Private AI Drafting: Map the Data Flow Before Making the Claim
Stefan Bogdanescu
Founder
Self-hosted is a starting point
Keeping the campaign database on your server is useful, but it does not prove that drafting stays there. Inspect the inference endpoint, selected model, cloud fallback, tool calls, application logs and backups. Each can create another copy of a prompt.
Choose a local-only path deliberately
Ollama documents local inference and optional cloud features. Its FAQ provides a local-only configuration. Test the deployed settings rather than inferring privacy from the model provider's name.
Send less data
A campaign draft often needs a segment description, offer and approved voice examples, not a subscriber's full record. Use synthetic examples during testing. Review prompt logs and access controls as well as the prompt itself.
Describe the boundary accurately
Prefer a narrow statement such as "this drafting task uses our configured local endpoint" after verifying it. Avoid "data never leaves the box" unless the entire path supports that statement. Self-hosting alone is not a compliance certificate or proof that no third party processes personal information.
Sources and scope
Configuration and test results determine the behavior of an individual installation. This guide does not certify a live deployment or promise a cost, deliverability or compliance result.