AI Meeting Notes Cost Calculator

The cost being displaced here is the labor of writing up a meeting after it ends. That is measured directly: minutes of manual note work per meeting, converted to hours, priced at a loaded hourly rate, and scaled to monthly volume using 4.33 weeks per month. Fifteen minutes of write-up after a 45-minute meeting is typical, and at 15 meetings a week that is roughly 16 hours a month of pure transcription and summarization labor. Against it sits per-seat software priced in a narrow band - Fireflies around 18 dollars, Fathom and Grain around 19, Otter around 20 per user per month - which means the comparison usually turns on how many seats you actually need rather than on the sticker price. The accuracy figures attached to each tool are word error rate estimates, the standard automatic speech recognition metric defined as substitutions plus insertions plus deletions divided by reference words. Modern systems reach single-digit word error rates on clean read speech but degrade substantially on real meeting audio, where overlapping speakers, accents, crosstalk and poor microphones push errors higher, and where speaker diarization errors compound them. Quoted accuracies of 90 to 93 percent are vendor-reported and should be treated as optimistic for technical vocabulary and proper nouns. Two further caveats: recording consent law varies, with California Penal Code section 632 and about ten other US states requiring all-party consent, and stored transcripts become discoverable records with retention and confidentiality obligations of their own.

Capturing speech verbatim was a trained human skill for most of recorded history. Isaac Pitman published his shorthand system in 1837 and John Robert Gregg his competing system in 1888; the stenotype machine, patented in the United States in the early twentieth century, let court reporters keep pace with natural speech. Corporate minute-taking inherited that tradition, and the secretarial write-up remained the standard method of meeting documentation well into the 1990s. Machine transcription advanced in slow steps and then very quickly. Bell Labs built Audrey in 1952, able to recognize spoken digits from a single speaker. IBM demonstrated the Shoebox in 1962. DARPA's Speech Understanding Research program from 1971 to 1976 produced Carnegie Mellon's Harpy system, and hidden Markov models became the dominant approach through the 1980s. Dragon NaturallySpeaking brought continuous dictation to consumer desktops in 1997, though it still required per-speaker training. Deep neural network acoustic models from around 2012 cut error rates sharply, and end-to-end sequence models removed much of the hand-built pipeline. Purpose-built meeting tools followed: Otter.ai was founded in 2016 under the name AISense, with Fireflies.ai emerging the same period. OpenAI's release of Whisper as open weights in September 2022 commoditized the underlying recognition step, shifting competition among meeting-notes vendors toward summarization, action-item extraction and calendar and CRM integration.