Security and privacy14:39
Recording meetings with an AI bot? Consent, terms of use, data retention
Clarifying when meeting recordings fall under legitimate interest versus consent. Includes opening disclaimers, Labor Code deadlines, retention periods, and cost ranges.

Listen to the article summary
Synthetic voice.We compile real deployment experience here and share it freely; it does not replace legal or accounting counsel for your specific case. A standard scenario in a 40-person company: sales has kept Teams transcription active for three months, HR purchased an AI note-taker for remote recruiting, and customer support records client calls to track decisions. No one signed a single document, no one audited where the files reside, and sales reps paste meeting summaries into the CRM alongside personal notes about clients. The issue is not theoretical: it takes one complaint from an unhired candidate, or a client demanding data access and a copy of the recording. At that point, it is not the board asking "did we have a legal basis?", but the supervisory authority.

Step 1: choose the legal basis separately for clients, candidates, and employees
In practice, a B2B meeting with a client typically relies on legitimate interest (Article 6(1)(f) of the GDPR), namely documenting commercial agreements and service quality. Consent is often problematic here: revoked after two months, it obliges you to delete the recording and transcript even during an active dispute over project scope. The Article 6(1)(f) path usually requires three components: a balancing test drafted on a single page, clear advance notice before recording starts, and honouring objections under Article 21 of the GDPR. When someone says "no", the bot leaves the call and the sales rep takes manual notes. A recruitment interview operates differently: recording the full interview exceeds the data catalogue in Article 22(1) of the Labour Code, which requires candidate consent (Article 6(1)(a)): voluntary, providing a genuine alternative, and without impact on the evaluation. Internal team meetings follow a distinct path detailed below. Which approach fits your company depends on your specific contracts and processes; we outline standard market practices.
Script to read at the start of a sales call (30 seconds): "Before we begin: our AI assistant has joined the call to record audio and prepare summary notes. We delete the recording after 14 days; the notes stay until the topic is closed. The content is not used to train the vendor's models. The data controller is [company name]; full details are in the calendar invite. If anyone prefers not to be recorded, I will turn it off now and take notes manually." Calendar invite footer version: "The AI assistant [tool name] participates in this meeting to record audio and generate a transcript with action items. Legal basis: Art. 6(1)(f) GDPR (documenting business agreements). Retention: 14 days. Objections and queries: rodo@[domain].com. Privacy notice: [link]." For recruitment calls, the footer reads differently: "We only record the interview with your consent given at the start of the call. Declining has zero impact on your evaluation." These are sample texts: a starting point for discussions with your legal counsel, not copy-paste templates. Adapting them to your internal policies will always require adjustments. One rule applies regardless of wording: when a partner hosts the meeting and the bot is yours, the duty to inform remains on your side, because you control the recording.
Step 2: employee track: announcement and two-week waiting period
Transcribing calls involving your employees constitutes a form of workplace monitoring under Art. 22(3) § 4 of the Labor Code, subject to monitoring rules in Art. 22(2) § 6–10. The sequence is defined strictly by law. Day 1: define the purpose, scope, and implementation method in workplace regulations; if your company has none (under 50 staff without a collective bargaining agreement), state it in an employer announcement. Day 2: announce the policy through standard internal channels (company-wide email plus a dated board notice). Day 16: activate the tool, as the law mandates a two-week notice period for the team. Every new hire must receive written notice before their first day, alongside their employment terms. Add permanent visual cues: the bot must appear on the participant list under a name explicitly stating it records, not just "Assistant". Supervisory authorities look skeptically at employee consent as a legal basis: subordination makes genuine voluntary consent difficult to prove.
Early warning signs before issues become costly: the first red flag is transcripts capturing unintended data: medical conditions, divorce proceedings, personal debt, or political views shared during small talk. These are special category data under Art. 9 GDPR, which no LIA can legitimize. Second red flag: summaries containing subjective personal evaluations ("caller seemed irritated", "candidate appeared insecure") pasted into your CRM, where they remain for years and resurface during data subject access requests. Third: managerial metrics such as "talk ratio", "call sentiment", or "sales rep engagement". Enforce two technical rules from day one: route the transcript exclusively to the meeting owner rather than a shared channel, and export only operational business outcomes to the CRM: scope, price, deadline, next steps.
Step 3: retention calendar and testing whether you need an impact assessment
A proposal to consider when drafting the policy: audio/video file: 14 days (common range is 7–30 days), automatic deletion configured in the tool, not manual. Transcript: until the deal closes, and with a signed contract up to 3 years from the date claims become due (Article 118 of the Civil Code for claims related to business activity). CRM summary: no quotes or employee evaluations, stored as long as the commercial relationship documentation. Recruitment: recording deleted upon completion of the hiring process, at the latest 30 days after the decision; longer retention only with separate consent for future recruitments, typically up to 12 months. Internal calls: 30 days, unless the material enters an employee dispute. Test for a data protection impact assessment (DPIA): the more "yes" answers, the more critical it is to perform before launch: (1) does the tool operate continuously on employee calls; (2) do health data or other Article 9 GDPR data appear in conversations; (3) does the scale exceed several dozen people or 200 meetings per month; (4) does the tool score or compare employees; (5) do data leave the EEA; (6) does it apply to individuals with no real choice, including candidates. The current list of operations requiring an impact assessment is published by the Personal Data Protection Office: check there for the current status.

Step 4: vendors, AI Act threshold, and local variant
Four tools actually seen in Polish companies with 10–250 employees. Microsoft Teams with Copilot: recording and transcript stay in your tenant's OneDrive/SharePoint, region lockable within the EU Data Boundary, content is not used to train models, data processing agreement covered under Microsoft Products and Services DPA; cost of Microsoft 365 Copilot is approx. 30 USD per user monthly, Teams Premium approx. 10 USD. Zoom with AI Companion: cloud recordings in Zoom with region selection, option to train on customer content is disabled by default and verifiable in the admin panel (Account Settings → AI Companion), paid plan approx. 13–19 EUR per user monthly. Google Meet with Gemini: recording on the organization's Drive, Workspace data are not used to train models, license from approx. 14 EUR per user monthly. Independent AI note-takers (Fireflies, tl;dv, Otter): requires downloading and signing a separate DPA, a "do not use my data to train models" switch in the Privacy section, and server location checks; a US-only vendor triggers standard contractual clauses and a transfer impact assessment; price 18–29 USD per user monthly. Check vendor pricing directly, as tiers change frequently. Boundary with the AI Act: talk time tracking, keyword detection, and meeting summaries are permitted conversation analytics. Evaluating employee or candidate sentiment, enthusiasm, or engagement constitutes workplace emotion inference: a prohibited practice starting 2 February 2025 (Article 5 of Regulation 2024/1689), subject to penalties up to 35M EUR or 7% of annual turnover. Starting the same date, Article 4 applies to AI literacy: in practice, one 60–90 minute session, documentation, and an attendance list are sufficient.
Local setup, when you do not want recordings leaving the company: a workstation with a single 16–24 GB VRAM GPU, faster-whisper large-v3 model for transcription, pyannote for speaker diarization, and an 8–14B model for summaries. One-off hardware cost: 9,000–20,000 PLN net; implementation and integration with the recordings folder plus automatic deletion after 14 days: 5–10 business days. Performance: 2–5 minutes of processing per hour of recording, plus 2–4 minutes for diarization and 1–3 minutes for the summary note. Polish transcription quality: 6–12% word error rate with a good microphone and a single speaker at a time, 15–25% with a laptop microphone and overlapping voices. This means the note requires proofreading and does not serve as verbatim evidence. The break-even point is calculable: 30 users at 100 PLN per month equals 3,000 PLN/month, so the hardware pays for itself in 4–6 months, but you must factor in about 0.5 days of administrator time per month.
How to verify it works. After 14 days from rollout, run three tests: ask the administrator to show that a recording from 15 days ago has actually disappeared (not scheduled for deletion, but non-existent); pick five summaries at random from the CRM and check how many contain opinions about individuals (the target is zero); ask a sales representative to recite the opening disclosure script from memory. After 30 days, check the logs: how many meetings were recorded, how many times someone refused, and whether that refusal was respected. After 90 days, look at a hard business metric: time elapsed from meeting to offer delivery, and the number of agreements lost between meetings. If it has not decreased by at least one business day, the tool does not justify its risk. It pays to know the scale of this risk: GDPR fines reach up to 20 million EUR or 4% of annual turnover, while a realistic scenario for a company of this size means a candidate complaint, a formal 14-day request for explanation, 3,000–8,000 PLN in legal fees, proceedings dragging on for 6–18 months, an order to delete the entire recording archive, plus civil claims under Article 82 of GDPR.
Documents in order with owners: management board decision on purpose and scope (Day 1), legitimate interests balancing test under point (f) by legal counsel or DPO (Days 2–4), updated privacy notices for clients, candidates, and employees (HR and legal), staff announcement or amendment to work regulations plus notification and a two-week waiting period (HR), data processing agreement (DPA) with the vendor and verification of transfers outside the EEA (IT lead with legal counsel), entry in the Article 30 GDPR record of processing activities, data protection impact assessment if the test requires it, recording clause in the framework agreement or NDA with the client, Article 4 AI Act training with attendance log, and automated retention setup. And the point you cannot offload to the system: a human decides whether the bot joins a specific call, a human turns it off at the counterparty's first request, and a human approves the note before sending it to the client. The role of data controller stays with your company, not the tool vendor: confirm this explicitly in the DPA. HEXART handles such implementations end to end: tool configuration, retention rules, local transcription setups on your hardware, and the full set of policy and clause drafts, typically within 2–3 weeks, working alongside your legal counsel. Book a meeting, email biuro@hexart.pl, or call +48 662 016 430.
Source materials
Sources consulted during research. The text above is our own; these third parties are not responsible for its content and have not authorised it.
