Data handling
Your data, and our AI models
The two questions every AI review asks — “do you train on our data?” and “do your model providers keep it?” — answered plainly. Training: never. Provider retention: zero wherever the provider offers it, on eligible plans — and never used for training.
Last updated: June 18, 2026
No training on your data
Neither Finn nor our model providers train, fine-tune, or improve models on your calls, transcripts, or contacts.
Zero data retention
Model APIs run in zero-data-retention mode wherever the provider offers it — otherwise only a short operational window, never used for training. Zero-retention passthrough is an Enterprise option.
Transcript, not audio
Language models receive transcript text, not raw call audio — with one exception, noted below.
No training
Your calls are not training data
Finn does not use Customer Data — call recordings, transcripts, prompts, or uploaded contacts — to train or improve any model, ours or a third party’s. Your data is processed only to deliver the service you configured, and for nothing else.
Every model provider in our pipeline is engaged under an API tier whose terms prohibit training on inputs and outputs. This is contractual, not best-effort.
Zero data retention
Model providers keep nothing to train on
Our language-model APIs are configured for zero data retention wherever the provider offers it, on eligible plans: the provider receives a request, returns a response, and does not store the prompt or the output. Where a provider does not offer a zero-retention tier, retention is limited to a short operational window (such as abuse monitoring), is never used for training, and is documented in our data-flow map. Zero-data-retention passthrough on third-party providers is available as part of our Enterprise data controls; on every plan, providers never train on your data and any retention is a short operational window.
OpenAI
Accessed via Microsoft Azure — call data is processed by Microsoft, not OpenAI. No training on inputs or outputs; zero-data-retention configuration where offered, on eligible plans.
Anthropic
Accessed via Google Cloud Vertex AI — call data is processed by Google, not Anthropic. No training on inputs or outputs; zero-data-retention configuration where offered, on eligible plans.
Served via Vertex AI. No training on inputs or outputs; zero-data-retention configuration where offered, on eligible plans.
Microsoft
Served via Azure OpenAI and Azure speech services. No training on inputs or outputs; zero-data-retention configuration where offered, on eligible plans.
What the models receive
Transcript in, response out
Language models receive transcript text and prompts — not raw call audio. The one exception is OpenAI’s realtime models (served via Azure), a speech-to-speech path that ingests audio directly; where it is used, that is disclosed in our data-flow map. Speech-to-text and text-to-speech providers process only the audio or text needed for their step and return the result.
The full path — which provider receives which data category — is documented in our data-flow map, and the providers themselves are listed on our subprocessors page.
Retention & deletion
How long data is kept, and when it goes
Storage window
Recordings and transcripts are held in your workspace region for a configurable cache window — 90 days by default — then aged out. You can shorten it.
On termination
When an agreement ends, Customer Data is returned or deleted within 30 days, as set out in our DPA.
Per provider
Model APIs retain nothing where zero-data-retention applies, and otherwise only a short operational window never used for training. Telephony carriers apply their own regulatory retention to call metadata. Only your workspace storage holds to the cache window above.
Export & erasure
You can request export or deletion of Customer Data at any time — email [email protected].
始める
Finnを導入し、安心してスケール。
アイデアから本番の音声自動化へ — 安全に、確実に、運用リスクなしで。
隔週リリース
私たちは素早く動き、必要なものをお届けします
強力なツール
ダッシュボード、レポート、自動化などをあらかじめ用意