AI Meeting Assistant
Glossary

Tasks and follow-ups automatically extracted from meeting discussions by AI analysis. Most meeting assistants identify who is responsible and what needs to be done, saving teams from manually reviewing recordings or notes after the fact.

Software that uses artificial intelligence to record, transcribe, summarize, and analyze meetings. These tools range from cloud-based services that join calls via bots to on-device apps that process audio locally. See our full rankings for a comparison of the leading options.

A recording method where a virtual participant (bot) joins your video call to capture audio and video. The bot appears in the attendee list and is visible to all participants. Tools like Otter.ai and Fireflies.ai use this approach.

Audio transcription and AI analysis performed on remote servers. Meeting audio is uploaded over the internet, processed by the provider's infrastructure, and results are returned to the user. This is the most common approach, but raises privacy considerations since your audio leaves your device.

AI-powered analysis that goes beyond transcription to extract insights like talk ratios, sentiment, coaching opportunities, and deal signals from meetings. This category is common in sales-focused tools and is a core strength of platforms like Fireflies.ai and Avoma.

Adherence to the European Union's General Data Protection Regulation, which governs how personal data (including meeting recordings) is collected, stored, and processed. For meeting assistants, GDPR compliance typically requires explicit consent from participants before recording, clear data retention policies, and the ability for users to request data deletion.

A system that connects insights across multiple meetings, building a searchable network of topics, decisions, and relationships over time. Rather than treating each meeting in isolation, a knowledge graph lets you trace how a project or topic has evolved across weeks of conversations.

See Bot-Based Recording. A software agent that joins video conferences as a participant to record the session. The bot is visible to all attendees and typically requires calendar integration or a meeting link to join.

A recording method where the app captures system audio directly from your device, without joining the call as a separate participant. No bot appears in the meeting. Tools like Hedy use this approach for more discreet, privacy-friendly recording. Also called system-level audio capture.

AI-powered removal of background sounds (typing, dogs, construction) from audio during calls or recordings. Some tools process noise cancellation on-device in real time, while others apply it during post-meeting transcription in the cloud.

Audio transcription and AI analysis performed entirely on the user's computer or phone, without sending data to external servers. This provides stronger privacy since audio never leaves the device. Hedy and MacWhisper are examples of tools that offer on-device transcription.

Converting speech to text as it happens during a live meeting, rather than processing a recording after the meeting ends. Real-time transcription lets participants follow along with a live text feed and is useful for accessibility, note-taking, and quick reference during calls.

The process of identifying and labeling which person said what in a multi-speaker recording. Accuracy varies significantly across products and depends on factors like the number of speakers, audio quality, and whether the tool has voice samples to learn from.

See Native Audio Capture. A method of recording that captures all audio playing through the device's audio system, without adding a bot or virtual participant to the meeting.

How correctly a tool converts spoken words to written text, typically measured as Word Error Rate (WER). Factors affecting accuracy include audio quality, accents, background noise, and specialized vocabulary. See our methodology for how we test transcription quality.

An open-source speech recognition model developed by OpenAI. Several on-device meeting assistants, including Hedy and MacWhisper, run Whisper locally for private transcription. The model is available in several sizes that trade off speed for accuracy.

The standard metric for measuring transcription accuracy. Calculated as the number of errors (insertions, deletions, substitutions) divided by the total number of words. Lower is better; professional human transcription typically achieves 4-5% WER.

A legal requirement that every participant in a conversation must agree before it can be recorded. California, Illinois, Pennsylvania and several other US states apply this standard, as opposed to one-party consent states where only the person recording needs to agree. Under GDPR, one participant's consent never covers the others. Bot-free tools make this harder to satisfy accidentally, because nothing announces the recording — see our privacy guide.

A meeting assistant that takes actions beyond producing notes — drafting follow-up emails, creating tasks in a project tool, updating CRM records, or answering questions across your meeting history without being asked each time. Otter.ai's Meeting Agent and Sales Agent, Read.ai's Ada, and Microsoft's Teams Facilitator are examples.

A contract required under HIPAA between a covered entity and any vendor that handles protected health information on its behalf. Without a signed BAA there is no lawful path to put PHI through a tool, regardless of its encryption or marketing claims. Several vendors gate BAAs to enterprise tiers — see our healthcare guide.

An Illinois statute governing the collection of biometric identifiers including voiceprints. BIPA is unusually consequential because it carries a private right of action with statutory damages, meaning individuals can sue directly. Claims under BIPA survived dismissal in the August 2026 Otter.ai ruling, and any tool performing voiceprint-based speaker identification on Illinois residents carries exposure.

Recording a meeting from the user's own device audio rather than sending a bot into the call. Nothing appears in the participant list and no platform permission is required, which became important after Microsoft and Google restricted third-party notetaker bots in 2026. It is not the same as on-device processing — most bot-free tools still send audio to a server — and it shifts the duty to disclose recording onto the user.

California's wiretapping and eavesdropping statute. Section 631 targets third parties who eavesdrop on a communication. In August 2026 a federal court held that an AI notetaker vendor can be treated as a third-party eavesdropper under section 631 when it independently collects, retains, and uses recordings for its own commercial purposes rather than simply passing them through for the customer.

The physical or jurisdictional location where data is stored and processed. Often the first question European buyers ask, and rarely sufficient on its own: a vendor can host in the EU while routing transcription through a model provider elsewhere, which moves personal data outside the EEA regardless of where the database sits. Check the subprocessor list — see our GDPR guide.

AI that infers a person's emotional state. Article 5(1)(f) of the EU AI Act prohibits inferring employees' emotions from biometric data in workplace and education settings, a prohibition that consent cannot cure. The European Commission's guidelines place emotion inferred from written text outside that scope, so transcript-based sentiment scoring is treated differently from inference drawn from vocal tone — see our analysis.

The European Union's regulation on artificial intelligence. Its Article 5 prohibitions have applied since 2 February 2025, and general applicability including penalties and enforcement arrived on 2 August 2026, with fines for prohibited practices reaching the higher of €35 million or 7% of worldwide annual turnover. A tool that transcribes and summarises meetings is not high-risk under the Act; its main duty is transparency under Article 50.

Content produced by an AI model that is fluent, plausible, and not supported by the source material. In meeting assistants this typically appears as a summary asserting a decision that was discussed but never made, or attributing a statement to the wrong speaker. It is the main reason summaries should be verified against the transcript before being circulated as a record.

An international standard for AI management systems, covering how an organisation governs the development and deployment of AI. It is beginning to appear alongside SOC 2 and ISO 27001 in vendor security documentation, and signals a more mature approach to AI governance than security certification alone.

An open protocol that lets AI models connect to external data sources and tools through a standard interface. Several meeting assistants now expose an MCP server so your meeting history can be queried by whichever model you prefer, rather than only through the vendor's own chat. Otter.ai, Wispr Flow, and Bluedot all offer MCP access.

A contractual undertaking that a vendor will not use customer conversations to train its models. This is distinct from encryption and from data residency, and it is the criterion that most often separates tools suitable for confidential work from those that are not. Several popular tools train by default unless the user finds an opt-out, and at least one gates that opt-out to enterprise plans.

Assistance delivered during a conversation rather than after it — surfacing a question worth asking, a follow-up worth making, or context on something just mentioned. Distinct from a live transcript, which shows what was said without interpreting it. Most of the category is post-meeting only; Hedy is the clearest example of the real-time approach.

The rules governing how long a vendor keeps your recordings, transcripts, and derived data, and whether you can configure that period. Indefinite retention is the default in much of this category and is the second half of the "collects, retains, and uses" formula that courts have begun to scrutinise. Configurable retention is frequently gated to higher pricing tiers.

An auditing standard covering a service organisation's controls for security, availability, processing integrity, confidentiality, and privacy. Type I assesses the design of controls at a point in time; Type II assesses their operating effectiveness over a period, typically six to twelve months, and is the stronger signal. SOC 2 says nothing about whether a vendor trains on your data.

A third party a vendor uses to process your data on its behalf — commonly a cloud host, a transcription engine, or a model API. Subprocessors are where data residency claims most often break down, because an EU-hosted product may still call a model hosted elsewhere. Every serious vendor publishes a current subprocessor list; ask for it in writing.

A legal characterisation of a party that intercepts a communication it is not a participant in. AI notetaker vendors have generally argued they are mere tools, extensions of the customer who deployed them. In August 2026 a federal court declined to accept that framing where the vendor independently collected, retained, and used recordings for its own commercial purposes.