There's a conversation happening in marketing right now that most people aren't ready to have yet. We've spent the better part of a decade building what I call the "data collection industrial complex"—massive CRMs, customer data platforms, analytics suites that promise 360-degree customer views. And we succeeded. We have the data.
But here's what nobody wants to admit at conferences: we still don't really know what to do with it in the moment that matters.
I watch marketing teams every day—brilliant people, well-funded teams—running campaigns that are still fundamentally based on batch-and-blast logic dressed up in personalization language. "Dear [First Name]" isn't personalization. Sending an abandoned cart email 24 hours later isn't real-time intelligence. It's just slower batch processing.
The infrastructure we built was designed for a different era. It collects, it stores, it reports. But it doesn't think.
Why generative AI didn't fix this (yet)
When GPT-3 arrived, then GPT-4, I had the same rush of excitement everyone else did. Finally—AI that understands context, that can write, that can engage. Marketing teams jumped in fast. We automated content creation, experimented with chatbots, generated endless variations of ad copy.
But six months in, something became clear: we'd automated the output without upgrading the decision-making.
We could generate better emails faster, but we still didn't know who should receive which message at what moment based on what they were actually doing right now.
Generative AI gave us better tools to execute. It didn't give us better judgment.
What marketing actually needs—what we've needed for years—is a system that can observe customer behavior across channels, understand intent in context, make decisions about next actions, and execute those decisions autonomously. Not a tool that waits for human input. A system that acts like a really good marketing director who's watching everything and making calls in real-time.
That's not generative AI. That's agentic AI.
Agentic ≠ Automated. Agentic systems decide; automated systems execute.
When architecture actually matters
I spend a lot of time reading patents. Most are incremental—slight improvements on existing systems, repackaging old ideas with new terminology. But every once in a while, you encounter something that makes you reconsider what's architecturally possible.
Aswajit Mohapatra's provisional patent on agentic AI for e-commerce is one of those moments.
What struck me wasn't just the technology stack—though it's sophisticated—it was the mental model. Mohapatra isn't thinking about marketing automation as a series of triggered workflows. He's thinking about it as an autonomous decision-making system that learns continuously.
The architecture breaks down into four layers, but what matters is how they work together:
The data foundation handles what most systems do poorly—transforming messy behavioral signals into coherent, decision-ready features. Frequency metrics, lifetime value predictions, engagement segmentation. This isn't novel, but it's done rigorously.
The preprocessing layer is where customer data becomes customer intelligence. This is the difference between knowing someone bought a blue shirt and understanding they're in an active exploration phase for a wardrobe refresh.
The agentic core is where it gets interesting. This isn't a single model—it's a hybrid system combining collaborative filtering, deep learning, and reinforcement learning for recommendations. Dynamic pricing models that respond to demand signals and inventory levels in real-time. A domain-specific LLM that's been fine-tuned specifically for shopping conversations, not general chat.
And here's the part that matters for privacy-conscious brands: virtual try-on technology that runs entirely on-device. No biometric data leaves the customer's phone. This isn't just good ethics—it's smart risk management.
The delivery layer handles orchestration across platforms—Shopify, WhatsApp, CRM systems, even AR mirrors in physical retail.
What makes this agentic rather than just automated is the reinforcement learning loop. The system doesn't just execute predefined rules. It learns which actions drive outcomes and adjusts its decision-making accordingly. Over time, it gets better at predicting what will work.
How the reinforcement learning actually works
Most marketing systems operate on fixed rules: if cart abandoned → send email. Agentic systems use RL to optimize which action produces the best outcome in this specific context.
The system maintains a policy network that maps states (customer context, behavioral signals, inventory levels) to actions (recommend product, adjust pricing, trigger support). Each action generates a reward signal based on outcomes (purchase, engagement, retention). The policy network updates continuously using these rewards to improve future decisions.
This means the system isn't following a decision tree—it's learning a strategy. In technical terms, it's closer to AlphaGo's approach than to traditional if-then automation.
Where theory meets execution: Shop-Intel
Mohapatra didn't just patent this—he built it. Shop-Intel is the commercial implementation, and it's one of the few real-world examples I've seen where this actually works at scale.
Full disclosure: I'm not affiliated with Shop-Intel or Mohapatra. I'm looking at this as a case study of what's architecturally possible when you design for agentic decision-making from the ground up.
The conversational AI layer handles natural language shopping queries with 83% better accuracy than general-purpose LLMs. That's significant. It means the system actually understands retail context—product attributes, inventory constraints, seasonal trends, category relationships. When someone asks "Do you have this in petite sizes?" the system knows what that means for fit, for inventory, for related products.
The personalization engine adjusts promotions dynamically. Not just "here's 10% off" but "here's 15% off this category because you abandoned this item type twice, you're a repeat customer, and we have excess inventory." That's merchandising intelligence embedded in the customer experience.
The demand forecasting reduces overstock by 40%. I've worked with enough retail operations teams to know what that means financially. Overstocking is one of the biggest margin killers in retail. A 40% reduction is material.
Cart abandonment dropped 28% in implementations I've reviewed. That's not from better email sequences—it's from real-time intervention. The right support at the right moment with the right context.
Let's be honest about the metrics
The reported numbers: 42% conversion lift, 35% increase in average order value.
I've been in marketing long enough to be skeptical of any vendor-reported metrics. We don't know the baseline conditions. We don't know what else changed during implementation. We don't know if these results hold across different verticals or customer segments.
But here's what I do know: even if the real-world impact is 60% of what's reported—say, 25% conversion lift and 20% AOV increase—that's still transformational for most retail operations. These aren't marginal improvements. They're structural changes to how the business performs.
What gives me confidence isn't the topline numbers. It's the mechanism. The system is doing things that demonstrably should improve performance: better targeting, more relevant recommendations, dynamic pricing, proactive support. The causality makes sense.
What this means for the marketing function
I've heard the replacement narrative too many times to take it seriously anymore. "AI will replace marketers." No, it won't. But it will change what we do.
What agentic AI replaces is low-value decision-making. Should this customer get an email today? Which product should we recommend? What discount level converts this segment? These are important questions, but they're not creative questions. They're optimization problems.
What humans should be doing—what we're actually good at—is setting strategy, defining brand positioning, understanding cultural context, creating narratives that resonate emotionally, knowing when to break the pattern.
The best marketing teams I know will use agentic AI to handle execution while they focus on the creative and strategic work that actually differentiates brands. The worst teams will try to do everything manually and wonder why their competitors are moving faster.
The privacy architecture matters more than people realize
One aspect of Mohapatra's work that doesn't get enough attention is the privacy-first design, particularly around the virtual try-on functionality.
Most AR and virtual try-on systems send your biometric data—face scans, body measurements—to central servers for processing. This creates enormous privacy risk. It's also increasingly legally problematic in jurisdictions with strong data protection laws.
Processing this data on-device, using federated learning models, solves this structurally. The AI model improves from aggregate learning without individual data ever leaving the device. This isn't just ethically better—it's architecturally sounder.
As privacy regulations tighten globally, brands that built privacy into their AI systems from the beginning will have a significant competitive advantage over those trying to retrofit it.
What's actually changing
The shift from "marketing automation" to "agentic marketing systems" represents a fundamental rethinking of what software should do.
Traditional automation executes predefined workflows. If this, then that. Even sophisticated systems are essentially decision trees with more branches.
Agentic systems make decisions. They evaluate multiple variables, predict outcomes, choose actions, execute, observe results, and adjust their decision-making. The marketer's role shifts from "execute this campaign" to "set these objectives and constraints."
This is the difference between a tool and a system. Tools require constant human direction. Systems operate autonomously within defined parameters.
Mohapatra's work represents some of the more sophisticated thinking I've seen in this direction. It's not just automation with better AI models. It's a different paradigm for how marketing systems should function.
Where this goes next
I think we're at the very beginning of understanding what's possible with agentic marketing systems. The applications I've seen so far are primarily in e-commerce, but the logic extends to B2B, to content marketing, to customer lifecycle management.
Imagine a system that doesn't just send renewal reminders but understands customer health signals, predicts churn risk, and automatically adjusts engagement strategies for at-risk accounts. That's not theoretical—it's the same agentic architecture applied to a different domain.
Or consider content marketing. A system that understands which content drives pipeline, which topics resonate with which segments, and automatically adjusts content strategy and distribution. Not scheduling tools—decision-making systems.
The constraint right now isn't the technology. It's organizational readiness. Most marketing teams aren't structured to work with autonomous systems. We're still organized around campaign execution, not continuous optimization.
The teams that figure this out first—that restructure around agentic systems rather than trying to fit them into existing workflows—will have an enormous advantage.
A final thought on empathy
There's something interesting happening in Mohapatra's approach that goes beyond the technical architecture. The system is designed to understand customers as dynamic, contextual beings rather than static profiles.
Traditional personalization is basically sophisticated stereotyping. "People like you bought this." Agentic personalization is contextual understanding. "Based on what you're doing right now, in this context, with this intent, here's what makes sense."
That's closer to how humans actually think about helping each other. We don't categorize people into segments. We respond to context.
I don't know if AI can truly be empathetic—that's probably the wrong framing. But systems that respond to context rather than category feel more human, not less. They feel like someone's actually paying attention.
Maybe that's the real insight here. The best technology doesn't replace human connection. It makes it possible at scale.
What to do about this
If you're building marketing systems or making technology decisions for a brand, here's what's worth exploring:
Audit your stack for decisions still driven by static rules. Look for places where you're using if-then logic when you could be using learning systems. Email triggers, promotion logic, content recommendations—these are all candidates for agentic upgrades.
Start experimenting with reinforcement learning loops in small workflows. You don't need to rebuild everything. Pick one decision point—like email send timing or product recommendations—and test whether an RL-based approach outperforms your current rules.
Think in terms of systems that learn, not tools that wait for instructions. When evaluating new technology, ask: "Does this make decisions, or does it just execute mine?" If it's the latter, you're buying automation, not intelligence.
Build privacy into the architecture from day one. Don't treat it as a compliance checkbox. The brands that figure out privacy-preserving AI architectures now will have structural advantages as regulations evolve.
The shift to agentic systems is happening whether marketing teams are ready or not. The question is whether you're designing for it or getting disrupted by it.
What are you seeing in your marketing stack? Are you running into the same gaps between data collection and intelligent action?
Ankita Sahoo writes about how AI is reshaping marketing from the inside out. She helps brands and builders bridge the gap between technology and human insight.