How to make AI-driven segments work in composable CDPs

AI-driven segments outperformed standard segments by up to 42% on a recent head-to-head test. This result is typical for brands shifting from a rule-based approach to AI-driven segmentation. The lift tends to be even greater if no segmentation was used previously.

Many “packaged” CDP offerings have bundled data science that performs critical predictive AI with relatively minimal configuration. However, if you adopt a composable approach to your CDP, you might question how to make AI-based segments work across myriad channels — given that a “composable” CDP is dependent on the data and attributes that reside in your data warehouse.  

What do CDPs packaged with data science provide?

This topic could be an article by itself, but I’d broadly categorize packaged CDP data science offerings into three categories:

• Behavioral enrichments.
• Custom data science builders.
• Bring your own.

Behavioral enrichments

Several CDPs have innovated with offerings that categorize user behavior around:

• Content affinity.
• Channel affinity.
• Behavioral scoring. 

These categorizations may be useful in isolation for rules-based segmentation or as valuable features for building custom models. 

Examples include:

• Lytics’ behavioral scoring and content affinity, which work nicely with its JavaScript tag. 
• BlueConic has a similar suite of behavioral scores. 
• Simon Data’s Simon Predict capability provides predictive analytics for specific marketing outcomes. 

Custom data science builders

Several packaged CDPs offer data science builders for configuring machine learning models that provide regular scoring through user-defined parameters.

Lytics, Blueshift, BlueConic and others have been early adopters. The giants, Adobe and Salesforce, have predictive capabilities. Even mParticle and Twilio Segment have introduced capabilities within the last 6-12 months after years of promoting data quality. 

These “build your own” solutions are powerful, but they force a lot of semi-technical decisions onto users of platforms that often have non-technical marketing users. The dissonance between the offering and the day-to-day end user results in adoption challenges.

Bring your own

All CDPs can onboard attributes to a given customer. Data science scores can be one of these. Many clients I’ve worked with have made significant investments in data science and seek to better connect the data science outputs to marketing activations. 

It’s been interesting to me that even in 2023, there are still marketing data science exercises that aren’t tied to a clear marketing use case. The CDP can solve for onboarding predictive scores and customer intelligence to marketing channels, but the in-house data science first needs to exist.

That’s the nice thing about packaged CDP. Data science actually exists there. Yet, the argument for going composable is strong. It offers theoretically faster time-to-value, simpler implementation, improved privacy and lower total cost of ownership. So, what is a company to do?

A framework to understand data science in composable

Let’s review three scenarios for where your enterprise is in its current data science maturity:

• Scenario 1: My company has pre-existing models.
• Scenario 2: My company has no pre-existing models or data science resources available.
• Scenario 3: My company has a desire to build custom models.

Scenario 1: My company has pre-existing models

If you’re a very mature or “born digital” organization that has made the requisite investments in data science to power predictive AI in your marketing segmentations, I have good news for you.

Composable architecture is a seamless way to take a “composable” CDP and make all of those data science enrichments connect to your marketing channels. All you need to do is ensure that those scores are updated regularly and that your composable CDP has visibility into the scores. (Read more about other pitfalls here.)

Scenario 2: My company has no pre-existing models or data science resources available

Building a data science practice from scratch is hard and expensive work. Making the case for using data scientists assigned to other organizational problems is another issue. 

For example, we have a CPG client with a sophisticated data science practice for predicting futures pricing and availability of ingredients to manufacture its products. However, those data scientists are not focused on marketing activations. 

I don’t have experience in buying billions of dollars of produce or chemicals. Still, I suspect the nuances of predicting tomato futures pricing are different than predicting if a customer will churn in the next 90 days. Each model would have its own unique features, and the experience of the data scientists would have a big impact on the success of the models.

So, what is a company left to do? Should they hire data engineers, data scientists and data analysts to build databases, engineer features, build models, interpret them and then explain them to drive adoption to a busy marketing team? 

More and more, organizations are looking to “rent” data science. They might set up an AI platform like Predictable or Ocurate with opinionated data science models for specific marketing use cases. These solutions have very fast time-to-value. 

Alternatively, the company may choose to go more custom. Platforms like Faraday promise data enrichment and highly flexible model configurations. But the user still needs the technical acumen to know what to predict and how to configure a model — even if it doesn’t require hand-coded Python.

Scenario 3: My company has a desire to build custom models

Before you go down this route, evaluate the cost. Truly building models that scale out requires involvement from several highly-compensated employees. 

To do it right, you’ll need:

• Data engineers to collect and curate the data.
• Data scientists to feature engineer and model the data.
• Analysts to interpret and make the case to use the data. 

You might find employees with a gift in two of these areas. But people who excel in two of those areas are rare. Usually, people are best in one of those three areas. 

If you’re committed to building marketing data science, think about tools that aid you in getting started. If you’re using Google Cloud Platform, for example, consider their Vertex offering and its “Model Garden.” 

If you only have access to GA data, think about learning more about iBQML which lets you leverage data in BigQuery to predict specific on-site outcomes that are additive to digital marketing efforts. 

If you have a more robust BigQuery buildout, leverage BQML, which can score data outside native GA data. The “starter” concepts in these capabilities can build organizational momentum to make further data science investments.

How do I use data science in composable CDP?

After deploying a CDP, a common question arises: how do we optimize data science when the CDP and connected marketing channel share overlapping capabilities? This might include audiences exported to channels with predictive capabilities such as Facebook, Google Ads, the brand’s ESP, etc. 

The answers I provide are specific to a client’s use cases. Your ad tools generally have data that the CDP and your data warehouse do not. I recommend highly-targeted seed audiences from your data warehouse or CDP while leveraging the best bidding from the ad platforms you’re using for acquisition and remarketing use cases. 

In my experience, well-chosen, AI-powered seed audiences outperform lookalikes from rules-driven audiences. For example, an advertiser recently performed a head-to-head test on Facebook between lookalikes of audiences using AI-based predictions and lookalikes of rules-based engaged customers. The conversion rate of the AI-powered seed audience outperformed the rules-based segment by 25%.

Your ESP may have knowledge about email engagement that your data warehouse lacks. If it does, use the adtech approach above. If you have collected the data your ESP has, use CDP/ data warehouse-driven segmentation and decisioning. This also makes you flexible to use multiple ESPs if you have geographic or brand-specific needs. But again, specific recommendations depend on specific use cases and data.

Key considerations when expanding AI usage in composable CDPs

Let’s say you’re convinced you want to start or expand AI usage in your composable CDP. Here’s a checklist of questions to ask yourself:

Do you have all marketing data available in your cloud data warehouse? 

This might include website data such as GA4, data from engagement with owned channels such as email and all transaction/loyalty history. 

It may include identity solutions or rules-based matching for resolution of the customer across channels. Consent data is critical to all usage of first-party data.

Do you have the necessary skills on your team to leverage AI? 

This includes access to data engineers, data scientists, marketing analysts and marketing operations practitioners.

Do you have a tactical plan to deploy the AI-based audiences? 

There is a strategy component to this. But the specific tactics often get overlooked in use case road mapping. There should be a marketing operations plan that determines the necessity of certain data in audience building and the practical application of that audience in each channel. 

Do you have a measurement plan for AI-based audiences in your CDP? 

The measurement plan should include specific test audiences and a way to measure lift and ROI. Ensure that success criteria are made clear upfront, and stakeholders are aligned on what a successful test means for future rollouts.

Good luck in your rollout of AI in your CDP efforts — composable or not. Likely, there is a path for you to adopt the capability in your workflows in a manner that is cost-effective and additive to your marketing team’s ROI.