The Bloomreach Engagement implementation playbook: Part 5
Part 5: Loomi AI and predictive attributes
Static segments answer the question “who fits this definition right now?” Predictive attributes answer a different question: “what is this customer likely to do next?” That shift, from descriptive to predictive, is what Bloomreach’s Loomi AI engine enables, and it’s the capability that most differentiates a modern CDP implementation from the email platforms of a decade ago.
This part covers what Loomi actually does, how to instrument it, and how to deploy its outputs without creating brittle dependencies on AI behavior.
What Loomi is and isn’t
Loomi is Bloomreach’s umbrella term for the platform’s AI capabilities. It’s a collection of models trained on your data (and broader Bloomreach benchmarks) to predict specific customer behaviors. The main capabilities:
Predictive attributes. Per-customer scores for specific outcomes (churn probability, next-purchase likelihood, lifetime value).
Optimal Send Time (OST). Per-customer best time of day to send a message.
Content recommendations. Per-customer product or content suggestions.
Smart segments. AI-assisted segment definition based on example customers.
Generative content tools. Subject line suggestions, body copy variants, personalization recommendations.
Loomi is a statistical layer that optimizes within the boundaries you set. It will not replace strategy, and it will not rescue a mediocre implementation. If your scenarios, segments, and content are mediocre, Loomi will make them slightly less mediocre but not actually good.
Predictive attributes in practice
A predictive attribute is a continuous score (usually 0–1) indicating the probability of a specific outcome. Once configured, it’s updated continuously as new data flows in, and it’s available in scenarios and segments like any other customer attribute.
The ones most clients deploy first:
Probability to purchase in the next N days. Used to target re-engagement campaigns at lapsing-but-recoverable customers.
Probability to churn (not purchase in next N days). The mirror of the above; used to prioritize retention efforts.
Predicted customer lifetime value. Used to tier marketing investment, higher CAC tolerance for higher-predicted-LTV customers.
A Loomi-powered win-back scenario might check churn_probability > 0.7 AND predicted_ltv > 500 before offering a 20% discount, and the same scenario might offer only a 5% discount (or no discount) to customers with the same churn risk but low predicted LTV. This is the kind of surgical decision-making that static segmentation struggles with.
Data requirements. Predictive attributes need enough data to train meaningful models. Rule of thumb: a predictive attribute becomes reliable around 10,000 customers and 6 months of event history, and substantially more accurate by 100,000 customers. Very small or very new implementations should wait to deploy predictions, the models will exist but they won’t be useful.
Validation. When a predictive attribute goes live, validate it against holdout data. Bloomreach provides a model performance view (precision, recall, ROC) that most teams never open, and reviewing it should be standard practice. If a “probability to purchase” model has 60% precision, you need to know that before you build scenarios that treat it as ground truth.
Optimal Send Time
The email node has a setting to send at a fixed time, at the customer’s local time, or at the Optimal Send Time computed per customer. OST uses the customer’s historical engagement pattern, when they’ve opened past emails, to predict when they’re most likely to be in their inbox for any given message.
OST produces measurable lift, typically 10–20% open rate improvement over fixed-time sends, but it has two preconditions most teams miss:
The customer needs engagement history. OST for a customer with zero past opens falls back to a population average, which is no better than a fixed send. Consider only enabling OST for customers with at least 5 past opens.
The scenario needs to tolerate variable send times. If your scenario’s next step depends on “send the second email 24 hours after the first,” and the first email goes out at OST, you no longer have a fixed 24-hour window, it becomes 24 hours after OST, which could be anywhere in a 48-hour spread. Design for that.
Smart content: recommendations, subject lines, and generation
Loomi’s content capabilities are divided into two types: selection (pick the best from existing options) and generation (create new content).
Selection is the more mature capability and the safer one to deploy. Recommendation blocks in emails pick personalized products per customer from your catalog. Subject line A/B tests with dynamic allocation shift traffic to the winning variant automatically. Content variant selection picks the best-performing hero image per customer segment.
Generation. AI-written subject lines, body copy, and design variants, is newer and requires a human in the loop.
💡Scalero’s recommendation: use generative tools for ideation and first drafts, not final copy. An AI-suggested subject line that a human approves is a productivity multiplier. An AI-generated subject line sent without review is a brand risk.
Ethical AI and customer trust
Two policies belong in every Loomi deployment:
Transparent opt-out. Customers should be able to turn off behavioral personalization if they want to. Most won’t, personalization is usually a feature, but the option matters for trust and, in some jurisdictions, for compliance.
Bounded automation. Don’t let AI-driven decisions affect pricing, eligibility, or access unilaterally. AI-picked product recommendations are fine. AI-driven dynamic pricing decisions that give different customers different prices based on predicted willingness-to-pay is a legal and reputational minefield, not just a technical feature. Decide explicitly where your AI authority stops.
Monitoring AI in production
AI features degrade silently. A model that was 75% accurate at launch can drift to 55% over a year of data changes without any error message. Three monitoring practices:
Track holdout accuracy weekly. Bloomreach surfaces this for predictive attributes; check it monthly at minimum.
Track behavioral lift, not AI outputs. The question to answer is whether customers targeted by Loomi are converting at higher rates than a matched control group. Run a 5% holdout that always receives the non-Loomi version of any AI-driven experience, and track the delta.
Review for bias. Periodically check whether predictions differ systematically by demographic attributes in ways that could create discriminatory outcomes. This is especially important for any model that affects offer eligibility or pricing.
With AI in place, the last question is how to keep all of this running cleanly for the long term. Part 6 covers governance.
