Skip to content
AI Workshop

AI Meeting Assistants 2026: Copilot, Gemini, Zoom and Granola — Time Saved, Accuracy and GDPR

In 2026 the AI note-taker became the most widely used AI tool most employees never chose. It joins calls as a bot, or silently listens through the laptop, and sends a summary before the next meeting starts. According to Fellow.ai's survey, 75% of professionals now use an AI note-taker in work meetings, about twice as many as in 2023. Granola, a bot-free notepad, raised $125 million at a $1.5 billion valuation in March. And in August a US federal court let wiretapping claims against Otter.ai go forward. For companies in Croatia, Germany and Austria the question is no longer whether people use these tools, but whether they use them in a way that is productive, accurate and legal.

75%
of professionals use an AI note-taker in work meetings (Fellow.ai survey)
47%
of users report accidental capture of content they did not mean to share
$1.5B
Granola's valuation after its $125M Series C (March 2026)

Adoption is high, and so is unease

The numbers behind the trend show a split. Most professionals use an AI note-taker, yet the same surveys find that half of non-users stay away because of privacy and security, almost half of users have seen the tool capture something it should not have, and 84% say they change how they speak when a note-taker is in the room. That last figure matters more than it looks: if people self-censor, the meeting record becomes less honest exactly when it becomes more complete.

AI Note-Takers: Adoption vs Trust Gap (% of respondents)

Sources: Fellow.ai survey (75% adoption); figures on behaviour change, privacy barrier and accidental capture as compiled in 2026 industry surveys (Laxis, Saner.ai). Mostly vendor research, so read the values as directional.

Adoption is also uneven. One 2026 dataset puts it at 72% in SaaS and tech companies but 43% at companies with more than 5,000 employees, where legal and IT reviews slow rollouts down. In practice that means many large companies have a policy gap: employees install free tools themselves, a textbook case of shadow AI.

A market nobody agrees on

Analysts cannot even agree on how big the market is, because they define it differently. Precedence Research sizes the narrow AI note-taking segment at about $0.74 billion in 2026 and the broader meeting-assistant category at $1.42 billion. Research and Markets says $3.91 billion, Grand View Research $4.31 billion. The forecasts agree on direction: roughly 19% to 24% annual growth.

AI Meeting Assistant Market Size: Four Estimates for 2026 (USD bn)

Sources: Precedence Research (AI note-taking: $0.74B in 2026 to $3.48B in 2035, 18.75% CAGR; AI meeting assistants: $1.42B), Research and Markets ($3.91B in 2026 to $9.33B in 2030, 24.3% CAGR), Grand View Research ($4.31B). Scope definitions differ.

Who offers what

For most companies the first choice is the assistant already built into the meeting platform they pay for. Independent tools win when teams switch between platforms or want notes that also cover in-person meetings.

ToolHow it capturesPricing modelWatch out for
Microsoft 365 Copilot in Teams (Intelligent Recap)Inside Teams, needs recording or transcriptionCopilot add-on; Copilot Business for under 300 users listed at $21 per user per monthTeams only; licence cost per seat
Gemini in Google Meet ("Take notes for me")Inside Meet, notes saved to Google DocsIncluded from Workspace Business StandardAdmins can require participant consent before notes start
Zoom AI notesZoom, plus Teams, Meet and in-person via My NotesIncluded with paid Zoom plansProduct naming changed in 2026; check the current plan
GranolaDevice audio, no bot joins the callPer user subscription; team workspaces and APIs since March 2026No visible bot means notice to participants is your job
Otter.ai and other botsA bot joins the call as a participantFreemium, per userOngoing US class action over consent

Sources: vendor pages and 2026 comparison guides (Fellow.ai, TechTarget, UC Today). Prices change often; verify before buying.

The more interesting shift is where the notes go next. Granola's MCP server is an official connector in Claude and ChatGPT, so an assistant can answer "what did we promise this client last month?" across all your meetings. Meeting notes are turning into company memory that agents can query, which makes them far more useful and far more sensitive. We explain the plumbing in our guide to the Model Context Protocol, and how Copilot and Claude compare for office work in AI office productivity.

How much time does it really save?

Time-saved figures span almost an order of magnitude. Zoom's own commissioned analysis estimates about 1.6 hours per week on meeting follow-up. Independent research lands at roughly 3 to 5 hours. Vendor marketing goes up to 8 to 10 hours. The gap comes from definitions: some count only note-writing, others add the time saved by skipping meetings and reading the summary instead.

Time Saved per Employee per Week: Claimed vs Measured

Sources: Zoom-commissioned ROI analysis (about 1.6 h); independent estimates of 3–5 h and vendor claims of 8–10 h as summarised by StealthAgents research (2026). Bars for ranges show the midpoint.

Our experience from client projects matches the middle of the range: the real gain is not the transcript but the action items that arrive in the task tracker without anyone retyping them. If the summary is not connected to where work happens, most of the value is lost.

Accuracy: good on clean English, weak in real meetings

Speech recognition has become very good under good conditions. The best system on one independent leaderboard (July 2026) reached a 3.5% word error rate on conversational audio, and leading engines average about 5.6% on English. Real meetings are harder. Speaker attribution adds its own errors: the best result in the MISP 2025 challenge still misattributed 8.1% of speech. In multi-speaker clinical conversations, error rates above 50% have been reported. Croatian, regional dialects and mixed-language meetings usually score worse than English; our article on voice AI and speech interfaces covers why.

Transcription and Speaker-Attribution Error Rates (%)

Sources: independent ASR leaderboard snapshot, July 2026 (3.5% WER); published averages for leading engines on English (5.6%); MISP 2025 challenge (8.09% diarization error rate); reported results for multi-speaker clinical speech (above 50%). Different test sets, not directly comparable.

The practical consequence: a summary can be fluent and wrong. A name assigned to the wrong speaker turns into an action item for the wrong person. Every summary that leads to a decision, a client commitment or an HR matter needs a human owner who checks it.

The legal side: consent is the product

The first wave of litigation has arrived in the US. In the consolidated class action In re Otter.AI Privacy Litigation, a federal judge ruled on 13 August 2026 that key claims, including wiretapping under the California Invasion of Privacy Act, can proceed. The core allegation is that the bot records other participants without asking them and leaves consent to the account holder. A separate suit against Granola, filed on 30 July 2026, also raises the question of whether speaker identification is biometric processing.

In Europe the rules are already stricter:

  • GDPR. Recording and transcribing a meeting is processing of personal data. You need a lawful basis (often legitimate interest with a documented balancing test, since consent from employees is rarely "freely given"), a clear notice to every participant, including external guests, a data processing agreement, EU or adequately protected hosting and a retention period. See our GDPR guide for generative AI.
  • EU AI Act. Since 2 February 2025, Article 5 bans AI systems that infer emotions of people in the workplace. "Sentiment" or "engagement" scores for meeting participants are exactly the kind of feature to switch off. More in our EU AI Act compliance guide.
  • Germany and Austria. In Germany, recording the non-public spoken word without consent is a criminal offence under section 201 StGB, and the works council has a co-determination right under section 87(1) no. 6 BetrVG over any technical system able to monitor employee behaviour or performance. Austrian works councils have comparable rights under the ArbVG. Plan for a works agreement before rollout.
  • Croatia. Unauthorised audio recording of a non-public conversation is an offence under Article 143 of the Criminal Code, so participants must know that the meeting is recorded. Croatian employers also have to inform and consult the works council on measures that affect employee privacy.

Our rollout checklist: pick one approved tool (usually the one in your existing Microsoft 365, Google Workspace or Zoom licence), block unapproved bots at the meeting platform level, announce recording at the start of every call and in the invite, never record HR, legal or health conversations by default, switch off emotion and sentiment analysis, set automatic deletion of recordings (for example after 30 days, keeping only the approved summary), and name a human owner for every summary that creates an action.

What this means for companies in Croatia and DACH

  • Your employees already use AI note-takers. If you have not approved one, they are using free tools on their own accounts. Decide and govern rather than ban.
  • Start with what you already pay for. Copilot, Gemini and Zoom notes cover most needs without a new vendor or data transfer.
  • Value comes from integration. Connect summaries to your CRM and task tracker, and later to agents via MCP, or the time savings stay small.
  • Consent design is not optional. GDPR, the AI Act, section 201 StGB and works council rights make notice, opt-out and retention part of the product, not paperwork.

The bottom line

AI meeting assistants are one of the few AI tools with near-universal adoption and an easy-to-explain benefit. The risks are just as concrete: wrong attributions, sensitive content in the wrong place and recording without consent. Companies that treat the note-taker as a governed system, with one approved tool, clear rules and integration into real workflows, get the three to five hours a week. Those that let every employee pick their own bot get the lawsuits.

Rolling out AI meeting assistants the right way?

AI Workshop helps you choose between Copilot, Gemini, Zoom and independent tools, writes the GDPR, AI Act and works council documentation, and connects meeting notes to your CRM, task tracker and Claude-based agents. Our Claude Certified Architects keep the data in the EU.

Talk to an AI consultant

Boris Agatić

Founder of AI Workshop and Claude Certified Architect. Over 10 years of building software, guiding companies from the first strategy call to production AI agents.

LinkedIn (opens in a new tab)