Is "the AI uses my conversation as context" just a polite way of saying my data goes to another company?
Largely, yes, and the technical framing should not obscure it. When a product generates a response using your conversation history, that history has to be included in the request sent to whichever provider runs the model — so the provider's infrastructure receives it, at least for the duration of processing.
Independent research comparing major AI platforms found meaningful, documented differences in what happens next: whether conversation content becomes training data, what categories of information get shared onward with other third parties, and how transparent each platform actually is about either.
Does that make every AI product equally risky?
No, and treating them as interchangeable hides a real difference. The same comparative research found platforms differ in what they share and with whom — some pass contact details, location or app interaction data to third parties by default, others are more restrictive.
That is why "the AI provider has my data" is too coarse a statement to act on. The useful question is what the specific provider's terms say about retention, training use and onward sharing — not whether data transits their servers at all, which for current architectures it necessarily does.
What is worth checking about any product that uses an LLM provider?
- > Does the provider train on this data by default, and can it be turned off? This is one of the most consequential differences between vendors, and it is usually specified on an API terms page rather than in the consumer-facing product terms.
- > How long does the provider retain the raw request, separate from whatever the product itself stores? Process-and-discard is a different risk profile from indefinite retention.
- > Is there an actual data processing agreement governing the relationship, rather than two general privacy policies pointed at each other? A specific contract is a stronger commitment than a statement of practice.
Who receives your data in Closer, specifically?
Naming it is the mitigation, so here is the shipped configuration:
- > DeepInfra receives the audio and the text derived from it. Transcription, meeting summaries, knowledge-graph extraction, prep briefs, Ask AI answers, pitch roleplay and text-to-speech all run through it.
- > OpenAI receives text for two narrower jobs: generating embedding vectors, and the memory engine's fact extraction — the step that turns a conversation into short stored memories.
- > Anthropic is declared in the provider registry but is not wired to any pipeline in the shipped setup. It receives nothing unless an operator deliberately reconfigures the app to use it.
- > What we are not claiming: we are not going to tell you here that these providers do not train on requests, or that a data processing agreement is in place, because those are contractual questions and this is a blog post. If your compliance process needs that in writing, ask us and we will answer specifically rather than generically.
What does Closer itself store, and how well protected is it?
The parts people usually assume, stated plainly instead:
- > Audio is not stored. Closer keeps a hash of the recording, used as a cache key so the same recording is not transcribed twice. The transcript and summary derived from it are stored.
- > Application-level encryption covers credentials only. Closer encrypts SMTP passwords and Google refresh tokens with AES-256-GCM. Transcripts, summaries, memories and graph facts are not given that layer — they get whatever the database provider offers underneath.
- > Tenant isolation is enforced in application code. Every tenant table carries an account identifier and ownership is checked on every request before any client data is read. There is no Postgres row-level security doing it at the database layer.
- > Transcript-derived text is treated as untrusted input. Prompt-injection defences run on it, and recipients for any drafted email always come from the client record, never from something said in a conversation.
- > Closer holds no security or privacy certification, and nothing on this page should be read as one.
Key takeaways
- > Using a conversation as LLM context means sending it to a third-party provider's servers — that is inherent to the architecture, not a shortcut.
- > Providers differ measurably in retention, training use and onward sharing, so "an AI company has my data" is too coarse to act on.
- > In Closer, DeepInfra receives audio and transcript-derived text; OpenAI receives text for embeddings and memory fact extraction.
- > Closer stores no audio, only a hash used as a cache key — but transcripts and summaries are stored.
- > Application-level encryption covers SMTP passwords and Google refresh tokens, not meeting content; tenant isolation is enforced in application code, not by database row-level security.