Solab: AI meeting notes where every claim shows its moment

In about two years the AI notetaker went from novelty to default attendee. Zoom told investors in May that paid users of its AI Companion had grown 184% year over year, and that its My Notes feature had reached 1.5 million licensed users within four months of launch; in March, Granola — an app built around meeting notes — raised $125 million at a $1.5 billion valuation. Then July 2026 became the month the industry started retrofitting consent onto what it had already shipped. On 16 July, Google added new controls for Meet’s note-taking, including starting notes automatically only when a meeting has three or more participants. Microsoft announced a mid-meeting switch to shut Copilot’s meeting AI off “when conversations shift to sensitive topics” — and on 9 July moved its arrival back to the second half of August. That same week an Associated Press feature, syndicated across the US press, asked whether the bots belong in meetings at all, quoting the CEO of the HR certification body HRCI: “I don’t think companies should use it at all.” Bloomberg had already named the genre in late June: the uninvited meeting guest.

The bot in the room is now a legal question

The lawsuits make the discomfort concrete. In a California federal court, a consolidated class action alleges that Otter’s notetaker records meeting participants who never consented and that conversations feed the company’s model training; Otter argues no unlawful interception occurred, and the motion to dismiss, argued on 20 May, remains undecided as this post goes out. Two Illinois suits allege that Fireflies’ speaker recognition builds voiceprints of participants — including people who never signed up — without the written consent the state’s biometric law requires. These are allegations, not findings; but institutions are not waiting: the University of Washington blocked Read AI’s integrations because the bot “can join, transcribe, and summarize meetings without the awareness or consent of other attendees.” And in a July 2026 survey of 500 US workers commissioned by the law firm Kolmogorov Law, a third had had an AI notetaker in their meetings — while only about a third of those said permission was always asked first.

The failure modes are not hypothetical either. Fortune spent February documenting what it called HR nightmares: everyone leaves the call, the bot stays, and the transcript of what was said next goes out to the full invite list. The canonical case is older — in 2024, a venture firm’s notetaker kept transcribing after a researcher left the meeting, then mailed him hours of the investors’ private conversation, their own “cooked metrics” included. He walked away from the deal.

The second failure: notes nobody can check

Consent is only half the problem. The other half is what the note claims. In a study presented at ACM FAccT in June 2024, researchers testing OpenAI’s Whisper speech-to-text found hallucinated content in 1.4% of transcription segments — and 38% of those hallucinations carried real harm: invented violence, fabricated names, false attributions. A later re-test found a model update had removed most of them, which is precisely the point: transcription sometimes invents, quietly, and is fixed just as quietly. An Associated Press investigation that autumn found a University of Michigan researcher hitting hallucinations in eight of the ten public-meeting transcripts he inspected — and reported that a widely deployed medical scribe built on Whisper erases the original audio, leaving nothing to check the transcript against.

Erased audio is the extreme case of a general design: most meeting AI hands you a fluent summary and no practical way to verify a single line of it short of re-listening to the whole recording. And people don’t. A 2025 CHI paper from Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers and found that the more confidence people place in generative AI, the less critical scrutiny they apply to its output. A fluent, unverifiable summary is not a note; it is a decision waiting to be made on faith.

Cited, or it isn’t in the note

Solab is our conversation-intelligence product, and this is the rule it is built around. Every claim in a Solab note links back to the exact, timestamped moment in the recording where it was said — click it and you are listening to the source. When the record does not contain an answer, the note says so: unknown stays unknown, instead of turning into a confident sentence nobody actually spoke. That one property changes what a note is. Checking a claim stops costing a re-listen and starts costing a click, and the summary you circulate becomes something colleagues can audit rather than something they have to take on faith.

A record you make on purpose

Solab’s answer to the consent question is structural: there is no calendar bot. Nothing joins your meeting on your behalf, lingers after you leave, or mails a transcript to the invite list on its own. You record in the browser — deliberately, visibly — or bring the material afterwards: audio, video, a media link, even a PDF. A record with an owner is a record you can be honest about: the person who starts it knows it exists, can say so at the top of the meeting, and decides who ever sees the note. And we concede the convenience honestly — an auto-joining bot is easier than pressing record. That trade is the design.

One conversation, five ways to read it

What comes back is not a wall of text. Solab transcribes — timestamped, multilingual and speaker-aware — then builds a note you can read five ways: Overview for the gist, Analysis for what mattered, Transcript for the verbatim record, Knowledge map for how the ideas connect, and Chat for questions grounded in what was actually said. Semantic search spans your whole workspace, so “what did we decide about the migration?” finds the moment across every conversation you have captured, and notes export cleanly to Markdown, PDF or calendar files.

• Capture anything: browser recording, audio and video upload, media-URL import, PDFs and documents
• Timestamped, multilingual, speaker-aware transcripts
• Five views per note: Overview, Analysis, Transcript, Knowledge map, Chat
• Cited AI — every claim links to its exact moment; unknown stays unknown
• Workspace-scoped semantic search, insights and export (Markdown / PDF / .ics)

Whose model hears the meeting

Where the record lives — and what it feeds — differs more between vendors than most buyers assume. Otter’s privacy policy, effective 16 June 2026, states that the company trains its proprietary AI on de-identified audio recordings and on transcriptions, which may contain personal information. Fireflies’ policy states the opposite: no personal information used for model training. Both are choices a vendor is allowed to make; the point is that they diverge this sharply, that the detail sits in policies most teams never read — and that either way, the audio of your meetings is typically processed and stored on infrastructure in the United States. Solab’s answer is the platform’s answer: transcription runs on our in-house Solab speech-to-text family — the model family the product is named after — through the Qevron gateway. Our models, our gateway: your meetings are not fanned out to a third-party transcription API whose terms can drift under you.

Cited, or it isn’t in the note. Known, or it stays unknown.

In Türkiye the frame is the KVKK. A recording of an identifiable person’s voice is personal data, and whatever legal basis the recording rests on, the Article 10 duty to inform stands. The Board’s decision 2023/1461 drew the line precisely: it found a school’s continuous audio recording disproportionate and unlawful even where the video could be justified, fined the audio and the missing notice separately, and ordered the recordings destroyed. The case law also accepts the other direction — a recording can be lawful without explicit consent where it is the only way to establish or protect a right. Add it up and the lesson is not “don’t record”: what is indefensible is the unannounced, unbounded record, while a deliberate one with an owner, a scope and a note trail is exactly what the law rewards. Solab, from Arpanet Bilişim A.Ş., was engineered for the KVKK from its first line. You can start free at solab.me — and for questions about deployment and scale, contact us and we will scope it with you.