At Adobe Summit 2026, Adobe rebranded its Experience Cloud solution to Adobe CX Enterprise and made it clear the future is agentic. Agent Orchestrator and the first wave of purpose-built agents are now generally available, including Audience, Journey, Data Insights and Experimentation. The new CX Enterprise Coworker tier is coming in the months ahead to orchestrate longer-running, goal-oriented workflows across them. The promise is compelling: AI that doesn’t just answer questions, but reasons through multi-step work and acts on the user’s behalf.
But here’s the reality: the quality of what these agents produce is directly proportional to the quality of the foundation they’re built on. Would a new hire understand your schema and data from a field name and description alone? If the answer is no, an AI agent won’t either. Adobe’s own 2026 AI and Digital Trends report makes the gap concrete, with 75% of organizations citing data integration and quality as their top challenge for agentic AI. Most organizations will turn on Adobe’s AI agents and be disappointed. Not because the agents don’t work, but because their foundation isn’t ready. Here are three questions worth asking before your first agent goes live.
1. Is updating schema metadata and data dictionary still a backlog item?
Adobe’s Data Dictionary feature in Customer Journey Analytics (CJA) and Adobe Analytics, along with schema metadata in Adobe Experience Platform (AEP), is consistently underutilized. The work to build and maintain it is often deprioritized, kept out of date, or duplicated in a separate tool outside Adobe. A Data Dictionary doesn’t directly deliver data or insights, so it’s typically treated as an ease-of-use feature. Deploying AI agents changes that calculus. The schema metadata and Data Dictionary become the source that teaches an AI agent specific language and data. Metadata and natural language descriptions are now what accelerate an organization’s ability to accomplish tasks efficiently and accurately through agents. This is the highest-impact area for AI agent readiness and the quickest win for most.
Agent Orchestrator’s knowledge base draws from customer metadata about business objects: schema field names, descriptions, dataset names, audience definitions, and journey configurations. An agent asked to ‘build an audience of high-value churned customers’ must map each of those terms to actual fields, datasets, and segments in your environment. Without descriptive metadata, it guesses.
In CJA specifically, the Data Insights Agent builds visualizations using components from your data view. If those components have clear, descriptive names and well-written descriptions, the agent can map natural language questions to the right dimensions and metrics. If components are cryptically named, the agent is flying blind.
The ROI calculation for metadata investment has fundamentally changed. Every description you write, whether on a schema field, a data view component, an audience, or a journey, is now an instruction to an AI agent about how to understand and use your data. Invest in metadata richness before enabling agents and you’ll see dramatically better results from day one.
2. Are users compensating for data quality issues that an AI agent can’t?
Experienced marketers and business users carry a mental model of data that took months or years to build. They know the data’s quirks, the migration artifacts from Adobe Analytics and manual steps required to extract real insights.
When an AI agent comes to work, it doesn’t get the same onboarding. Agents work entirely from existing schema, unfiltered data and an existing identity graph. This is the biggest risk with Agents: they will confidently generate insights from a system as it currently exists, without knowing what’s wrong with it. Business users asking agents questions can get plausible-looking answers without the context to know they’re wrong. Schemas and data need to be approached with an understanding of who’s consuming them and how. Natural language removes the friction that used to surface data problems. With agents, those problems don’t disappear; they just become harder to see.
Schema design for AI readability
The AI models powering Agent Orchestrator process natural language, which means they perform best when schema fields are human-readable and well-described. Fields with cryptic names and no descriptions force agents to guess meaning, while fields with clear display names, comprehensive descriptions and standard structures give agents the context they need to interpret prompts correctly.
Organizations should audit their schemas with a new lens: would a new team member who has never seen the data understand what each field captures from its name and description alone? If not, the agent won’t understand it either. Leveraging standard Adobe field groups wherever possible also helps, since agents are trained on Adobe’s standard XDM components and inherently understand them better than custom equivalents.
Data quality as an agent dependency
AI agents amplify both the value of good data and the consequences of bad data. A human analyst encountering a null value in a critical field can intuitively work around it; an AI agent may silently exclude millions of profiles from an audience without flagging the gap. Before agents go live, organizations should invest in auditing data quality across key fields: null rates, inconsistent enum values, timestamp accuracy, and duplication.
Identity resolution maturity
The Audience Agent’s ability to detect duplicates, provide accurate sizing and create actionable segments depends entirely on the quality of the identity resolution. If the identity graph is noisy (shared devices causing profile collapse, or missing identity stitching between channels), agent-created audiences will inherit those issues at scale. Validating merge policies and identity graph hygiene is a prerequisite for trusting agent-driven audience work.
3. Are governance controls ready to work without a human in the loop?
When a human analyst builds an audience, they bring institutional knowledge about which data fields are sensitive, which datasets have contractual restrictions and which customer segments require consent. An AI agent doesn’t carry that institutional memory. It relies entirely on the governance framework that is built into the platform. The labels, consent policies and access controls are the guardrails that keep agents from crossing lines they don’t know exist. Additionally, as agents increasingly operate across third-party platforms, those guardrails need to extend beyond Adobe’s boundary. What is configured in Adobe’s systems today becomes the governance foundation for the entire ecosystem.
Data usage labels, policies and privacy
AEP’s Data Usage Labeling and Enforcement (DULE) framework is the primary mechanism for ensuring AI agents respect privacy and contractual constraints. Labels categorize data by restriction type (contractual obligations, identity sensitivity and data sensitivity), and governance policies link those labels to specific marketing actions per channel. When an agent attempts to activate data that violates a policy, the platform blocks it automatically. This only works if labels are comprehensively applied, and policies are actively enforced, not just drafted. Consent policies also need to be configured for channels where customer opt-in is required, or an agent could unknowingly activate restricted data at scale.
Equally important is Attribute-Based Access Control (ABAC), which governs what data is visible to users within the platform. Since agents operate under the permissions of the authenticated user, ABAC labels on schema fields directly determine what data an agent can see and act on. A layered model (DULE for outbound usage, consent policies for customer preferences, and ABAC for internal visibility) creates the governance depth that responsible agent deployment requires.
Permissions and administrative readiness
Agent Orchestrator introduces a new permissions and metering layer that administrators need to plan for. Agent jobs consume AI Credits on a usage-based metering model, which means cost forecasting and credit allocation become governance decisions in their own right. Beyond the budgeting question, each agent operates under the authenticated user’s product-level access controls, so the permissions you assign to a role directly determine what an agent can see and do.
The key consideration isn’t just who gets access but designing a tiered model that aligns agent capabilities with user roles. A campaign manager using Journey Agent to build multi-channel journeys needs a different permission set than an analyst using Audience Agent to explore segment overlap. Before enabling agents in production, establish a dedicated sandbox for agent testing. This gives teams a safe space to learn how agents interpret data and governance configurations before any of those interpretations reach a customer.
Cross-platform governance
At Summit 2026, Adobe expanded its agentic AI partnerships to include Anthropic, AWS, Google Cloud, IBM, Microsoft, NVIDIA, and OpenAI, with cross-platform interoperability built on open standards including Model Context Protocol (MCP) and the Agent2Agent (A2A) protocol. This means the governance perimeter no longer ends at Adobe’s platform boundary. When a third-party agent queries EP data via A2A, any DULE labels and ABAC restrictions should travel with that request. The guardrails built inside Adobe today are the foundation for governing a broader agent ecosystem tomorrow.
Final thoughts
The gap between “we have AI agents” and “our AI agents are producing trustworthy, compliant and actionable outputs” is entirely determined by the metadata, data quality and governance foundation underneath them. The organizations that will get the most value from agentic AI will be those that invest in governance, data quality and metadata richness first, and then let agents amplify that solid foundation.
Like a new hire, new AI agents will act on exactly what they have been given. Protiviti Digital helps Adobe CX Enterprise customers build the foundation that holds up when agents are making decisions at scale and across organizational boundaries, making agents’ first day at work better than most.
To learn more about our Adobe consulting services, contact us.
