Every year, Adobe Summit brings announcements that promise to change how marketers and data teams operate. This year, the conversation centers on Agent Orchestration and how the hype is becoming real for Adobe.

We've spent the last several weeks close to the product training, heading into Summit with a clear-eyed view: some of this is deployable today, some of it is 12+ months from being enterprise-ready, and nearly all of it requires a data foundation most organizations haven't finished building. 

Adobe's Agent Orchestrator isn't a single AI model. It's a coordination layer—a modular architecture that routes tasks to specialized agents, manages shared memory across those agents, and connects to both Adobe and third-party systems. That architecture is more mature than we expected. 

What stood out:

• Shared memory is real and live within AEP. Agents read from and write back to a unified customer profile in real time, changing what multi-step workflows can do. 
• The reasoning engine is more deliberate than a chatbot. It builds a plan, selects the right agents, and adjusts dynamically as conditions change, making complex task delegation actually possible. 
• The out-of-the-box agent suite is more impressive than most people realize. Ten purpose-built agents cover audience segmentation, journey optimization, experimentation, data engineering, site optimization, content production, and more.  

What isn't there yet: 

• Robust error handling when a mid-chain agent fails. If one agent in a multi-step sequence hits a problem, recovery is inconsistent. For production deployments involving high-stakes customer interactions, this matters. 
• Fully autonomous campaign management. End-to-end autonomous campaign execution—without human checkpoints—is still a roadmap item. 

Deployable today (or very close): 

• Audience Agent — Natural language audience segmentation within Real-Time CDP is ready now. The ability to prompt for high-value segments without writing manual queries unlocks productivity. 
• Journey Agent — Automated journey optimization within Adobe Journey Optimizer is ready for teams with mature AJO implementations. You can identify drop-off points and adjust in-flight journeys in real time, with high-volume, well-instrumented journeys driving the best results.  
• Data Insights Agent in CJA — Natural language queries automatically produce the right visualization from the right data view, and the agent can turn an entire Analysis Workspace project into a stakeholder PowerPoint. The March 2026 Copilot integration brings CJA data into Teams, so business leaders never have to open the tool. Output quality depends on your data view hygiene, but this is a noteworthy agent that most analytics teams will touch first. 

12+ months out for most organizations: 

• Brand Concierge at Scale — The conversational commerce experience is compelling in demos. But deploying it as a primary customer-facing front door requires a fully unified profile, governed content, and multi-agent coordination that most organizations aren't ready to support. 
• Fully Autonomous Campaign Management — The Experimentation Agent and Content Production Agent work well in assisted mode. Full autonomy—where agents hypothesize, execute, and optimize without human review—needs more guardrails before it’s enterprise-appropriate. 
• Multi-Agent Orchestration Across Third-Party Systems — Agent to Agent (A2A) and Model Context Protocol (MCP) enable agents from different vendors to collaborate, but production-grade implementations crossing Adobe and non-Adobe stacks are still early.  

The B2B case worth watching: For B2B revenue teams, Adobe's Account Qualification Agent automatically builds buying groups, surfaces missing contacts in target accounts, and prepares engagement summaries before sales calls. For organizations using account-based marketing (ABM), it aligns marketing and sales around a shared account view without requiring a full agentic stack.

Every agent Adobe will announce at Summit shares one characteristic: it makes decisions on behalf of a practitioner—selecting the right metrics, choosing visualizations, identifying drop-off points, building audience segments. That's powerful, and we're excited about what it unlocks. 

But from our work building and evaluating conversational AI systems across Snowflake, Databricks, and other platforms, we've learned that the hardest part isn't getting these systems to work. It's knowing how well they're working, and whether that's changing over time. 

The challenges are consistent across platforms:  

• Natural language is ambiguous. The same question phrased in different ways can produce different results.  
• Data Configuration matters enormously. The quality of your data views, naming conventions, and semantic context directly determines output accuracy.  
• CJA quality is dependent on workspace hygiene. The failure modes that matter most are the ones you can't see: answers that look right but aren't. 

To address these challenges, Adobe has published rigorous research on severity-based error taxonomies and continual improvement frameworks for AI Assistant. The reasoning panel in Agent Orchestrator provides useful transparency into how agents arrive at their responses. 

What we’re watching for: how that rigor translates into capabilities that practitioners can use in their own environments. When the Data Insights Agent answers a question about campaign performance, how does an analytics team systematically verify accuracy across hundreds of queries rather than spot-checking a handful? As the Journey Agent moves from assisted recommendations to autonomous optimization, what gives a marketing leader confidence that agent decisions are improving rather than drifting? 

These are the questions every organization deploying AI agents needs to answer early in their implementation journey. Evaluation is what separates AI systems that earn lasting trust from those that plateau after the initial excitement fades. 

Here's the part most Summit sessions will skip: agent orchestration is only as good as the data foundation underneath it. 

Most organizations have not fully deployed Real-Time CDP. That matters when agents are making decisions autonomously. An agent orchestrating journeys on incomplete or poorly resolved customer profiles accelerates bad decisions at scale. 

Three things need to be in place before agents deliver real value: 

• A Strong AEP Identity Graph for Platform Scale: The agents that matter most to customer experience (Journey, Audience, and Experimentation) perform better using a well-resolved identity graph. You don't need perfect identity resolution to get started, but an incomplete graph does limit upside for high-value agents.
• Real-Time Data Access: The Journey Agent and Audience Agent respond to behavioral signals as they happen. If your data pipeline is batch-based, agents can’t personalize experiences in real time. 
• Governance and Consent Built into the Activation Layer. Agent orchestration operating on customer data is a compliance question, not just a marketing one. Adobe's architecture includes role-based access, AES 256-bit encryption, and regional data residency controls, and does not use customer data to train its models. But your organization still needs consent logic embedded in how data gets activated. One gap creates downstream risk across every agent in the chain. 

In our experience, most organizations don't yet have a mature data foundation. If that’s you, invest there before you invest in agents. The reason is practical: organizations without that foundation typically don’t drive significant ROI from AEP, and most of these agents require AEP Prime or Ultimate editions, AJO B2B add-ons, or specific CJA tier access. It’s hard to justify that additional investment in Adobe technology when the underlying platform hasn’t yet returned its expected value. 

The ROI framing that resonates most with business leaders: this isn't about replacing people, it's about removing the operational bottleneck between insight and action. Creative teams reclaim hours lost to manual content production. Analysts move from hours-long reporting cycles to on-demand answers. Journey teams catch drop-off in real time rather than the next campaign cycle.  

The gap between knowing what a customer needs and being able to respond to it at scale is where revenue leaks. Agents close that gap, but only if the data foundation is there first. 

We're bullish on where this is headed. The architecture is real, the agent suite has expanded, and the use cases (particularly in analytics, audience, and B2B account qualification) are deployable for organizations that have done the foundational work. 

If you're at Summit and want to cut through the demo layer to understand what's actually applicable to your organization's stack right now, let's connect.