How Marketing agent (Affinity) uses AI

This guide is a compliance reference on Marketing agent's AI architecture, data handling, and safeguards.

Overview

Marketing agent uses AI to build and optimize email campaigns. You describe a goal in natural language, and Marketing agent builds the campaign — audience, content, and send logic — then keeps optimizing it after launch based on real performance when you prompt it. Every campaign stays a draft until you review and approve.

EU AI Act risk classification

Limited risk. Marketing agent generates campaign content, audience selection, and send logic for human review.

Bloomreach's role

Bloomreach acts as both provider and deployer of this AI feature.

AI system and models

Architecture

Marketing agent is built as one agent with capabilities that cover the full campaign lifecycle. It uses base GenAI/LLM models for reasoning and authoring, and standard Bloomreach AI features and ML models for scoring, ranking, and personalization inside the campaigns it builds.

Base GenAI/LLM models

The system uses Gemini or Anthropic models hosted on Google Cloud's Vertex AI, which enforces its own built-in safety, moderation, and bias-mitigation guardrails.

Bloomreach doesn't train its own foundation models — it integrates provider-hosted models into a Bloomreach-managed workflow.

Model inference runs in the provider's infrastructure, on the provider's own training data. Core project data and campaign orchestration stay in Bloomreach-controlled infrastructure. The system transmits only the data necessary to fulfill a specific user intent — preventing raw or batch data oversharing. Every payload sent to an external provider for inference is contractually barred from being retained, tracked, or used for model training.

Standard Bloomreach AI features and ML models

Alongside the agent, campaigns use Bloomreach AI/ML features already available in the platform — including Recommendations, Predictions, Contextual Personalization, and Loomi Analytics. Marketing agent wires these into campaign flows when they improve the outcome, without asking the marketer to configure each model by hand.

These features run in Bloomreach-controlled infrastructure. They score, rank, and personalize using only that project’s own data — nothing learned carries over between customers.

Data handling and security

Data sources

Marketing agent reads your project's own data at runtime — your full data schema, events, segments, predictions, and recommendation definitions, plus your existing campaigns and assets. You can point it at an existing campaign by name or ID, and ask it to build something similar.

Customer data use

By default, Marketing agent doesn't access PII connected to customer profiles. The model never reads individual customer records or values — only schema and field names.

For example, a gender segmentation request returns an aggregate count, never individual identities. If you include personal data in a prompt, the model processes it as part of that request — you're responsible for that data complying with relevant laws.

Your data is never used to train, fine-tune, or improve any AI model — Bloomreach's or the third-party models Marketing agent calls. Models query your data in real time and retain nothing between sessions or across customers; that's a hard system-level constraint, not a setting.

Since no training happens, there's no separate opt-out beyond not using the feature. Only Personal Data Viewer permission holders can see personal data — the same model as email campaigns.

Data retention

Staging and intermediate data expire within about a day; model-artifact storage has no fixed expiry. A terminated account's data is deleted within 30 days platform-wide. No customer-configurable retention period exists today.

Privacy protections

Marketing agent works with anonymized, aggregated data wherever it touches customer data — access is scoped to schema and field names, not values, so PII is excluded by design rather than masked after the fact.

Underlying PII access stays gated behind the Personal Data Viewer permission across Marketing. Marketing agent inherits Marketing's existing GDPR compliance and Data Processing Agreement.

Risk management and guardrails

Marketing agent requires human review of all generated content — nothing is sent or published without your approval. After launch, it can adjust parameters like wait times based on performance, but only within rules you've already set.

Known limitations

Marketing agent works from the brief, project data, and context you give it — a vague brief produces vague results. It tries a request and tells you when something's out of reach, rather than working from a fixed can't-do list. Mandatory human review before every send is the primary safeguard against unexpected output changes.

Guardrails

During the initial setup, you and Bloomreach document your business rules — prohibited claims, excluded audiences, quiet hours, discount authorization — through a shared questionnaire, and Marketing agent applies them.

Beyond that, you can add more by telling the agent directly; it remembers rules within a session.

Bias and fairness

Bias can come from 3 places: training data behind the language models, which can reflect societal bias; your own customer datasets, which may include sensitive attributes; and how you phrase a prompt, which can skew tone or fairness.

Gemini's built-in safety filters run during generation. Gemini, the primary provider, screens 4 categories — harassment, hate speech, sexually explicit content, dangerous content — each rated by probability, and blocks anything medium or higher by default.

Bias-risk awareness

Reviewing content for bias is part of the mandatory human-review step. Marketing agent also surfaces information about these bias risks to you directly, alongside the importance of reviewing outputs.

No AI system guarantees complete neutrality — apply human oversight before publishing.

Transparency and user responsibility

Bloomreach labels Marketing agent's output as AI-generated inside the product. Before publishing, review outputs for bias, avoid prompts that could reinforce stereotypes, and handle any personal data you include in a prompt appropriately.

The business rules documented with Bloomreach at onboarding (see Guardrails, above) are the closest equivalent to a prohibited-uses list.

Agent access management

Marketing agent inherits your own Marketing permissions — both module access and role. If your project doesn't have the email module enabled, it can't create an email; if you have the module but not the Campaigns Admin role, it can't create a campaign.

Personal data access is further gated behind the Personal Data Viewer permission, and no output reaches customers without your approval.

AI compliance at Bloomreach

Reach the Data Protection Officer at [email protected]. Learn more about Bloomreach's AI development principles.