No training on your data
AI that doesn't train on your company data
Your team conversations do not feed third-party models. And it does not depend on each employee remembering to switch it off.
Updated
What it means exactly
Model providers offer usage modes in which the content you send is not used for training and is not retained beyond what is needed to answer. Acotia operates in those modes and pins the setting at company level, not at each user's.
The problem it solves
In a personal account the default is usually the opposite, and nobody in the company knows who changed it or when. Centralising access is what lets one decision apply to the whole workforce instead of person by person.
And our own systems
We do not train anything on the content of your conversations either. What we need to operate is telemetry: which model, how many tokens, which policy was applied. Acotia does not store chat content: it lives encrypted on each person's device; the gateway processes it in memory and does not log it.
What a security lead checks before approving an AI tool
Three things: whether the provider uses content for training, how long it keeps it and who can change those settings. With personal accounts, the answer to all three depends on each employee. With centralised access, it depends on one company configuration that can be reviewed.
Telemetry yes, content no
To measure usage and cost we keep metadata: model, tokens, detected data categories, action applied, team and cost. The content of messages is not part of that record and Acotia does not log it (the gateway processes it in memory).
Official sources
Each provider's data-use terms, worth reading in their current version: OpenAI, Anthropic, Google (Gemini API) and Mistral.
Frequently asked questions
Do AI providers train on what my team writes?
It depends on the plan and the provider. On personal accounts it is usually a setting each user changes; in business or API tiers, providers offer terms under which submitted content is not used for training. It is worth reading each provider's current terms.
Does Acotia train models on my employees' conversations?
No. All we need to operate is telemetry: which model, how many tokens, which policy was applied. Acotia does not store chat content: it lives encrypted on each person's device; the gateway processes it in memory and does not log it.
How do I show a client that their data is not used for training?
With three things: the company configuration, which is one and not each user's; each provider's terms; and the evidence log, which shows for every request which model received it and what was applied. None of it depends on anyone's memory.