Cada termo que você precisa para navegar reuniões com IA, da transcrição em tempo real ao consentimento GDPR.
A specific, owned task that emerges from a meeting and requires follow-up work afterwards. A well-formed action item names a single owner, describes the outcome (not the activity), and includes a due date. AI meeting copilots extract action items by listening for commitment language — "I'll send that over by Friday," "Can you ping the design team?" — and surface them in the post-call recap so nothing is lost between the call and the follow-up email. Action items differ from decisions (which record what was agreed) and from open questions (which capture what's still unresolved). The most common failure mode is action items with no owner, or owners with no due date — both of which guarantee the item will resurface in next week's meeting with no progress.
A communication discipline in which the listener focuses fully on the speaker, signals understanding through verbal and non-verbal cues, and reflects back what they heard before responding. In meetings, active listening looks like: paraphrasing the speaker's point before pushing back, asking clarifying questions instead of jumping to solutions, and resisting the urge to draft your reply while the other person is still talking. For non-native English speakers, active listening also includes asking the speaker to slow down or rephrase when needed — a habit many professionals avoid because they fear losing credibility, but which actually signals confidence and reduces costly misunderstandings. AI meeting copilots support active listening by handling the cognitive load of note-taking, freeing participants to engage with the conversation in real time.
A company-wide meeting at which leadership shares strategic updates, financial performance, hiring news, and major decisions with all employees. All-hands meetings typically run weekly, biweekly, or monthly depending on company size, and reserve a portion for live Q&A from employees. The format is part-broadcast, part-dialogue: leadership speaks to a passive audience for most of the call, then opens the floor. Distributed and async-first teams often pre-record the broadcast portion and reserve the live time entirely for questions, which respects time zones and lets employees watch on their own schedule. AI meeting copilots are particularly useful here: they transcribe the call, generate a recap that can be shared with absent employees, and capture the Q&A in searchable form so anyone can revisit specific answers later.
An operating principle where written, asynchronous communication is the default mode of work, and synchronous meetings are the exception reserved for high-bandwidth conversations (debates, brainstorming, urgent decisions). Async-first teams write more than they talk: they document decisions in shared docs, raise issues in long-form threads, and reserve real-time calls for moments when written words genuinely won't suffice. The benefit is time-zone independence, deeper thinking, and a permanent record of why decisions were made. The risk is over-correcting into excessive documentation overhead. Meeting copilots help bridge the two modes: they convert the meetings you do hold into searchable async artifacts (transcripts, recaps, action items) that the rest of the team can consume on their own schedule rather than having to be in the room.
Any communication where the sender and receiver are not engaged at the same time — emails, Slack threads, Loom videos, GitHub issues, recorded standup updates. Async communication respects the receiver's time and attention by letting them respond when they have the bandwidth to think clearly, rather than interrupting their flow. Done well, it reduces meeting load, improves decision quality (because thoughts are written down), and accommodates global teams. Done poorly, it produces endless threads with no resolution and leaves urgent items buried. Effective async communication requires explicit norms: a Service Level Agreement on response times, a clear escalation path for urgent items, and the discipline to write self-contained messages with full context rather than "can we hop on a call?"
An immutable, timestamped record of every action taken in a system — who accessed what, when, and from where. Audit logs are foundational to security and compliance: they let auditors verify that access controls work, let security teams investigate incidents after the fact, and demonstrate to regulators that the organization can answer "who did this and when." For a meeting copilot, the audit log typically records transcript access (every time a user views or exports a transcript), retention events (when transcripts were deleted), permission changes, and admin actions. SOC 2 and HIPAA both require audit logs with specific retention periods and tamper-evidence. The log itself must be append-only — a system that lets users edit or delete audit entries is, by definition, no longer an audit log.
The practice of shifting between different languages, dialects, registers, or communication styles depending on the audience. In multinational workplaces, code-switching often means moving between casual English with peers, formal English in customer calls, technical English in engineering reviews, and one's native language at home — sometimes all in the same day. The cognitive load is real and often invisible: non-native English speakers report exhaustion after a day of meetings that native speakers don't experience. Awareness of code-switching helps managers run inclusive meetings: slow down for non-native speakers, share agendas in writing ahead of time, and use AI transcription so participants can re-read what was said rather than relying on real-time comprehension under pressure.
The maximum amount of text an LLM can process in a single request, measured in tokens (roughly, words plus punctuation). A model with a 200,000-token context window can consider about 150,000 English words at once — enough to fit an entire meeting transcript, the participants' calendars, and the team's project documentation in a single call. Larger context windows let meeting copilots ground their suggestions in more material: not just "what was said in the last five minutes" but "what was said five meetings ago and what's in the linked spec." The trade-off is latency and cost: bigger contexts take longer to process and consume more tokens per request. Modern copilots increasingly rely on RAG to avoid stuffing irrelevant context into the window when only a small slice actually matters.
The requirement that customer data be stored and processed within a specific geographic jurisdiction. EU customers under GDPR often require data to stay within the EU; Canadian healthcare systems often require data to stay within Canada; some defense customers require data to stay within the US. Data residency is not the same as data sovereignty (which adds requirements about whose laws apply) but the two are often confused. For meeting copilots, data residency typically requires running infrastructure in the customer's region (EU-hosted clusters for EU customers), keeping transcripts and recordings in regional object storage, and ensuring that the model inference itself happens within the region — which can be challenging when the LLM provider only has US-region endpoints.
The policy that defines how long an organization keeps customer data before deleting it. Retention policies balance regulatory requirements (some industries mandate minimum retention of audit logs and communication records) against privacy principles (GDPR's data minimization says keep no more than necessary). For meeting copilots, retention typically covers raw audio recordings (often deleted within days), transcripts (often kept months to years for searchability), generated recaps (kept indefinitely as user-owned artifacts), and embeddings (kept as long as the underlying transcript). A good retention policy specifies the trigger (date created, last accessed, user-requested deletion), the holding period, and what "deletion" actually means — soft delete that can be restored, or hard cryptographic erasure.
A persistent, append-only record of decisions made in meetings, with enough context that someone joining six months later can understand both what was decided and why. A good decision log entry captures: the decision itself in one sentence, the alternatives that were considered, the trade-offs that were discussed, the people involved, and the date. Decision logs are the institutional memory of a team — without one, teams relitigate the same trade-offs every few months as new members join. AI meeting copilots extract decisions from transcripts by detecting phrases like "let's go with," "we're decided," "final call is." The output of every meeting where any non-trivial choice was made should include a decision log entry in the recap.
The process of segmenting an audio recording by speaker — answering the question "who spoke when?" Diarization is what lets a meeting transcript show "Alice: Let's ship it. Bob: Not yet." rather than an undifferentiated stream of words. Good diarization is harder than it looks: speakers interrupt each other, audio quality varies across mic setups, and accents and prosody throw off the speaker-embedding models that drive most diarization systems. State-of-the-art systems combine acoustic embeddings (what does this voice sound like) with speech-recognition signals to assign each utterance to a speaker label. Once labels are stable across the meeting, they can be mapped to real names using the participant list from the calendar invite or a one-time enrollment voice sample.
A team whose members work from different physical locations, time zones, or both. Distributed is broader than "remote" — a team can be remote (all working from home) but still co-located in one city, while a distributed team spans multiple regions even if some members share an office. Distributed teams require deliberate practices: overlapping working-hours windows for sync work, written-first communication, recorded meetings for those who can't attend live, and explicit norms about response times. Meeting copilots are core infrastructure for distributed teams because they convert the synchronous meetings (which by definition exclude some time zones) into searchable async artifacts that everyone can consume on their own schedule.
A numerical representation of text (or audio, or images) in a high-dimensional vector space, such that semantically similar inputs end up close together in the space. The sentence "let's reschedule" and "can we move this?" have very different surface words but produce embeddings that are close neighbors. Embeddings power semantic search in meeting copilots: instead of asking users to remember the exact words they used in a meeting, you embed both the query and every transcript chunk, then return the chunks whose embeddings are closest to the query's. Embeddings are also the substrate of RAG — the retrieval step is typically a vector-similarity search over an embedding index. Modern embeddings (1024 to 3072 dimensions) capture nuance well enough to distinguish related-but-distinct concepts.
The practice of storing data in encrypted form on disk, so that an attacker who gains physical access to the storage medium (or, more commonly, a cloud-storage misconfiguration) cannot read the data without the encryption keys. For a meeting copilot, encryption at rest typically covers raw audio files, transcripts in the database, and generated recaps. Industry-standard practice is AES-256 with keys managed by a dedicated Key Management Service (AWS KMS, Google Cloud KMS) and rotated on a regular schedule. Encryption at rest is distinct from encryption in transit (TLS for data moving over the network) and from application-level encryption (where the application encrypts each record with a user-specific key before the database ever sees it). SOC 2 and most enterprise security questionnaires consider encryption at rest a baseline requirement.
Under the EU General Data Protection Regulation, consent to process personal data must be freely given, specific, informed, and unambiguous, and the user must be able to withdraw it as easily as they gave it. For meeting copilots operating in the EU, GDPR consent typically requires: a clear disclosure that the meeting will be recorded and transcribed, the identity of the data controller, the legal basis for processing, the retention period, the user's rights (access, deletion, portability), and an explicit opt-in action — pre-checked boxes don't count. Implicit consent ("by attending this meeting, you consent...") is on shaky legal ground; best practice is an explicit join-time prompt that participants must actively dismiss to enter the call.
The practice of giving an LLM access to verified source material at inference time, so its outputs cite specific evidence rather than relying on parametric knowledge that may be wrong, stale, or fabricated. In meeting copilots, grounding typically means: when the user asks "what did Alice commit to in the standup?" the system retrieves the relevant transcript chunks (via RAG), feeds them into the prompt, and instructs the model to answer only from the provided context — with citations back to the timestamp. Grounding is the primary defense against hallucination: a grounded model that doesn't find supporting evidence should say "I don't see Alice committing to anything in the available transcripts" rather than inventing a plausible-sounding commitment.
A failure mode where an LLM generates output that is fluent and confident but factually wrong or entirely fabricated. In meeting copilots, hallucinations show up as: invented action items that nobody committed to, misattributed quotes, summaries that contradict the transcript, and made-up participant names. Hallucinations are particularly dangerous in meeting contexts because the output is plausible and the consumer often won't catch the error — a recap that says "Alice agreed to ship by Friday" looks authoritative even if Alice never said it. Defenses against hallucination include grounding (force the model to cite specific transcript chunks), explicit "I don't know" instructions in the system prompt, and post-generation verification where each claim in the output is checked against the source transcript before the recap is delivered.
Softening language used to indicate uncertainty, openness to other views, or politeness — "I think," "perhaps," "it might be the case that," "correct me if I'm wrong, but…" Hedging is essential in collaborative discussion: it lowers the social cost of being wrong and invites correction. Over-hedging, however, undermines authority and signals lack of conviction; an engineer who hedges every technical claim sounds uncertain even when they're right. The skill is calibration: hedge when you're actually unsure, state plainly when you're not. Non-native English speakers often over-hedge as a politeness habit imported from their first language; their contributions get discounted as a result. Meeting copilots that analyze post-call communication patterns can help speakers see their own hedging frequency and decide whether to recalibrate.
A decision-rights framework that assigns four roles to every task or decision: Responsible (does the work), Accountable (owns the outcome and has final say), Consulted (must be asked before the decision), and Informed (must be told after the decision). A common failure pattern is having multiple Accountables, which guarantees deadlock — by definition exactly one person is Accountable. RACI is most useful when introduced lightly: a one-line annotation in a project plan or in the meeting notes for a contentious decision. Overapplied, it becomes bureaucratic theater. Underapplied, teams discover three weeks later that nobody actually owned the migration and it's now blocking the launch.
An architecture where an LLM, before generating its response, first retrieves relevant documents from an external knowledge base and includes them in the prompt. Instead of relying solely on what the model learned during training, RAG lets the model answer using up-to-date, customer-specific, or proprietary information. In meeting copilots, RAG is what makes "what did we decide about the API redesign last quarter?" actually work: the system embeds the query, finds the most relevant transcript chunks via vector search, and passes them to the LLM with an instruction to answer only from the retrieved context. RAG is also the primary defense against hallucination in long-tail factual questions — if the right transcript isn't retrieved, a well-designed RAG system should say so rather than guess.
AI features that operate during a live meeting, with sub-second latency, rather than after the call is over. Examples: a private prompt-coach that suggests follow-up questions to the user as their counterpart speaks, on-the-fly definitions of unfamiliar acronyms, live translation overlays, and instant pull-up of relevant prior-meeting context when a name or topic is mentioned. Real-time assist is technically harder than post-call processing because the model has only milliseconds of context, must run continuously, and must not distract the user. The user-experience design is at least as important as the model — a real-time assist that pings the user every ten seconds is worse than no assist at all. The bar for shipping is high; the value when done well is also high.
A concise summary produced at the end of a meeting that captures the key decisions, action items, and open questions discussed. In an AI meeting copilot, recaps are typically generated automatically from the transcript within seconds of the call ending and sent to participants via email or a shared channel. A good recap is short enough to be read in under a minute, lists action items with owners and due dates, and links back to the relevant timestamps in the transcript for anyone who needs more context. Recaps are the single most-used artifact from a meeting copilot because they convert an hour of talk into a few paragraphs that absent stakeholders can absorb asynchronously — and because they let the participants themselves verify, days later, what was actually decided.
The legal and ethical practice of informing all participants on a call that the meeting is being recorded or transcribed. Recording-disclosure laws vary by jurisdiction: many US states require only one-party consent (the person initiating the recording), while California, Florida, and most of the EU require all-party consent. In practice, responsible meeting copilots disclose recording at join time with a visible banner, an audio chime, or both — even in jurisdictions where one-party consent is sufficient, because the trust cost of a participant feeling surreptitiously recorded is far higher than the friction cost of disclosure. The disclosure should name what's being captured (audio, transcript, AI-generated summaries) and give participants a clear path to object.
A team meeting held at the end of a project, sprint, or incident, where participants reflect on what worked, what didn't, and what to change next time. The classic format is three columns: keep doing, stop doing, start doing — though variants abound (sailboat, 4Ls, mad/sad/glad). Retrospectives only work if participants feel safe being candid, which means the meeting must be structured to surface uncomfortable truths without punishing the messenger. Action items from retrospectives should be specific and few — three concrete commitments are worth more than fifteen vague intentions. AI meeting copilots help retrospectives by capturing the conversation itself (so participants can listen back to themes they noticed in the moment) and by extracting and tracking the resulting commitments through to the next retro.
Under GDPR Article 17, the right of an individual to have their personal data deleted by a data controller upon request, in defined circumstances. For meeting copilots, the right to erasure typically means: when a participant asks for their data to be removed, the vendor must delete the audio recording, the transcript, any embeddings derived from that transcript, any AI-generated artifacts (recaps, action items) that quote the participant, and any backups — and must do so within the legally required timeline (usually 30 days). This is more complex than it sounds because erasure must propagate through caches, search indexes, vector databases, and analytics warehouses, none of which are typically designed for selective deletion of individual records.
A class of NLP techniques that classify text or speech along an emotional dimension — positive, negative, neutral — and increasingly along finer-grained axes like confidence, frustration, or engagement. In meeting copilots, sentiment analysis powers features like "the customer's tone shifted negative at minute 23" or "this stakeholder seemed disengaged for the second half of the call." Sentiment is harder than it looks: sarcasm, deadpan humor, and cultural communication norms (a Dutch participant's bluntness is not anger) all confound naive models. The most useful applications of sentiment analysis in meetings are private and reflective — surfacing patterns to the speaker themselves rather than scoring participants for managers — because the latter creates a chilling effect where people perform emotions for the algorithm.
Explicit verbal cues that tell listeners where you are in your argument and where you're going — "there are three things I want to cover; first…" or "let me give you the bottom line, then the reasoning." Signposting is one of the highest-leverage communication skills in meetings: it costs the speaker a few seconds and saves the listener significant cognitive load. Without signposting, listeners have to infer your structure from context; with it, they can budget attention and ask better questions. For non-native English speakers, signposting is especially valuable — it makes your message robust to imperfect pronunciation or grammar, because listeners know what to listen for. The opposite of signposting is stream-of-consciousness, where the speaker thinks out loud and the listener has to assemble the structure themselves.
A meeting between an employee and their manager's manager (skipping the direct manager). Skip-levels serve two functions: they give senior leaders a less-filtered view of what's happening on the ground, and they give employees a channel to surface issues their direct manager can't or won't escalate. Skip-levels work best when they're routine (not crisis-driven), explicitly not for performance feedback about the direct manager (which would poison the dynamic), and short — 20-30 minutes is usually plenty. Many organizations institutionalize skip-levels as a quarterly or biannual rhythm. AI meeting copilots help skip-levels by producing a low-friction record of recurring themes across many one-off conversations — patterns that are invisible in any single meeting but obvious across twenty.
A widely-recognized auditing standard developed by the AICPA that evaluates a service organization's controls related to security, availability, processing integrity, confidentiality, and privacy. A SOC 2 Type II report covers an extended observation period (typically six to twelve months) and is the de-facto baseline that enterprise buyers ask for before signing meeting-copilot contracts. The report is a several-hundred-page narrative of the vendor's controls and the auditor's tests; in practice buyers care about the executive summary and the absence of significant exceptions. SOC 2 is not a one-time event — it requires ongoing operation of the documented controls and an annual re-audit. For a meeting copilot, the most-scrutinized controls usually involve encryption, access management, employee background checks, vendor risk management, and incident response.
The human-readable names attached to each utterance in a transcript — "Alice," "Bob," "Customer" — as opposed to the raw "Speaker 1, Speaker 2" labels that diarization produces by default. Mapping anonymous speaker clusters to real names is a separate step from diarization itself, and there are several approaches: matching speakers to the calendar invite (works when the participant list is short and known), one-time voice enrollment (a participant speaks a brief sample once, and their voice is recognized in future meetings), or post-hoc manual labeling. Speaker labels make transcripts dramatically more useful — a recap that says "Alice committed to the timeline" is far more actionable than "Speaker 3 committed to the timeline."
A short, daily team meeting — traditionally held standing up to keep it short — where each participant covers what they did yesterday, what they're doing today, and what's blocking them. Standups originated in agile software development but have spread broadly. Done well, they take ten to fifteen minutes and surface blockers before they become emergencies. Done poorly, they balloon into thirty-minute status meetings that ignore the third question entirely. The async-first version replaces the live call with a written update in a shared channel, which respects time zones and tends to produce more thoughtful blockers. AI meeting copilots that listen to the live version can produce a written digest after the fact, giving you both modes for the cost of one.
The discipline of allocating a fixed amount of time to a topic or activity, and stopping when the time is up regardless of completeness. In meetings, time-boxing means agendas with per-topic minutes ("15 min: API redesign, 10 min: hiring plan, 5 min: open Q") and a facilitator who actually enforces the boundaries. Time-boxing forces prioritization and prevents the most-vocal topic from absorbing the whole meeting. It also makes meetings feel respectful of participants' time, which is more than aesthetic — it materially affects engagement. The hard part is the enforcement: when a topic runs over, the facilitator must either close it (with an action to follow up async) or explicitly trade time from another topic, not just let the agenda silently slip.
A written record of everything spoken in a meeting, attributed to individual speakers and timestamped. Transcripts are the foundational artifact that every other meeting-copilot feature is built on: recaps, action-item extraction, search, sentiment analysis, and Q&A over past meetings all run against the transcript rather than the raw audio. Transcript quality depends on speech-recognition accuracy, diarization quality, and the meeting environment (clear audio, single speakers, minimal cross-talk). A transcript with 95% word-accuracy and clean speaker labels is far more useful than one with 99% word-accuracy but mis-attributed speakers — the second creates plausible-but-wrong quotes that consumers can't easily detect. Transcripts are also the highest-PII artifact a copilot produces, which makes their storage, retention, and access controls a primary security concern.