What Adobe says the agent is
In Adobe's partner webinars, the Data Insights Agent is described as an agent that automates insight generation: a marketer asks a business question in natural language and the agent builds the visualization in Analysis Workspace, the analysis interface of Customer Journey Analytics, without the user needing to know what a metric or a dimension is.
It is reached through the AI Assistant in Customer Journey Analytics; a data question typed there is routed by a reasoning engine to the agent. It is built on Adobe Experience Platform's Agent Orchestrator, like Adobe's other agents. Adobe's presenters stress that it is not text in, text out: the visualization is built in the workspace so the user can see how the number was produced.
Adobe announced the agent as generally available, first inside Customer Journey Analytics, later also in a single conversational interface across Adobe's experience applications.
What it reads: the data view as knowledge base
This is the fact that decides the grading.
Adobe's product manager explains that the data view becomes a knowledge base for the agent, and that the knowledge base is built out of the component names in the data view, not out of the data points within those components. When a user asks for profits in October, the agent scans that knowledge base of component names for profit and for the time range, then pulls the matching date range. Adobe presents this as a privacy decision: the agent sees names, not values, and customers control which data views and which users it can reach.
The consequence for accuracy is direct. The agent finds the metric by its name. If the data view names a metric "orders" and the metric is built on an event that fires twice per checkout, the agent's answer is well-formed, well-visualized and wrong by a factor of two. It cannot know, because it never saw the values, and even if it had, the values would agree with each other.
On one national retailer's site, begin checkout fired twice and a checkout progress event fired six times in one pass. The whole Adobe stack on that checkout ran inside a child frame, so a reading in the top window saw none of it. 279 of 280 tag rules carried no consent gate. Every one of those facts is upstream of the data view. The agent inherits all of them.
Skills shipped and skills announced
Adobe describes the capability at launch as one skill, data visualization, and names three more on the roadmap: data summarization, root cause analysis and recommendation. A related feature, Data Storytelling, generates slides from a workspace project through a button rather than a conversation, using the same engine.
Adobe also describes the agent as a service reachable through more than one interface: the AI Assistant in Customer Journey Analytics today, and a Microsoft Copilot surface described as the Adobe marketing agent, tapping the same agent underneath.
Each new skill is a new class of answer. Root cause analysis on a data view whose consent field is a constant will find causes that do not exist. The re-audit tier on /answer-audit exists for exactly this: the same question set, re-run when the vendor ships a skill.
Access and enablement, as Adobe describes it
Adobe's partner session describes a limited-time access program making the agent available to all Customer Journey Analytics customers regardless of package, without a contract change, with enablement through the product profile: the AI Assistant selected under reporting tools, the agent enabled, and the data views the agent may use enabled for the instance. The two sessions give different end dates for the program; the transcripts do not settle which. Adobe points customers to its Experience League documentation for the current detail.
The enablement detail matters for the audit: which data views the agent can reach is a configuration choice, and the audit records it as part of the row.
How we grade a Data Insights Agent answer
The row: the question as a marketer asks it, the agent's answer with the freeform table and visualization it built, the measured truth, the evidence, one of six classes. Rubric on /how-we-grade.
The truth for a Customer Journey Analytics question comes from two places: the same data queried directly, to check that the agent chose the right components; and the site, measured on real devices per consent state, to check that the components hold what their names claim. The second check is the one the agent cannot do for itself, by design. Method on /tracking-audit.
The Adobe Analytics side of that measurement, Launch rules, AppMeasurement beacons, the Web SDK, is the work of adb.webclat.com. This page grades the answers; that site audits the implementation the answers come from. The method for the category is on /ai-agent-evaluation; the one case we can show is on /cases.