Adobe Experience Platform ships two predictive services, and organizations routinely buy the wrong one for the question they have — not because the products are unclear, but because the choice gets made from the wrong end: pick a service, then look for a question it fits.

The reliable way round starts from the answer. Customer AI produces a statement about a person, forward in time. Attribution AI produces a statement about a path, backward in time. Which one to license, configure and staff follows from which shape the business asked for. What both services are lives on Customer AI and Attribution AI.

Table of contents

1. Two services, two shapes of answer
2. Which question yields which service
3. If it is Attribution AI, the next choice is the score family
4. Integration across Adobe Experience Cloud
5. What to plan for: history, volume and two easy misreadings
6. Conclusion: describe the table before choosing the model

Two services, two shapes of answer

Customer AI generates customer predictions at the individual level with explanations — propensity scores such as churn and conversion, at scale. One row of output is one profile, carrying a score from 0 to 100 relative to the population, a probability between 0 and 1, the propensity type, and the influential factors behind the number with their importance.

That shape is what makes it operational: a number and its reasons sit on a profile, and anything reading profiles can act on it — an audience, a journey, a campaign.

Attribution AI is a multi-channel algorithmic attribution service that calculates the influence and incremental impact of customer interactions against specified outcomes. It credits the touchpoints leading to conversion events — display impressions, email sends and opens, paid search clicks — across Adobe and non-Adobe data alike.

Here one row of output is one conversion, with its touchpoints stored alongside it, each carrying its name, media channel and scores. Its scores can be synced to the Real-Time Customer Profile, but the shape stays analytical: unlike Customer AI, this is not the service you build an audience from — its consumer is a report or a budget conversation.

Which question yields which service

There is a one-sentence test: write the business question down and look at its subject.

If the subject is a person — which customers, who, how many — the answer attaches to a profile, and that is Customer AI. If it is a touchpoint, channel, campaign or creative — which of these earned it, what did that spend do — the answer attaches to a conversion, and that is Attribution AI.

Then ask the more decisive question: where is this answer going next? An answer acted on per person has to reach Real-Time Customer Profile, because that is what allows a score to be used in segmentation. An answer defended in a meeting needs a row per conversion, sliceable by channel, geography, media type and product. Destination is a harder constraint than phrasing.

Sentences that stall here have two subjects. Which channels bring us customers who then churn is two questions, answered by two models whose outputs are joined afterward.

If it is Attribution AI, the next choice is the score family

Attribution AI produces two categories of score from the same run, and the difference is what you are claiming.

Algorithmic scores come from your own history: influenced is the fraction of the conversion each touchpoint is responsible for, incremental the marginal impact directly caused by it. Rule-based scores apply a fixed convention regardless of the data — first touch, last touch, linear, U-shaped (40 percent to the first touch, 40 to the last, the remaining 20 shared) and time decay, which credits touchpoints closer to the conversion more.

Having both is the point: a rule-based number restates a convention stakeholders recognize, a bridge from however they measure today, while an algorithmic number is a claim about influence derived from your data. Reporting one in the language of the other is how attribution work loses its audience.

Integration across Adobe Experience Cloud

The two services share prerequisites, which is the commercially useful part: one data preparation effort serves both. Every Intelligent Services dataset must conform to the Consumer ExperienceEvent schema or arrive through the Adobe Analytics source connector, the history must contain both positive and negative customer journeys, and combining datasets requires a shared identity type. That work is done once for the estate — which is why "which one first" is a sequencing question, not an either/or.

Downstream they diverge as their output shapes suggest. Customer AI scores become profile attributes and flow into audiences, activation and journeys — configuration detail in Customer AI scores. Attribution AI results are read in its insights interface, which takes several attribution models at once for side-by-side comparison. Both sit inside the wider Adobe AI services picture.

What to plan for: history, volume and two easy misreadings

Both services convert their configuration into a demand for days of data, by different arithmetic. Customer AI needs at least 30 days to build features, then the eligibility lookback plus the outcome window. Attribution AI needs its training window plus its lookback, and the defaults are large: the most recent two quarters plus 56 days, which is 236 days of history before anything runs. This constraint decides more cases than the question does.

Where Adobe Analytics is the source, add the backfill. The source connector can take up to four weeks to fill in history, so a new connector may not yet hold what the calculation assumes — a schedule risk that looks like a modeling failure.

The Attribution AI usage metric spans more than the sandbox in front of you. Total conversion events scored tracks the total for the current calendar year including all sandbox environments and any deleted service models. Deleting a model does not return its consumption, and a development sandbox counts against the same total as production.

Position breakdown is not an attribution model. Attribution AI also reports touchpoints by position — starter for the first touch, player for anything in between, closer for the last before a conversion. That is a dimension of the report; reading it as a credit rule produces confident nonsense.

A useful arithmetic check: the sum of percentage contribution for an attribution model across all touchpoints and positions should equal 100 — worth running on any breakdown handed to you.

Conclusion: describe the table before choosing the model

The cheapest way to get this right costs nothing and happens before any configuration exists. Describe the intended output as a table, out loud, to the person who asked. Say what one row is. "A customer, with a number and the reasons behind it" is Customer AI. "A conversion, with the touchpoints that led to it and what each earned" is Attribution AI. And if they expected the other one, you found that out in a conversation rather than a quarterly review.

Softwhale runs that conversation first and the configuration second, then builds the foundation both services need — schema, identity, and a history containing failures as well as successes. Where both questions are live, both models get built: they are not competing implementations of one idea.