security
    Updated

    What data does an AI SDR train on and is my data safe?

    The two data flows

    1. Retrieval (safe by design): your docs are indexed into a vector database that only you can query. When a visitor asks a question, the AI retrieves the relevant chunks and generates a grounded response. No training happens.

    2. Training (opt-in only, ideally never): some vendors offer to fine-tune the model on your transcripts for higher accuracy. This is genuinely useful but requires strict data handling: dedicated model, no cross-customer leakage, deletion on churn.

    What to require in the DPA

    • Explicit training opt-out as the default.
    • Named LLM provider (OpenAI, Anthropic, Google, or open-source) with sub-processor status.
    • Data retention window (30–90 days for transcripts is standard; longer is a red flag).
    • Deletion on request within 30 days.
    • No cross-customer aggregation of prompts, transcripts, or retrieval indices.

    See the security and compliance deep-dive.

    Related questions

    Which LLMs power AI SDRs?

    OpenAI GPT-5 and Anthropic Claude are the two most common in 2026. A minority of vendors use open-source (Llama, Mistral) for cost or data-sovereignty reasons. Ask which one; the answer changes the risk profile.

    Can prompts be injected to leak my data?

    Yes, unless the vendor has invested in prompt-injection defenses. Ask for red-team results from a third-party auditor and require adversarial testing coverage in the SOC 2 report.