Chatbot data privacy: the questions to ask before you install

Updated August 6, 2026

A chat widget sees names, emails, order numbers, and the occasional pasted password. What to ask any vendor about storage, AI training, retention, and GDPR before the script tag goes in.

A chat widget is a data pipe you invite onto every page of your site. Visitors type names, emails, order numbers, medical details, and — more often than anyone admits — full passwords into chat boxes. Before evaluating features, it's worth being precise about what happens to that text: where it's stored, who processes it, whether an AI model trains on it, and how long it lives. Most vendors answer these questions somewhere; the good ones answer them on a page you can find.

The AI-era question that didn't exist five years ago: does the model train on my customers' conversations? When AI answers are involved, conversation text flows to a model provider (OpenAI, Anthropic, or the vendor's own models). The distinctions that matter: API-based processing with no training retention (the mainstream posture — providers commit contractually not to train on API traffic), vendor-side fine-tuning on your transcripts (ask explicitly), and anonymized aggregate training (read the definition of 'anonymized' skeptically). If the vendor can't say which of these applies, that's your answer.

GDPR mechanics you'll actually be asked about: a Data Processing Agreement you can countersign; a subprocessor list (the model provider belongs on it); data residency (where transcripts physically live); export and deletion — can you pull a workspace's data out, and can you honor a visitor's erasure request without filing a support ticket? Retention defaults matter more than retention options: a tool that keeps transcripts forever unless configured otherwise will keep yours forever, because nobody configures otherwise.

Two implementation details separate careful products from careless ones. First, secret hygiene: visitors paste credentials into chat; a well-built product masks or refuses to store obvious secrets, and its AI never echoes them back. Second, key custody: if you bring your own OpenAI key, where does it live, who can read it, and is it encrypted at rest? A vendor that displays your API key back to you in plaintext settings is telling you how the rest of the plumbing looks.

Your own obligations don't disappear because the vendor's do. Disclose chat in your privacy policy, including the AI processing. Don't feed chat transcripts into your marketing stack without a lawful basis. Set retention to what you actually need — support history has diminishing value after months, and stored text is pure liability after that. And train your team that transcripts are customer data, not internal chatter: the transcript email feature and the CSV export both leave your compliance boundary the moment someone clicks them.

For calibration, Dchat's answers to its own checklist: transcripts live in your workspace's database (self-hosted deployments keep them entirely on your infrastructure); AI processing goes through provider APIs without training retention, and bring-your-own-key mode means the data relationship with the model provider is contractually yours; and workspace data export and deletion are built-in. Whatever tool you pick, the vendor's willingness to answer this page of questions in writing is itself the strongest signal you'll get.

Ask which of three AI postures applies: API no-training, vendor fine-tuning, or 'anonymized' aggregates.

Retention defaults beat retention options — nobody reconfigures the default.

A vendor's willingness to answer in writing is the strongest signal available.