How Campaign Agent uses AI

Campaign Agent uses multiple AI models and agents to generate campaigns, content, and personalization. Transparency and human oversight are built-in requirements — no campaign goes live without your review.

AI models and technologies

Multi-agent architecture

Campaign Agent uses multiple AI agents that collaborate to deliver campaign creation capabilities. These agents work together with other Bloomreach features such as:

• Recommendations
• Predictions
• Contextual Personalization
• Loomi Analytics

Base models

The system primarily relies on Gemini models from Google DeepMind, with certain components using OpenAI models. Both providers apply model-level safety and bias-mitigation mechanisms, which serve as the foundation for responsible AI use within Campaign Agent.

Self-optimization functionality

Campaign Agent includes self-optimization functionality that uses AI and data-driven algorithms to automatically adjust the duration of wait nodes within a campaign for each customer. This dynamic optimization analyzes campaign performance to maximize engagement or conversions without requiring manual input, and only uses data collected within a single project.

Data usage and privacy

Default PII handling

By default, Campaign Agent doesn't access or process personally identifiable information (PII) in connection with customer profiles. It can process PII in anonymized and aggregated form, but never in direct connection with a specific customer profile.

For example, if Loomi Analytics is prompted to create a gender segmentation, it accesses the list of values for customer attributes and returns the aggregated number of customers matching the segmentation. This data is anonymized, meaning individual identities are never exposed.

User-submitted data

If you include personal data in a prompt, the large language models (LLM) will process that information as part of fulfilling the request. It's your responsibility to ensure that any personal information you share is managed properly and in accordance with relevant laws and regulations.

Data isolation

User-provided data isn't used to train or improve the model for other clients, and no cross-client data sharing occurs.

Permissions for personal data

Campaign Agent follows the same data access and privacy model as existing Bloomreach email campaign functionality. Only users with Personal Data Viewer permissions can see personal data.

Potential sources of bias

Bias may arise from:

• Model training data: Can reflect societal or cultural biases.
• Customer datasets: May include attributes such as skin tone, gender, or audience segmentation.
• User prompts: Phrasing or context can influence the tone or fairness of generated content.

Outputs may occasionally contain unintentional bias or unbalanced language.

Mitigation measures

Built-in safety features

Campaign Agent relies on Gemini and OpenAI's built-in safety and fairness mechanisms, which filter harmful content during generation.

Gemini safety filters

Campaign Agent's primary AI provider, Gemini, includes built-in safety filters across four categories:

• Harassment: Filters negative or harmful comments targeting identity or protected attributes.
• Hate speech: Blocks content that's rude, disrespectful, or profane.
• Sexually explicit content: Restricts references to sexual acts or lewd content.
• Dangerous content: Prevents promotion or facilitation of harmful acts.

Each category assigns content a probability rating (high, medium, low, or negligible). By default, Gemini blocks content with medium or higher probability of being unsafe.

Learn more about Gemini safety settings.

Mandatory human review

Campaign Agent requires human review of all generated content before it goes live. No campaign is sent or published without your approval.

User education

Campaign Agent makes information about potential bias risks and the importance of reviewing outputs available to users.

Your responsibilities

You're responsible for ensuring that generated content is fair, accurate, and compliant with applicable laws and ethical standards. Before publishing or sending:

• Review outputs for potential bias or inappropriate content.
• Avoid prompts or inputs that could reinforce stereotypes or discriminatory language.
• Ensure any personal data included in prompts is handled appropriately.

Disclaimer

While Bloomreach and its model providers take measures to promote fairness, safety, and data protection, no AI system can guarantee complete neutrality or accuracy. Always apply human oversight before publishing or using AI-generated content.