An Optimove AI Offer Decisioning Agent campaign is a specialized type of recurring A/B/n campaign (comprising two or more competing actions and a control group) that learns and improves over time by analyzing its results. Unlike standard A/B/n campaigns, an AI Offer Decisioning Agent campaign automatically adjusts and optimizes its performance based on the response patterns of customers receiving it, as compared to those of the control group. The result is consistently greater marketing performance with minimal effort on the part of the marketer.
An AI Offer Decisioning Agent campaign aims to maximize its overall success in two ways:
-
Changing the mix of actions to favor better-performing actions over others.
If, for example, action A consistently performs better than action B, the AI Offer Decisioning Agent will gradually increase the number of customers receiving action A, while reducing the number of customers receiving action B. -
Personalizing the mix of actions delivered for each granular sub-group of customers within the campaign’s overall target group.
If, for example, it is observed that high-spending European men respond better to action A and budget-shopping American women respond better to action B, then the mix of actions sent to each of these subgroups will automatically and independently adapt for maximum campaign response.
Some additional notes about self-learning campaigns before diving deeper into how they work:
- They must be recurring campaigns (so that learning and adjustments can be made).
- They currently support 2-4 competing actions (besides the control group).
- Self-optimization will only begin after enough campaign response data has been accumulated (i.e., when enough customers respond). Until that time, the campaign will continue running as a standard campaign. There are some thresholds required for this optimization to take place: For the Test Group (per action), at least 50 customers must be targeted & at least 15 responders are required.
For the Control group, at least 30 customers were targeted & at least five responders - The more customers respond to campaigns, the faster and better the AI Offer Decisioning Agent adjustments will occur.
- Greater differences in response rates to the different actions will result in faster and better AI Offer Decisioning Agent adjustments. In other words, if customer response patterns to the various actions are very similar, the standard random A/B mechanism will continue for the duration of the campaign recurrence schedule.
- The optimization process is performed against the single KPI selected when setting up a campaign.
How it Works
Optimove AI Offer Decisioning Agent campaigns are based on a mechanism that involves three main steps:
- Calculating uplift per action (by comparing each action’s test group with the control group)
- Using uplift results to adjust the action mix
- Identifying uplift results for granular subgroups and adjusting action mixes per subgroup
Over time, each marketing action sent to customers will generate some measurable uplift. For example, sending action A may generate an average long-term uplift of $7 per customer, while sending action B generates an average long-term uplift of $8 per customer. However, before sending the campaign many times to many customers over a long period, you don’t yet know how it will perform, and it’s dangerous to reach conclusions based on too little data. At the same time, you want to maximize the uplift of every campaign, even in the short term.
(This type of challenge is sometimes referred to as the “multi-armed bandit” problem in probability theory and machine learning jargon. The name comes from the analogy of a slot-machine player facing a set of identical-looking slot machines – he wants to maximize his chances of winning by playing the right machines in the correct order. He starts by observing jackpot patterns while playing the various machines, but quickly faces a trade-off between dropping his coins in certain machines that seem to be paying out better than others versus dropping coins in other machines to discover if, perhaps, others are even better.
As more data accumulates, the uncertainty embedded within the response data of the set of actions decreases. The AI Offer Decisioning Agent uses these cumulative results to estimate the “real” long-term average uplift of each action, which in turn allows them to be translated into probabilities that reflect their relative impact on customer behavior. Returning to our example, Optimove converts the action A uplift of $7 and the action B uplift of $8 into some number between 0 and 1, which will serve as the new action mix for future campaign occurrences.
For each action, Optimove calculates both average KPI values and response rates, from which an average uplift per customer is derived.
To achieve robust KPI values, Optimove AI Offer Decisioning Agent uses a trimmed mean, instead of a regular mean, for KPI values. This prevents individual customers with unusual behaviors from affecting the understanding of the core behavior of the action’s effect across all customers. For example, an extreme outlier may have dramatic impact on a subgroup’s average uplift. Yet, as an outlier, his KPI figures may represent “noise” rather than the core long-term behavior of his subgroup. Using a trimmed mean also has the added benefit of increasing robustness and tempering the effect of short-term fluctuations in response data.
The second relevant metric, response rate, is robust by nature, as it contains no outliers (a customer may either respond, 1, or not, 0). Optimove combines these two metrics to calculate the average uplift per action per customer (i.e., the average increase as compared with the control group), which in turn serves as the basis for deriving probabilities per campaign, as described in the next section.
From Uplift to Probability
Returning to our earlier example, we have action A, which generates an average uplift of $7 per customer, and action B, which generates an average uplift of $8 per customer. A first thought might be to transform the uplift amounts into probabilities via a formula like this:
for campaign A and for campaign B, where UA stands for the uplift of campaign A, and UB for the uplift of campaign B.
for campaign A and 
for campaign B, where UA stands for the uplift of campaign A, and UB for the uplift of campaign B.So, taking our example, this would give us:
Probability of action A == 47%
= 47%Probability of action B == 53%
= 53%Using this formula, action A would be sent to 47% of the Target Group and action B would be sent to 53% of the Target Group, instead of the initial 50%-50% split. (In actual practice, as stated, these numbers would be changed to account for the control group as well.) This demonstrates how the advantage action B has over action A in terms of uplift gets translated into some difference between their respective probabilities, such that campaign B will be targeted to a larger proportion of the Target Group than action A.
While the above formula conveys the basic idea of how one might transform uplift measurements into probabilities, the actual formula Optimove uses in its AI Offer Decisioning Agent is a variation of the following function, sometimes referred to as a “softmax” function. Roughly speaking, it replaces the actual uplift figures with their exponents:
for campaign A and campaign B, where e is the exponential constant (about 2.71), and UA and UB are each campaign’s uplift.
for campaign A and
campaign B, where e is the exponential constant (about 2.71), and UA and UB are each campaign’s uplift.This type of formula is often used in machine learning and statistics to model probabilities. Its main advantage over the first formula is its ability to handle negative uplifts if the control group’s performance exceeds that of the test group.
Personalizing Action Mixes among Subgroups
The process described above is conducted on the Target Group as a whole, and simultaneously on any number of subgroups of customers identified within the Target Group. Subgroups consist of customers that share demographic and/or behavioral characteristics as indicated by their lifecycle stages, segmentation layers, and micro-segment clusters.
The primary metric used to select a subgroup for independent optimization is the number of respondents. Only once the number of customers responding to a campaign’s actions passes a certain threshold will Optimove begin analyzing opportunities for improving the action mix delivered to that sub-group. All subgroups for which insufficient data have been accumulated are grouped and treated within the top-level Target Group.
Once enough data has been collected for a particular sub-group to allow the derivation of confident conclusions, Optimove will automatically adjust the action mix for that specific sub-group. This cyclical process effectively cascades the optimal action mixes down from the Target Group as a whole, through the various demographic and behavioral subgroups identified within it.
Note: In cases of highly heterogeneous Target Groups, where uplift scales vary dramatically between its subgroups (e.g., when the average uplift of Active customers is measured in dollars and the average uplift of Churn customers is measured in cents), the AI Offer Decisioning Agent will automatically adjust the action mix in each subgroup given its uplift scale.
Take, for example, a campaign in which VIP customers had an average uplift of $5.00 and $2.00 for actions A and B, respectively, while Active customers had an average uplift of $0.50 and $0.20, respectively. In this case, the scales differ greatly in absolute terms, yet the relative differences are effectively the same in both cases.
Understanding the Benefits of the AI Offer Decisioning Agent
Optimove’s AI Offer Decisioning Agent takes you a big step forward toward the goal of one-to-one marketing – reaching every customer with the right message. For each campaign you run, somewhere out there is an ideal customer-promotion-channel mix that will get you an optimal response rate – but there are simply too many possibilities to test out and only so much you can do. Tweaking the action mix for every sub-group iteratively is simply impractical.
This is where the AI Offer Decisioning Agent comes in. By continuously adjusting the action mix for every sub-group according to customers’ preferences it helps you automatically deliver your messaging more accurately, thus achieving a better response from your customers – and optimal results for every campaign.
Putting the AI Offer Decisioning Agent to Work for You
Now that you understand what the Optimove AI Offer Decisioning Agent is and how it can significantly improve your marketing performance, it’s time to put it to work for you! Continue to the next article in this series, Using the AI Offer Decisioning Agent.