Fireflies.ai co-founder Sam Udotong says the company’s first “AI” note-taker was, at times, two humans silently dialing into customers’ meetings and writing notes by hand. The admission, describing practices in 2017 with early beta users paying $100 per month, spotlights rising concerns over consent, privacy, and marketing claims during the rush to sell AI-powered tools.
Udotong shared that the approach helped test demand before building a product at scale. He maintains early users knew a person could participate. Critics argue the practice blurred the line between prototyping and misrepresentation, raising legal and ethical questions that now face the wider AI sector.
The Early Product and the Pitch
“We told our customers there’s an ‘AI that’ll join a meeting.’ In reality it was just me and my co-founder calling in to the meeting sitting there silently and taking notes by hand.” — Sam Udotong
Fireflies.ai launched in 2017, promoting automated meeting notes. According to Udotong, the team joined calls to simulate the service and learn from real scenarios. Users were charged for access during this beta period.
Udotong says it was known to these early customers that “there was a human in the loop.”
The account highlights how startups sometimes run “Wizard of Oz” tests, where humans perform tasks a future system is meant to automate. Supporters say this can validate a product faster. But it also risks overselling capabilities and mishandling sensitive data.
Privacy and Disclosure at Stake
The practice raises a core issue: did every person on each call know a third party was listening? Meetings often include confidential plans, personal data, or protected information. Even if a buyer agreed, other participants might not have given informed consent.
Privacy specialists warn that silent note-takers can create exposure if participants are not told who is present. Companies also face duties under contracts and data protection laws to safeguard meeting content and limit access to need-to-know parties.
Marketing claims are another risk. Selling an “AI that’ll join a meeting” while using people behind the scenes can be deemed misleading if not clearly disclosed. Regulators have increased scrutiny of AI advertising and the accuracy of automation claims.
Common Tactic, Uncommon Transparency
Human-in-the-loop methods are common in AI development. They improve accuracy, train models, and provide quality control. But the method depends on clear consent and data hygiene. Without that, companies may lose trust and invite penalties.
Experts say responsible use requires three basics:
- Plain disclosure of human involvement.
- Consent from all participants on a call.
- Strict data handling and deletion policies.
Enterprise buyers now ask detailed questions about how AI tools gather and store meeting content. They also seek audit logs and encryption standards. Those demands have hardened since 2017, as AI meeting assistants have spread across workplaces.
Industry Impact and Legal Risks
The episode highlights the pressure on AI startups to show traction early. It also shows how easily trust can erode. If customers later learn that people listened to meetings, they may question every feature and every claim.
Consumer protection laws can apply when advertising suggests a level of automation that does not exist. Data protection rules can apply when third parties access personal or confidential information without proper notice. Class actions over AI training data and scraping have already signaled a tougher climate.
At the same time, investors and customers still prize speed. Companies that balance rapid testing with strong disclosure may avoid backlash while building better products.
What to Watch Now
As meeting assistants spread, policy trends point to tighter rules. Regulators are drafting guidance on AI claims and transparency. Large clients are adding clauses on human access, data residency, and retention limits.
For vendors, the playbook is changing. Clear labels for human involvement, call-wide notices, and opt-out tools are becoming standard. So are third-party audits and strict deletion timelines.
For buyers, due diligence now includes asking who or what joins a call, how notes are stored, and whether training uses client data. The bar for trust is rising across the market.
Udotong’s account captures a moment from AI’s scrappier phase. The lesson for today’s builders is simple: be clear, get consent, and protect the data. The companies that do so will be best placed to win long-term confidence as automated note-taking becomes routine.
