The AI agent pitch is everywhere in martech right now. Autonomous campaign agents. Agents that write copy, pick audiences, decide send times, and optimise bids. If you have been to a marketing technology conference in the last 18 months, you have heard this pitch on repeat.
The technology is real. The problem is the order in which companies are buying it.
The Pattern Playing Out Right Now
Most companies buy the agent layer first. Then they discover the data layer is broken. Then comes an expensive, unbudgeted data infrastructure project while the agent sits on the shelf waiting for something usable to run on.
This is not a new failure mode. The mid-2010s big data wave followed the same shape: everyone bought Hadoop clusters and dashboards, then realised their data pipelines were too messy to actually use. The tools were fine. The foundation was not ready.
What an AI Agent Actually Needs
A marketing agent needs to see what a specific customer is doing right now, not six hours ago. It needs to make a decision about what message to send, what channel to use, and when. Then it needs to act. That loop has to close in seconds.
What most companies have instead is customer data spread across five to eight systems that do not talk to each other cleanly. CRM in one place. CDP in another. Data warehouse somewhere else. Event streams from mobile, web, and product landing in yet another silo. Campaign history buried in a fourth system. The agent needs all of that in one place, at the same time, structured consistently.

Why Stale Data Is Worse Than No Agent
Consider how a real customer behaves. They browse three products, add one to cart, and leave. They return after a price-drop notification, ignore a follow-up email, then respond to a push notification two days later. A purchase means something very different depending on whether it came after three abandoned carts, a support complaint, a loyalty offer, or a referral from a friend.
An agent flying on stale or incomplete data does not just underperform. It sends the wrong message at the wrong moment based on outdated context. That actively damages customer trust. A bad automated decision is worse than no decision at all.
The Two Questions to Ask Before Buying
- Do you have a single, consistent view of each customer that updates in real time, not a report that refreshes overnight, but a live record of activity across every touchpoint?
- Can you act on that data while the customer is still in the moment? Is it clean and connected enough to be trusted?
If the answer to either question is no, the agent conversation is premature.
The Honest Order of Operations
The intelligence layer is increasingly a commodity. Models are improving. Agent frameworks are multiplying. What remains genuinely scarce is a trusted, real-time, unified view of your customer that an agent can actually rely on. That does not come bundled with an agent subscription.
For some companies the right move is an internal data infrastructure investment. For others it means choosing a platform where the behavioral data layer is already built in. Either way, the foundation has to come before the agent. Agents will not fix broken marketing foundations. They will expose them.

