive can look like luck but is partly made through product execution. They also discuss the “thousand little things” that make horizontal SaaS defensible, the advantage of startup velocity, and how incumbents like Google face more scrutiny—such as the glue-pizza search incident—than a smaller company like Fireflies or Perplexity would.