Skip to content

AI Transparency

AI Data Policy

How DialerBee uses AI, what data is processed, how models are scoped, and your controls as a customer.

Quick answer

What is the DialerBee AI data policy? DialerBee uses AI primarily for answering machine detection: when a call connects, the system transcribes a short initial segment of audio, typically the first 3 seconds after pickup, and classifies the transcript as a live human or an answering machine. Raw audio segments used for AMD are not stored long-term; the classification result, confidence score and model version are retained for analytics, and full call recordings are a separate system governed by your own recording policy and retention settings. Customer call data is not used to train shared cross-tenant models by default, and agent override data stays within the tenant boundary unless explicitly configured otherwise. Tenants can disable all AI features with a single configuration change, and agents can override any AI classification at any time.

The Meet Layla live demo on this website works differently: during the call your voice goes to Deepgram or Munsit for speech recognition, the recognised words go to Google Gemini to write Layla's replies, and ElevenLabs turns her replies into speech. The demo does not record calls or keep transcripts; what it keeps, and for how long, is in the Live demo section below.

1. What AI Does in DialerBee

DialerBee uses AI for one primary purpose: Answering Machine Detection (AMD). When a call connects, the system transcribes the first few seconds of audio and classifies the transcript as a live human or an answering machine. This classification determines whether the call is routed to an agent or handled as voicemail.

Predictive dialing pacing also uses statistical models to optimize call-to-agent ratios, but this operates on aggregate campaign metrics — not individual call content.

Separately, the Meet Layla live demo on this website lets you talk to Layla, our AI voice agent. Section 7 explains what that demo processes and keeps.

2. What Audio Data Is Processed

  • AMD processes a short initial segment of call audio (typically the first 3 seconds after pickup)
  • Audio is transcribed in real time and the transcript is classified immediately
  • Raw audio segments used for AMD are not stored long-term
  • Classification results (human/machine, confidence score, model version) are retained for analytics
  • Full call recordings are a separate system and governed by your recording policy and retention settings

3. Model Training & Tenant Isolation

  • AMD feedback (agent corrections) can be tenant-scoped and governed by customer retention and AI settings
  • Customer call data is not used to train shared cross-tenant models by default
  • Agent override data stays within the tenant boundary unless explicitly configured otherwise
  • Base AMD models are trained on licensed, non-customer datasets and internal test data

4. Your Controls

AI Feature Flags
AI features can be controlled by tenant-level feature flags and opt-in settings.
Full Opt-Out
Tenants can disable all AI features with a single configuration change.
Agent Override
Agents can override any AI classification at any time. AI assists; humans decide.
Feedback Controls
You control whether agent corrections are used for tenant-scoped tuning.

5. Third-Party AI Services

AMD transcription may use Google Cloud Speech-to-Text. Only the short initial audio segment is sent for processing. Google's data processing terms apply to this service. See our subprocessor list for details. The Meet Layla live demo uses other services, described in section 7.

6. AI Decision Audit Trail

AI classifications can be logged with confidence score, model version, and human override status where audit logging is configured. This audit trail supports transparency, debugging, and compliance reporting.

7. Live demo (Meet Layla)

The Meet Layla page on this website (dialerbee.com/talk/) lets you talk to Layla, our AI voice agent, from your browser. It is a public demo, separate from the platform features above, and it processes data differently:

  • What is processed: your voice for the length of the call (the page shows the time limit), the words recognised from it, and Layla's replies. Before the call you give us your name and email address.
  • Who processes it: speech recognition by Deepgram (for English) or Munsit (for Arabic); Google Gemini, through Google's Gemini API, writes Layla's replies from the recognised words; ElevenLabs turns her replies into speech. The call's audio travels between your browser and our demo server, which runs on a Vultr server, sometimes through our relay (TURN) server on the same machine. Cloudflare runs the sign-up step and its bot check (Turnstile). The providers' own terms apply to the data they receive.
  • Not recorded: we do not record the call's audio and we do not keep a transcript. The demo server is set up to leave what you and Layla say out of its logs before they are written. An error message in those logs can, rarely, include a short fragment of what was said.
  • What we keep, and for how long:
    • The demo server's technical logs (call start and end, usage figures, errors; an error line can include a network address): up to 15 days.
    • The relay (TURN) server's log, which holds IP addresses: up to 4 days.
    • Abuse-prevention counters on this website, keyed to your IP address: deleted within 2 days.
    • Your sign-up (name, email address, the scenario you chose, the time, your IP address and its country): until you ask us to delete it. We use your email to follow up about DialerBee; see our Privacy Policy.
    • Like every page of this website, the demo page also records visit analytics, covered by the Cookies & Tracking and Data Retention sections of our Privacy Policy.
  • Your choices: the demo is optional, and you can hang up at any time. To have your sign-up deleted, contact privacy@dialerbee.com.

Questions about our AI data practices? Contact security@dialerbee.com

Last updated: November 2026