"Which customers are about to leave?" is where every retention program starts, and most organizations answer with a rule someone invented in a meeting — no purchase in 90 days, say. Cheap, easy to explain, and usually wrong about the customers who matter most.

Customer AI, part of the Intelligent Services family in Adobe Experience Platform, replaces the rule with a model: individual-level predictions of churn or conversion, produced without choosing an algorithm, tuning it or deploying it. The configuration takes an afternoon. Whether a useful answer exists is decided earlier, by what your event history contains.

Table of contents

1. What is Customer AI?
2. What the scores actually contain
3. What your data has to look like first
4. From an instance to an audience
5. Integration across Adobe Experience Cloud
6. What to plan for: quality, staleness and five hard limits
7. Conclusion: why propensity is a data engagement, not a configuration

What is Customer AI?

It is a propensity engine that scores individual profiles and explains itself, reporting which attributes moved each score rather than handing over a black-box ranking. That separates a list from a brief: a score sorts customers, a score with a reason tells a campaign team what to say. Two propensity types exist — churn and conversion — and the type sets the score’s intent and polarity.

What the scores actually contain

Each scored profile carries more than one number, and using the wrong one is a common mistake. The score, from 0 to 100, is a relative likelihood against the whole population — not a probability percentage, however much it looks like one. The probability, between 0 and 1, is the one Adobe recommends when comparing across goals or averaging across a population.

Alongside them sit the score date and the influential factors — the attributes that pushed the score, each with an importance of low, medium or high. Those factors are the interesting output: they turn "these people are at risk" into "here is what they have in common".

What your data has to look like first

Three conditions decide whether a propensity program is possible at all.

The data has to arrive in an accepted shape. Every dataset used by an Intelligent Service must either conform to the Consumer ExperienceEvent schema or come through the Adobe Analytics source connector; Customer AI also accepts the Adobe Audience Manager connector and Experience Event datasets on an Experience Platform schema. There is no fifth option.

The history has to contain failures. Data prepared for Intelligent Services must include both positive and negative customer journeys: to predict propensity to buy, the model needs paths that ended in a purchase and paths that did not. The abandoned cart is the training signal, not noise.

Multiple datasets need one shared identity. Combining sources is allowed only where each shares the same identity type, which makes the identity model a prerequisite — groundwork in XDM schemas and Identity Service.

There is arithmetic to do first, too. Thirty days is the minimum needed to build features at all, and where the eligibility lookback is longer, the requirement becomes that lookback plus the outcome window. Leave the eligible population undefined and it defaults to profiles active in the last 45 days — still a choice, just not yours.

From an instance to an audience

Configuration is four passes: name the instance and pick the propensity type, select datasets, define the goal as conditions on events, set the options. The goal can be narrowed with an eligible population and extended with custom events — events you consider influential beyond the standard fields, which is where a well-modeled schema pays for itself.

The options pass holds the switch that decides whether any of this reaches marketing. Enabling scores for Profile is what allows the model output to be used in segmentation; left off, the run completes, the scores sit in a dataset, and no audience can qualify on them. Adobe suggests leaving it off for a first model while you calibrate, then cloning the instance with it on.

Integration across Adobe Experience Cloud

A propensity score is only worth producing if something acts on it. Because the scores become profile attributes in Adobe Experience Platform, they can define an audience for activation through Adobe Real-Time Customer Data Platform, trigger a retention journey, and be analyzed against actual outcomes — with no separate scoring pipeline to maintain.

Verification comes from the service itself. Insights exist only for instances with a successful scoring run, and the performance summary then shows the actual churn or conversion rate for each propensity bucket against the expected rate. That is the honest test: a model whose high bucket does not out-convert its low bucket is not working, whatever its score distribution looks like.

If the real question is which touchpoints earned a conversion rather than who is likely to act, that is a different service — see Attribution AI against Customer AI.

What to plan for: quality, staleness and five hard limits

A poor model is reported, not hidden. The error model quality is poor means the area under the ROC curve came in below 0.65, and Adobe’s advice is to change a parameter and retrain rather than ship it. The usual causes are gaps inside the prediction and eligibility windows, stale dataset dates, or sparse commerce, application, web and search fields.

Deleting an instance does not delete published scores. Deletion removes the instance, its run history and the output dataset, but scores already synced to Real-Time Customer Profile stay on the profiles — an audience can keep qualifying on a prediction whose model no longer exists.

Scores age at the cadence you scheduled. Weekly or monthly are the options, and an audience reading a stale score behaves as though it were fresh.

Five things the service will not do, listed plainly by Adobe. It is not a product recommendations tool, and with thousands of SKUs should not stand in for one such as Adobe Target. It cannot say which stage of the buying journey someone is in, nor predict dynamic pricing or a purchase price point. It cannot tell you whether an offer will make a customer more likely to buy — you might still send discounts by propensity, but Adobe cautions this is not necessarily the best way to convert those customers. And it cannot find buyers for a product launching in the future, because training requires the success events to already exist.

Conclusion: why propensity is a data engagement, not a configuration

Standing up an instance is short work. Knowing whether your history can train a model, sizing the lookback a goal implies, choosing a target that matches a real decision, and getting the scores into the audiences that consume them — that is the work, and it sits upstream of the service.

Softwhale treats propensity as exactly that: schema and identity first, then a goal tied to a real retention decision, then scores wired into activation and measurement — the same chain we deliver across Adobe Experience Platform. If Customer AI is already running and nobody trusts the numbers, the performance summary usually explains why.