AssemblyAI

Standard terms allow indefinite retention in some modes and permit training and broad improvement unless excluded or opted out.

Plan Scope

AssemblyAI async and streaming speech APIs plus LLM Gateway; training opt-out, BAA, EU, TTL, and upstream-model choices materially change handling.

Prompt Retention

Without a configured TTL or BAA, async transcripts and text prompts can be retained indefinitely; uploaded audio deletion starts after 24 hours and completes within 48 hours. Streaming offers content ZDR only when training is opted out.

Response Retention

Async transcription artifacts follow the configured TTL, which can be as low as one hour; without TTL or BAA they are indefinite. LLM Gateway content follows customer TTL and upstream-provider rules.

Ordinary Logging

Transcript metadata is retained for logging and billing even where content ZDR applies; Usage Data can be used broadly for business purposes.

Model Training

The standard terms permit training and product improvement with Customer Data. Customers can opt out; BAA customers and EU-server use are excluded from AssemblyAI model training. All LLM Gateway upstream providers are configured no-training.

Product Improvement

Unless excluded or opted out, Customer Data and de-identified derivatives can be used for development, testing, benchmarking, marketing, and model improvement.

Human Or Operator Access

Retained production and training data can be accessed under security controls for support, model development, safety, and legal purposes; training and production environments are separate.

Subprocessors And Routing

LLM Gateway routes to Anthropic, Google, OpenAI, and Bedrock paths with provider-specific retention; OpenAI routes retain abuse logs for 30 days, while named Anthropic and Google routes can provide ZDR.

Deletion Controls

Customers can set an async TTL down to one hour, delete transcripts by API, opt out of model training, select EU processing, and use eligible Streaming or LLM Gateway ZDR configurations.

Caveat

Free/default use should not be assumed private: training is contractually permitted and async transcripts can persist indefinitely until the user selects an opt-out and retention control.