How your agents use your data

Where models run

Inside Arzana’s AWS account in the United States, through Amazon Bedrock

Training

No model provider trains on your prompts or outputs

Other customers

Never see, use, or benefit from your data

Approval

Every draft waits for a person on your team

When you leave

Your model and learned context are deleted with your data

We pick the model that does each job best from the major model families, including Claude from Anthropic and models from OpenAI. We reach them through Amazon Bedrock, inside our own AWS account, never through anyone’s consumer apps.

1. An email, PO, RFQ, or ERP record arrives in your Arzana environment on AWS.
2. Your agent builds a prompt from it, using the rules and context it has learned for your company.
3. The prompt goes to a model through Amazon Bedrock in our AWS account, in a U.S. region.
4. The answer comes back to your environment, where your agent turns it into a draft.
5. A person on your team approves, edits, or rejects the draft before anything posts or sends.

Prompts built from your data, and the answers that come back, stay within AWS. They are not sent to the companies that make the models and are never used to train a third-party or public model. Amazon is on the subprocessor list in our Trust Center, with what it does and where.

Corrections, edits, and feedback from your team make your agents better at your work. That learning is stored in your environment as rules, examples, and your own model, and it serves only your company. It never trains, tunes, or informs a model or output for any other customer, and it is never pooled or benchmarked against anyone else’s data.

Agents draft and people decide. Nothing writes to your ERP or sends to a customer until someone on your team approves it, and every approval, edit, and rejection is logged with the user and the time.

Language models can make mistakes. That is why every draft waits for review, why agents check each order against its quote and flag mismatches, and why your team’s corrections feed straight back into your agent. If an agent keeps making the same mistake, tell your named engineer and we adjust it.

Your model, rules, and learned context are deleted with the rest of your data, within 30 days after your 90 day export window ends, and we confirm it in writing.