At Cannes Lions last week, Adobe announced something that barely registered as news: Omnicom, WPP, Accenture, and Stagwell's Code & Theory are now running agentic AI workflows through Adobe CX Enterprise. Rachel Thornton, Adobe's CMO for CX Orchestration, said the thing everyone's been saying for two years: "Agentic AI is no longer something brands experiment with, but what they run on."
That's true. And it's a problem.
The infrastructure is real. The partnerships are real. The agents are shipping. But I've watched enough CMO conversations to know what's actually happening in most organizations: they're bolting Claude Agents or OpenAI's agents onto measurement stacks that were already broken, then wondering why ROI still looks like a guessing game.
When agents move faster than your data layer
Here's the shape of the problem. An agentic workflow,say, a content operations system that Code & Theory is launching for sports organizations,can generate, approve, and distribute personalized creative across 50 owned channels in minutes. Real-time, intelligent, adaptive.
But most brands can't tell you whether that creativity actually moved revenue. Not because the agents are bad. Because their attribution layer is five years old and was built for a different era of media.
Agents are good at executing decisions faster than humans. They're terrible at operating in a vacuum. The moment an agent needs to decide what to do next,should it double down on this audience segment? Should it shift budget? Should it change creative direction?,it needs data that most organizations simply don't have in a format it can actually use.
Adobe's agents have access to Adobe's own CDP and media data. That's the exception, not the rule. If you're running agents on Anthropic's Claude or Microsoft's platform, they see your data through whatever integration layer you built. And if that layer was designed for human dashboards and quarterly reviews, it's already obsolete.
The three-part gap
1. Real-time data freshness
Most marketing tech stacks refresh data on a cadence: hourly, sometimes daily. Agents operating on yesterday's data make decisions optimized for yesterday's conditions. The agents themselves are 10x faster than humans, but they're running on stale intelligence. You've just created faster decision-making on incorrect premises.
2. Feedback loops that actually close
An agent needs to know: did that decision work? Most attribution models still can't answer that question with enough speed and certainty for an agent to adapt mid-campaign. You end up with agents that execute intelligently but learn slowly, which defeats the whole point.
3. Brand safety and guardrails that match agent speed
Humans reviewing every output is the bottleneck that kills agent value. But automated brand safety systems running on keyword lists and old rules miss context that matters. An agent writing about cannabis products needs to understand SB 243 compliance, local inventory rules, and medical vs. recreational nuance. Generic guardrails don't cut it.
What CMOs actually need to do first
Before you deploy agents into your marketing operations, answer these three questions honestly:
Can your data layer keep up? Your CDP, analytics stack, and attribution model need real-time or near-real-time refresh. If you're still waiting for T+2 reporting, agents will outrun your ability to course-correct. Budget 30% of your agentic project timeline to fixing your data foundation.
Do you have a ground truth for performance? Not a model. Not a proxy. An actual way to say, "This agent's decision led to this outcome." That might be incrementality testing, holdout groups, or econometric modeling. But you need it before agents start making autonomous decisions about spend allocation.
Are your compliance and brand rules codified? Generic LLM guardrails are theater. If your brand has specific rules,compliance rules, tone rules, media mix constraints,those need to be in the agent's instructions, version-controlled, and auditable. Not in a manual review process that slows everything down.
The Cannes moment was real, but incomplete
Adobe's partnerships prove the infrastructure is ready. Accenture can now deploy agentic workflows at Fortune 500 scale. Code & Theory has a template for content operations. WPP is unifying paid and owned data.
But none of that matters if your measurement, your data freshness, and your governance are still built for the 2023 version of marketing.
The agents aren't the bottleneck anymore. Your data layer is.
FAQ
Q: Do we need to wait for perfect attribution before deploying agents? No. But you do need a framework for measuring whether agents are working. Start with a single, isolated workflow (content creation, email segmentation, bid management) with clear KPIs and holdout groups. Expand from there.
Q: Can smaller teams use agentic AI without enterprise data infrastructure? Yes, but with limits. Agents work best when they have real-time feedback. Smaller teams can start with constrained domains (one channel, one product line) where you can manually validate agent decisions and close the loop quickly. Scale after you prove the pattern.
Q: Will Adobe agents work better because Adobe owns the data stack? Yes, structurally. Adobe's CX Enterprise agents have native access to real-time customer data. If you're on Salesforce, Anthropic, or Microsoft's platform, you're building data bridges that might lag. Plan for that latency in your expectations.
Q: How fast do our systems actually need to refresh? For real-time personalization: sub-minute. For campaign optimization: hourly minimum. For strategic recommendations: daily is acceptable. Don't upgrade your entire CDP for real-time if agents only need daily refreshes.
Q: Are there compliance risks with autonomous agents making marketing decisions? Absolutely. Cannabis, healthcare, financial services, and regulated industries require agents to understand context-specific rules. Generic compliance guardrails won't work. Budget for custom rule-sets and ongoing audits.