Configure a Customer AI instance

Last update: 2024-01-22
  • Created for:
  • User

Customer AI, as part of AI/ML Services enables you to generate custom propensity scores without having to worry about machine learning.

AI/ML Services provide Customer AI as a simple-to-use Adobe Sensei service that can be configured for different use cases. The following sections provide steps for configuring an instance of Customer AI.

Create an instance

In the Platform UI, select Services in the left navigation. The Services browser appears and displays all available services at your disposal. In the container for Customer AI, select Open.

The Customer AI UI appears and displays all your service instances.

  • You can find the Total profiles scored metric located in the bottom-right side of the Create instance container. This metric tracks the total number of profiles scored by Customer AI for the current calendar year including all sandbox environments and any deleted service instances.

Service instances can be edited, cloned, and deleted by using the controls on the right-hand side of the UI. To display these controls, select an instance from your existing Service instances. The controls contain the following:

  • Edit: Selecting Edit allows you to modify an existing service instance. You can edit the name, description, and scoring frequency of the instance.
  • Clone: Selecting Clone copies the currently selected service instance setup. You can then modify the workflow to make minor tweaks and rename it as a new instance.
  • Delete: You can delete a service instance including any historical runs. The corresponding output dataset will be deleted from Platform. However, scores that were synced to Real-Time Customer Profile are not deleted.
  • Data source: A link to the dataset used by this instance. If multiple datasets are being used, selecting the hyperlink text opens the dataset preview popover.
  • Last run details: This is only displayed when a run fails. Information on why the run failed, such as error codes are displayed here.
  • Score definition: A quick overview of the goal you configured for this instance.

To create a new instance, select Create instance.

Set up

The instance creation workflow appears, starting on the Set up step.

Below is important information on values that you must provide the instance with:

  • Name: The instance’s name is used in all places where Customer AI scores are displayed. Hence, names should describe what the prediction scores represent. For example, “Likelihood to cancel magazine subscription”.

  • Description: A description indicating what you are trying to predict.

  • Propensity type: The propensity type determines the intent of the score and metric polarity. You can either choose Churn or Conversion. Please see the note under scoring summary in the discovering insights document for more information on how the propensity type affects your instance.

Setup screen

Provide the required values and then select Next to continue.

Select data

By design, Customer AI uses Adobe Analytics, Adobe Audience Manager, Experience Events in general, and Consumer Experience Event data to calculate propensity scores. When selecting a dataset, only ones that are compatible with Customer AI are listed. To select a dataset, select the (+) symbol next to the dataset name or select the checkbox to add multiple datasets at once. Use the search option to quickly find the datasets you’re interested in.

Select and search for dataset

After selecting the datasets you wish to use, select the Add button to add the datasets to the dataset preview pane.

Select datasets

Selecting the info icon info icon next to the dataset opens the dataset preview popover.

Select and search for dataset

The dataset preview contains data such as the last update time, source schema, and a preview of the first ten columns.

Select Save to save your drafts as you move along the workflow. You can also save draft model configurations and move to the next step in the workflow. Use Save and continue to create and save drafts during model configurations. The feature enables you to create and save drafts of the model configuration and is particularly useful when you have to define many fields in the configuration workflow.

The Create workflow of the Data Science Services Customer AI tab with Save and Save and continue highlighted.

Dataset completeness

There is a dataset completeness percentage value in the dataset preview. This value provides a quick snapshot of how many columns in your dataset are empty/null. If a dataset contains a lot of missing values and these values are captured elsewhere, it is highly recommended you include the dataset containing the missing values. In this example Person ID is empty, however, Person ID is captured in a separate dataset that can be included.


Dataset completeness is calculated using the maximum training window for Customer AI (one year). This means data that is more than one year old is not taken into account when displaying your dataset completeness value.

Dataset completeness

Select an identity

You can now join multiple datasets to one another based on the identity map (field). You must select an identity type (also known as an “identity namespace”) and an identity value within that namespace. If you have assigned more than one field as an identity within your schema under the same namespace, all the assigned identity values appear in the identity dropdown prepended by the namespace such as EMAIL (personalEmail.address) or EMAIL (workEmail.address).

select same namespace


The same identity type (namespace) must be used for every dataset you select. A green checkmark appears next to the identity type within the identity column indicating datasets are compatible. For example, when using the Phone namespace and mobilePhone.number as the identifier, all identifiers for the remaining datasets must contain and use the Phone namespace.

To select an identity, select the underlined value located in the identity column. The select an identity popover appears.

select same namespace

In the event that more than one identity is available within a namespace, make sure to select the correct identity field for your use case. For example, two email identities are available within the email namespace, a work and personal email. Depending on the use case, a personal email is more likely to be filled in and be more useful in individual predictions. This means that EMAIL (personalEmail.address) would be selected as the identity.

Dataset key not selected


If no valid identity type (namespace) exists for a dataset, you must set a primary identity and assign it to an identity namespace using the schema editor. To learn more about namespaces and identities, visit the Identity Service namespaces documentation.

Define goal

The Define goal step appears and it provides an interactive environment for you to visually define a prediction goal. A goal is composed of one or more events, where each event’s occurrence is based on the condition it holds. The objective of a Customer AI instance is to determine the likeliness of achieving its goal within a given time frame.

To create a goal, select Enter Field Name and followed by a field from the dropdown list. Select the second input, a clause for the event’s condition, then optionally provide the target value to complete the event. Additional events can be configured by selecting Add event. Lastly, complete the goal by applying a prediction time frame in number of days, then select Next.

Will occur and will not occur

While defining your goal, you have the option to select Will occur or Will not occur. Selecting Will occur means that the event conditions you define need to be met for a customer’s event data to be included in the insights UI.

For example, if you would like to set up an app to predict whether a customer will make a purchase, you can select Will occur followed by All of and then enter (or a similar field) and exists as the operator.

will occur

However, there may be cases when you are interested in predicting whether some event will not happen in a certain timeframe. To configure a goal with this option, select Will not occur from the top-level dropdown.

For example, if you are interested in predicting which customers become less engaged and do not visit your account login page in the next month. Select Will not occur followed by All of and then enter web.webInteraction.URL (or a similar field) and equals as the operator with account-login as the value.

will not occur

All of and any of

In some cases, you may want to predict whether a combination of events will occur and in other cases, you may want to predict the occurrence of any event from a pre-defined set. In order to predict whether a customer will have a combination of events, select the All of option from the second-level drop-down on the Define Goal page.

For example, you may want to predict whether a customer purchases a particular product. This prediction goal is defined by two conditions: a commerce.order.purchaseID exists and the productListItems.SKU equals some specific value.

All of example

In order to predict whether a customer will have any event from a given set, you can use the Any of option.

For example, you may want to predict whether a customer visits a certain URL or a web page with a particular name. This prediction goal is defined by two conditions: web.webPageDetails.URL starts with a particular value and starts with a particular value.

Any of example

Eligible population (optional)

By default, propensity scores are generated for all profiles unless an eligible population is specified. You can specify an eligible population by defining conditions to include or exclude profiles based on events.

eligible population

Custom events (optional)

If you have additional information in addition to the standard event fields used by Customer AI to generate propensity scores, a custom events option is provided. Using this option allows you add additional events that you deem influential which may improve the quality of your model and help to provide more accurate results. If the dataset you selected includes custom events defined in your schema, you can add them to your instance.


For an in depth explanation on how custom events effect Customer AI scoring results, visit the Custom event example section.

event feature

To add a custom event, select Add custom event. Next, input a custom event name then map it to the event field in your schema. Custom event names are displayed in place of the fields value when looking at influential factors and other insights. This means that the custom event name will be used instead of the ID/value of the event. For more information on how custom events are displayed, see the custom event example section. These additional custom events are used by Customer AI to improve the quality of your model and provide more accurate results.

Custom Event field

Next, select the operator you wish to use from the available operators drop-down. Only operators compatible with the event are listed.

Custom Event operator

Lastly, enter the field value(s) if the operator selected requires one. In this example, we only need to see if a hotel or restaurant reservation exists. However, if we wanted to be more exact we could use the equals operator and enter an exact value in the value prompt.

Custom Event field value

Once complete, select Next in the top-right to continue.

Custom profile attributes (optional)

You can define important Profile dataset fields (with timestamps) in your data in addition to the standard event fields used by Customer AI to generate propensity scores. Using this option allows you to add additional profile attributes that you deem influential which may improve the quality of your model and provide more accurate results. Additionally, adding custom profile attributes allows Customer AI to better showcase how particular profiles ended up in a propensity bucket.


Adding a custom Profile attribute follows the same workflow as adding a custom event. Similar to custom events, custom profile attributes affect your model scoring in the same way. For an in-depth explanation, visit the Custom event example section.

add a custom profile attribute

Select profile attributes from the Profile snapshot export

You can also choose to include profile attributes from the daily Profile snapshot export. These attributes are synced to the Profile snapshot export and display the most recently available value. They automatically show up and do not require a dataset to be selected in the configuration step.


Do not select a profile attribute that has been updated as a result of the prediction goal or is highly-correlated with prediction goal. This results in data leakage and over-fitting of the model. For example, total_purchases_in_the_last_3_months is an attribute that predicts purchase conversion…

Adding a custom event example

In the following example, a custom event and profile attribute is added to a Customer AI instance. The goal of the Customer AI instance is to predict the likelihood of a customer to buy another Luma product in the next 60 days. Normally, product data is linked to a product SKU. In this case, the SKU is prd1013. After the Customer AI model is trained/scored, this SKU can be linked to an event and displayed as an influential factor for a propensity bucket.

Customer AI automatically applies feature generation such as “Days since” or “Counts of” against custom events such as Watch purchase. If this event was considered an influential factor on why customers are high, medium, or low propensity, Customer AI displays it as Days since prd1013 purchase or Count of prd1013 purchase. By creating this as a Custom event, you can give the event a new name making the results much easier to read. For example, Days since Watch purchase. Additionally, Customer AI uses this event in its training and scoring even if the event is not a standard event. This means you can add multiple events that you think might be influential and customize your model further by including data such as reservations, visitor logs, and other events. Adding these data points further increases the accuracy and precision of your Customer AI model.

example of a custom event

Set options

The set options step allows you to configure a schedule to automate prediction runs, define prediction exclusions to filter certain events, and toggle Profile on/off.

Configure a schedule (optional)

To set up a scoring schedule, start by configuring the Scoring Frequency. Automated prediction runs can be scheduled to run on either a weekly or a monthly basis.

Prediction exclusions (optional)

If your dataset contained any columns added as test data, you can add that column or event to an exclusion list by selecting Add Exclusion followed by entering the field you wish to exclude. This prevents events that meet certain conditions from being evaluated when generating scores. This feature can be used to filter out irrelevant data inputs or promotions.

To exclude an event, select Add exclusion and define the event. To remove an exclusion, select the ellipses () to the top-right of the event container, then select Remove Container.

Profile toggle

The Profile toggle allows Customer AI to export the scoring results into Real-Time Customer Profile. Disabling this toggle prevents the models scoring results from being added to Profile. Customer AI scoring results are still available with this feature disabled.

When using Customer AI for the first time you can toggle this feature off until you are satisfied with the model output results. This prevents you from uploading multiple scoring datasets to your Customer Profiles while fine tuning your model. Once you have finished calibrating your model, you can clone the model using the clone option from the Service instances page. This allows you to create a copy of your model and toggle profile on.

Profile toggle

Once you have your scoring schedule set, prediction exclusions included, and the profile toggle where you want it to be, select Finish in the top-right to create your Customer AI instance.

If the instance is created successfully, a prediction run is immediately triggered and subsequent runs execute according to your defined schedule.


Depending on the size of the input data, prediction runs can take up to 24 hours to complete.

By following this section, you have configured an instance of Customer AI and executed a prediction run. Upon the run’s successful completion, scored insights automatically populate profiles with predicted scores if the profile toggle is enabled. Please wait up to 24 hours before continuing to the next section of this tutorial.

Next steps

By following this tutorial, you have successfully configured an instance of Customer AI and generated propensity scores. You can now choose to use the Segment builder to create customer segments with predicted scores or discover insights with Customer AI.

Additional resources

The following video is designed to support your understanding of the configuration workflow for Customer AI. Additionally, best practices and use case examples are provided.


The following video is out of date. For the most up-to-date information refer to the documentation.


In this video, I’ll show you how to create an instance of Customer AI, an AI EMO service that’s part of Adobe Sensei. After watching this video, you should be ready to dive in and start using this powerful AI driven service to create your own predictions without the need for data science expertise. First, let’s define a problem. Luma, an athletic apparel retailer wants to increase sales of watches. They want to know which customers are more likely to purchase a watch so they can optimize their marketing spend. Customer AI will generate prediction scores, which will be added and synchronized with real-time customer profiles so you can use them in marketing actions across Adobe Experience platform. Before we dive into the interface, let’s understand the basic requirements for a Customer AI instance. First, Customer AI is a predictive algorithm, which predicts the likelihood of and rationale behind a future action. Before you can create an instance, you need to know what customer action you want to predict. And in our case, we’d like to predict whether a visitor to the website would purchase a watch. Next, we need to know what the timeframe is for the action. For example, do we want to predict the likelihood a customer will purchase a watch in the next 30 days or in the next week? Lastly, by default, Customer AI scores all profiles. Need to determine whether you want to narrow your eligible population, for example, do you want to predict that recent website visitors within a specific timeframe will purchase a watch? Also, Customer AI requires a minimum amount of historical data. You can learn more about how to prepare your data and what kind of historical data is needed in the documentation. Now let’s go to the Customer AI interface. And Adobe Experience platform navigate to the data science tab using the left rail. From here, select services, and then select open from within the Customer AI card. The Customer AI screen appears and displays all your service instances. You can select any existing service instance to edit, clone, delete, and view details about that instance. We’re interested in creating a new instance though, so let’s select the create instance button. The instance creation workflow appears starting on the setup step. First, we enter the instance name. This name is used in every place where a Customer AI score is displayed. So let’s give it a good descriptive name. For propensity type, we can choose between churn or conversion. In this case, we want conversion. Select next to proceed to the select data step. For this step, we select the datasets which will be used to predict scores. By design, only datasets compatible with Customer AI are listed. Currently, those are experience event datasets, data coming from the Adobe Analytics source connector or data coming from the Audience Manager source connector. If you don’t see your dataset, double check that it’s compatible with Customer AI. If you’re not sure which datasets you need, click the info icon to load a preview. In the modal, you can see various details of the dataset as well as some sample records. You can also select multiple datasets to use in a model. The one requirement is that the datasets share a common identity. The interface will display the primary identity namespace of the datasets underlying schema. However, you can change this to use any of the other identity namespaces defined in this schema. The main requirement when you use multiple datasets is that you select the same identity namespace for each dataset. When you use multiple datasets, all columns with the same name will be merged, even if the data is different. The dataset preview area is a great way to confirm consistency across your datasets. Each dataset will display 10 rows in the preview. Once you’re done selecting your datasets, select next. Now we define our goal, the action we want to predict by selecting the corresponding event from our dataset. We’re given a lot of options in the UI that we can use to customize our goal. To learn more about the available options, visit the documentation. In this example, we want to use multiple events joined by and conditions. Notice that depending on the data type of the event field, we’ll see relevant options in the next dropdown. Note that there are two operators above the conditions. The first, all of means both of the conditions must be true. There must be a purchase ID and this queue must contain watch. Alternatively, I can create an or condition by selecting any of. Just above, I can choose between will occur and will not occur. Will occur will predict that the customer will buy a watch, and will not occur will predict that they won’t buy a watch. You can then specify the timeframe for the prediction. In this example, we’re predicting the likelihood that the customer will purchase a watch in the next 30 days. Customer AI looks for a list of standard events as defined in the documentation. If there are additional events that you’d like the model to take into consideration, you can specify that in the custom events section. This could improve model quality and provide more accurate results. Similar to custom event definitions, you can specify profile attributes that you’d like the model to take into account. For example, there are two profile attributes that you’d like the model to consider such as city and membership type. You can specify that here. By default, propensity scores are generated for all profiles. However, you’re welcome to limit this eligible population to a smaller subset. If we want to get predictive scores for only those who have visited the website in the last 30 days, this is specified under eligible population. You can also include or exclude profiles based on the conditions you set. Select next to continue to the final step, set options. We need to set up our scoring schedule depending on how often we want our prediction scores to be generated. Prediction runs can be scheduled to run on either a weekly or monthly basis. If we want, we can exclude any event from this prediction that may be an anomaly and skew our results, such as an event that was configured for testing purposes.

Once the setup is complete, Adobe Sensei will begin scoring each customer on their likelihood to purchase a watch. The profile toggle allows Customer AI to disable syncing to the real-time customer profile. By default, scores are synced to profile. Click finish to create your Customer AI instance. Initially, the hyperlink on our instance name is disabled since the scores haven’t been generated. Customer AI will immediately start to train and score, but the entire process can take up to 24 hours to complete. After the initial training and scoring, scores will be updated according to the scoring schedule we specified when creating our instance. If there were any issues with our training or scoring, the status will show as failed. We can click anywhere on the row and the right rail will give us additional details on what went wrong. So we walked through some of the basics of creating a Customer AI instance. You should now feel comfortable creating your own instances of Customer AI to predict customer behavior. -

On this page