Leveraging Propensity Models: Marketing that Thinks Like Your Customer
In today’s digital landscape, personalized marketing is crucial for maximizing customer engagement and conversion rates. Propensity scores help marketers understand consumer behaviours, determine the likelihood of performing certain actions, or even the likelihood of interacting with specific communication channels. By leveraging Adobe Real-Time Customer Data Platform and its Customer AI feature, businesses can enhance targeting and optimize marketing campaigns. This helps in achieving the following outcomes:
- Reduced costs by optimizing the number of campaigns
- Increased engagement rates by delivering messages through customers' preferred channels
- Enhanced customer experience with personalized communication
- Improved conversion rates by aligning marketing efforts with user behaviour
- ROI improvement of 20-30% reported by customers via Customer AI
Traditional marketing strategies often rely on event/intent-based targeting without considering a customer's preferred channel. This can lead to overcommunication across multiple channels, resulting in customer fatigue and lower engagement rates.
As part of the digital journey, users often switch between different browsers, and each browser generates its own Experience Cloud ID (ECID). This creates multiple ECIDs for the same customer, making its interactions stitched across different touchpoints.
Customer AI models operate at the touchpoint level, capturing and analyzing data from various customer touchpoints. These models assess customer interactions across multiple ECIDs to generate propensity scores that predict the likelihood of engagement through specific channels.
It is possible that a single profile can have multiple ECIDs, and it is important to unify the touchpoint of customer-to-single ECID. The Data Distiller offering, available in Adobe Experience Platform, enables users to execute PostgreSQL queries to curate data within the platform’s data lake and prepare it for use in platform-based applications. It helps resolve fragmentation issues caused by multiple Experience Cloud IDs (ECIDs) per user, which leads to incomplete customer profiles, inconsistent propensity scores, and inefficient marketing efforts.
Introduction to Adobe Real-Time CDP’s Customer AI feature & Adobe Experience Platform Data Distiller
Adobe Real-Time CDP integrates data from multiple sources to create a unified customer profile. The Data Distiller offering consolidates customer events into one ECID by running Distiller scheduled SQL jobs. The Customer AI feature enhances this by generating propensity scores, predicting customer behavior, and enabling personalized marketing.
Key Features of Customer AI in Adobe Real-Time CDP:
- AI-powered insights for predicting customer engagement behavior
- Automated and scalable processing of high-volume data
- Dynamic updating of propensity scores based on real-time interactions
Other key technical capabilities of Customer AI include:
- Built on XDM-compliant data structures for seamless integration
- RESTful API access to propensity scores for online decisioning
- Configurable ML model training frequency (weekly/monthly)
Key technical challenges in propensity generation
Key technical challenges are:
- Fragmented customer identities: Customers generate multiple ECIDs across devices and sessions, making it difficult to unify data.
- Complex data transformation: Consolidating event data from multiple sources requires extensive processing and mapping.
Additional technical challenges include:
- Handling data latency in daily/monthly decisioning scenarios
- Managing computational resources for ML model training
- Ensuring data privacy compliance during identity resolution; balancing prediction accuracy with processing time requirements
The solution lies with Adobe Experience Platform Data Distiller, which standardizes and unifies data in the datalake, allowing Customer AI to generate a single propensity score per customer profile, ensuring accurate and efficient targeting.
How does Customer AI in Adobe Real-Time CDP generate marketing preference?
Customer AI leverages machine learning to predict the most effective communication channel for each customer.
Process Flow:
- Data collection: Customer interactions across SMS, Email, and WhatsApp are gathered using Web SDK, Mobile SDK, and Adobe Analytics.
- Data processing: Data Distiller consolidates and structures event data into unified customer profiles in the datalake. The following SQL snippet is used to combine the events into 1 ECID and load into one derived dataset and this dataset will be used as input in the Customer AI model:
from
(
select
identitymap['ecid'] [0].id as PrimeECID,
explode (identitymap['ecid']),
identitymap
from
profile_snapshot_export_e13c0e86_da84_4f2a_abe3_73554c45f0b9
)
) profilesnapshot
inner join customer_digital_platform_dataset on profilesnapshot.ecid = identitymap['ECID'] [0].id
where
timestamp <= current_date and timestamp >= (current_date + interval '-
- Customer AI modelling: Customer AI models are designed to assess the likelihood of customer engagement across various communication channels (SMS, Email, WhatsApp) and schedules based on their interactions.
Channel propensity: The following parameters are set while creating the Customer AI models from an event dataset, with eligible population being records where utmmedium contains "sms”, “email”, “whatsapp”:
- Channel Propensity SMS:
- Goal: utmmedium includes "sms”
- Channel Propensity WhatsApp:
- Goal: utmmedium includes "whatsapp”
- Channel Propensity Email:
- Goal: utmmedium includes " email”
These models generate datasets with propensity scores for each channel:
- customer_ai_scores_channel_propensity_sms
- customer_ai_scores_channel_propensity_email
- customer_ai_scores_channel_propensity_whatsapp
Technical deep dive: The machine learning models employ gradient boosting with XGBoost and handle class imbalance through various sampling techniques. Feature importance is calculated using SHAP values, enabling explainable AI outcomes.
- Score Consolidation: Using Data Distiller offering, we consolidate these scores into a unified dataset. New schema and dataset will store the calculated propensity scores, serving as the core repository for future analysis.
- Marketing Classification: With Data Distiller offering, customers are categorized into:
- Channel Classification:
- Preferred Channel: The channel with the highest propensity score.
- Dominant Channel: A channel with at least 50% higher propensity than others.
- Non-Preferred Channel: A channel with at least 50% lower propensity than others.
- Channel Classification:
- Audience Creation: Based on the marketing propensity data, we create targeted audiences and map them to the respective campaigns.
This classification helps marketers deliver targeted campaigns through the most effective channels.
Data flow from Adobe Real-Time CDP to Adobe Campaign
Seamless integration between Adobe Real-Time CDP and Adobe Campaign ensures propensity scores are leveraged effectively for targeted marketing campaigns. The enriched profiles with propensity scores are sent to Adobe Campaign, which uses these scores to deliver messages via preferred channels, ensuring higher engagement and conversion rates.
Architecture diagram
The high-level architecture of the Channel Propensity Model ensures a seamless flow from data collection to campaign execution, enabling real-time decision-making and personalized marketing.
These channel propensity-based audiences can be further strategically extended across other Adobe platforms and third-party platforms to enhance personalization, reduce marketing waste, and improve return on investment.
- In-Adobe solutions (such as Adobe Journey Optimizer, Adobe Target, Adobe Experience Manager), allow marketers to tailor journeys, personalize content, and optimize campaigns and communications based on a customer’s preferred channels.
- With non-Adobe platforms such as email/SMS platforms (Salesforce, Braze, Resulticks, ZineOne), paid media (Google, Meta, LinkedIn), CRM systems, and call centers, businesses can activate high-propensity segments to focus spending where it’s most effective—avoiding low-performing channels and improving engagement.
This data-driven approach not only enhances the relevance of communication but also helps reduce campaign costs by minimizing inefficiencies, ultimately driving higher conversion rates and greater marketing ROI.
Troubleshooting Common Issues: Low model accuracy (AUC < 0.65)
- Check for sufficient positive examples in training data (min 500 recommended)
- Ensure lookback window captures relevant customer journey events
- Validate that eligible population filtering isn't too restrictive
Future scope & enhancements
Personalized marketing is evolving fast. With AI, brands can tailor not just timing and channels, but also tone, visuals, and content to match individual customer traits and life stages. By orchestrating messages across inbound, outbound, and offline touchpoints—using both behavioural and contextual data—marketers can drive deeper engagement. Adobe Real-Time CDP empowers this next-gen personalization. Reach out for a demo.
For more insights, visit our Business Newsletter
Aradhana Padhiary and Parth Hetamsaria also contributed to this article.
Special thanks to Pawan Sevak, Arava Sai Kumar, Sameeksha Arora, and Alyssa Espiritu for their valuable review and contributions to the tech blog.