Generative AI arrives inside enterprise software quietly — in a content editor, an analytics workspace, a journey canvas — and by the time anyone asks the governance questions, three teams are already using it.

Those questions are answerable, and Adobe’s own documentation answers them. Three matter most: what a user consents to before the first prompt, what Adobe documents about the data behind each capability, and what stays attached to the content afterwards. Each answer comes with a scope — because a true statement about one feature becomes false the moment the scope is dropped.

Turning those answers into a working policy is what our Adobe AI services page describes.

Table of contents

1. What is the generative AI agreement in Adobe CX Enterprise?
2. Why three separate gates rather than one
3. What Adobe documents about the data
4. Scope: AI Assistant is a name, not one system
5. Provenance: C2PA metadata travels with the content
6. What to plan for: credits, consent and review
7. Conclusion: why the agreement and the data trail matter

What is the generative AI agreement in Adobe CX Enterprise?

Adobe’s documentation for the AI layer of its enterprise applications is explicit: before using CX Enterprise generative AI features, you must understand and follow the Adobe CX Enterprise Generative AI User Guidelines. It is a condition of use, not a recommendation, and it applies to the layer as a whole.

Underneath that sits a product-level step. In Adobe Journey Optimizer a user agreement is displayed the first time you use generative AI in the application. And underneath that sits a third thing that is not consent at all: permission. Generating content there requires the Generate Content permission, and Adobe states that the application’s AI agents all require both access to AI Assistant and agreement to the guidelines.

Why three separate gates rather than one

Because they fail differently. Consent not given blocks a person; permission not granted blocks a role; guidelines nobody read block nothing at all, and surface later in the wrong meeting.

That is why enablement is a project rather than a switch, and why the honest answer to "can we use it" is usually "yes, once someone grants the permission".

What Adobe documents about the data

The scope has to be named, so: what follows is what Adobe documents for AI Assistant in Adobe Experience Platform, on its privacy, security and governance page. It is not a statement about every Adobe application.

Adobe states that no personal data is being used by AI Assistant today, even for training purposes, and that it is unaware of consumer data. Existing access control policies are honored — with one caveat: new attribute-based access control policies are reflected after a maximum of 24 hours, and during that window users with newly restricted access still reach those fields and objects. A log of previous interactions is viewable with a 30-day retention policy; the assistant is grounded in sandbox-specific data and public Adobe documentation, with no data shared across sandboxes and prompts not shared to other customers. With Adobe Experience Platform Healthcare Shield it is described as HIPAA-ready.

The same documentation carries an obligation running the other way. Adobe’s legal disclaimer for AI Assistant states that using the chatbot constitutes consent that what you type is collected, used, disclosed and retained by Adobe and its service providers under your organization’s agreement with Adobe — and advises adding personal data only where necessary and where you have the right to use it. That is a rule for the people typing.

Scope: AI Assistant is a name, not one system

Adobe uses AI Assistant as a feature name in many of its applications, and its own note is the clearest warning available: the feature pulls information only for the application you are using — AI Assistant in AEM answers about AEM. Real-Time CDP shows the same boundary from the data side: the assistant reads metadata about objects such as audiences, datasets and destinations, and does not access the data inside the sandbox. It can tell you an audience’s name, not who is in it.

Practically, one enablement decision does not generalize: each application’s assistant has its own prerequisites, permissions and answer to "what can it see". That is why every product page in our Adobe AI services practice owns its own AI facts.

Provenance: C2PA metadata travels with the content

The second half of the question is about output. Adobe attaches machine-readable C2PA metadata to content generated or edited with generative AI using Adobe technologies, including supported third-party models running inside Adobe workflows, and the rollout across Creative Cloud, Document Cloud, Firefly and CX Enterprise applications ran through August 2026. Images, audio, video, documents and text are in scope, the metadata is preserved as content moves through supported Adobe workflows, no action is required to attach it, and — the fact that changes planning — it is enabled by default and cannot be turned off.

Adobe Marketo Engage shows what preservation means in practice. Generating an image always attaches fresh metadata. Cropping it, or rendering a text overlay onto it, would normally destroy that metadata because the pixels are recreated — so Marketo Engage reads it from the source first and re-attaches it to the result, and Adobe is careful to say the edit does not add a new generative AI action. A cropped stock photo with no generative history gets none. The counterpart limit is external: Adobe does not control how platforms outside its applications interpret provenance, and content created before the rollout gets none retroactively. Our Adobe Marketo Engage page covers that product.

What to plan for: credits, consent and review

The commercial condition is the AI credit — a usage-based metric that quantifies the execution of actions or jobs. It does not apply to everything: Adobe lists CX Enterprise Coworker and Adobe Experience Platform agents as the services consuming credits, with agent jobs costing more when they involve advanced reasoning, validation, multi-agent coordination or integration. Coworker’s published rate is explicitly introductory, available for a limited time and subject to change. We read those terms with you before anything is switched on: the entitlement, the credit consumption model and the review cadence around them belong in the two-week Solution Blueprint, not in the first invoice.

Two operational conditions complete it, and both are documented for AI Assistant in Adobe Experience Platform: that assistant is supported in English only, and non-English input may produce inconsistent or erroneous results. Adobe’s own instruction on its output is to verify it — check the sources, review the reasoning, submit feedback when something looks wrong. A review step is not caution; it is the documented way to use the feature.

Conclusion: why the agreement and the data trail matter

None of this argues against generative AI in an Adobe stack. It argues for knowing three things before a team depends on it: which agreement was accepted and by whom, what each capability does with the data it can see, and what provenance is attached to what it produces. All three are documented, answerable per feature, and none generalizes across products.

Establishing them for your estate — entitlements, permissions, guidelines, review workflow — is what our Adobe AI services page is there to start.