AgentControl and information privacy

This topic describes how the LaunchDarkly AgentControl product treats personally identifiable information (PII).

Privacy and PII in LaunchDarkly

End-user data is information about your customers that your application sends to LaunchDarkly as part of a context. End-user data can include PII, including names, email addresses, or other unique identifiers, depending on how you define your context attributes.

AgentControl works similarly to LaunchDarkly feature flags with respect to PII in end-user data. Specifically:

• We recommend that you use private attributes when you target on PII or other sensitive data. To learn more, read Minimizing LaunchDarkly’s access to end-user data.
• If your organization prohibits sending PII to AI providers, then you must take care not to include PII in any of the context attributes that you reference in the messages in your config variations. LaunchDarkly does not have guardrails that prevent you from sending PII from your application to LaunchDarkly, or from LaunchDarkly to a third party, such as an AI model.

LaunchDarkly and AI models

How AgentControl interacts with an AI provider depends on where the model call occurs:

• Model calls from your application using an AgentControl config, including online evaluations: Your application calls the model provider using your provider credentials. Depending on which SDK functions you use, either your own code or the LaunchDarkly AI SDK running inside your application makes the request. In both cases the call happens within your application, on your own infrastructure, and LaunchDarkly does not act as an intermediary or proxy model requests and responses. LaunchDarkly supplies the configuration and records evaluation results. For online evaluations, LaunchDarkly receives only the numeric score. The judge’s reasoning is produced inside your application and is not transmitted.
• Playgrounds and offline evaluations: LaunchDarkly calls the model provider and sends the prompts, variable values, model parameters, and evaluation criteria required for the run. LaunchDarkly uses either provider credentials configured in your project or LaunchDarkly-provided model access. If you use your own credentials, LaunchDarkly calls the provider under the terms of your agreement with that provider. If you use LaunchDarkly-provided model access, your prompts reach Anthropic under an agreement between LaunchDarkly and Anthropic rather than yours. To learn more, read Model provider access.

LaunchDarkly does not use information you provide to fine-tune or train models.

If you use variables within your config variation messages, LaunchDarkly substitutes their values into the message. If your organization prohibits sending PII to AI providers, do not include PII in variables used by playgrounds or offline evaluations.